Image-based, pooled phenotyping reveals multidimensional, disease-specific variant effects

SUMMARY

Genetic variants produce complex phenotypic effects that confound current assays and predictive models. We developed Variant in situ sequencing (VIS-seq), a pooled, image-based method measuring variant effects on molecular and cellular phenotypes in diverse cell types. Applying VIS-seq to ~3,000 LMNA and PTEN variants yielded high-dimensional morphological profiles capturing changes in protein abundance, localization, activity and cell architecture. VIS-seq identified a subset of linker subdomain LMNA variants that increase nuclear circularity, in contrast to aggregating or low abundance rod subdomain variants that decrease circularity. VIS-seq also identified autism-associated PTEN variants that mislocalize, and accurately distinguished autism-linked from tumor syndrome-linked and gnomAD control variants. Most variants impacted a multidimensional phenotypic continuum not recapitulated by any single functional readout. By linking variants to cell images at scale, VIS-seq illuminates how variant effects cascade from molecules to subcellular structures to cells, providing a framework for resolving the complexity of variant function.

In Brief:

Variant in situ sequencing (VIS-seq) links genetic variants to cell images, revealing how variants affect molecules, subcellular structures, and cells at scale. Applied to thousands of LMNA and PTEN variants, VIS-seq illuminated how variants impacted a multidimensional phenotypic continuum not recapitulated by any single functional readout.

Graphical Abstract

INTRODUCTION

Genetic variants change the sequence of transcripts and peptides deterministically, but their effect on molecular or cellular phenotypes is difficult to predict. For example, protein-coding variants can alter protein structure, function or localization; disrupt macromolecular complexes and subcellular processes; or alter cell internal structure, morphology and behavior. Deep mutational scanning enabled the systematic, pooled evaluation of thousands of protein-coding variants in a single experiment for some of these molecular or cellular phenotypes¹. However, the vast majority of deep mutational scans and other multiplexed experiments have focused on simple, one-dimensional phenotypes² like cell growth^3,4 or protein abundance^5,6 in cultured human cell lines or non-human models. These experiments generally provide limited insight into variant pathomechanism, largely fail to parse complex gene-disease relationships and cannot inform on cell-type specific effects. Consequently, genetic variants are often viewed as affecting one phenotype and classified as either benign or pathogenic. In reality, variants have complex, multidimensional phenotypic consequences. Thus, scalable methods are needed that can enable variant effect measurements on generalizable, interpretable, and information-dense molecular and cellular phenotypes in specialized cell types.

Building on in situ sequencing methods for image-based, pooled CRISPR screens^7–13, we massively expanded the scope of deep mutational scanning by enabling pooled measurement of variant effects using cell imaging. Because imaging can spatially resolve biomolecules, subcellular structures and cell morphology using fluorescent proteins, in situ hybridization, antibodies, and cell-staining dyes, it yields insights into the effects of diverse perturbations. For example, imaging of cells with dyes that stain subcellular compartments¹⁴ revealed the effects of chemical¹⁵, CRISPR¹⁶, and ORF overexpression¹⁷ perturbations and illustrated that image-based profiling is at least as powerful as RNA-based profiling for clustering these perturbations¹⁸. Arrayed, image-based measurement of the localization of ~3,000 human pathogenic variant proteins expressed in a cell line also demonstrated that mislocalization is a common defect¹⁹. We overcame longstanding technical challenges to develop Variant in situ sequencing (VIS-seq), a method based on a circular RNA barcode that yields a thirteen-fold increase in reads per cell compared to a linear construct. We also created a transgene expression cassette that prevents silencing, enabling a ten-fold increase in scale compared to previous multiplexed experiments in human induced pluripotent stem (iPS) cells^20–23 and their derivatives. Thus, VIS-seq is a scalable platform that uses imaging of cells to quantify the effect of variants on diverse phenotypes in disease-relevant cell types.

We used VIS-seq to derive morphological profiles for ~3,000 synonymous, missense, nonsense and frameshift variants of LMNA and PTEN from ~11.4 million cell images. We chose these genes because variants in each are associated with multiple diseases, and the underlying molecular and cellular mechanisms linking variants to each disease remain poorly understood. Morphological profiles of LMNA variants expressed in U2OS²⁴ cells revealed extensive effects on the lamin A protein’s abundance, localization, aggregation, and ability to maintain nuclear shape. LMNA profiles enabled us to discover gain-of-function variants that increase nuclear circularity and revealed a direct connection between lamin A’s three-dimensional protein structure, localization and effect on cells. Morphological profiles of PTEN variants expressed in both iPS cells and iPS-derived, NGN2-induced neuron-like cells comprehensively mapped the interdependence of the PTEN protein’s abundance, activity, and localization, and revealed a nuclear localization defect enriched among autism spectrum disorder-associated variants. LMNA and PTEN morphological profiles identified known pathogenic variants with high accuracy and, in the case of PTEN, enabled us to discriminate between autism and tumor syndrome-associated variants. Thus, VIS-seq expands our ability to deeply phenotype genetic variants in different cell contexts, illuminating how variant effects propagate from molecules to subcellular structures to cells. More fundamentally, we show that variants perturb a phenotype space larger and more complex than previously explored, spanning a multidimensional continuum rather than a pathogenic-benign binary and highlighting the limitations of current one-dimensional experimental or predictive methods.

RESULTS

Barcoded circular RNAs and a universal expression cassette enable Variant in situ Sequencing

Variant in situ Sequencing (VIS-seq) is a platform for pooled, image-based profiling of thousands of transgenically expressed protein-coding variants (Figure 1A). VIS-seq comprises a cassette with a promoter expressing the protein variant and a second promoter expressing an abundant circular RNA^25,26 containing one or more barcodes that are sequenced in situ to reveal the identity of the variant expressed in each cell (Figure 1B). The VIS-seq expression cassette includes flanking insulators²⁷, a ubiquitous chromatin opening element²⁸, and an intron^29,30, along with a viral regulatory element³¹. Together, these elements prevent transgene silencing in iPS cells before and after differentiation, eliminating a longstanding barrier to transgene-based experiments in stem cells and their derivatives³². The cassette sustained expression of WT mEGFP³³-tagged histone H1.4 over 14 days of passaging in human iPS cells and after 56 days in differentiated iPS-derived cardiomyocytes, whereas expression from a lentiviral construct was silenced (Figure 1C). A linear RNA barcode yielded zero or one in situ sequencing reads in 56% of all cells and 28% of non-silenced cells, preventing effective barcode calling (Figure 1D), so we developed a circularizing, in situ sequencing-compatible RNA barcode^25,26. 93% of cells expressing this circularizing RNA barcode had two or more reads, and reads per cell increased ~13-fold in all cells and ~5-fold in non-silenced cells.

Figure 1. Transgenic variant expression and in situ sequencing with VIS-seq.

(A) VIS-seq uses fluorescent in situ sequencing of abundant circular RNA barcodes to genotype cells expressing protein variants. (1) A variant library in the VIS-seq expression cassette is integrated into cells via piggyBac-ase. (2) Cells are fixed; barcodes are reverse transcribed, captured with a padlock probe and amplified; (3) cells are stained and imaged; (4) barcodes are sequenced in situ; (5) single cell phenotype-genotype pairs are determined using STARCall; and (6) features for each cell are extracted using CellProfiler.

(B) VIS-seq expression cassette (PB = piggyBac³⁴; PD = padlock probe¹³⁰ binding site; UCOE = universal chromatin opening element²⁸; HBB IVS2 = hemoglobin subunit beta intervening sequence 2^29,30; IRES = internal ribosome entry site¹³¹; WPRE = woodchuck hepatitis virus post-transcriptional regulatory element³¹; cHS4 = chicken beta globin locus control region hypersensitive site 4²⁷).

(C) Time-series plot of mEGFP positivity of human WTC11 iPS cells and derived cardiomyocytes. mEGFP-tagged WT histone H1.4 is expressed via the VIS-seq cassette (blue) or lentivirus (red). Two replicates are plotted.

(D) Violin plot of the number of rolling circle amplification colonies per cell after one cycle of sequencing for VIS-seq expression cassette (blue) and lentivirus (red) cells from (C) at day 14. Silenced cells were excluded. Two replicates are shown.

(E) Matched images of mEGFP-tagged lamin A variants in U2OS cells, stained with DAPI, phalloidin, WGA, and mitoprobe (left), with corresponding first base in situ sequencing; magenta=guanine, blue=thymidine, yellow=adenine, and red=cytosine (right). Silver dashed lines depict the borders of cells. Scale bar indicates 10 μm.

(F) Matched images of mEGFP-tagged PTEN library in human WTC11 PTEN-KO inducible-NGN2 iPS cells, stained with DAPI and phalloidin-CF568 (left, top row) and first base in situ sequencing (right, top row). Matched images of mEGFP-tagged PTEN library in neuron-like cells, seven days after induction of NGN2, stained with DAPI and anti-MAP2 AF568 (left, bottom row) and first base in situ sequencing (right, bottom row). Silver dashed lines depict the borders of iPS cells (top) or cell bodies of NGN2-induced neuron-like cells (bottom). Nucleobase coloring identical to (E). Scale bar indicates 10 μm.

VIS-seq starts with integration of a library-bearing expression cassette via PiggyBac transposition³⁴ at low multiplicity of integration (MOI) (Figures 1A and S1A). Cells are fixed and barcodes are reverse transcribed, captured with a padlock probe and amplified^7,35. Cells are stained with the desired combination of fluorescent antibodies, RNA FISH probes, and dyes¹⁴. Images are collected and stains and probes are removed or bleached^10,36,37. A sequencing primer is hybridized to the amplified barcodes which are sequenced in situ⁷. Images are aligned and stitched in a single step using STitching, Alignment and Read Calling for in situ sequencing (STARCall), software we developed to reduce the need to maintain precise registration between rounds of phenotyping and in situ sequencing images³⁸ (available at https://github.com/FowlerLab/starcall-workflow). Barcode sequences are extracted and mapped to protein variants, then cell and nuclear borders are segmented^39,40. Lastly, morphological features are extracted using CellProfiler⁴¹, generating a single cell matrix of image features (Figure 1A). We demonstrated the VIS-seq workflow on four genes with variant mislocalization phenotypes,^19,42 LMNA, PTEN, RPS19, and HIST1H1E, in U2OS or human iPS or iPS-derived cells (Figures 1E, 1F, S1B, and S1C).

VIS-seq reveals morphological effects of thousands of LMNA variants

We used VIS-seq to phenotype variants in the rod domain of the intermediate filament protein lamin A, encoded by LMNA. Lamin A assembles into a poorly understood meshwork lining the inner nuclear membrane, maintaining nuclear shape and mechanical stability^43,44, connecting the nucleus to the cytoskeleton^45,46, and regulating gene expression^47,48. LMNA missense variants cause disparate diseases, collectively called laminopathies^47–53. Immunostaining of lamin A in patient fibroblasts⁵⁴ and transient transfection of lamin A variants into cell lines^42,55 revealed altered localization and morphology, with some pathogenic rod domain variants leading to lamin A aggregates, changes in nuclear shape, and other nuclear defects. However, not all pathogenic LMNA variants form aggregates, and the relationship between these molecular and cellular phenotypes and clinical phenotypes remains poorly understood⁴².

We created a library of >1,700 mEGFP-tagged LMNA synonymous, missense and frameshift variants of the central rod domain between amino acid positions 178 and 273, a hotspot for pathogenic variants. We integrated this library into U2OS LMNA-KO cells and imaged ~10 million fixed, library-expressing cells after staining chromatin (DAPI), microfilaments (phalloidin), plasma membrane (wheat germ agglutinin, WGA)¹⁴, and mitochondria (anti-12S and 16S mitochondrial rRNA FISH probes¹⁰, Figures 2A and S1D). In situ sequencing of the 12-base variant barcode followed by removal of cells with multiple barcodes yielded thousands of cells for most variants as well as ~900,000 cells expressing wild type (WT) lamin A (Figures 2B and S2A–S2C). We used CellProfiler⁴¹ to extract 1,077 replicable morphological features for each cell that measured the intensity, distribution, texture, and granularity of the mEGFP-lamin A and stain channels, as well as correlations between channels (Figures 2C and 2D, Table S1). Features that reflected biologically irrelevant information (e.g., pixel coordinates, cell orientation)⁵⁶ were removed. Thus, we obtained image and feature information for 1,767 variants expressed in ~6.6 million cells (Figures 2A, S2D).

Figure 2. Morphological profiles of 1,767 LMNA variants identify variants that alter function, including aggregating and pathogenic variants.

(A) Morphological profile analysis workflow. (1) A matrix of ~6.6 million single cells each with 1,077 CellProfiler-derived features, a sequenced barcode and associated lamin A variant was generated. (2) Feature medians and earth-mover distances were computed among all cells expressing each variant; feature selection removed features that were highly correlated, had low variance, or were biologically irrelevant. (3) Two metrics were computed to quantify the overall effect of each variant. Morphological impact score was defined as the cosine distance between a variant morphological profile and the median synonymous profile. Distinguishability score was defined as the area under a receiver operating characteristic curve (AUROC) of a classifier trained to distinguish a variant from WT using single cell feature profiles. (4) Variant embeddings were visualized using UMAP following dimensionality reduction with PCA. Heatmaps of position by substitution effects were visualized for each feature.

(B) Boxplots of the number of profiled single cells per variant of the indicated class across both replicates.

(C) Median over cells of variant mean nuclear intensity in the mEGFP-lamin A channel for variants in both replicates colored by variant type as in (B), with Pearson’s r shown.

(D) Stacked barplot showing proportion of all features, selected features which generate variant profiles, and hit features (defined in (E)) by imaging channel, cellular compartment, or method of computing the feature. The number of features in each set is indicated. Color legends are shown below.

(E) Volcano plot of the median of LMNA feature median z-scores across all 1,767 variants (x-axis) against geometric-mean p-values (y-axis, Komologorov-Smirnov test against the WT distribution of feature values). Both the median effect (x-axis) and p-values (y-axis) are computed over all profiled variants. Red dashed lines show the Bonferroni-corrected p<0.01 threshold and median effect thresholds used to define hit features. Point area is proportional to the number of variants that pass thresholds. Coloring is by imaging channel, see (D) for color legend.

(F) Boxplots of morphological impact scores for variants of the indicated class. Aggregating controls indicate independently-measured variants that aggregated in >15% of HEK 293T cells⁴² (ClinVar variants of uncertain significance=VUS and likely pathogenic/pathogenic variants=P/LP). *** indicates Mann-Whitney p<0.001.

(G) Boxplots of distinguishability scores for variant classes from (F). *** indicates Mann-Whitney p<0.001.

(H) UMAP visualization of PCA-transformed morphological profiles for each variant. Circles indicate synonymous (green), missense (grey) and frameshift (purple) variants, triangles indicate ClinVar likely pathogenic/pathogenic (P/LP, red).

(I) UMAP visualization from (H). Colored symbols indicate independently measured fraction of cells (“+” symbols, >15%; “x” symbols, <15%) with lamin A aggregates in HEK 293T cells.

To summarize feature heterogeneity across cells, we computed the median and earth-mover distance (EMD)^57,58 from WT for each feature for every variant (Figure 2A). We removed 745 features that had low variance, high correlation to other features, or had EMDs with low reproducibility⁵⁸, leaving 332 “selected features.” Selected features were more balanced than all features across channels, compartments, and types of measurement; they comprise each variant’s “morphological profile” used to score, visualize, and cluster variants (Figure 2D). Independently, we identified 407 “hit” features that were the most consistently perturbed across variants (Figures 2E and S2E, Table S1). Hit features were overwhelmingly derived from the mEGFP-lamin A channel and were used to interpret variant effects.

Morphological profiles detect variant aggregation and summarize impact and pathogenicity

We summarized the dissimilarity between each variant’s profile and WT via two metrics: first, a morphological impact score describing the magnitude of a variant’s effect on its morphological profile (Figure 2F, Table S2); and second, a distinguishability score describing the performance of a gradient-boosted decision tree model trained for every variant. Each model is trained to discriminate between single cells expressing the variant and cells expressing WT based on all features. The distinguishability score for a variant is defined as the area under the receiver operating characteristic curve (AUROC) computed on a test set of variant and WT cell features (Figures 2G and S2F, Table S2). Morphological impact (Pearson’s r=0.88) and distinguishability scores (r=0.92) were highly concordant between replicates (Figures S2G and S2H). Both scores were significantly lower for synonymous variants than for missense and frameshift variants, with distinguishability scores better discriminating different classes of variants (Figures 2F and 2G). LMNA variant features, morphological profiles, and scores can be explored at https://visseq.gs.washington.edu/.

Next, we validated the morphological profiles and summary scores using control aggregating and pathogenic variants. The 14 known aggregating variants⁴² segregated from synonymous variants on the UMAP visualization of variant profiles, and had significantly higher morphological impact and distinguishability scores, illustrating the power of unbiased morphological profiling to detect this known lamin A phenotype (Mann-Whitney U p<0.001, Figures 2F–2I). All pathogenic (P) or likely pathogenic (LP) ClinVar⁵⁹ variants except p.Leu204Arg, a single-submitter variant in an individual with dilated cardiomyopathy, occupied regions of the UMAP distinct from nearly all synonymous variants and had markedly elevated morphological impact and distinguishability scores compared to synonymous variants and variants of uncertain significance (VUS, Mann-Whitney U p<0.001). Impact scores computed with both median and EMD features modestly outperformed EMDs alone but substantially outperformed medians alone at separating aggregating or clinical control variants (Figures S2I and S2J). Thus, unbiased summarization of morphological profiles at both variant- and single-cell resolution accurately differentiated aggregating and pathogenic control variants from synonymous variants (distinguishability score AUC=0.99 for aggregating controls and 1 for ClinVar controls, impact score AUC=0.94 and 0.88, Figure S2J).

