Live-cell transcriptomics with engineered virus-like particles

SUMMARY

Transcriptomic profiling is widely applied to characterize cellular gene expression, yet existing approaches lyse cells and preclude direct analysis of transcriptional dynamics in the same sample over time. We addressed this limitation by engineering mammalian cells to “self-report” their transcriptional states via mRNA export in virus-like particles (VLPs). Repeated sampling of culture media from VLP-producing cell populations faithfully captured evolving transcriptional states in complex biological settings, including acute inflammatory stimulation of primary cell spheroids and multi-day differentiation of pluripotent stem cells. We engineered VLP components for multiplexed readouts from distinct cell-types in co-culture and to tune self-reported RNA profiles. Finally, we demonstrated the unique utility of self-reporting for selective longitudinal tracking of endothelial cell dynamics within the enclosed architecture of a microphysiological co-culture system to identify perivascular stroma-dependent temporal gene programs underlying vasculogenesis. Altogether, this work establishes cellular self-reporting as a broadly enabling technology for live-cell transcriptome-wide gene expression profiling.

Graphical Abstract

In brief:

Najia, Le, Borrajo et al. establish “cellular self-reporting,” a technology to export molecular analytes from cells in virus-like particles generated by the cells themselves. The authors demonstrate how mRNA export enables longitudinal transcriptome-wide profiling of living cells by sampling cell media. This platform transcends the destructive nature of conventional transcriptomic analysis methods and reveals gene expression dynamics in complex biological systems in situ.

INTRODUCTION

The dynamic regulation of gene expression is integral to the function of cells and tissues in development, homeostasis, stress response, and disease. Unbiased gene expression profiling via high-throughput RNA sequencing (RNA-seq) has become a cornerstone of modern biomedical science.^1-3 However, RNA-seq almost exclusively entails destruction of cells in the biological sample, as cells are lysed for RNA extraction or transcripts are retrieved from a few cells using invasive micromechanical procedures at each time point.^4,5 The destructive nature of available gene expression profiling technologies prevents the integration of multiple functional readouts from the same biological sample at different times, hindering studies of dynamic cellular phenomena. In addition, the requirement for physical sampling of cells prevents analysis of inaccessible cells, such as those in structured in vitro research models and natural tissues. The development of computational “pseudotime” algorithms reflects broad interest to study the dynamic evolution of transcriptional states, yet in the absence of methods for direct time series measurements, they remain limited as inferential techniques.^6-11

Targeted approaches to study time-dependent gene expression are an active area of technology development. The advent of genetically encodable fluorescent proteins provided an early framework for enabling true time-series analysis of individual gene products using optical microscopy. Several groups have further pioneered optical methods that use endogenous or exogenous probes to monitor a few native or tagged target transcripts at a time.^12-22 High-content imaging of live cells promises dynamic resolution of cell states^23,24 and potential for inference on gene expression.²⁵ Transcriptome in vivo analysis (TIVA) enables transcriptome-wide RNA capture in living cells, but has not yet been reported for longitudinal profiling.²⁶ More recently, other molecular recording methods have also allowed information about transcript age or prior transcriptional states to be encoded in the DNA or RNA of living systems for readout at a single end point.^27-31 While useful, these methods variously require a prior transcriptional hypothesis, have limited multiplexing capacity and dynamic range, and/or destroy the living biological sample in the act of measurement. Overall, the limitations of these approaches motivate the need for technologies that allow direct live-cell longitudinal monitoring of transcriptome-wide gene expression to study dynamic cellular phenomena.