LMNA variant morphological profiles form clusters with explainable effects

We used morphological profile similarity to group variants into ten clusters (Figures 3A and S3A). Within-cluster profiles were more similar than between-cluster profiles, and clusters were robust to resampling and reclustering (Figures S3B and S3C). To understand the effects of variants in each cluster, we focused on four hit features that are interpretable and highly perturbed. These “landmark features” are: lamin A intensity in the nucleus, lamin A intensity at the nuclear boundary, lamin A nuclear granularity, and nuclear circularity (Figures S2E, S3D, and S4A, Tables S1 and S2). 76 (90%) synonymous variants occupied clusters 5, 6, and 7, which, along with 519 (31%) missense variants, had little effect on landmark features and were similar to WT (Figures 3A, 3B, S3A, and S3B). The known aggregating variants were exclusively present in clusters 8, 9, and 10, along with 380 (24%) other missense variants⁴² (Figure 3A). Variants in these clusters had low nuclear and boundary lamin A abundance and high granularity, presumably because of relocalization from the nuclear boundary to aggregates (Figures 3B and S4B). These variants also exhibited the largest decrease in nuclear circularity, leading to both more elliptical and more concave nuclei and indicating that these variants abrogate lamin A’s ability to maintain nuclear shape^43,46,60 (Figures 3B and S4A–S4C). The pathogenic variant p.Asn195Lys⁵⁰, along with six other pathogenic variants (47% total), were also found in these clusters (Figures 2H and 3A). To better understand the effects of variants on nuclear boundary abundance, we interrogated lamin A intensity as a function of distance from the nuclear center (Figures 3C and S4D). In clusters 8 and 9, variants consistently reduced laminar abundance, but variants in cluster 10 had more heterogeneous effects.

Figure 3. Clustering LMNA variant morphological profiles yields rich, interpretable phenotypes.

(A) Center: UMAP of LMNA variant profiles (Figure 2H) colored by Louvain¹²⁸-derived clusters. Periphery: images of variants and description of nuclear morphology, lamin A abundance and lamin A localization associated with each cluster. Silver dashed lines depict the borders of cells, blue=DAPI, green=lamin A-mEGFP. Scale bar indicates 5 μm. Graphical representations of each cluster indicate nuclear envelope morphology (black), lamin A localization and abundance (green) and nuclear circularization (red arrows). UMAP point and text color indicate cluster identity, depicted in (B).

(B) Heatmap of landmark feature medians and EMDs for each cluster, colored by z-score versus the synonymous variant distribution. The nuclear mEGFP-lamin A intensity feature measures the average lamin A pixel intensity in the nucleus, the boundary mEGFP-lamin A intensity feature measures the average lamin A pixel intensity within two pixels of the nuclear boundary. The nuclear mEGFP-lamin A granularity 1 feature is related to the presence of lamin A aggregates, with high values indicating aggregates. The nuclear circularity feature is the normalized ratio of nuclear perimeter squared to nuclear area, with high values indicating a circular nucleus.

(C) Plot of the mEGFP-lamin A total intensity in radial deciles of distance from the centroid of the nucleus, normalized to the DAPI total intensity in each decile. Bin 1 corresponds to the innermost radial decile and bin 10 to the outermost (see diagram below plot). mEGFP-laminA/DAPI median intensity for clusters 1, 3, 4, and 8 z-scored versus the synonymous variant distribution are shown. The remaining clusters are shown in Figure S4C.

The 454 (28%) missense variants in clusters 2, 3, and 4 had smaller overall decreases in nuclear and boundary abundance. Cluster 3 variants produced aggregates of varied sizes and were lower in abundance across the nucleus, whereas cluster 4 variants did not aggregate and were lower in abundance at the nuclear boundary (Figures 3B, 3C, S4B, and S4D). The pathogenic and known partially-aggregating⁴² variant p.Arg190Gln, along with five other pathogenic variants (40% total), were also found in these clusters (Figures 2H and 3A). Surprisingly, the 292 (18%) missense variants grouped in cluster 1 drove large increases in nuclear circularity and an increase in lamin abundance in the nuclear interior, suggesting that these variants, including p.Glu228Val, a pathogenic variant associated with multiple overlapping clinical phenotypes⁶¹, may exert their effects via a gain-of-function mechanism (Figures 3B and 3C).

In addition to defining clusters based on variant morphological profiles, we explored the relationship between landmark features (Figure S3D). 44% of profiled missense variants perturbed nuclear circularity whereas only 20% drove changes in nuclear lamin A granularity, despite aggregation being readily appreciable in images of patient-derived cells^42,55,62. Thus, VIS-seq grouped variants by their effects on lamin A intensity, localization, aggregation, nuclear shape and other features, and revealed a class of variants that increase nuclear circularization.

VIS-seq reveals the structural basis of LMNA variant effects

Next, we compared lamin A morphological profiles to prior crystal structures^63–65 and models⁶⁶ to understand the structural basis for the segregation of variants into clusters. Lamin A forms a soluble dimer in the cytoplasm that assembles into a multimeric filament and creates a meshwork at the lamina. The region of lamin A we mutagenized spans the rod domain comprising two α-helical coiled coil subdomains, coils 1B and 2A, separated by the linker 12 subdomain (Figure 4A). In other intermediate filament protein structures, linker 12 is flexible and adopts non-helical conformations, but in crystal structures of partial lamin A filaments, linker 12 is α-helical^63,64,67. An alternate model suggests that the functional state of linker 12 is flexible and non-α-helical^66,68. To explore the relationship between these lamin A structures and function, we created variant effect maps for each landmark feature alongside UMAPs colored by feature scores (Figures 4A, 4B, and S4E–S4G). These heatmaps revealed differences between variant effects at proposed coil and linker positions and profound effects of proline substitutions, which disrupt α-helices. Moreover, coil and linker substitutions had similar morphological impact scores but different morphological effects (Figures S5A–S5C). At coil positions, missense substitutions were enriched in clusters 3 and 8, characterized by aggregation, and cluster 4, characterized by low laminar abundance, (Fisher exact test q<0.01–0.001, Figure S5B). However, at linker 12 positions, missense substitutions were strongly enriched in cluster 1, characterized by increased circularity and preserved lamin A abundance (Fisher exact test q<0.001, Figure S5B). Proline substitutions had the highest morphological impact scores of all, but their morphological effects depended on their location (Mann-Whitney U p<0.001, Figures 4A, 4C, S4E–S4G, and S5D). To further explore the effects of proline substitutions, we clustered them by their morphological profiles, yielding linker, linker-proximal, and α-helical position groups (Figure 5E, Table S2). Of linker 12 positions, 17 of 19 (89%) were in the linker group, and all positions in the α-helical group were in the coil subdomains (Figure 4C). By revealing that variants at flexible linker positions, linker-proximal positions, and α-helical coil positions can have distinct effects, our results support the alternate structural model⁶⁶.

Figure 4. Disruption of lamin A multimerization dictates effects on aggregation and abundance.

(A) Map of lamin A missense variant effects on the nuclear circularity feature (grey boxes=missing variants, black dots=synonymous substitutions). Blue to red coloring indicates feature z-score versus the synonymous variant distribution. Above the map, positions are annotated by their participation in multimer contacts (first row, defined below), α-helical average substitution profile (second row, defined below and in Figure S5F), and proline-substitution profile (third row, defined below and in Figure S5E). Linker or coil subdomains are shown directly above the heatmap⁶³. Position-averaged scores are below and amino acid substitution marginal feature z-score distributions are on the right (points depict variants; black lines depict medians).

(B) UMAP of variant profiles (Figure 2H) colored by nuclear circularity score as in (A).

(C) Proline substitutions are shown as triangles on the UMAP shown in (Figure 2H), colored by clustering of the effects of proline substitutions as defined in Figure S5E with legend above.

(D) The number of proline substitutions is plotted for each subdomain⁶³ colored as in (C) with legend above. *** indicates χ² p<0.001.

(E) The number of α-helical positions (defined in Figure S5E) plotted by subdomain⁶³ and colored based on the effect of variants on aggregation or abundance at the position (defined in Figure S5F), with legend above. ** indicates χ² p<0.01.

(F) The number of α-helical positions (defined in Figure S5E) plotted by their participation in structural contacts ((G), see Methods), colored as in (E). Color legend shown above. ** indicates χ² p<0.01.

(G) A11 tetramer structure 6JLB⁶³ (bottom) and the A22 tetramer structure⁶⁴ (top) with positions colored as in (E) (magenta=aggregation-sensitive, light blue=abundance-sensitive, dark blue=low impact, grey=position not profiled, black=position not clustered).

(H) Graphical representation of lamin A processing, localization, aggregation, and degradation, annotated with their relationship to dimerization, multimerization and phenotypic effects as well as variant clusters.

Figure 5. Morphological profiles of >1,200 PTEN variants in iPS cells and NGN2-induced neuron-like cells.

(A) Image of PTEN library at positions 112–172 expressed in human PTEN-KO iPS cells carrying dox-inducible NGN2 (left) and neuron-like cells 7 days after induction of NGN2 (right). MAP2=anti-MAP2 antibody, ConA=concanavalin A. Scale bar indicates 10 μm.

(B) UMAP visualization of 1,228 PTEN variant morphological profiles collected in iPS cells (left) and NGN2-induced neurons (right). Point color indicates variant type: synonymous (green), missense (grey), 3-nt deletions (blue), nonsense (red), and frameshift (purple).

(C,E,G) Missense variant effect maps for iPS cell PTEN landmark features (grey boxes=missing variants, black dots=synonymous substitutions). Blue to red coloring indicates feature z-score versus the synonymous variant distribution. Linker or coil subdomains are shown directly above the heatmap. Position-averaged scores are below and amino acid substitution marginal feature z-score distributions are on the right (points depict variants; black lines depict medians).

(D,F,H) UMAPs of PTEN variant morphological profiles in iPS cells (left) and NGN2-induced neurons (right) colored by landmark features. Blue indicates low and red indicates high z-score versus the synonymous variant distribution, as in (C,E,G).

(I) Barplot of fraction of variants that perturb landmark features in iPS cells and NGN2-induced neurons. For each variant, a feature was considered perturbed if |z-score|>2.5 and KS-test Bonferroni-corrected p<0.01. *** indicates χ² p<0.001.

(J) Graphic depicting the relationship between PTEN localization (cell membrane, cytoplasm, or nucleus), activity and protein abundance, annotated with changes to landmark features.

We further investigated coil positions. Helices in the coiled coils can make contacts with adjacent helices across a dimer-forming interface in coil 1B and multimer-forming interfaces in both coil 1B and 2A^63,64. Prior work has linked aggregating or pathogenic variants to dimer-forming contacts^42,69, or examined the effect of a few variants on multimer formation⁶³. To systematically probe the nature of these contacts, we clustered α-helical positions based on the average morphological profile across all substitutions at each position (Figure S5F, Table S2). We identified three groups of positions where substitutions were either aggregation-sensitive, abundance-sensitive or low impact, with corresponding significant differences in morphological impact scores and variant cluster segregation (Mann-Whitney U p<0.001, Figures S5G–S5J). The disparate morphologies characterizing substitutions in each of the three groups of positions were dictated by whether variants occurred at dimer or multimer interfaces (Figures 4E–4G and S5H). For example, 67% of dimer–facing positions, which occur only in coil 1B, were in the abundance-sensitive group, whereas 100% of multimer-facing positions, which occur in both coil 1B and 2A, were in the aggregation-sensitive group. 86% of non-interacting residues, as well as the synonymous variant average profile, were in the low impact group (Figures 4E–4G). The proportion of amino acid substitutions also differed significantly across clusters. For example, substitutions to acidic residues, enriched in variant clusters 8 and 9, drove high impact and distinguishability scores and low abundance and circularity at α-helical positions. Substitutions to aromatic amino acids, enriched in variant clusters 1, 3, 9, and 10, drove aggregation especially at aggregation-sensitive α-helical positions (Fisher’s exact test, q<0.001–0.05, Figures S5K and S5L).

Thus, unsupervised clustering of LMNA variant morphological profiles yielded insight into lamin A structure. The effects of variants in linker 12, enriched in the high-circularity and preserved nuclear abundance variant cluster 1, suggest that linker 12 is not a static coiled coil but, as postulated by the alternate structural model, is flexible⁶⁶ (Fisher’s exact test q<0.001, Figures 4H and S5J). The effects of variants at α-helical positions were largely dictated by their location in dimer or multimer interfaces. Here, disruption of dimer interfaces at abundance-sensitive positions, disproportionately by low abundance variants in cluster 4, apparently did not prevent formation of lamin A multimers needed to circularize the nucleus (Fisher’s exact test q<0.001). Disruption of multimer interfaces at aggregation-sensitive positions, disproportionately by variants in clusters 3, 8, and 9, was characterized by lamin A aggregation and profound loss of nuclear circularity (Fisher’s exact test q<0.01–0.001). Rather than a simple continuum of loss of function based on protein folding or stability, disruption of different aspects of lamin A structure and assembly led to diverse effects. VIS-seq enabled us to trace these effects as they cascade from molecular phenotypes like protein abundance and localization to cellular phenotypes like nuclear morphology.

Morphological profiling of PTEN variants in iPS cells and NGN2-induced neuron-like cells

We next focused on Phosphatase and TENsin homolog (PTEN), a lipid and protein phosphatase that negatively regulates the PI3K-AKT-mTOR pathway and plays roles in cell growth, cell migration, and genome stability^70–75. Pathogenic germline PTEN variants give rise to a spectrum of autosomal dominant tumor predisposition phenotypes⁷⁶ and are a cause of autism spectrum disorder, often co-occurring with macrocephaly and developmental delay⁷⁷. Somatic PTEN mutations occur frequently in cancer^70,78. Patients with PTEN variants can present with a complex mix of these phenotypes, and the relationship between changes in PTEN molecular function and disease remains unclear^79,80. To explore this relationship, we collected morphological profiles for 1,228 PTEN catalytic domain synonymous, missense, nonsense, 3-nt deletion and frameshift variants in >2.8 million human iPS cells and >2 million NGN2-induced neuron-like cells (hereafter NGN2-induced neurons) in two replicate PTEN-knockout clonal cell lines⁸¹ (Figures 5A, S1E, and S6A–S6E). iPS cells were imaged with dyes staining chromatin (DAPI), microfilaments (phalloidin), and endoplasmic reticulum (concanavalin-A)¹⁴ as well as mEGFP-tagged PTEN and an antibody staining phospho-Thr308-AKT (pAKT) which reports PTEN lipid phosphatase activity⁸². NGN2-induced neurons were imaged with the same dyes and antibodies except that a marker of neuronal differentiation⁸³ (anti-MAP2 antibody) replaced phalloidin.

Feature selection led to variant morphological profiles comprising 1,174 feature medians or EMDs for iPS cells and 1,036 feature medians or EMDs^57,58 for NGN2-induced neurons. Due to the difficulty of robustly segmenting neurites, all NGN2-induced neuron features were derived from segmented nuclei and cell bodies. Hit features in both cell types were disproportionately from mEGFP-PTEN and pAKT channels (Figures S6F–S6I, Table S3). UMAP visualization of the iPS cell and NGN2-induced neuron variant profiles demonstrated clear separation of synonymous from 3-nt deletion and nonsense variants (Figure 5B). Likewise, PTEN variant morphological impact and distinguishability scores were higher for 3-nt deletion, frameshift, and nonsense variants than for synonymous and missense variants in both iPS cells and NGN2-induced neurons (Mann-Whitney U p<0.001, Figures S6J and S6K), and replicate impact scores were strongly correlated (Pearson’s r=0.88, 0.89, Figure S6L, Table S4). Impact (Pearson’s r=0.95) and distinguishability scores (r=0.90) were highly concordant between iPS cells and NGN2-induced neurons, suggesting that variants tended to be similarly disruptive in both cell types (Figure S6M). PTEN variant features and morphological profiles can be explored at https://visseq.gs.washington.edu/.

PTEN variants alter abundance, lipid phosphatase activity, and localization in a cell type-specific manner

To explore variant-specific differences between cell types, we focused on three interpretable, highly-perturbed landmark features: Steady-state PTEN abundance as reflected by mEGFP-PTEN intensity, lipid phosphatase activity as reflected by pAKT intensity, and nucleocytoplasmic localization as reflected by pixel-level DAPI-PTEN correlation (Figures 5C–5H, S6F–S6I, and S7A–S7D, Tables S3 and S4). Both PTEN abundance and lipid phosphatase activity exhibited a steep gradient across variants in the UMAP in both iPS cells and NGN2-induced neurons (Figures 5D and 5F). Our PTEN abundance and activity measurements were correlated with previous measurements of activity in a yeast fitness assay (Pearson’s r=−0.84 versus iPS cell pAKT intensity and −0.86 versus NGN2-induced neuron) and abundance in a human cell line (r=0.78 versus iPS cell mEGFP-PTEN intensity, 0.77 versus NGN2-induced neuron, Figure S7E)^5,84. As expected, abundance-sensitive PTEN positions were concentrated in the core of the protein⁷¹ (e.g., β-sheet positions 118–121, α-helix positions 137–141, Figures 5C and S7F–S7H). Some variants had cell type-specific abundance effects; 43% of variants altered PTEN abundance in iPS cells versus 52% in NGN2-induced neurons (χ² p<0.001, Figures 5I and S8A). Activity-sensitive positions were concentrated at the catalytic surface (e.g., P-loop⁷¹ positions 123–130 and 171) and in second/third-shell positions (e.g., 155 and 160, Figures 5E and S7G). Unlike for abundance, variants appeared to have similar effects on activity in each cell type (Figures 5I and S8A).