To overcome the constraints of destructive profiling, we reasoned that approaches whereby mammalian cells themselves externalized RNA molecules could lend temporal transcriptional insights across time. Indeed, cells are known to export RNA through natural biological processes that produce RNA-containing extracellular vesicles (EVs). However, EV biogenesis is complex, context-dependent, and integral to the natural function of mammalian systems, making it less attractive as an engineering platform.³² By contrast, retroviruses such as the Moloney murine leukemia virus (MMLV) have evolved a streamlined and orthogonal mechanism to package and export their RNA genomes within membrane-bound virions. Central to this process is Gag, a structural polyprotein that is sufficient for the formation of replication-incompetent, virus-like particles (VLPs).³³ VLPs already support a broad range of technological applications, including vaccines³⁴, measurement of protein-protein interactions^35,36, and cellular export and delivery of proteins or synthetic nucleic acids.^37-40 While VLPs can be engineered to package a range of analytes, RNA is an essential structural component of retroviral virions.^41,42 In the absence of the viral genome or viral packaging signals, MMLV Gag maintains the ability to assemble and produce VLPs by interacting with and packaging cellular RNA molecules.^40,43-45 We hypothesized that the inherent affinity of MMLV Gag to cellular RNA could be harnessed to create a genetically-encoded and host-orthogonal system, enabling molecular engineering to tailor the properties of VLPs for control of their production and collection from host cells.

Here, we describe “cellular self-reporting,” a technology platform that repurposes viral machinery to package and export molecular analytes from living cells via VLPs. We demonstrated the utility of this technology for non-destructive, transcriptome-wide profiling of cellular RNAs. Repeatedly sampling VLP-containing cell culture media enabled direct insights into evolving transcriptional states from the same population of cells over time, preserving biological integrity of the sample and affording repeated-measures statistical analysis. We then applied cellular self-reporting to primary endothelial cells to monitor transcriptional dynamics of vasculogenesis in situ within enclosed microphysiological devices, resolving complex temporal phenomena that are otherwise inaccessible to traditional destructive sampling approaches.

RESULTS

MMLV Gag packages, exports and stabilizes cellular RNAs in VLPs

To enable live-cell transcriptional profiling, we repurposed MMLV Gag to non-specifically package and export cellular RNAs in VLPs (Figure 1A). The cellular RNA contained in these VLPs can then be collected from the culture media and processed for RNA-seq to enable longitudinal transcriptional profiling of the same population of living cells (Figure 1B). We first evaluated the utility of MMLV Gag – consisting of the N-terminal matrix domain (MA, p15), the central capsid domain (CA, p30), and the C-terminal nucleocapsid domain (NC, p10) – as the basis of a synthetic RNA export pathway, by transfecting the full coding sequence (Table S1) into HEK293T cells. Since Gag localization and assembly on lipid membranes is mediated through myristoylation of the MA domain^46,47, we generated a Gag truncation mutant lacking the MA domain (GagΔMA) as a negative control for VLP formation. We detected MMLV Gag p30 protein in the cell media of cells transfected with full-length Gag, but not in the media of cells expressing GagΔMA (Figure S1A), as expected. Gag C-terminally fused to mCherry formed punctate foci within cells (Figure S1B), consistent with Gag multimerization during VLP assembly, and scanning electron microscopy further revealed the formation of VLPs approximately 160 nm in diameter in the culture media of Gag-expressing cells (Figure S1C).

Fig. 1. Cellular self-reporting with MMLV Gag-enabled live-cell transcriptomics.

(A) Schematic of MMLV Gag-mediated export of cellular RNAs in VLPs.

(B) Experimental approach for live-cell transcriptional time-course experiments by sampling cell media from cells engineered to self-report transcriptional states via VLPs.

(C) TaqMan qPCR assay to quantify GAPDH transcripts from free RNA or cell media of Gag-expressing cells +/−RNase treatment. Data are normalized to the -RNase control. Error bars reflect standard deviation.

(D) TaqMan qPCR assay to quantify GAPDH expression from VLP-containing cell media of Gag-expressing cells incubated at 37°C over four days.

(E) Quantification of the number of genes detected by RNA-seq in cell lysate and VLP-containing cell media samples from HEK293T cells transfected with H2B-GFP, GagΔMA or Gag expression constructs.

(F) Scatterplot of normalized gene expression detected from cell lysate and VLP-containing cell media. HEK293T cells were engineered to express MMLV Gag via plasmid transfection (left) or lentiviral transduction (right). Each point represents a gene in mutually exclusive gene categories (protein-coding and non-coding mitochondrial genes are collectively labeled as “Mitochondrial,” whereas “Protein-coding” or “Non-coding” labels reflect genes in the nuclear genome). Pearson correlation coefficients are noted for each gene category. TPM = transcripts per million.

(G) SHAP feature importance plot for gene-level features from a gradient boosting regression model to predict the log₂(fold-change) in expression of genes detected from VLP-containing cell media relative to cell lysate (positive log₂(fold-changes) reflect enrichment in cell media). Each point represents a gene. All features are continuous variables, except for gene class features, which are binary. RNA cytoplasm:nucleus ratio represents the log₂(fold-change) of genes between subcellular compartments (positive log₂(fold-changes) reflect enrichment in the cytoplasm relative to the nucleus). Splice ratio is the ratio of cDNA length to gene length (quantified from TSS to TES). Features are ranked by the mean absolute SHAP value (right).

(H) Heatmap of global gene expression Pearson correlation coefficients between cell media samples derived from H2B-GFP, GagΔMA or Gag expressing cells and associated cell lysate samples.

(I-J) CellNet classification heatmap of cell lysate and cell media from cell lines (I), and pluripotent stem cells and primary cells (J). The CellNet classifier was retrained in (I) to include cell lysate samples of HEK293T and HT1080 cells, whereas the default classifier was used in (J). The gene regulatory network (GRN) score represents the degree to which the sample recapitulates the GRN of the known cell types within the CellNet classifier.

We next characterized the cellular RNAs that are packaged in VLPs. Expression of a ubiquitously expressed gene, GAPDH, was detectable in the culture media of HEK293T cells transfected with full-length Gag and resistant to degradation following RNase treatment of the culture media (Figures 1C, S1D), consistent with prior reports.⁴⁰ We further verified that GAPDH transcripts were stable in VLPs and detected in cell media following four days of incubation at 37°C (Figure 1D). To more extensively characterize RNA exported in VLPs, we performed RNA-seq⁴⁸ on HEK293T cells transfected with Gag and the associated culture media following 48 hours of VLP production. On average, 12,931 unique genes were detected in cell lysate transcriptomes, and of those genes, 11,598 genes were also detected in the cell media (Figure 1E). Significantly fewer genes were detected in the media of cells expressing H2B-GFP or GagΔMA (Figure 1E), suggesting that the RNA signal in the media of Gag-expressing cells is largely VLP-dependent. Moreover, the expression of genes detected from cell media was correlated with conventional RNA-seq measurements on cell lysate (Pearson, r = 0.797, Figures 1F, S1E-F), suggesting that high-dimensional transcriptional information can be derived from self-reported RNA. While protein-coding and non-coding genes were represented in the cell media, mitochondrial genes were notably depleted compared to cell lysate (Figure 1F). This observation is consistent with the expected sequestration of mitochondrial transcripts within mitochondria and away from the cytosol where Gag multimerization is initiated. These data established that MMLV Gag can broadly package, export, and stabilize cellular RNA transcripts in VLPs for non-destructive transcriptome-scale gene expression profiling by RNA-seq.

Dissecting biases associated with self-reported RNAs

The observation that transcripts of certain gene classes are differentially represented in VLP-containing cell media samples compared to the cell lysate transcriptome prompted us to dissect biases associated with self-reported RNA measurements. We trained a gradient boosting regression model on gene-level and k-mer sequence features (Table S2A) to predict the fold change in gene expression between cell media and associated cell lysate samples. The global importance and directionality of each feature to the model’s predictions were interpreted using Shapley additive explanations (SHAP) (Figures 1G, S1G, Table S2B). The model readily identified that mitochondrial genes were depleted in the cell media, in-line with our prior observations. The top feature by SHAP analysis was the subcellular localization of transcripts, wherein genes with higher expression in the cytoplasm relative to the nucleus were likely to be over-represented in cell media measurements relative to cell lysate measurements (Figure 1G). We further verified that gene expression detected in the cell media of Gag-expressing cells was globally more concordant with cytoplasmic lysate compared to nuclear lysate and subtly associated with gene length (Figures 1H, S1H-I). The frequency of specific 3-mers in the gene sequence (Figure S1G) as well as GC content in untranslated regions (Figure 1G) were associated with differential representation in cell media and cell lysate, suggesting that sequence content and/or RNA secondary structure could also influence preferential packaging of RNA in VLPs.