The third landmark feature, DAPI-PTEN correlation, is a sensitive readout of nucleocytoplasmic localization (Figures 5G and S8B). PTEN is found in the nucleus, where it dephosphorylates nuclear proteins and may have non-enzymatic functions^72,74,75,80. The DAPI-PTEN correlation feature formed a gradient across variants in the UMAP, with less consistency than abundance or activity (Figure 5H). 67% of variants altered DAPI-PTEN correlation in iPS cells versus 48% in NGN2-induced neurons (χ² p<0.001, Figures 5G, 5I, S7C, and S8C). However, PTEN nucleocytoplasmic localization was related to changes in both abundance and lipid phosphatase activity: Low abundance variants had low DAPI-PTEN correlation levels and thus cytoplasmic localization compared to WT, whereas low lipid phosphatase activity variants had high DAPI-PTEN correlation levels and thus nuclear localization (Figures 5C–5H and S8D). We used kernel ridge regression to model DAPI–PTEN correlation as a function of abundance and activity and computed the residual as a mislocalization score (see Methods). This score quantified each variant’s effect on localization relative to WT independent of abundance and activity, and was reproducible (Pearson’s r=0.68, Figures S8E and S8F). The mislocalization score revealed variants that shifted PTEN into the nucleus without altering pAKT levels. We hypothesized that such variants would occur at the membrane-binding positions 163 and 164⁸⁵ and the substrate specificity-determining positions 167 and 168⁸⁶. Indeed, at these positions, 43 (72%) profiled variants had elevated mislocalization scores yet WT-like lipid phosphatase activity. By contrast, only 18 (27%) variants at four adjacent activity-sensitive positions (165, 166, 170, and 171) had elevated mislocalization scores (Figures S8E, S8G, and S8H). Thus, analyzing the relationships between multiple interpretable VIS-seq features enabled us to isolate variant effects on specific phenotypes.

Overall, PTEN missense variants impacted diverse combinations of landmark features in iPS cells and NGN2-induced neurons, most frequently altering all three features in both cell types (20%, Figure S8I). Variants that reduced lipid phosphatase activity, concentrated at catalytically important positions, were more likely to be distinguishable in iPS cells whereas variants that reduced abundance, concentrated in buried positions, were more likely to be distinguishable in NGN2-induced neurons (Figures S8J and S8K). Thus, morphological profiles enabled us to dissect the relationship between PTEN abundance, lipid phosphatase activity, and nucleocytoplasmic localization in a cell-type-specific manner (Figure 5J).

PTEN variant morphological profiles are validated by clinical phenotypes and reveal disease-specific variant effects

We curated 5 likely benign (LB) and 62 likely pathogenic or pathogenic (LP/P) missense variants from ClinVar⁵⁹, and these clinical control variants were separated on the UMAP in both iPS cells and NGN2-induced neurons (Figure S9A). LP/P variant morphological impact scores were significantly elevated as compared to both synonymous and LB/B variant impact scores (Mann-Whitney U p<0.001, Figure S9B). To explore the relationship between PTEN variant morphological profiles and autism spectrum disorder/developmental delay (ASD/DD), PTEN Hamartoma Tumor Syndrome (PHTS), or somatic cancer, we curated 66 germline variants in 541 probands for PHTS and ASD/DD phenotypes (Table S5) as well as 22 variants from the COSMIC database⁸⁷ (Figure 6A, Table S4). 31 variants in this combined set were associated with only ASD/DD, PHTS or somatic cancer, and 41 were associated with multiple clinical phenotypes. Distinguishability scores separated P/LP variants or curated variants associated with all three clinical phenotypes from LB and gnomAD⁸⁸ controls better than yeast fitness or human cell abundance scores (Figures 6B and S9C). Additionally, morphological impact and distinguishability scores were significantly increased for variants associated with each clinical phenotype when compared to synonymous or gnomAD⁸⁸ variants (Mann-Whitney U p<0.001, Figures 6C, S9D, and S9E). However, variants associated with ASD/DD had significantly lower morphological impact scores compared to variants associated with PHTS or somatic cancer (Mann-Whitney U p<0.01–0.001). Thus, morphological profiles were superior to single phenotype measurements in capturing overall PTEN missense variant pathogenicity and in relating variants to clinical phenotypes.

Figure 6. PTEN variant morphological profiles predict clinical phenotypes and reveal disease-specific variant effects.

(A) Venn diagram of clinical phenotypes associated with curated variants (see Methods and Table S5). Each clinical phenotype was associated with a variant if it occurred in at least one proband with that variant.

(B) Receiver operating characteristic (ROC) curves for univariate zero-shot models predicting ClinVar pathogenicity or each clinical phenotype (ASD = autism spectrum disorder, DD = developmental delay, PHTS = PTEN hamartoma tumor syndrome; see Methods for curation criteria) from iPS cell and NGN2-induced neuron-like cell distinguishability scores (this publication, solid lines), yeast fitness scores⁸⁴ (dashed line), VAMPseq scores⁵ (dashed line), or AlphaMissense⁹¹ or EVE⁹² scores(dot-dashed lines). Area under the curve (AUC) scores are shown for each model.

(C) Morphological impact scores for PTEN variants in iPS cells plotted by association with clinical phenotypes. gnomAD v4.1 (blue) and synonymous (green) variants are plotted for comparison. *** indicates Mann-Whitney U p<0.001 and ** indicates p<0.01.

(D) UMAPs of iPS cell (left) and NGN2-induced neurons (right) PTEN variant profiles highlighting curated variants from (B). Triangles indicate association with clinical phenotypes (colors match (D)). gnomAD v4.1 (blue) variants are also plotted. Circles represent other variants, colored green (synonymous) or grey (non-synonymous).

(E) Mislocalization scores (see Methods) for PTEN variants in iPS cells plotted by association with clinical phenotypes. gnomAD v4.1 (blue) and synonymous (green) variants are plotted for comparison. *** indicates Mann-Whitney U p<0.001 and ** indicates p<0.01.

(F) Volcano plots of median PTEN variant effects on iPS cell feature median z-scores (x-axis) against geometric-mean KS-test p-values (y-axis, tested against WT feature distributions) for variants associated with different clinical phenotypes (see (B)). Points are features with area proportional to the number of variants that pass thresholds and colored by imaging channel (see right for legend). iPS cell landmark features are highlighted. Red dashed lines show the Bonferroni-corrected p<0.01 threshold and median effect thresholds.

(G) Macro-averaged multiclass ROC curve for PHTS-associated, ASD/DD-associated, and gnomAD or ClinVar LB/B variants. Curves show one-vs-rest performance for univariant zero-shot models from (B) as well as multivariate support vector classifier models trained on landmark features measured in iPS cells or NGN2-induced neurons (this publication, dotted line). Corresponding macro-averaged AUC are shown.

We hypothesized that, rather than being points on a single dimension of pathogenicity, ASD/DD, PHTS, and cancer-associated variants might instead perturb distinct molecular and cellular phenotypes. Indeed, variants associated with each clinical phenotype occupied distinct regions of the UMAP, implying that they had disparate molecular and cellular effects (Figure 6D). ASD/DD-associated variants have been previously described as hypomorphic^89,90, and they do have lower overall iPS cell morphological impact as well as milder abundance and lipid phosphatase activity phenotypes compared to PHTS and somatic cancer-associated variants (Figures S9F and S9G). However, ASD/DD-associated variants also shifted aberrantly into the nucleus in iPS cells unlike PHTS and somatic cancer-associated variants (Mann-Whitney U p<0.01, Figure 6E). DAPI-PTEN correlation was the hit feature with the largest effect across ASD/DD-associated variants, unlike for PHTS-associated variants (Figures 6F and S9H). ASD/DD-associated variants perturbed mEGFP-PTEN channel features most strongly, whereas PHTS-associated variants perturbed pAKT channel features most strongly (Figures 6F and S9H, Table S3). Based on these observations, we trained a linear support vector classifier that could accurately discriminate between gnomAD control, PHTS-associated, and ASD/DD-associated variants using the landmark features (iPS cell landmarks macro-averaged five-fold cross-validated AUC=0.92 versus 0.76 for yeast fitness assay, 0.62 human cell abundance assay, Figure 6G). Thus, we dissected the relationship between PTEN variant molecular, cellular, and clinical phenotypes, identifying molecular phenotypes that differentiate ASD/DD- from PHTS-associated variants and revealing that ASD/DD-associated variants cannot be understood solely as hypomorphic.

Variant effect predictors perform poorly on many molecular and cellular phenotypes

State-of-the-art variant effect predictors leverage conservation and protein structure to make accurate inferences of overall variant effects. We compared three leading predictors, AlphaMissense⁹¹, EVE⁹², and REVEL⁹³, to our morphological profiles. Variant effect predictions correlated poorly with lamin A morphological impact scores as compared to impact score replicability (Spearman’s ρ=0.27 to 0.35, Figures S2G, S10A, and S10B) and more strongly with PTEN impact scores (Spearman’s ρ=0.69 to 0.84 in iPS cells and 0.64 to 0.81 in NGN2-induced neurons, Figures S6L and S10C–S10F). Similar correlations were observed between other variant effect data and predictions^91–94.

The correlation between individual lamin A or PTEN landmark features and predictions was more variable. For example, predictions were moderately correlated with lamin A variant abundance (Spearman’s ρ=−0.35 to −0.46) but poorly correlated with lamin A variant aggregation (ρ=0.23 to 0.25) and nuclear circularity (ρ=−0.26 to −0.37, Figure S10A). Similarly, predictions were moderately correlated with PTEN lipid phosphatase activity (ρ=0.57 to 0.79) but poorly correlated with PTEN abundance (ρ=−0.19 to −0.52) or iPS cell PTEN localization (ρ=0.05 to 0.32, Figures S10C and S10E). Predictions performed similarly to VIS-seq distinguishability scores in separating PTEN pathogenic variants from gnomAD controls (Figures 6B and S9C). However, one-dimensional predictions or assays performed poorly compared to multidimensional VIS-seq landmark features at discriminating between gnomAD control, PHTS-associated, and ASD/DD-associated variants (VIS-seq macro-averaged five-fold cross-validated AUC=0.92 versus 0.76 for AlphaMissense and EVE, Figure 6G). These results highlight the strengths and limitations of current predictors and illustrate how the multimodal data provided by VIS-seq could be used to develop or constrain a next generation of phenotype-aware predictors.

DISCUSSION

We introduce Variant in situ Sequencing (VIS-seq), a scalable single-cell method that uses in situ sequencing of circular RNA barcodes to simultaneously measure the effects of thousands of protein variants on features derived from cell images. VIS-seq includes an expression cassette engineered to prevent silencing and enable transgene assays in specialized or hard-to-transduce cell types where silencing and low expression limit current approaches³². Thus, VIS-seq expands multiplexed assays of variant effect beyond cell growth, protein abundance or reporter readouts in cancer cell lines or non-human cells by enabling morphological profiling that captures many biologically relevant, spatially-resolved molecular and cellular phenotypes.

We used VIS-seq to create morphological profiles for ~3,000 variants in LMNA and PTEN, genes with complex genotype-disease relationships. Variant morphological profiles for these genes comprise a high-dimensional space which can be interpreted using multiple landmark features with biological meaning. Prior work linked individual LMNA variants to aggregation⁴² and nuclear shape defects^54,55 but without a clear link to rod-domain structure and assembly. VIS-seq demonstrated that rather than generic loss-of-function, variants at linker 12, dimer-forming interfaces, and multimer-forming interfaces drove distinct cell morphologies defined by increased nuclear circularity, low abundance, or aggregation, respectively. Prior multiplexed datasets have measured PTEN variant effects on abundance and lipid phosphatase activity. VIS-seq in multiple cell types disentangled abundance, pAKT signaling, and localization, revealing a mislocalization phenotype enriched among autism-associated variants. Morphological profiles, which comprise a large set of simple measurements of the intensity, distribution and shape of different markers in cells, outperformed single phenotype measurements at discerning variant pathogenicity. They also illuminated how variant effects cascade from molecules to subcellular structures to cells in a fashion not captured by current experimental or predictive methods. Thus, we show that variant effects span a multidimensional continuum of function that transcends a pathogenic versus benign binary or a one-dimensional spectrum, highlighting the power of multimodal variant effect assays like VIS-seq.

Because VIS-seq can measure many imageable phenotypes in human cell lines and iPS-derived cell types, it will be broadly applicable, particularly to the ~5,400 disease-associated human genes⁹⁵. While other multiplexed assays of variant effect can measure cell growth^3,4,96, protein abundance^5,6 or transcriptomic changes^97,98, none offer the spatial information or multimodality of VIS-seq. For example, most genes are incompatible with cell growth assays^99–101, which do not generally provide mechanistic insight. Loss of protein abundance is a primary mechanism of missense variant pathogenicity^5,6,102,103, but fails to capture mislocalization or loss of catalytic activity¹⁰⁴. Lastly, missense variants do not always produce large transcriptomic changes, particularly in utilitarian cancer cell lines⁹⁷, instead exerting their primary effect on proteins or cells. VIS-seq leverages the extraordinary catalog of imaging-compatible stains and affinity reagents to overcome these limitations by enabling simultaneous measurement of a vast swath of molecular phenotypes like protein abundance, mislocalization, and protein function as well as organelle and cell phenotypes. Thus, because many aspects of variant dysfunction can be made visible in cells, VIS-seq can be applied broadly.

Limitations of the study

Despite its promise, VIS-seq has important limitations. Although our platform provides robust expression in many cell types, such exogenous expression systems create the possibility of expression-related artifacts. We used fluorescent protein fusions to visualize lamin A and PTEN, which changes lamin A function and can generally alter stability and other phenotypes¹⁰⁵. Although VIS-seq reagent costs are low (~$0.001/cell), a barrier to scaling is manual in situ sequencing which requires intensive hands-on time (~1 hour for imaging and ~1 hour for reagent exchange per base). Thus, sequencing a 12 base pair barcode takes at least 3 days. Two replicates of a typical human protein with ~8,000 missense variants, performed in a tiled fashion, would require ~4 weeks. Automation of reagent exchange and heating would reduce labor and speed data acquisition considerably. Customization of markers and stains to the target protein’s function creates informative morphological profiles but requires optimization of each new marker. Lastly, although image-derived features are useful for clustering and variant scoring, mapping profiles to cellular phenotypes requires manual feature and cell image interpretation. And unlike genes, proteins, or genomic elements which are quantified in single-cell -omics, most image features do not have straightforward biological relevance.

These limitations suggest productive avenues for future work. VIS-seq could be combined with time-lapse imaging or 3D organoid models¹⁰⁶ to capture dynamic or tissue-like phenotypes. Incorporating protein-tagging strategies for endogenous loci^107,108 with barcodes in the 3’-UTR would mitigate expression-related artifacts. Untaggable proteins could be immunostained for visualization. And, with rapidly advancing deep-learning-based pipelines for embedding image phenotypes^109,110, the bottleneck of measuring and interpreting morphology at scale will continue to diminish. In summary, VIS-seq represents a significant step forward in our capacity to assess variant effects on complex phenotypes. By uniting multiplexed assays with the power of imaging, our work serves as a blueprint for mapping how variant effects propagate from molecules to cells.

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Douglas Fowler (dfowler@uw.edu).

Materials availability

Plasmids generated in this study are publicly available through Addgene (252243, 252244, and 252245). Variant libraries and other unique reagents generated in this study are available from the lead contact upon reasonable request.

Data and code availability

• Variant morphological impact scores, distinguishability scores and landmark feature values can be found on MaveDB¹¹¹ (LMNA: https://mavedb.org/experiments/urn:mavedb:00001243-a, PTEN iPS cells: https://mavedb.org/experiments/urn:mavedb:00001244-b, and PTEN NGN2-induced neurons: https://mavedb.org/experiments/urn:mavedb:00001244-a).

• Tabular data is provided as a supplement or, for larger files, is available via Zenodo at doi.org/10.5281/zenodo.15787684.

• Image data, single-cell feature profiles, and CellProfiler pipelines are publicly available via the BioImage Archive at doi.org/10.6019/S-BIAD3095.

• All code necessary to reproduce our analyses, starting at the cell by features matrix or from summary data available via Zenodo, can be found at https://github.com/FowlerLab/visseq. STARCall code used to generate the cells by features matrix from phenotyping and genotyping images can be found at https://github.com/FowlerLab/starcall-workflow. Code used to generate trained XGBoost models and variant-level distinguishability scores can be found at https://github.com/FowlerLab/fisseqtools.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

General cell culture

U2OS cells, derived from an osteosarcoma in a female (ATCC, HTB-96), and HEK 293T cells, derived from the kidney of a female fetus (ATCC, CRL-3216), were cultured in media composed of DMEM (Gibco, 11965–092) with 10% fetal bovine serum (FBS; Hyclone, SH30071.03) and 1% Pen/Strep (Gibco, 15140–122). Cells were passaged in the following way: lifted by washing with DPBS (Gibco, 14190–144), incubating with Trypsin 0.05% (Gibco, 25300–054) for 5 minutes (HEK 293T) or 15 minutes (U2OS) then quenching with media, spun down at 300g for 5 minutes, and lastly resuspended in media and replated. Cells were tested routinely for Mycoplasma. For VIS-seq experiments, U2OS cells were plated at a density of 800,000 per well on all wells of a glass-bottom 6-well plate (Cellvis, P06–1.5H-N) and fixed ~12 hours after plating.

iPS cell culture

WTC11 iPS cells, derived from healthy adult male fibroblasts (Coriell, GM25256) were cultured on matrigel-coated six-well plates in complete mTeSR Plus (STEMCELL Technologies, 100–0276) media. The media was changed at least every other day. The plates were coated by diluting matrigel (Corning, 354234) to a final concentration of 80 μg/mL in DMEM-F12 (Gibco, 11320–033), adding 1 mL to each well (or 5mL to a 10-cm plate), and incubating the plates at room temperature for 1 hour. iPS cells were passaged as follows: washed with DPBS then incubated cells with Versene (Gibco, 15400–054) for 3–5 minutes until cells dissociated, then resuspended the cells in mTeSR Plus with 10 μM Y-27632 (SelleckChem, S1049) and replated. Media changes without re-plating used mTeSR Plus only. Cells were tested routinely for Mycoplasma. For VIS-seq experiments, iPS cells were plated on all wells of a matrigel-coated glass-bottom 6-well plate (Cellvis, P06–1.5H-N) at 600,000 per well and fixed using 4% paraformaldehyde, 6 hours after they were plated.

NGN2-induced neural differentiation of iPS cells

NWTC11.G3-WT¹¹² iPS cells, derived from WTC11 by Dr. Li Gan (Gladstone Institute, obtained with permission), were differentiated into NGN2-induced neuron-like cells using doxycycline-inducible NGN2 expression following a protocol reported previously⁸³. On day −3, iPS cells were dissociated using Versene and then resuspended in Pre-Differentiation medium with 10 μM Y-27632. Pre-Differentiation medium contains: KnockOut DMEM/F-12 (Gibco 12660–012), 1X MEM Non-Essential Amino Acids (Gibco 11140–050), 1X N-2 Supplement (Gibco 17502–048), 10ng/mL NT-3 (Peprotech, 450–03) and BDNF (Peprotech, 450–02), 1 μg/mL Laminin protein (Gibco 23017–015), and 2 μg/mL doxycycline (Sigma, D9891). Cells were plated at 800k/well using 1mL of media per well in 12-well plates coated in matrigel, and media was changed to Pre-Differentiation medium on days −2 and −1. On day −1, a glass-bottom 6-well plate was coated using 15 μg/mL poly-L-ornithine (Sigma, P3655) overnight and then washed with DPBS 3 times the following day followed by air drying at room temperature. On day 0, pre-differentiated cells were dissociated using Accutase (STEMCELL Technologies, 07922) and then pelleted at 200 g for 5 minutes, followed by resuspension in Maturation medium. Maturation medium contains: 50% Neurobasal-A medium base (Gibco 12349–015), 50% DMEM/F-12 base, 1X MEM Non-Essential Amino Acids, 0.5X GlutaMAX Supplement (Gibco 35050–061), 0.5X N-2 Supplement, 0.5X B-27 Supplement (Gibco 17504–044), 10 ng/mL NT-3 and BDNF, 1 μg/mL Laminin protein, and 2 μg/mL doxycycline. Cells were plated at 600k/well into poly-L-ornithine coated 6-well plates in 2 mL of Maturation medium per well. NGN2-induced neurons were fixed and processed for VIS-seq on day 4 (day 7 after doxycycline addition).

To confirm differentiation, NGN2-induced neurons were fixed (see below, VIS-seq NGN2-induced neuron fixation protocol) and stained using DAPI (Fisher Scientific, D1306) and rabbit anti-NCAM1 primary antibody conjugated to Alexa647 (CellSignal, 50831S) using the VIS-seq staining mix (see below, Cell Painting Buffer + 0.25% Triton-X 100) and imaged using the same settings as used in the VIS-seq phenotype imaging (Figure S1B).

Cardiomyocyte differentiation of iPS cells

WTC11 iPS cells were cultured on 10 μg/ml Vitronectin XF (STEMCELL Technologies, 100–0763) prior to cardiomyocyte differentiation. Small molecule directed differentiation was performed as previously described with modifications²². iPS cells were plated in 15 μg/ml Vitronectin XF-coated 24 well plates at 7.5×10⁴ cells per well in mTeSR Plus supplemented with 10 μM Y-27632. Media was changed to fresh mTeSR Plus without 10 μM Y-27632 the next day. Directed differentiation was initiated (D0) when cells reached 60–70% confluency by aspirating mTeSR Plus, washing with DPBS, and replacing media with RBA media: RPMI (Invitrogen, 11875135), 0.5 mg/mL bovine serum albumin (Sigma, A9418), 0.213 mg/mL ascorbic acid (Sigma, A8960) supplemented with 4.5 μM CHIR-99021 (TOCRIS, 4423). After two days (D2), CHIR-99021 containing media was removed and replaced with RBA supplemented with 2 μM Wnt-C59 (Selleck, S7037). On D4, Wnt-C59-containing media was replaced with RBA. On D6 (and every other day afterwards), media was replaced with cardiomyocyte media: RPMI, B27 plus insulin (Invitrogen, 17504044). After D12, cardiomyocyte media was removed and cardiomyocytes were dissociated to single cells with TrypLE Select 10X (Gibco, A12177–01) and replated on Matrigel-coated 6 well plates at 3×10⁶ cells per well in cardiomyocyte media supplemented with 10% FBS and 10 μM Y-27632. At D20, cardiomyocytes were purified by replacing cardiomyocyte media with lactate selection media for four days: DMEM, no glucose (Gibco, 11966–025) with 4mM sodium-L-lactate (Sigma, 71718). Lastly, cardiomyocytes were dissociated to single cells using TrypLE Select 10X on D56 and then mEGFP-positivity was measured by flow cytometry (see end of next section).

METHOD DETAILS

CRISPR/Cas9 knockout and cell-line preparation

For knockout of LMNA/C, U2OS C11 cells¹¹³ were electroporated on the Lonza 4D Nucleofector using protocol CM104 in a single well of a Lonza 4D-Nucleofector SE Kit S (Lonza, V4XC-1032). The electroporation mix contained 1 million cells resuspended in 30 μL of supplemented Lonza SE buffer with 1 μL of SpCas9 (Synthego) and 1 μL of LMNA-KO gRNA (Synthego). Cells were then recovered and seeded into 96-well plates at a dilution of 1 cell per 5 wells. Clones were grown over a month, passaged, and LMNA-knockout was confirmed by Western blot (see below). U2OS C3 LMNA-KO was the clone selected for both replicates of VIS-seq on LMNA variants.

For knockout of PTEN, NWTC11.G3-WT¹¹² iPS cells were electroporated on the Lonza 4D Nucleofector using protocol CM137 in a single well of a Lonza 4D-Nucleofector P3 Kit S (Lonza, V4XP-3032). The electroporation mix contained 200,000 cells per well resuspended in 30 μL of supplemented Lonza P3 buffer with 1 μL of SpCas9 and 1 μL of PTEN-KO gRNA (Synthego). Cells were then recovered and seeded into 96-well plates at a dilution of 1 cell per 5 wells. Clones were grown over 10 days, passaged, and then PTEN-knockout was confirmed by Western blot. PTEN-KO C18 was used for the first replicate of iPS cell and neuron VIS-seq on PTEN variants, and PTEN-KO C2 was used for second replicates.

VIS-seq expression construct cloning

VIS-seq expression constructs for expressing N-terminal mEGFP-tagged prelamin A (PBv2b-laminA) and PTEN (PBv2c-NSapI-PTEN) libraries were prepared by Gibson assembly¹¹⁴ using NEBuilder HiFi DNA Assembly Master Mix (New England Biolabs, E2621L) followed by Golden Gate¹¹⁵ insertion of barcodes and then variant libraries.

First, an intermediate plasmid was cloned which contains Esp3I/BsmBI sites for installation of a padlock-BC sequence in a future step. We cloned this plasmid by digesting pJBL_051²⁵ (gift from the Jay Shendure lab) with NotI-HF (New England Biolabs, R3189S) and SacII (New England Biolabs, R0157S) and performing Gibson assembly with the PD-BsmBIsites gene fragment (Twist, see Table S6 for DNA sequences). Then, the PBv2b expression plasmid was prepared by Gibson assembly of the following five fragments: intermediate backbone from the previous step digested with NheI-HF (New England Biolabs, R3131L) and MluI-HF (New England Biolabs, R3198S), UCOE and EF1a promoter/intron amplified from pHR-UCOE-EF1a-Zim3-dCas9-BFP¹¹⁶ (Addgene, 188775), and gene fragments containing mEGFP, MluI and EcoRI restriction sites, a truncated portion of HBB intron 2, an IRES, a puromycin resistance gene, and WPRE (mEGFP-HBBIVS2trunc-v1 and IRES-PuroR-WPRE-v1, Twist). All PCR amplifications were performed using Q5 Master Mix (New England Biolabs, M0492L), using 98C for the denaturation temperature, 72C for the extension temperature, and 60C for the annealing temperature and were stopped prior to saturation. The PBv2b expression plasmid was subsequently prepped and sequence-confirmed.

PBv2b-NSapI, PBv2b-CSapI, and PBv2c-NSapI expression plasmids were cloned using similar Gibson assemblies. The corresponding gene fragment was replaced with one containing SapI sites either upstream of mEGFP (mEGFP-HBBIVS2trunc-v2C, Twist) for PBv2b-CSapI or downstream of mEGFP (mEGFP-HBBIVS2trunc-v2N, Twist) for PBv2b-NSapI and PBv2c-NSapI. For PBv2c-NSapI, we also substituted a CAG promoter amplified from pCAG-NLS-HA-Bxb1^117,118 for the EF1a promoter.

All VIS-seq backbone plasmids (PBv2b, PBv2b-NSapI, PBv2b-CSapI, and PBv2c-NSapI) were barcoded using Golden Gate assembly. Klenow fill-in (New England Biolabs, M0212L) of VISseq_GG_BColigo (IDT) using primer VS_01 (IDT) yielded a double-stranded oligo with a 12-bp degenerate barcode used to barcode PBv2b, PBv2b-NSapI, or PBv2b-CSapI, and low-cycle PCR of VISseq_GG_BColigoX2 (IDT) with primers VS_02 (IDT) and VS_03 (IDT) yielded a double-stranded oligo with two 8-bp barcodes used to barcode PBv2c-NSapI. The barcoding reaction was performed in eight simultaneous reactions each with 500ng of input plasmid, a 2:1 molar ratio of barcode, 1.5 μL of Esp3I (New England Biolabs, R0734L), 1 μL of T4 ligase (New England Biolabs, M0202M), and 2.5 μL of 10X T4 ligase buffer with water added to 25 μL. These reactions were cycled between 37 °C and 16 °C 45 times, post-digested with 1 μL of BsmBI-v2 (New England Biolabs, R0739S) and then pooled, concentrated, and electroporated into 25 μL of NEB 10-beta electrocompetent E. coli (New England Biolabs, C3020K). Bacteria were plated onto BioAssay dishes (Fisher Scientific, 07–200-600) containing LB + 100 μg/mL carbenicillin (GoldBio, C-103–250) and grown at 30 °C overnight. Colonies were scraped and plasmid DNA was prepped the following day. Three electroporations were used for each plate. Each barcoded VIS-seq expression vector had a minimum of 25 million colonies per plate for 2 plates, which were pooled 1:1.

VIS-seq library cloning

The LMNA (prelamin A sequence, derived from NM_170707.4) variant cDNA library was purchased from Twist as an SSVL library and delivered in an arrayed format; positions 178 to 273 were pooled and restriction cloned into the VIS-seq PBv2b expression plasmid. PTEN (positions 112–172, derived from NM_000314.8), RPS19 (positions 52–109, derived from NM_001022.4), and HIST1H1E (positions 54–113, derived from NM_005321.3) variant cDNA libraries were cloned in-house using BsaI Golden Gate assembly, and then inserted into the VIS-seq expression plasmids with SapI Golden Gate assembly, similar to a previously published method¹¹⁹.

First, the plasmid pHSG299 (Takara Bio, 3299) was modified to contain two BsaI sites. Oligo libraries containing tiles of gene variants were amplified from the purchased Twist oligo pool and then assembled with the flanking gene blocks coding for the remainder of the cDNA (PTEN_tile3_s1/2 or RPS19_s1/2 or H1.4_s1 from Twist) and the modified pHSG299-BsaI. This Golden Gate reaction was performed using 100 ng of pHSG299-BsaI, a 2:1 molar ratio of gene blocks to backbone, a 2:1 molar ratio of the amplified oligo library to backbone, 1 μL of BsaI-HF-v2 (New England Biolabs, R3733L), 0.5 μL of T4 ligase, and 2.5 μL of 10X T4 ligase buffer with water added to 25 μL total volume. For Figure 1C, WT H1.4 cDNA was cloned by substituting the amplified oligo tile with a wildtype gene block (Twist). These reactions were incubated at 37 °C for 18 hours, then concentrated and electroporated into NEB 10-beta electrocompetent E. coli, followed by plasmid prep and sequence-confirmation.

Full cDNA variant libraries were cloned into corresponding VIS-seq expression plasmids. The prelamin A variant library (positions 178 to 273; Twist) and barcoded PBv2b VIS-seq expression plasmid were digested with MluI-HF and EcoRI-HF (New England Biolabs, R3101S) and then ligated at 16 °C overnight using 0.5 μL T4 ligase and T4 ligase buffer at a final 1X concentration in a 20 μL reaction. Other genes were assembled into barcoded PBv2b-NSapI (HIST1H1E), PBv2b-CSapI (RPS19), or PBv2c-NSapI (PTEN) using Golden Gate assembly: 100 ng of barcoded plasmid and a 2:1 molar ratio of variant library plasmid, 1.5 μL of SapI (New England Biolabs, R0569L), 0.5 μL of T4 ligase, and 2.5 μL of 10X T4 ligase buffer in a 25 μL reaction cycled between 16 °C and 37 °C for 3 hours. After assembly, reactions were post-digested with an additional 1 μL of SapI. All reactions were concentrated and then electroporated into NEB 10-beta electrocompetent E. coli followed by growth of dilutions in LB overnight, prep, and sequence confirmation. For all libraries, the dilution that corresponded most closely to ~20-fold barcode coverage over variants was selected the following day for plasmid preparation.

PacBio sequencing

Barcoded PBv2b-Prelamin A and PBv2c-NSapI-PTEN libraries were digested with NotI-HF. Then, the SmrtBell Prep Kit (SmrtBell Prep Kit 3.0, Pacific Biosciences, 102–182-700) instructions were followed to generate PacBio libraries. Libraries were sequenced using the Pacific Biosciences Revio sequencer. Pacybara¹²⁰ was used to process reads to generate variant-barcode association tables (LMNA) or variant-double barcode association tables (PTEN).

Low-MOI piggyBac integration of VIS-seq libraries

For low-MOI integration of the VIS-seq LMNA library, U2OS C3 LMNA-KO cells were electroporated using the Lonza 4D Nucleofector with protocol CM104 in 6 separate wells of a Lonza 4D-Nucleofector SE Kit S. Each cuvette contained 1 million cells per well resuspended in 30 μL of supplemented Lonza SE buffer with 200 ng of VIS-seq LMNA variant library and 50ng of hyPBase¹²¹ expression plasmid (gift from the Jay Shendure lab). Cells were then recovered in a 6-well plate in media (DMEM + 10% FBS + 1% Pen/Strep), using one well per electroporation well, for 4 days and then pooled and integrants were selected with 2ug/mL puromycin (Gibco, A11138–03) for 7 days. After selection, cells were recovered for at least 2 days prior to freezing in media with 5% v/v added DMSO (Sigma, D8418–50 mL). Library-expressing cells were thawed, washed and resuspended in media and recovered for 2 days prior to VIS-seq experiments.

For low-MOI integration of VIS-seq libraries into iPS cells, cells were electroporated in 2 cuvettes of Lonza 4D-Nucleofector P3 Kit X (Lonza, V4XP-3024) using the Lonza 4D Nucleofector with protocol CM104 (U2OS) or CM137 (iPS cells). The PTEN library was integrated into PTEN-null clones of NWTC11.G3-WT¹¹² iPS cells and other libraries were integrated into WTC11 iPS cells. Each cuvette contained 3 million cells resuspended in 100 μL of supplemented Lonza P3 buffer with 400 ng of VIS-seq expression plasmid and 100 ng of hyPBase¹²¹ expression plasmid. Cells were then recovered for 4 days in a 10-cm plate per cuvette and then pooled and integrants were selected with 1ug/mL puromycin for 3 days or until cells were >90% mEGFP expressing. After selection, cells were recovered for at least 2 days in mTeSR Plus and then frozen in liquid nitrogen after resuspension in CryoStor CS10 Cell Freezing Medium (STEMCELL Technologies, 100–1061). Library-expressing cells were thawed, washed and resuspended in mTeSR Plus with 10 μM Y-27632 and recovered for 2 days prior to VIS-seq experiments or NGN2-induced neuronal differentiation.

VIS-seq expression cassette silencing and in situ sequencing compared to lentiviral construct

The lentiviral plasmid pLenti_EF1a_mEGFP-H1.4 was cloned via a Golden Gate assembly using Esp3I of four fragments: backbone derived from PCR of lentiGuide-Puro¹²², EF1a-mEGFP-H1.4 fragment from amplifying the assembled PBv2b-NSapI-H1.4(WT) plasmid, a double stranded 12-bp barcode flanked by padlock sequences generated by Klenow fill-in of Lenti_GG_BColigo (IDT) with primer VS_04 (IDT), and an IRES-Puro-WPRE-SV40 terminator fragment derived from amplifying the assembled PBv2b-NSapI-H1.4(WT) plasmid. The plasmid was subsequently prepped and sequence-confirmed.

To generate lentivirus, HEK293T cells were seeded into a 6 well plate at 900,000 cells per well in the evening. The next morning, TransIT-293 (Mirus, MIR 2700) suspensions were made by adding 0.625 μg pMD2.G (Addgene, 12259), 1.25 μg psPAX2 (Addgene, 12260), and 0.625 μg pLenti_EF1a_mEGFP-H1.4 to 250 μL OPTI-MEM (Gibco, 31985–070) for every well to be transfected. After vortexing, 7.5 μL Transit reagent was added per well and mixed by gently flicking the tube, followed by a 15 minute incubation at room temperature. After the incubation, the suspension was distributed dropwise onto the plated cells and allowed to sit for 24 hours. The following day the media was replaced with fresh DMEM and placed back in the incubator for 48 hours. Following the incubation, the media was removed from the cells and spun at 1,000g for 5 minutes. The supernatant was filtered through a 0.45 μm syringe filter (VWR, 76479–012) and 1 mL aliquots frozen at −80 °C until use.

To generate iPS cells with integrated lentivirus expressing mEGFP-H1.4 and linear RNA barcodes, 10 μL of filtered lentiviral supernatant was added to a 6-well plate of WTC11 iPS cells plated at a concentration of 100,000 per well for two replicates, followed by puromycin selection after 4 days until >90% of cells were GFP-positive. To generate iPS cells with integrated VIS-seq cassette expressing mEGFP-H1.4 and circular RNA barcodes, the protocol for low-MOI integration of the VIS-seq plasmid PBv2b-NSapI-H1.4(WT) was followed (see above) for two replicates.