Many cell types can constitutively self-report their transcriptional state via VLPs

While our initial data were based on transient transfection of Gag, we sought to engineer mammalian cells to constitutively express Gag for longer-term live-cell studies of transcriptional dynamics. Gene expression detected by RNA-seq from the cell media of lentivirally transduced HEK293T cells correlated with cell lysate measurements (Pearson r = 0.714), indicating that stable integration of Gag resulted in comparable performance as plasmid transfection (Figure 1F). We extended these findings to other cell types by engineering HT1080 fibrosarcoma cells via lentiviral transduction and induced pluripotent stem cells (iPSCs) via piggyBac transposition of Gag. In both cases, gene expression measured from cell media was also correlated with cell lysate measurements (Figure S1J, Table S3). Importantly, constitutive expression of Gag was minimally perturbative for each cell type, evidenced by normal morphology, growth rates, and transcriptomic profiles, in which the transgenes themselves were the most prominent differential genes (Figures S1K-M).

We then assessed whether the transcriptional data derived from sampling VLP-containing cell media reflected the host transcriptome sufficiently to classify cell type. We applied CellNet⁴⁹ to cell media samples containing VLPs produced over 48 hours from each cell line to ascertain the activity of cell-type-specific gene regulatory networks (GRNs) and predict the biological identity of each sample. The self-reported transcriptomes supported classification to the expected cellular identities, mirroring the classification of cell lysate samples (Figure 1I). Finally, we observed similar classification performance from Gag-expressing primary endothelial cells and primary stromal fibroblasts across multiple donors (Figure 1J), demonstrating that self-reporting of transcriptional states in VLPs is applicable beyond immortalized cell lines.

Self-reported transcriptomes are detected across time and culture scales

Self-reporting performance depends on physical and temporal parameters related to VLP production and collection. For example, while destructive transcriptomic measurements represent an instantaneous census of transcripts, non-destructive self-reported measurements reflect RNA accumulated over a time interval within sampled VLPs. We therefore sought to assess self-reported transcriptional profiles as a function of the amount of time allowed for VLP production using cells constitutively expressing Gag (Figure S2A). VLP production in as little as three hours was sufficient to classify the cellular identity of the host transcriptome relative to negative controls (Figure S2B). However, cell type classification performance, global concordance with the cellular transcriptome, and the number of genes detected improved with longer durations of VLP production, generally plateauing at 48 hours (Figures S2C-E).

To explore the limit on sensitivity for small populations of self-reporting cells, we assessed the number of VLP-producing cells per sample by FACS-isolating single cells and groups of 100 cells (Figure S2F, Methods). After culturing these cells for 24 hours, we prepared RNA-seq libraries from the cell lysates and associated cell media. We used microscopy as part of the sample collection process and conservative computational filtering to ensure that cells were not inadvertently collected in cell media samples (see Methods). We found that gene expression in the media of single HT1080 cells correlated with the corresponding single-cell lysate measurements (Pearson’s r = 0.61, Figure S2G). We measured 589±63 unique genes in the media of single cells and 1,866±22 unique genes in the media of 100-cell cultures (Figure S2H). Although we could detect exported RNA from small populations and even individual self-reporting cells, given the low gene counts per cell, we prioritized larger cell numbers (≥100,000) in subsequent applications to ensure sensitive and reproducible detection of subtle gene expression dynamics in self-reported transcriptional profiles.

Repeated sampling of culture media captures transcriptional dynamics of neurogenesis

Building upon single time-point experiments, we asked whether self-reporting is powered to capture transcriptional phenomena in living cells over time. We leveraged Gag-expressing iPSCs as an in vitro model to study the dynamics of multi-day cellular differentiation and compare the performance of self-reported transcriptomes to conventional cell lysate measurements. We focused specifically on neurogenesis since differentiating cells progressively attenuate their proliferative capacity en route to becoming post-mitotic neurons, limiting the degree to which cells can be subsampled with destructive endpoint techniques.