After selection, both iPS cell populations were passaged and monitored for mEGFP-positivity using the BD FACSSymphony A3 Analyzer over 14 days. Cells were taken to flow cytometry on days 0 (immediately after selection), 4, 9 and 14. mEGFP expression was measured for at least 10,000 cells using the 488 nm laser and the BB515 filter. Nontransduced iPS cells were used as a control to set the mEGFP-positive gate. Day 14 transduced and VIS-seq cassette-expressing iPS cell populations were fixed, in situ libraries were generated, and the first base of in situ sequencing was performed (see below for protocol). Only mEGFP-expressing cells were analyzed and STARCall³⁸ was used to count the number of rolling circle colonies (rolonies) for each condition.

Cardiomyocytes were differentiated from transduced and VIS-seq-expressing iPS cell populations after selection (see above for protocol). On day 56 after start of differentiation, mEGFP-positivity of cardiomyocytes was measured using the BD Symphony S6 Cell Sorter with 488 nm laser and BB515 filter.

Western blots

Approximately 1 million cells per sample were incubated with 50 μL 1x RIPA buffer (Abcam, ab156034) supplemented with Halt protease inhibitor cocktail (ThermoFisher, 1860932) for 30 minutes on ice. Samples were spun down at 10,000g and the supernatant moved to new chilled tubes. The supernatants were assayed by BCA assay (ThermoFisher, Pierce BCA protein assay kit 23225) to determine the total protein concentration. Lysates were normalized and mixed with 2x Laemmeli sample buffer (BioRad, 1610737) supplemented with fresh β-mercaptoethanol (BioRad, 1610710) and boiled for 5 minutes at 95 °C. Boiled samples are spun at 10,000g for 1 minute right before loading 30 μg total protein per well for each sample into 8–16% Criterion TGX Stain-Free Protein Gel (BioRad, 5678103). Gels were run with 10x Tris/Glycine/SDS Running Buffer (BioRad, 1610732) diluted to 1x with type I ultrapure water for 35 minutes at 200V. After running, the gels were imaged for total protein on a BioRad Chemidoc MP. Transfer was done using Trans-Blot Turbo Midi 0.2 μm PVDF Transfer Packs (BioRad, 1704157) on the Trans-Blot Turbo Transfer System using the TURBO settings. Membrane blocking was done in 5% BSA (Sigma-Aldrich, A9418) in Tris buffered saline with Tween 20 (TBS-T; 50mM Tris-HCL, 150mM NaCl, 0.05% Tween-20) for rocking for 1 hour at room temperature. Primary antibody staining was done in TBS-T with 1% BSA rocking overnight at 4°C. Primary antibodies used were anti-phospho-(Thr308)-AKT rabbit mAb (Cell Signaling, 13038), anti-pan-AKT mouse mAb (Cell Signaling,2920), anti-PTEN rabbit mAb (Cell Signaling, 9559), and anti-lamin A/C mouse mAb (BioLegend, 600001) all used at a 1:1000 dilution from stock. Following primary, blots were washed three times for 10 minutes with TBS-T rocking at room temperature. Secondary staining was done in TBS-T with 0.01% SDS and 1% BSA. Secondary antibodies used were StarBright blue 520 goat anti-mouse IgG (BioRad, 12005866) and StarBright blue 700 goat anti-rabbit IgG (BioRad, 12004162) both used at a 1:5000 dilution. hFAB rhodamine anti-actin primary antibody (BioRad, 12004164) was used to stain for β-actin as a loading control in the secondary solution at a dilution of 1:5000. Secondary staining was performed on a rocker for one hour at room temperature followed by three 10 minute washes in TBS-T and developed on the BioRad Chemidoc MP.

VIS-seq molecular biology

In situ sequencing protocols are similar to those published previously³⁵. However, we detail them here, documenting the changes we made. All in situ sequencing experiments were performed using all wells of a glass-bottom #1.5H 6-well plate (Cellvis, P06–1.5H-N) with cells plated at 800,000 cells per well (U2OS) or 600,000 cells per well (iPS cell/NGN2-induced neuron), and unless indicated otherwise wash steps contain 2mL per well.

U2OS and iPS cells were first washed 3 times with DPBS prior to fixation. Then 1 mL of 4% PFA (Electron Microscopy Sciences, 15713-S) in DPBS was added to each well. NGN2-induced neurons were fixed in media, by removing 50% of media (leaving 1 mL remaining in well) and adding 250 μL of 20% PFA dropwise for a final concentration of 4% PFA, followed by swirling gently to mix PFA into media. All cells were incubated during fixation in the dark for 30 minutes at room temperature and then washed 3X with DPBS prior to enzyme steps.

Cells were permeabilized with 1 mL per well of 0.1% Triton-X 100 (Sigma, Y8787–50ML) in DPBS for 15 minutes at room temperature. This was followed by 2 washes of 0.1% Tween-20 (Sigma, P1279–500ML) in DPBS (henceforth called PBS-T). A 5’-amino-linked RT primer containing linked nucleic acid bases was then hybridized by incubation with 600 μL per well of 1 μM CROP_RT2_LNA_5Am^7,9 (IDT) for 15 minutes at room temperature followed by 3 PBS-T washes. Then cells were fixed with 1 mL per well of 3% PFA and 0.1% glutaraldehyde (Electron Microscopy Sciences, 16210) in PBS-T for 30 minutes at room temperature, followed by 3 PBS-T washes. 750 μL per well of reverse transcription reaction mix containing final concentrations of 1x Revertaid RT buffer, 250 μM dNTPs (New England Biolabs, N0447L), 400 μg/mL Ultrapure BSA (Thermo Scientific, AM2616), 0.5 U/μL RiboLock (Thermo Scientific, EO0381), and 4.8 U/μL Revertaid RT enzyme (Thermo Scientific, EP0451) in Nuclease-free water (Thomas Scientific, C001X09) and then added to each well. Water was added to the spaces between wells and the plate was sealed with foil. The plate was incubated at 37 °C while shaking at 80 RPM overnight.

Following reverse transcription, cells were washed three times with PBS-T and then fixed with 1 mL per well of 3% PFA and 0.1% glutaraldehyde in PBS-T for 30 minutes at room temperature. Cells were then washed five times with PBS-T to reduce the dNTP concentration. A previously reported optimized padlock probe⁹ was annealed onto the sample, the gap in the padlock probe was filled in, and the circle was ligated shut in a single reaction. This reaction was performed using 600 μL per well of the gap-fill mix containing final concentrations of 1X Ampligase buffer (Biosearch Technologies, SS000015-D3), 200 μg/mL Ultrapure BSA, 100 nM padlock probe (BESTMIP1 for PTEN, BESTMIP2_3ntdel for LMNA), 0.02 U/μL Taq-IT polymerase (Qiagen, P7620L), 50 nM dNTP, 0.1 U/μL RNaseH (New England Biolabs, M0297L) and 0.25 U/μL Ampligase (Biosearch Technologies, E0001–100D2) in Nuclease-free water. The low dNTP concentration is crucial and dNTP dilutions were re-prepped each experiment. The plate was incubated at 37 °C for 5 minutes and then 45 °C for 90 minutes.

The cells were washed twice with PBS-T and then 600 μL of RCA mix containing final concentrations of 1X Phi29 buffer (Thermo Scientific, EP0094), 250 μM dNTPs, 200 μg/mL Ultrapure BSA, 5% glycerol (Fisher Scientific, G33–1), and 0.25 U/μL Phi29 enzyme (Thermo Scientific, EP0094) in Nuclease-free water was added to each well. The plate was sealed with foil. The plate was incubated at 30 °C while shaking at 80 RPM overnight.

The cells were washed twice with PBS-T for the LMNA experiments and three times for the PTEN experiments and then stained. Different staining mixes were used for U2OS cells, iPS cells, and NGN2-induced neurons. U2OS cells were first stained with 600 μL per well of Mitoprobe¹⁰ staining mix containing final concentrations of 10% formamide (BP227–500), 250 nM 12S and 16S Mitoprobe¹⁰ (IDT), and 2X SSC (Invitrogen, 15557036) in Nuclease-free water. They were incubated at 37 °C for 20 minutes.

Following Mitoprobe staining, Cell Painting¹⁴ was performed on U2OS cells in a manner similar to previously described. Cell Painting buffer was prepared as 1X HBSS (Gibco, 14065–056), 1% BSA and 0.01% sodium azide (Sigma, S-8032). U2OS cells were then washed twice with PBS-T and 600 μL of Cell Painting mix containing 0.2 ng/μL DAPI (Fisher Scientific, D1306), 0.9 μL of WGA-CF555 (Biotium 29076–1), 0.75 μL of Phalloidin-CF568 (Biotium 00044-T), and 0.3 U/μL RiboLock in Cell Painting buffer was added. Dye concentrations were individually titrated per lot. Cells were incubated at room temperature for 30 minutes. Then U2OS cells were washed five times with PBS-T.

iPS cells and NGN2-induced neurons were first stained with 750 μL of a modified Call Painting/immunostaining mix that contained 0.2 ng/μL DAPI, 0.75 μL of Concanavalin A-AF750 (Thermo Scientific, C56127), 0.94 μL of Phalloidin-CF568 (iPS cells only), 0.25% Triton-X 100, 1.5 μL of anti-(Thr308)-pAKT Antibody (CellSignal, 13038T), and 1.5 μL of anti-MAP-2 Alexa594-labeled primary antibody (NGN2-induced neurons only; BioLegend, 801802) in Cell Painting buffer. As with U2OS cells, dye and antibody concentrations were individually titrated per lot. Both primary and secondary staining were incubated at room temperature while shaking at 80 RPM for 1 hour.

After painting and primary staining, iPS cells or NGN2-induced neurons were washed three times with PBS-T. Then they were incubated with 750 μL per well of Goat anti-Rabbit Alexa-647-labeled secondary antibody (Thermo Scientific, A-21244) at a concentration of 0.5 U/μL in Cell Painting buffer with added 0.25% Triton-X 100 for 1 hour while shaking at 80 RPM. Then cells were washed four times with PBS-T. Cells were kept in the dark during and after staining and prior to imaging.

Once the U2OS cells, iPS cells, and NGN2-induced neurons had been stained, 2 mL of Imaging mix containing 2X SSC, 10 mM ascorbic acid (Sigma, A92902–100G) and 0.2 ng/μL DAPI in Nuclease-free water were added to each well prior to imaging.

After phenotype imaging, dyes were removed differently for U2OS cells (LMNA) or iPS cells and NGN2-induced neurons (PTEN). U2OS cells were permeabilized with 2 mL 70% ethanol per well for 15 minutes at room temperature, which was sufficient to remove phalloidin; mitoprobe was removed later during sequencing steps by heating¹⁰. 1 mL PBS-T was added and 2 mL was removed followed by adding and removing 1 mL PBS-T four times to titrate the ethanol concentration below 1%. Then U2OS cells were washed twice with 2 mL PBS-T per well.

iPS cells and NGN2-induced neurons were washed twice with 2 mL PBS-T per well prior to chemical bleaching with 1 mL of 1.74 mg/mL sodium borohydride (Sigma, 452882) in DPBS incubated for 15 minutes at room temperature. The borohydride solution was used immediately after mixing. Then, cells were washed three times with PBS-T. After phenotyping dyes were removed or chemically bleached, we proceeded with in situ sequencing.

We then added 600 μL of 0.5 μM sequencing primer (using previously reported ISS1 (IDT) for PTEN experiments and ISS2 (IDT)⁹ for LMNA experiments) diluted in hybridization buffer of 2X SSC + 10% formamide to each well and incubated at room temperature for 30 minutes. We washed cells twice with PBS-T and then once with MiSeq Incorporation Buffer (IB, Illumina, MS-103–1003). We then proceeded to the base cycles: 600 μL of Illumina MiSeq reagent #1 (Illumina, MS-103–1003) was added to each well and the 6-well plate was incubated at 60 °C for 3 minutes. We then added and removed 1 mL IB four times until the incorporation mix was diluted >50-fold. We washed each well with 5X SSC + 0.5% Tween-20 (a wash-step replacement for Incorporation Buffer) and incubated wash steps at 55C (reduced temperature to prevent amplicon unbinding) for 5 minutes while shaking at 300 RPM. We repeated this wash and incubation four more times for a total of 5 washes to remove background base incorporation. We then performed base (genotyping) imaging (see below) in Imaging mix (2X SSC, 10 mM ascorbic acid, 0.2 ng/μL DAPI in Nuclease-free water).

After imaging was completed, we added 700 μL Illumina MiSeq reagent #4 (Illumina, MS-103–1003) to each well and incubated at 60 °C for 6 minutes, washed once with 2 mL IB per well. We then added 2 mL 5X SSC with 0.5% Tween-20 to each well and incubated at 55C for 1 minute while shaking at 300 RPM. We repeated this wash and incubation two more times for a total of 3 washes. We then washed once with 2 mL IB before adding Illumina MiSeq reagent #1 to incorporate another base. We repeated this process until 8 (PTEN) or 12 (LMNA) bases were imaged.

VIS-seq imaging

All VIS-seq imaging was performed using a Nikon Ti2-E inverted epifluorescence microscope with a motorized stage and motorized objective and filter cube turrets, similar to previously reported configurations for in situ sequencing³⁵. A Hamamatsu ORCA-Fusion scientific CMOS camera was used to collect images. A Lumencor Celesta light engine with 7 laser lines (408, 445, 473, 518, 545, 635, and 750 nm) was used for fluorescence illumination. We also used a Finger Lakes Instrumentation HS-625 high speed emission filter wheel with 5 bandpass filters (499–529, 553–577, 604–644, 662–691, and 698–766 nm), and enabled hardware-level triggering between these filters. The phenotype imaging was performed using a 0.75 NA Plan Apochromat Lambda 20X objective, whereas base cycles were performed using a 0.45 NA Plan Apochromat Lambda 10X objective. Laser power was set to 1% for 405nm laser (DAPI), 20% for other phenotyping lasers, and 100% for in situ sequencing lasers, and exposure time was maintained at 50ms for both phenotyping and in situ sequencing imaging.

Confocal imaging of protein variant libraries in iPS cells and NGN2-induced neurons

Imaging of PTEN, HIST1H1E, and RPS19 libraries in iPS cells or NGN2-induced neurons for Figure S1C (top level) was performed using a Nikon A1R HD25 laser scanning confocal microscope. 405, 488, and 561 nm Nikon LU-N4 lasers were used to image DAPI-stained, mEGFP-tagged protein libraries, Phalloidin-CF568-stained cells, respectively. Imaging was performed using an Nikon Apochromat Lambda S LWD 40x water-immersion objective (NA 1.15). A Nikon galvano scanner and Nikon A1-DUG-2 GaAsP photomultiplier tubes were used to acquire 1024×1024 12-bit color images of single z-planes.

Stitching and alignment of VIS-seq images

Stitching, alignment, and read calling were performed using STARCall³⁸. Stitching and alignment is required for processing of VIS-seq data due to the small (<3 pixel) features being detected and the need for them to align to one another across all imaging cycles. STARCall stitches the raw microscope images across positions and cycles simultaneously using an extended MIST algorithm¹²³ incorporating improvements from ASHLAR¹²⁴.

Using stage positional information, a graph was constructed representing images as nodes with edges connecting every overlapping image pair. This graph contained all imaging cycles, with edges between the same tile in one cycle to the respective tile in all other cycles. For every edge on this graph, both inter-cycle and intra-cycle, we performed the phase cross-correlation algorithm¹²⁵ to find the alignment that maximizes the correlation between the two images. The alignment for each edge was scored by calculating the zero-normalized cross-correlation (ZNCC) on the overlapping region of the two images. The ZNCC was computed by normalizing the overlapping pixel intensities to unit vectors with mean zero and computing their dot product. To remove erroneous offsets, we calculated a lower-bound for the ZNCC scores of offsets, using the 95th percentile of ZNCC scores calculated on a random selection of non-overlapping images. Any pair of images that had a ZNCC score below this threshold was removed from the graph. Removing these pairs caused some portions of the graph to become disconnected, especially if cells in some regions of the well were sparsely plated thus yielding few features for alignment. To reconnect these portions, we trained a linear model on the remaining offsets, predicting their offset given their stage positions. This model was then used to replace edges on disconnected images.

We then constructed a linear system of equations on the global (x,y) positions of the images, where each offset in the graph creates two equations, relating the difference in position of the two images to the offset. This system was then solved using least squares, finding the optimal global position for each image. In particular, the system of equations was solved minimizing the mean absolute error (MAE), the solution for which can be found quickly with linear programming, with an additional benefit being that the influence of outliers is limited when compared to minimizing the mean squared error. All images were then combined to form composite images for each cycle, with overlapping regions between images combined using a weighted average. The code for STARCall and instructions on how to use it can be found on GitHub (see Data and code availability section).

Read calling of genotyping images

We defined the base reads as bright, high frequency features present in only one of the channels and changing frequently across bases, best characterized by a small point spread function. In addition, we characterized common image features that should be filtered out: low-intensity cell background, diffuse features that are present in most channels and increase monotonically across cycles, and cell debris, which present as extremely bright, high frequency features present in all channels and not fluctuating between cycles. With the objects of interest characterized, we developed a procedure for selecting only the fluorescent reads similar to previous approaches used for in situ sequencing datasets⁷: First we subtracted a small Gaussian blur (σ=4), which left only small, high frequency features, filtering out general and cell background. We then z-score normalized the resulting image, which corrected for differences between channels in intensity. Next, we subtracted the second-maximal channel on a per-pixel basis, which did not affect the fluorescent dots present in any one channel but greatly reduced the intensity of cell background and cell debris present in all channels. After these filtering steps we applied a small (σ=1) Gaussian blur to smooth out noise we introduced, then clipped any negative values to zero. Finally, we calculated the standard deviation on a per-pixel level across imaging cycles, which resulted in higher values for features that change between bases, filtering out background and debris that doesn’t change between cycles. After summing across channels we obtained a grayscale image containing the features we wanted to extract and believed to only be fluorescent reads. The Laplacian of Gaussian blob detection algorithm was applied to find these features at the expected size (1≤σ≤3). The positions of the blobs were used across all cycles to extract the base reads, which are defined using the highest channel intensity in each cycle for that position. We extracted these values from the images after we subtracted the small Gaussian blur and z-score normalized them, additionally applying a dilation morphological filter with a footprint of 2 pixels to ensure the maximum value of each dot is sampled. We associated barcodes (LMNA) or pairs of barcodes (PTEN) to the nearest genotype, making no call if two genotypes are equally close by edit distance. Later, we filtered cells on their least edit distance barcode.

Segmentation of cells and nuclei

Cell segmentation was performed using Cellpose version 2.2.1³⁹ for cell border segmentation and StarDist version 0.8.5⁴⁰ for nuclear segmentation. CellPose requires the estimated diameter of cells to be specified to rescale images before inference. The method for cell segmentation may be adjusted depending on the cell type and imaging setup. We used diameter values of d=50 for U2OS cells and d=35 for iPS cells and NGN2-induced neurons. For U2OS cells we provided DAPI as the nuclear channel and wheat germ agglutinin (WGA) and phalloidin as the cytoplasmic channel to CellPose. For iPS cells we provided DAPI as the nuclear channel and WGA as the cytoplasmic channel. Due to difficulties segmenting NGN2-induced neurons, we used only nuclear segmentation, converting it into an estimated cell segmentation by expanding the nuclear labels out to neighboring cells, with a maximum expansion distance of 25 pixels.

Intensity normalization and feature extraction

We performed intensity normalization for each well’s phenotype imaging separately to adjust for uneven illumination of the field of view by subtracting the 5th percentile value of that pixel across all such fields. After stitching and registering phenotype images with genotypes, we then partitioned each well’s phenotype image into a 20-by-20 grid and passed images, cell segmentations from Cellpose³⁹, and nuclear segmentations from StarDist⁴⁰ into CellProfiler version 4.2.6⁴¹ for all grid positions that have genotyped cells. We then merged all of these grid positions to generate a well-level cells-by-features matrix. In the case of the PTEN VIS-seq experiments, we adjusted the pAKT channel (stained with AF647) for bleed-through from the ConA channel (stained with AF750) by subtraction. The CellProfiler⁴¹ pipelines for feature extraction used in PTEN and LMNA VIS-seq experiments are available in the STARCall GitHub (see Data and code availability section).

Generation of variant-level morphological profiles

We filtered out cells whose consensus barcodes were >1 (LMNA) or >2 (PTEN) edit distance from a barcode in the barcode-variant lookup table, had no variant called, contained barcodes (LMNA) or combinations of 2 barcodes (PTEN) that were present in <10 cells, or contained variants with <5 barcodes. We then extracted variant-level feature median and earth-mover distance (EMD) values^57,58 from WT for all variants and features. We filtered out feature EMDs which were non-reproducible in a manner similar to Pearson et al⁵⁸. Briefly, we randomly partitioned WT cells into two batches and computed scores equal to EMD values between the batches for each feature normalized to median absolute deviation. After repeating this 25 times, we removed features above an average score threshold of 1.5 times the interquartile range plus the lower quartile average score.

We then used Pycytominer version 1.2.1⁵⁶ to perform variant-level median aggregation and variant-level median/EMD normalization and selection, for generating morphological profiles. This normalization was a z-scoring for each feature median using synonymous feature means and standard deviation in the LMNA VIS-seq experiment, and a z-scoring over all variants for the PTEN VIS-seq experiments. We first removed non-biologically-relevant features like orientational or positional features, yielding 1077 features for LMNA and 3222 for PTEN. We ran pycytominer⁵⁶ feature_select() using both “correlation threshold” and “variance threshold” to generate the variant-level morphological profiles.

Clinical variant curation

All clinical variant analysis excluded variants with predicted splicing impact defined by SpliceAI¹²⁶ scores>0.2, since these assays do not detect splicing effects. LMNA ClinVar⁵⁹ variants were accessed on 6/18/2024 and LP/P variants were curated. Conflicting variants were labeled as variants of uncertain significance (VUS). PTEN ClinVar variants were accessed on 12/19/2024; conflicting variants were treated as their most recent classification to generate a putative list of profiled VUS, LB/B and LP/P variants.

PTEN missense variants on gnomAD v4.1⁸⁸ were accessed on 5/14/2025 and used as a control set. PTEN missense variants present in 10 or more patient cancers on the Catalogue Of Somatic Mutations In Cancer (COSMIC)⁸⁷, accessed on 12/12/2024, were used to define somatic PTEN cancer-associated variants.

For curation of PTEN clinical variant-phenotype associations for PHTS and ASD/DD, 170 publications containing a total of 541 probands with PTEN missense variants were identified using HGMD and ClinVar (accessed 5/13/2025). The clinical features for each individual with a PTEN missense variant were curated manually by a molecular genetic pathologist blinded to the functional assay results (AEM) and included demographics, clinical diagnosis (Cowden Syndrome (including Cowden-like Syndrome), Bannayan-Riley-Ruvalcaba syndrome, or PTEN hamartoma tumor syndrome, and the presence of autism spectrum disorder (ASD) or developmental delay/intellectual disability (DD/ID)). For the analysis, probands reported to have ASD, DD, or ID were considered to be affected by the neurodevelopmental phenotype. These phenotypes were combined due to inconsistencies in reporting practices. A large language model (Claude Sonnet 4, Anthropic) was used to assemble the data table (Table S5), with the output verified manually. Any variant linked to a clinical phenotype in at least one affected proband was then associated with that clinical phenotype. See Figure 6A for the number of curated variants in each clinical category.

Variant effect predictor curation and comparison

We curated scores from three leading in silico variant effect predictors: AlphaMissense⁹¹, EVE⁹², and REVEL⁹³. We compared these predictor scores to VIS-seq readouts by plotting VIS-seq-derived impact or distinguishability scores (or landmark feature z-scores; x-axis) against predictor scores (y-axis), and by visualizing VIS-seq-derived UMAP embeddings colored by predictor scores (Figure S10). Spearman correlation coefficients between predictor scores and VIS-seq readouts are shown. ClinVar variants are shown on the plot for comparison.

QUANTIFICATION AND STATISTICAL ANALYSIS

Impact scoring and clustering of variant-level morphological profiles

We performed principal component analysis (PCA) for dimensionality reduction and uniform manifold approximation and projection (UMAP)¹²⁷ for visualization of morphological profiles. For PCA, we used the minimum number of PCA dimensions that could explain 60% (PTEN) or 70% (LMNA) of the dataset variance. For PTEN NGN2-induced neuron profiles, we noticed that three variants with poor inter-replicate correlation were outlying in PC1, and they were removed prior to rerunning PCA and further analysis. We then performed UMAP on PCA-reduced variant profiles using cosine similarity to visualize the local structure of the phenotype manifold.

To generate morphological impact scores, we computed the cosine similarity between a variant’s morphological profile and the median synonymous variant morphological profile, computed by taking the median over each selected feature. We then defined the morphological impact score as the cosine distance, defined as ½ × (1 minus the cosine similarity). Morphological impact scores were compared between groups of variants using boxplots with the median and inter-quartile range (IQR) shown, with n shown below the label indicating the number of variants in each group, and Mann-Whitney U p-value cutoffs reported (Figures 2F, 6C, S5C, S5D, S5I, S6J, S9B, and S9D).

To perform Louvain clustering¹²⁸ on the LMNA profiles, we generated an unweighted k-nearest-neighbor graph using k=25 of the PCA-reduced variant profiles using cosine distance. Then, we used the Louvain community detection algorithm to partition variants into clusters. To demonstrate separation by profiles, we plotted variants organized by cluster against selected features used to generate profiles (Figure S3A). We plotted cosine similarities between pairs of variants within a cluster and between different clusters to demonstrate that clusters were distinct, with Mann-Whitney U p-value cutoffs reported (Figure S3B). To demonstrate cluster robustness we randomly removed 10% of variants, performed clustering 100 times, and plotted the fraction of cluster co-occurrence (Figure S3C). For each cluster, the z-scored ratio between lamin A and DAPI intensity across radial deciles is shown with Mann-Whitney U p-value cutoffs reported (Figures 3C, S3D). Lastly, the distribution of variants across clusters by subdomain, position group, and amino acid substitution are shown with Fisher’s exact test p-value cutoffs reported (Figures S5B, S5J, and S5I).

Training of variant-specific classifiers and distinguishability scoring

To characterize the morphological distinguishability of each variant from WT using single-cell image data, we trained a collection of binary classifiers, each tasked with distinguishing a single variant from the wild type (Figure S2F). The input to each variant-specific model was the set of CellProfiler-derived⁴¹ features (1,077 for LMNA, 3,222 for PTEN iPS cells, and 3,227 for PTEN neurons), extracted from wild type cells and the cells corresponding to that variant. To avoid any data leakage, no feature selection was performed. Features were not z-scored. The output was a binary label indicating whether each cell is wild type or variant. In order to prepare the training data, we first removed any variants that had fewer than 150 cells between both replicates, after which we divided the filtered training data into a 8:1:1 train/validation/test split, stratified by variant type. For each variant, we used the train and validation sets to train a single binary classifier using the Extreme Gradient Boosting (XGBoost) algorithm¹²⁹, where the loss on the validation set was used to determine when to stop training. Finally, after training variant-specific classifiers, we defined each variant’s distinguishability score as the area under the receiver operating characteristic curve (AUROC) on the held-out test set (Tables S2 and S4). This test set was not used during training, providing an unbiased estimate of how well each model generalizes to unseen data. Distinguishability scores were compared between groups of variants with n reporting the number of variants in each group and Mann-Whitney U p-value cutoffs reported (Figures 2G, S5I, S6K, and S9E).

For PTEN variants, the distinguishability score difference was defined as the percentile difference in distinguishability scores between iPS cells and NGN2-induced neurons (Figure S8J). iPS cell or NGN2-induced neuron features were ranked by Spearman correlation with the distinguishability score difference. Spearman correlation values were converted to Student’s t-test statistics and then Bonferroni-corrected p-values were reported (Figure S8K).

Volcano analysis of features

To generate variant-level feature p-values for each replicate, we performed a Komolgorov-Smirnov (KS) test between each variant and WT for each feature to generate a matrix of p-values with dimensions of (number of variants: 1,767 for LMNA and 1,228 for PTEN) by (number of features: 1,077 for LMNA, 3,222 for PTEN iPS cells, and 3,227 for PTEN neurons). Distributions of feature values over single cells were tested. See Tables S2 and S4 for the number of cells in each replicate for each variant. P-values from each replicate were combined using Fisher’s method. Next, we generated a volcano plot in which each feature was summarized by the median z-score effect across all variants (x-axis) and the geometric mean, across variants, of the KS-test p-values for that feature (y-axis). This analysis was repeated for both feature medians and feature EMDs. See Tables S1 and S3 for the geometric mean p-values and median effects on each feature. We then called features as hits if they met both criteria: Bonferroni-corrected p-value < 0.01 and |median z-score| > 0.25 for LMNA (or > 1.0 for PTEN) (Figures 2D, S2E, S6G, and S6I). For both LMNA and PTEN, landmark features were interpretable hit features that were highly significant and strongly perturbed. For curated variant-specific volcano plots (Figure 6F), we summarized each feature by the median z-score effect across variants in the curated group (x-axis) and by the geometric mean of the corresponding p-values across variants in that same group (y-axis).

Structural analysis

We used the lamin A structures of A11 tetramer 6JLB⁶³ and recently published A22 four-coil interaction⁶⁴, and PTEN structure 1D5R⁷¹ to perform structural analysis of morphological profiles. Structures were visualized and residues colored with position-averaged feature z-scores using PyMOL v3.1. Lamin A rod subdomains were defined using annotations in Ahn et al.⁶³ and PTEN P-loop and TI-loop annotations were defined using annotations in Lee et al.⁷¹

To call lamin A dimer-facing and multimer-facing positions, we plotted the inter-chain minimum β-carbon to β-carbon distance for each lamin A position (except glycines) on a given chain. We identified residues with inter-chain local minima distances, and any residues within 1.5 Angstroms of this inter-chain local minimum distance, as interacting with an opposing chain (Figure S6H). Interactions were labeled as dimer-facing or multimer-facing depending on the geometry of the chains: coil 1B forms the A11 lateral structure and therefore contains both types of interacting residues whereas coil 2A forms the A22 four-helix bundle and therefore only contains multimer-facing residues.

Groups of LMNA positions defined by clustering (Figures S5E and S5F) were compared by lamin A subdomain (Figures 4D and 4E) or by participation in dimer- or multimer- contacts (Figure 4F) with n equal to the number of such positions shown on the y-axis and χ² goodness-of-fit p-value cutoffs reported.

Mislocalization scoring

We noticed that PTEN variant localization, as measured by DAPI-PTEN correlation, was related to variant abundance as measured by mEGFP-PTEN intensity and lipid phosphatase activity as measured by pAKT intensity (Figure S8D). To isolate the effect of variants on localization, we computed a mislocalization score for each variant. For each variant, we fit a kernel ridge regression model with a radial basis function kernel (α = 1.0, γ = 1.0) to predict DAPI–PTEN correlation from abundance and activity features. The residual from this model, defined as the observed DAPI–PTEN correlation minus the model-predicted value, was used as the mislocalization score. Positive residuals indicate greater nuclear localization than expected given abundance and activity, whereas negative residuals indicate less nuclear localization than expected. iPS cell mislocalization scores are reported in this publication (Figure S8E). Mislocalization scores are shown using boxplots illustrating median and IQR for curated clinical PTEN variant categories with n indicating the number of variants and Mann-Whitney U p-value cutoffs reported (Figure 6E).

PTEN VIS-seq score or feature effects for curated clinical variants

For comparisons of VIS-seq-derived scores or features between curated PTEN variant categories, gnomAD⁸⁸ variants that were not ClinVar LP or P were used as a control set. Scores or features are shown using boxplots illustrating median and IQR with n indicating the number of variants in each clinical category and Mann-Whitney U p-value cutoffs reported (Figures 6C, 6E, and S9D–S9G).

Assay or predictor performance on curated clinical variants

For LMNA, we used VIS-seq-derived variant-level impact scores (computed from feature medians, feature EMDs, or both) and distinguishability scores for zero-shot binary classification of either ClinVar pathogenicity or aggregation⁴² versus synonymous controls (Figure S2J). High-frequency LMNA variants in gnomAD are often pathogenic and are not an appropriate control set. Since these models are zero-shot, no training was performed and the test set is the entire dataset of variants. Because protein-level variant effect predictors do not provide meaningful scores for synonymous variants, they were also excluded from this analysis.

For PTEN, we used VIS-seq-derived variant-level distinguishability scores (from iPS cells and NGN2-induced neurons), together with yeast fitness⁸⁴ scores, VAMP-seq⁵ scores, and in silico predictors (AlphaMissense⁹¹ and EVE⁹²), for zero-shot binary classification of ClinVar pathogenicity or curated disease-associated variant classes versus ClinVar likely benign or gnomAD variants. gnomAD variants that were not ClinVar LP or P or present in the clinical variant-phenotype associations we curated were used as a control set. REVEL⁹³ was trained and validated on ClinVar and therefore was excluded from this analysis. Since these models are zero-shot, no training was performed and the test set is the entire dataset of variants. ROC curves and corresponding AUC values for these zero-shot models are shown in Figures 6B and S9C.

For three-class prediction among (i) gnomAD controls, (ii) PHTS variants not associated with ASD/DD, and (iii) ASD/DD variants not associated with PHTS, we first excluded variants not belonging to exactly one of these classes. Then, we trained linear support vector classifiers (SVC) using either all three VIS-seq landmark features from iPS cells or NGN2-induced neurons, or one of the one-dimensional variant-level scores used in the binary analyses mentioned above. Classes were reweighted to address class imbalance. We used stratified five-fold cross-validation for generating the out-of-fold probabilities used for ROC and AUC. For each model, multiclass ROC curves were computed by macro-averaging one-vs-rest sensitivity and specificity across the three classes, weighing each class equally (Figure 6G).

Supplementary Material

1

Figure S1. piggyBac MOI titration and knockout Western blots, related to Figure 1

(A) Fraction of cells with 0, 1, 2, or 3 integrations, determined by in situ sequencing, after co-transfection into U2OS cells of LMNA VIS-seq library DNA at different quantities (in nanograms) with plasmid encoding Piggybac-ase (at a 4-fold lower mass) and puromycin selection.

(B) Imaging of day-7 neuron-like cells derived from clonal NGN2-inducible PTEN knockout lines stained for DAPI (blue) and NCAM1 (red). Scale bar indicates 20 μm.

(C) Unmatched images of mEGFP-tagged H1.4, RPS19, and PTEN libraries in human WTC11 iPS cells or PTEN library in NGN2-induced neuron-like cells (right), stained with DAPI and phalloidin-CF568 (top row) and first base in situ sequencing (bottom row). Some cells express variants with localization defects (HIST1H1E=chromatin binding, RPS19=nucleolar-cytoplasmic, PTEN=nucleo-cytoplasmic). Nucleobase coloring identical to Figure 1E. Scale bar indicates 20 μm.

(D) Western blot of U2OS cells showing parental line (left) and LMNA knockout line clone 3 (right) stained for lamin A protein. Two replicate lanes are shown. Clone 3 was used in subsequent LMNA VIS-seq experiments.

(E) Western blot of NGN2-inducible iPS cells showing parental line (left) and PTEN knockout clonal lines (right) stained for PTEN and phospho-AKT protein. Clonal lines C2 and C18 were used for PTEN VIS-seq replicate 1 and 2, respectively.

2

Figure S2. LMNA VIS-seq replication, scoring, and feature analysis, related to Figure 2

(A) 12-basepair barcode sequences on circular RNAs were read by in situ sequencing-by-synthesis in each cell and then mapped to corresponding LMNA variants using a barcode-to-variant dictionary made by long-read sequencing. Example cells with reads in all 12 cycles are shown.

(B) Histogram of the total number of 12-base pair reads per sequenced cell in a single well of LMNA replicate 2 experiment.

(C) Histogram of the edit distance between consensus cell-level 12-base pair read and nearest library barcode in a single well of LMNA replicate 2 experiment. Red line indicates that cells with edit distance of 0 or 1 were used if they matched to a unique barcode.

(D) Number of cells genotyped for each LMNA variant in both replicates of VIS-seq screen colored by variant type, with Pearson’s r shown.

(E) Volcano plots of median LMNA variant effect on feature median z-scores (x-axis, left) or feature EMD z-scores (x-axis, right) against geometric-mean KS-test p-values (y-axis, tested against WT feature distributions). Both the feature effect size (x-axis) and p-values (y-axis) are computed over all profiled variants. Points are features with area proportional to the number of variants that pass thresholds and coloring is by imaging channel (top) or cellular compartment (bottom) (see right for legend). Red dashed lines show the Bonferroni-corrected p<0.01 threshold and effect thresholds used to define hit features.

(F) Flowchart describing training a binary classifier for each variant to distinguish images of cells expressing that variant from corresponding WT cell images. For each variant, a distinguishability score summarizing classifier performance was computed as the area under a ROC curve on a test set of single cells. 0.5 indicates random classifier performance and 1 indicates perfect discrimination between variant and WT single cells. For a full description, see Methods.

(G) Morphological impact score of variants in both replicates of the LMNA VIS-seq experiment colored by variant type as in (D), with Pearson’s r shown.

(H) Distinguishability score of variants in both replicates of the LMNA VIS-seq experiment colored by variant type as in (D), with Pearson’s r shown.

(I) Boxplots of morphological impact scores computed from only median, only EMD, or median and EMD-based variant profiles for the same groups of variants as Figure 2F. Points represent individual variant impact scores.

(J) Receiver operating characteristic (ROC) curves are plotted for univariate zero-shot models predicting LMNA aggregating control variants (left) or ClinVar P/LP variants (right) from synonymous variants, using impact scores shown in (I) or distinguishability scores. AUC is shown on the right for each model.

3

Figure S3. LMNA variant profile clustering is robust and separates variants by selected and landmark features, related to Figure 3

(A) Heatmap of LMNA variants organized their Louvain-derived cluster (y-axis, colored according to Figure 3A) versus selected feature medians (left portion of heatmap, x-axis) or selected feature EMDs (right portion of heatmap, x-axis) organized by hierarchical clustering. Each feature medians and EMDs is z-scored using the synonymous distribution of the feature. Imaging channel and compartment are annotated for each feature on the top, with clustering dendrogram shown (see legend on the right). Landmark feature medians, EMDs, and impact and distinguishability scores are shown separately on the right. Nuclear mEGFP-lamin A intensity and granularity features, nuclear shape features, and nuclear correlation features between DAPI and mEGFP-lamin A channels are annotated below.

(B) Boxplot depicting cosine similarity computed for pairs of variants using either selected feature medians or EMDs. Pairs of variants were chosen in different ways to assess the robustness of the Louvain-derived clusters (see Figure 3A). To assess the baseline cosine similarity between the morphological profiles of variants, pairs were chosen from all variants (“global”, left column, grey boxes). To assess the similarity between variants in each cluster, pairs were drawn from the cluster (“within”, right columns, blue boxes). To assess the similarity between variants within a cluster compared to variants outside the cluster, one variant was drawn from the cluster and one was drawn from outside the cluster (“between”, right columns, orange). Each boxplot represents the distribution of cosine similarities computed for all pairs of the indicated type. Louvain cluster colors are shown above. *** indicates Mann-Whitney U FDR-corrected q<0.001.

(C) To assess the stability of our Louvain clustering (see Figure 3A), the clustering was performed 100 times, each time omitting a randomly selected 10% of variants. The fraction of samples in which each pair of variants was located in their original Louvain cluster is shown. Variants are organized by their original Louvain cluster labels shown on the left.

(D) UpSet plot showing the number of missense LMNA variants that impact each combination of landmark features defined in Figure 3B. Features were considered impacted if EMD z-score>2.5 and KS-test Bonferroni-corrected p<0.01). Sets, represented by bars and dots in the plot, are colored by the most frequently impacted feature in the set (no impacted features=black, circularity=orange, granularity=purple, intensity/boundary intensity=teal).

4

Figure S4. LMNA VIS-seq landmark feature analysis, related to Figure 3

(A) Randomly-selected cells from the 10th and 90th percentile in the landmark feature distributions are shown. mEGFP-tagged lamin A channel is shown in green, and DAPI in blue. Silver dashed lines depict the borders of cells. Scale bar indicates 5 μm.

(B) Plot showing variant effects on nuclear mEGFP-lamin A granularity 1, 2 and 3 features by Louvain cluster of residence. Z-scores are versus the synonymous distribution. Synonymous variants are shown on the left. Granularity scale by pixel size is shown in the legend.

(C) LMNA feature median z-scores for mEGFP-lamin A boundary intensity, nuclear eccentricity, nuclear solidity, and nuclear eccentricity are plotted against each other for all profiled variants. Z-scores are versus the synonymous variant distribution. Variant type is colored according to: synonymous variants (green), missense (grey), and frameshift variants (purple).

(D) Plot over nuclear radial deciles of the mEGFP-lamin A total intensity in each decile normalized to the DAPI total intensity in that decile. Bin 1 corresponds to the innermost radial decile and bin 10 to the outermost. The data for the remainder of the clusters and a diagram of nuclear radial deciles are shown in Figure 3C.

(E) Map of lamin A missense variant effects on the lamin A nuclear intensity feature (left; grey boxes=missing variants, black dots=synonymous substitutions) and on the UMAP visualization of variant profiles (right, see Figure 2H). Blue to red coloring indicates feature z-score versus the synonymous variant distribution. Above the map, positions are annotated by their participation in multimer contacts (first row, defined below), α-helical average substitution profile (second row, defined below and in Figure S5F), and proline-substitution profile (third row, defined below and in Figure S5E). Linker or coil subdomains are shown in the thicker annotation bar directly above the heatmap⁶³. Shown below the heatmap are position-averaged scores, and on the right are amino acid substitution marginal feature z-score distributions (points depict variants; black lines depict medians).

(F) Heatmap and UMAP for missense substitution effects on the lamin A nuclear boundary intensity feature, colored and annotated as in (E).

(G) Heatmap and UMAP for missense substitution effects on the lamin A nuclear granularity 1 feature, colored and annotated as in (E).

5

Figure S5. VIS-seq profiles separate lamin A residues by structural and functional properties, related to Figure 4

(A) UMAP visualization of LMNA synonymous (dark green) or missense variant profiles colored by lamin A subdomain.

(B) Heatmap of proportion of missense variants in coil or linker 12 subdomains present in each Louvain cluster (see Figure 3A). Proportion is indicated inside each box. Color scale for heatmap shown on the right. *** indicates Fisher’s exact test FDR-corrected q<0.001 and ** indicates q<0.01.

(C) Morphological impact score of missense substitutions plotted by domain, compared with synonymous variants (green). *** indicates Mann-Whitney U-test p<0.001.

(D) Morphological impact score of lamin A missense substitutions plotted by amino acid, compared with synonymous variants (green). *** indicates Mann-Whitney U-test p<0.001.

(E) Pearson’s correlation coefficients between LMNA proline substitution morphological profiles in PCA space (light color=positive correlation, dark=negative correlation). Bars on top indicate each position’s subdomain and participation in multimer contacts (see (G), Methods, color legend shown above and (E)). Bars on right indicate dendrogram-derived grouping into three clusters: linker, linker-proximal and α-helical.

(F) Pearson’s correlation coefficients between the α-helical (defined in (D)) PCA space, position-averaged variant morphological profiles (light color=positive correlation, dark=negative correlation). Synonymous variants were averaged into a single PCA vector and included (green annotation, top). Bars on top indicate each position’s subdomain⁶³ (color legend above and (D)) and participation in multimer contacts^63,64 (see (G), Methods). Bars on right indicate dendrogram-derived grouping into three clusters: strongly aggregating, low abundance, and low impact.

(G) Contour plots indicate location on the UMAP shown in (Figure 2H) for missense substitutions at each group of positions defined in (E), labeled according to their effects on lamin A.

(H) Intra-dimer (left) and inter-dimer coil 1B (middle) and coil 2A (right) minimum β-carbon distances derived from A11⁶³ and A22⁶⁴ multimer structures. Dimer-facing residues are indicated in blue (left) and multimer-facing residues are indicated in red for A11⁶³ and A22⁶⁴ structures (middle, right, respectively). See Methods for how these residues are defined.

(I) Morphological impact score (left) or distinguishability score (right) of missense substitutions at α-helical position groups as defined in (E) by clustering amino acid positions. *** indicates Mann-Whitney U-test p<0.001.

(J) Heatmap of proportion of missense variants in each position cluster from (E) and (F) present in each Louvain cluster (see Figure 3A). Proportion is indicated inside each box. Color scale for heatmap shown on the right. *** indicates Fisher’s exact test FDR-corrected q<0.001, ** indicates q<0.01, and * indicates q<0.05.

(K) Heatmap of average effects on scores and landmark features of substitutions to amino acids in proline substitution profile-defined position groups (see (E)) and α-helical position groups (see (F)). Lprox indicates linker-proximal, αH indicates α-helical, Agg indicates aggregation-sensitive, Abund indicates abundance-sensitive, and LI indicates low impact position groups. Color scales for each heatmap shown on the bottom.

(L) Heatmap of proportion of missense variants to groups of amino acids present in each Louvain cluster (see Figure 3A). Proportion is indicated inside each box. Color scale for heatmap shown on the right. *** indicates Fisher’s exact test FDR-corrected q<0.001, ** indicates q<0.01, and * indicates q<0.05.

6

Figure S6. PTEN VIS-seq replication, scoring, and feature analysis in iPS cells and NGN2-induced neurons, related to Figure 5

(A) Histogram of total read count per genotyped cell in single well of PTEN iPS cell (blue) or NGN2-induced neuron-like cell (green) replicate 2 experiments. Each genotyped cell has two 8-bp barcodes.

(B) Total (summed) edit distance between consensus cell-level 16-base pair double barcode reads and nearest library double barcode in a single well of PTEN iPS cell (blue) or NGN2-induced neuron (green) replicate 2 experiments. Red line indicates the edit distance ≤ 2 filtering threshold we used for matching of sequenced barcodes to our barcode library.

(C) Total edit distance between consensus cell-level 16-base pair double barcode reads and nearest library double barcode (x) or second nearest library double barcode (y) in a single well of PTEN iPS cell replicate 2 experiment. Red line indicates that cells with total edit distance < 3 were used if they matched to a unique library double barcode.

(D) Boxplots of the number of profiled single cells per variant of the indicated class for iPS cells (left) or NGN2-induced neurons (right) containing single PTEN variants over both replicates.

(E) Number of cells genotyped for each PTEN variant in iPS cell (x) and NGN2-induced neuron (y) of independently transfected and differentiated replicate 1 (left) and replicate 2 (right) of VIS-seq experiment, with Pearson’s r shown.

(F) Stacked barplot showing proportion of all features measured in iPS cells, selected features which generate variant profiles, and hit features (defined in (G)) by imaging channel, cellular compartment, or method of computation of feature. The number of features in each set is indicated. Color legends are shown below.

(G) Volcano plots of median variant effect on iPS cell feature median z-scores (x-axis, left) or feature EMD z-scores (x-axis, right) against geometric-mean KS-test p-values (y-axis). Both the feature effect size (x-axis) and p-values (y-axis) are computed over all profiled variants. Points are features with area proportional to the number of variants that pass thresholds and coloring is by imaging channel (top) or cellular compartment (bottom) (see (F) for legend). Red dashed lines show the Bonferroni-corrected p<0.01 threshold and effect thresholds. Hit features are defined by median effect thresholds (left).

(H) Stacked barplot showing proportion of all features measured in NGN2-induced neurons as in (F).

(I) Volcano plots of NGN2-induced neuron features as in (G). See (H) for legend.

(J) Morphological impact score for PTEN variants in iPS cells (left) and NGN2-induced neurons (right), plotted by variant type. *** indicates Mann-Whitney p<0.001.

(K) Distinguishability scores for PTEN variants in iPS cells (left) or NGN2-induced neurons (right), plotted by variant type. Shown comparisons (top) have Mann-Whitney p<0.001.

(L) Morphological impact scores for variants in each replicate of PTEN VIS-seq experiment in iPS cell (left) and NGN2-induced neurons (right), colored by variant type as in (E), with Pearson’s r shown.

(M) Morphological impact scores (top) and distinguishability scores (bottom) for PTEN variants in both iPS cells and NGN2-induced neurons are compared, colored by variant type as in (B). Pearson’s r shown.

7

Figure S7. Landmark features in VIS-seq PTEN experiments, related to Figure 5

(A,B,C) Missense variant effect maps for NGN2-induced neuron-like cell PTEN landmark features (grey boxes=missing variants, black dots=synonymous substitutions). Blue to red coloring indicates feature z-score versus the synonymous variant distribution. Linker or coil subdomains are shown in the thicker annotation bar directly above the heatmap. Shown below the heatmap are position-averaged scores, and on the right are amino acid substitution marginal feature z-score distributions (points depict variants; black lines depict medians).

(D) Two randomly-selected iPS cells (top) or NGN2-induced neurons (bottom) expressing PTEN WT or variants are shown. mEGFP-tagged PTEN channel is shown in green, DAPI in blue, phalloidin (top) or MAP2 (bottom) in yellow, and pAKT in red. Silver dashed lines depict the borders of cells. Scale bar indicates 5 μm.

(E) PTEN abundance score for PTEN variants in iPS cells (upper left) and NGN2-induced neurons (upper right) plotted against VAMP-seq⁵ score, with best-fit line and Pearson’s r shown. pAKT intensity score for PTEN variants in iPS cell (bottom left) and NGN2-induced neuron-like cells (bottom right) plotted against yeast fitness⁸⁴ score, with Pearson’s r shown.

(F) iPS cell mEGFP-PTEN intensity score positional averages are used to color the PTEN crystal structure (1D5R)⁷¹. Solvent-exposed surface is colored by the exposed residue’s z-score. Blue indicates low-intensity positions. The bound tartrate molecule is shown in green. Residue color scale shown to the right of (H).

(G) iPS cell pAKT intensity score positional averages are used to color the PTEN crystal structure (1D5R)⁷¹.

(H) iPS cell DAPI-PTEN correlation score positional averages are used to color the PTEN crystal structure (1D5R)⁷¹.

8

Figure S8. Relationships between PTEN landmark features and scores in iPS cells and NGN2-induced neuron-like cells, related to Figure 5

(A) iPS cell landmark feature z-scores are plotted against NGN2-induced neuron-like cell landmark feature z-scores for all profiled variants, colored by variant type. Z-scores are versus the synonymous variant distribution.

(B) iPS cell nucleus to cytoplasm PTEN intensity ratio z-scores are plotted against iPS cell DAPI-PTEN correlation z-scores for all profiled variants, colored by variant type as in (A).

(C) Violin plots of iPS cell and NGN2-induced neuron PTEN abundance (left) and DAPI-PTEN correlation z-scores (right). iPS cell-specific nuclear-localized variants are indicated. *** indicates KS p<0.001.

(D) iPS cell PTEN landmark feature z-scores are plotted against each other for all profiled variants. Variant type is colored according as in (A). iPS cell-specific lipid phosphatase-active nuclear-localized variants are indicated.

(E) The mislocalization score is defined as the variant-level residuals when DAPI-PTEN correlation is regressed against PTEN intensity and pAKT intensity (see Methods). Positional heatmap of iPS cell mislocalization scores for missense, stop gain, or 3-nucleotide deletion variant profiles in iPS cell. Positional average of scores shown on the bottom and amino acid substitution marginal distribution shown on the right.

(F) Mislocalization score in iPS cells computed independently for replicates 1 and 2 are plotted. Synonymous (green) and missense (grey) variants are shown. Pearson’s r shown on the top left.

(G) PTEN structure (1D5R)⁷¹ with spheres centered at α-carbon atoms with radii indicating the number of lipid-phosphatase active (defined as pAKT z-score<2.5) and nuclear-localized (defined as DAPI-PTEN correlation z-score>2.5) variants at that position. The top four positions indicating nuclear localization are highlighted, with the number of variants at each of these positions indicated.

(H) iPS cell mislocalization score positional averages are used to color the PTEN crystal structure (1D5R)⁷¹. Blue indicates aberrantly cytoplasmic-localizing positions and red indicates aberrantly nuclear-localizing positions. Color scale shown on the right.

(I) UpSet plot showing the number of PTEN missense variants that perturb each combination of landmark features in iPS cells and NGN2-induced neurons. Landmark features for each variant were considered perturbed if feature |z-score|>2.5 and KS-test Bonferroni-corrected p<0.01. Sets, represented by bars and dots in the plot, are colored by combinations of impacted features (neither pAKT nor PTEN abundance impacted in both cell types = black, pAKT impacted in both cell types = orange, PTEN abundance impacted in both cell types = teal, both pAKT and PTEN abundance impacted in both cell types = purple).

(J) The distinguishability score difference is defined as the variant-level difference in distinguishability score percentile between iPS cells and neurons. Positional heatmap of distinguishability score difference for missense, stop gain, or 3-nucleotide deletion variant profiles in iPS cells. Positional average of scores shown on the bottom and amino acid substitution marginal distribution shown on the right.

(K) Waterfall plots of ranking all iPS cell features (top) or neuron features (bottom) ranked by Spearman’s correlation with distinguishability score difference. Positive values indicate larger values of feature are associated with distinguishability in iPS cells compared to neurons. Landmark features are indicated and tracks are shown below the plot indicating feature imaging channel and cellular compartment. Lines above and below the plot labeled by *** indicate which features have Student’s t-test Bonferroni-corrected p<0.001.

9

Figure S9. PTEN VIS-seq profiles predict pathogenicity and clinical phenotype, related to Figure 6

(A) UMAP visualization of iPS cell (left) and NGN2-induced, neuron-like cell (right) PTEN variant profiles with triangles indicating variants in ClinVar classified as likely benign (LB, blue), likely pathogenic (LP, red), pathogenic (P, dark red), or variant of uncertain significance (VUS, yellow). All profiled variants are plotted in the background colored green (synonymous) or grey (otherwise) for comparison.

(B) Morphological impact scores for PTEN iPS cell (left) or NGN2-induced neuron (right) profiles are plotted by ClinVar label. Synonymous variants (green) are included for comparison. *** indicates Mann-Whiney U p<0.001.

(C) Receiver operating characteristic (ROC) curves are plotted for univariate zero-shot models predicting ClinVar pathogenicity or each clinical phenotype (ASD = autism spectrum disorder, DD = developmental delay, PHTS = PTEN hamartoma tumor syndrome; see Methods for curation criteria) from iPS cell or NGN2-induced neuron distinguishability score (this publication, solid lines), yeast fitness scores⁸⁴ (dashed line), VAMPseq scores⁵ (dashed line), AlphaMissense⁹¹ or EVE⁹² scores (dot-dashed lines). Area under the curve (AUC) scores are shown in the box for each model.

(D) Morphological impact scores for PTEN variants in NGN2-induced neurons are plotted by variant association with clinical phenotypes. gnomAD v4.1 (light blue) variants and synonymous variants (green) are plotted for comparison.

(E) Distinguishability scores for PTEN variants in iPS cells (top) or NGN2-induced neurons (bottom) are plotted by variant association with clinical phenotypes. gnomAD v4.1 (light blue) variants and synonymous variants (green) are plotted for comparison.

(F) Feature scores for PTEN variants in iPS cells are plotted by variant association with clinical phenotypes. gnomAD v4.1 (light blue) variants and synonymous variants (green) are plotted for comparison. Features plotted include PTEN intensity (left) or pAKT intensity (right). *** indicates Mann-Whitney U-test p<0.001.

(G) Feature scores for PTEN variants in NGN2-induced neurons are plotted by variant association with clinical phenotypes. gnomAD v4.1 (light blue) variants and synonymous variants (green) are plotted for comparison. Features plotted include PTEN intensity (left), pAKT intensity (middle), or mislocalization score (right). *** indicates Mann-Whitney U-test p<0.001.

(H) Plot of PTEN median effects on feature z-scores in the ASD/DD-only variants (x-axis) against PHTS-only variants (y-axis), in iPS cells (left) or NGN2-induced neurons (right). Points are features colored by imaging channel (see left for legend). Landmark features are highlighted. Red dashed lines show median effect thresholds at synonymous z-scores of 1 or −1.

10

Figure S10. VIS-seq measurements elaborate computational variant effect predictions, related to Figure 6

(A) Predictor scores for AlphaMissense⁹¹, EVE⁹², and REVEL⁹³ are plotted against LMNA VIS-seq variant impact and distinguishability scores as well as landmark feature medians. ClinVar P/LP variants are shown with red triangles. Predictor-specified pathogenic and benign thresholds are shown by horizontal dotted lines. Pearson’s r between logit-transformed predictor scores and VIS-seq scores are shown on the bottom left.

(B) Predictor scores for AlphaMissense⁹¹, EVE⁹², and REVEL⁹³ are colored on the UMAP visualization of LMNA VIS-seq profiles.

(C) Predictor scores for AlphaMissense⁹¹, EVE⁹², and REVEL⁹³ are plotted against PTEN VIS-seq iPS cell variant impact and distinguishability scores as well as landmark feature medians. ClinVar P/LP variants are shown with red triangles and B/LB variants are shown with blue triangles. Predictor-specified pathogenic and benign thresholds are shown by horizontal dotted lines. Pearson’s r between logit-transformed predictor scores and VIS-seq scores are shown on the bottom left.

(D) Predictor scores for AlphaMissense⁹¹, EVE⁹², and REVEL⁹³ are colored on the UMAP visualization of PTEN iPS cell VIS-seq profiles.

(E) Predictor scores for AlphaMissense⁹¹, EVE⁹², and REVEL⁹³ are plotted against PTEN VIS-seq NGN2-induced neuron-like cell variant scores and landmark features as in (C).

(F) Predictor scores for AlphaMissense⁹¹, EVE⁹², and REVEL⁹³ are colored on the UMAP visualization of PTEN NGN2-induced neuron VIS-seq profiles as in (D).

11

Table S1. LMNA feature-level volcano analysis, related to Figure 2

12

Table S2. LMNA variant-level landmark features and scores, related to Figures 2–4

13

Table S3. PTEN feature-level volcano analysis, related to Figures 5 and 6

14

Table S4. PTEN variant-level landmark features, scores, and clinical annotations, related to Figures 5 and 6

15

Table S5. Curated PTEN missense variant–phenotype associations from the literature (PHTS and ASD/DD), related to Figure 6

16

Table S6. DNA sequences used in this publication, related to STAR Methods

(A) Primers and oligos used for cloning

(B) Libraries

(C) Gene fragments used for cloning

(D) Plasmids

(E) Guide RNAs

(F) Oligos for in situ sequencing

KEY RESOURCES TABLE.

REAGENT OR RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-NCAM1 Rabbit mAb, Alexa Fluor 647 conjugate	Cell Signaling Technology	Cat#50831S; RRID: AB_3717968
Anti-MAP2 Mouse mAb, Alexa Fluor 594 conjugate	BioLegend	Cat#801802; RRID: AB_2721382
Anti-phospho-AKT (Thr308) Rabbit mAb	Cell Signaling Technology	Cat#13038T; RRID: AB_2629447
Goat Anti-Rabbit IgG H&L cross-adsorbed secondary antibody, Alexa Fluor 647 conjugate	Thermo Scientific	Cat#A-21244; RRID: AB_2535812
Anti-pan-AKT Mouse mAb	Cell Signaling Technology	Cat#2920; RRID: AB_1147620
Anti-PTEN Rabbit mAb	Cell Signaling Technology	Cat#9559; RRID: AB_390810
Anti-lamin A/C Mouse mAb	BioLegend	Cat#600001; RRID: AB_2810655
StarBright Blue 520 Goat Anti-Mouse IgG	Bio-Rad	Cat#12005866; RRID: AB_2934034
StarBright Blue 700 Goat Anti-Rabbit IgG	Bio-Rad	Cat#12004162; RRID: AB_3712168
hFAB Rhodamine Anti-Actin Primary Antibody	Bio-Rad	Cat#12004164; RRID: AB_2861334
Bacterial and virus strains
NEB 10-beta electrocompetent cells	New England Biolabs	Cat#C3020K
Chemicals, peptides, and recombinant proteins
DMEM	Gibco	Cat#11965-092
DMEM/F-12	Gibco	Cat#11320-033
Fetal Bovine Serum (FBS)	Hyclone	Cat#SH30071.03
Penicillin/Streptomycin	Gibco	Cat#15140-122
OPTI-MEM	Gibco	Cat#31985-070
DPBS	Gibco	Cat#14190-144
Versene	Gibco	Cat#15400-054
Trypsin-EDTA 0.05%, Phenol Red	Gibco	Cat#25300-054
TransIT-293	Mirus	Cat#MIR 2700
Dimethyl Sulfoxide (DMSO)	Sigma-Aldrich	Cat#D8418-50mL
CryoStor CS10 Cell Freezing Medium	STEMCELL Technologies	Cat#100-1061
Matrigel Matrix	Corning	Cat#354234
mTeSR Plus	STEMCELL Technologies	Cat#100-0276
ROCK Inhibitor Y-27632	SelleckChem	Cat#S1049
Puromycin Dihydrochloride	Gibco	Cat#A11138-03
0.45 um Syringe Filter	VWR	Cat#76479-012
SpCas9 2NLS Nuclease (300 pmol)	Synthego	N/A
Poly-L-ornithine	Sigma-Aldrich	Cat#P3655
Accutase	STEMCELL Technologies	Cat#07922
Doxycycline hyclate	Sigma-Aldrich	Cat#D9891
Recombinant Human NT-3	Peprotech	Cat#450-03
Recombinant Human BDNF	Peprotech	Cat#450-02
GlutaMAX Supplement 100X	Gibco	Cat#35050-061
MEM Non-Essential Amino Acid Solution 100X	Gibco	Cat#11140-050
B-27 Supplement 50X	Gibco	Cat#17504-044
N-2 Supplement 100X	Gibco	Cat#17502-048
KnockOut DMEM/F-12	Gibco	Cat#12660-012
Neurobasal-A Medium	Gibco	Cat#12349-015
Laminin Mouse Protein	Gibco	Cat#23017-015
RPMI 1640 Medium	Invitrogen	Cat#11875135
DMEM, no glucose	Gibco	Cat#11966-025
Sodium-L-Lactate	Sigma-Aldrich	Cat#71718
Vitronectin XF	STEMCELL Technologies	Cat#100-0763
CHIR-99021	TOCRIS	Cat#4423
Wnt-C59	Selleck Chemicals	Cat#S7037
Bovine Serum Albumin (BSA)	Sigma-Aldrich	Cat#A9418
Ascorbic Acid	Sigma-Aldrich	Cat#A8960
B27 + Insulin	Invitrogen	Cat#17504044
TrypLE Select 10X	Gibco	Cat#A12177-01
Paraformaldehyde 20% Solution	Electron Microscopy Sciences	Cat#15713-S
Triton X-100	Sigma-Aldrich	Cat#Y8787-50ML
Tween-20	Sigma-Aldrich	Cat#P1279-500ML
Glutaraldehyde 25% Solution	Electron Microscopy Sciences	Cat#16210
Deoxynucleotide (dNTP) Solution Mix 10 mM	New England Biolabs	Cat#N0447L
Ultrapure BSA	Thermo Scientific	Cat#AM2616
Ribolock RNase Inhibitor	Thermo Scientific	Cat#EO0381
RevertAid H Minus Reverse Transcriptase	Thermo Scientific	Cat#EP0451
Ampligase Buffer	Biosearch Technologies	Cat#SS000015-D3
Taq-IT Polymerase	Qiagen	Cat#P7620L
Ampligase	Biosearch Technologies	Cat#E0001-100D2
RNase H	New England Biolabs	Cat#M0297L
Phi29 DNA Polymerase	Thermo Scientific	Cat#EP0094
Glycerol	Fisher Scientific	Cat#G33-1
Formamide	Fisher Scientific	Cat#BP227-500
Nuclease-Free Water	Thomas Scientific	Cat#C001X09
20X SSC	Invitrogen	Cat#15557036
10X HBSS	Gibco	Cat#14065-056
Bovine Serum Albumin (fraction V)	Sigma-Aldrich	Cat#A2153-100G
Sodium azide	Sigma-Aldrich	Cat#S-8032
Ascorbic Acid	Sigma-Aldrich	Cat#A92902-100G
4’,6-Diamidino-2-phenylindole (DAPI)	Fisher Scientific	Cat#D1306
Sodium borohydride	Sigma-Aldrich	Cat#452882
Phalloidin-CF568 conjugate	Biotium	Cat#00044-T
Wheat Germ Agglutinin (WGA)-CF555 conjugate	Biotium	Cat#29076-1
Concanavalin A-Alexa Fluor 750 conjugate	Thermo Scientific	Cat#C56127
Carbenicillin Disodium, USP Grade	GoldBio	Cat#C-103-250
Kanamycin	Sigma-Aldrich	Cat#K1537-25G
Q5 High-Fidelity 2X Master Mix	New England Biolabs	Cat#M0492L
Klenow Fragment, 3’-5’ exo-minus	New England Biolabs	Cat#M0212L
NEBuilder HiFi DNA Assembly Master Mix	New England Biolabs	Cat#E2621L
T4 DNA Ligase	New England Biolabs	Cat#M0202M
BsaI-HFv2	New England Biolabs	Cat#R3733L
SapI	New England Biolabs	Cat#R0569L
Esp3I	New England Biolabs	Cat#R0734L
BsmBI-v2	New England Biolabs	Cat#R0739S
NheI-HF	New England Biolabs	Cat#R3131L
MluI-HF	New England Biolabs	Cat#R3198S
EcoRI-HF	New England Biolabs	Cat#R3101S
NotI-HF	New England Biolabs	Cat#R3189S
SacII	New England Biolabs	Cat#R0157S
RIPA Buffer	Abcam	Cat#ab156034
Halt Protease Inhibitor Cocktail	ThermoFisher Scientific	Cat#1860932
Laemmli Sample Buffer	Bio-Rad	Cat#1610737
β-Mercaptoethanol	Bio-Rad	Cat#1610710
8–16% Criterion TGX Stain-Free Protein Gel	Bio-Rad	Cat#5678103
Tris/Glycine/SDS Running Buffer	Bio-Rad	Cat#1610732
Trans-Blot Turbo Midi 0.2 um PVDF Transfer Packs	Bio-Rad	Cat#1704157
Critical commercial assays
Glass 6-well Plate	Cellvis	Cat#P06-1.5H-N
Corning Untreated 245 mm Square BioAssay Dishes	Fisher Scientific	Cat#07-200-600
Lonza 4D Nucleofector SE Kit S	Lonza	Cat#V4XC-1032
Lonza 4D Nucleofector P3 Kit S	Lonza	Cat#V4XP-3032
Lonza 4D Nucleofector P3 Kit X	Lonza	Cat#V4XP-3024
Illumina MiSeq v2 Nano 500 Kit	Illumina	Cat#MS-103-1003
Pierce BCA Protein Assay Kit	ThermoFisher Scientific	Cat#23225
SMRTbell Prep Kit 3.0	Pacific Biosciences	Cat#102-182-700
Deposited Data
Impact scores, distinguishability scores, landmark features	MaveDB; Rubin et al.	URN#s: 00001243-a, 00001244-a, 00001244-b
Variant-level features, EMDs, p-values, scores	Zenodo	Cat#15787685; https://zenodo.org/records/15787685
Phenotype image data, single-cell features and genotypes, CellProfiler pipelines	Biolmage Archive	Accession#S-BIAD3095; https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD3095
Experimental models: Cell lines
Human embryonic kidney 293T (HEK 293T)	ATCC	Cat#CRL-3216; RRID: CVCL_0063
U2OS cells	ATCC	Cat#HTB-96; RRID: CVCL_0042
U2OS C11 cells	Hasle et al.	N/A
WTC11 iPS cells	Coriell	Cat#GM25256; RRID: CVCL_Y803
NGN2-inducible WTC11 iPS cells (NWTC11.G3-WT)	Gladstone Institute; gift from Li Gan lab	N/A
Oligonucleotides
See Table S6
Recombinant DNA
pHSG299	Takara Bio	Cat#3299
pHR-UCOE-EF1a-Zim3-dCas9-BFP	Addgene; Replogle et al.	Cat#188775; RRID: Addgene_188775
pJBL_051	Gift from Jay Shendure lab; Lalanne et al.	N/A
lentiGuide-Puro	Addgene; Sanjana et al.	Cat#52963; RRID: Addgene_52963
pCAG-NLS-HA-Bxb1	Addgene; Hermann et al.	Cat#51271; RRID: Addgene_51271
Hyperactive piggyBac transposase expression vector (hyPBase)	Gift from Jay Shendure lab; Yusa et al.	N/A
pMD2.G	Addgene	Cat#12259; RRID: Addgene_12259
psPAX2	Addgene	Cat#12260; RRID: Addgene_12260
PBv2b	This publication	VIS-seq expression vector. N/A
PBv2b-NSapI	This publication	VIS-seq expression vector. Cat#252243; RRID: Addgene_252243
PBv2b-CSapI	This publication	VIS-seq expression vector. Cat#252244; RRID: Addgene_252244
PBv2c-NSapI	This publication	VIS-seq expression vector. Cat#252245; RRID: Addgene_252245
Software and algorithms
Cellpose v2.2.1	Stringer et al.	https://www.cellpose.org/; RRID: SCR_021716
StarDist v0.8.5	Schmidt et al.	https://github.com/stardist/stardist
STARCall	Bradley et al.	https://github.com/FowlerLab/starcall-workflow
CellProfiler v4.2.6	Carpenter et al.	https://cellprofiler.org/; RRID: SCR_007358
Pycytominer v1.2.1	Serrano et al.	https://github.com/cytomining/pycytominer
SciPy v1.14.1	Virtanen et al.	https://scipy.org/; RRID: SCR_008058
NumPy v1.24.4	Harris et al.	https://numpy.org/; RRID: SCR_008633
Scikit-learn v1.5.2	Pedregosa et al.	https://scikit-learn.org/; RRID: SCR_002577
Scikit-image v0.24.0	van der Walt et al.	https://scikit-image.org/; RRID: SCR_021142
Matplotlib v3.9.2	Hunter et al.	https://matplotlib.org/; RRID: SCR_008624
NetworkX v3.4.2	Hagberg et al.	https://networkx.org/; RRID: SCR_016864
Snakemake v7.32.4	Molder et al.	https://snakemake.readthedocs.io/; RRID: SCR_003475
Pandas v2.2.3	McKinney et al.	https://pandas.pydata.org/; RRID: SCR_018214
Python v3.11	Python Software Foundation	https://www.python.org/; RRID: SCR_008394
XGBoost v3.0.1	Chen and Guestrin	https://xgboost.ai/; RRID: SCR_021361
PyMOL v3.1	Schrodinger LLC	https://www.pymol.org; RRID: SCR_000305
Pacybara	Weile et al.	https://github.com/rothlab/pacybara

Highlights.

• VIS-seq enables optical measurement of variant effects in diverse cell types at scale

• LMNA variants can drive aggregation or low abundance, modulating nuclear circularity

• PTEN localization distinguishes autism from tumor syndrome variants

• VIS-seq reveals how complex variant effects can cascade from molecules to cells