We stably introduced Gag via piggyBac transposition into iNGN iPSCs featuring doxycycline (dox)-inducible control of Neurogenin-1 and Neurogenin-2 transcription factors, which drove the development of neurons with bipolar morphology within four days (Figures S3A-C), and verified that the cells self-reported their transcriptional state within VLPs (Figure 1J). Constitutive Gag expression did not impede the differentiation process, as confirmed by staining with the neuronal marker NCAM1 (Figure S3D).

We captured transcriptional profiles over the course of the differentiation, starting when the iPSCs were still in a pluripotent state (Day -1) and then every 24 hours for four days after dox induction (Figure 2A). Cell media samples were collected from the same population over multiple timepoints and subjected to RNA-seq. We found that gene expression profiles obtained from VLP-containing cell media were correlated with measurements from cell lysates and reproducible among replicates throughout the four-day time course (Figures 2B, S3E-F). Furthermore, principal component analysis (PCA) revealed progressive changes in gene expression profiles across time for both cell lysate and cell media samples, suggesting that self-reporting can detect transcriptional changes that occur during differentiation (Figures S3G-H).

Fig. 2. Repeated sampling of VLPs from media of live cells captures transcriptional dynamics during neurogenesis.

(A) Experimental workflow to longitudinally profile transcriptional dynamics during neuronal differentiation via repeated sampling of cell media from Gag-expressing iPS cells. iNGN iPS cells constitutively expressing Gag-2A-mNeon were differentiated into bi-polar-like neurons by doxycycline induction of Neurogenin-1 and Neurogenin-2 over four days. Samples for RNA-seq were collected either destructively by lysing cells or non-destructively by sampling the culture media. Cell media timepoint measurements reflect the cumulative VLP-exported transcripts since the prior media collection. Brightfield and mNeon (green) images show cells at days zero and four of differentiation.

(B) Overlap in differentially expressed genes (DEGs) detected in cellular lysate and cell media samples. DEGs were defined as genes with log₂(fold change) > 1 or log₂(fold change) < −1 and false discovery rate (FDR)-adjusted P value < 0.01. Statistical significance of the overlap was calculated using a hypergeometric test with N = 10,436 total genes detected in the cell lysate (P < 0.0001).

(C) Scatterplot of fold changes in gene expression between day zero and day four for all detected genes (light purple) and DEGs detected in both cell lysates and cell media (dark purple). Pearson correlation coefficients (r) are presented in each plot.

(D) Heatmap of z-scored log-transformed TPMs. Genes ordered by decreasing log₂ fold change in cellular lysate.

(E) Gene expression of stem cell (top) and neuronal (bottom) markers across 4-day differentiation in cell lysate (top, mean ± SEM) and cell media (bottom).

(F) Line plots of normalized enrichment scores (NES) for select gene ontology (GO) terms describing stem cell states (left) and neuron development (right) across four days of differentiation. Gene set enrichment analysis (GSEA) was performed on ranked log₂ fold changes in gene expression between each time point and day zero from cell media.

We next performed differential gene expression analysis to identify temporally dynamic genes during iNGN differentiation. Comparing day zero and day four, 772 differentially expressed genes detected in cell lysate samples were equivalently identified in cell media samples from live cultures (Figure 2B, Table S4A-B). Additionally, the gene expression fold changes from cell media samples were concordant with those of cell lysate measurements (Figure 2C). Among the shared differentially expressed genes, we observed similar temporal progression of gene expression across all four days of differentiation in the cell lysate and cell media samples (Figure 2D). We then specifically compared the expression of stem cell and neuronal marker genes in cell lysate and media samples over the course of neurogenesis. As expected, we detected time-dependent attenuation in the expression of pluripotent stem cell marker genes DNMT3B, LIN28A, NANOG and PRDM14, and an increase in the expression of neuronal marker genes MAP2, RBFOX3 and TUBB3 throughout the differentiation (Figures 2E, S3I). The temporal dynamics of these individual pluripotent and neuronal genes were further reflected in global transcriptional changes of cell type-specific gene expression programs. Specifically, cell media samples captured gene programs involved in nucleic acid metabolism and stem cell division that were downregulated, while neural development-related gene expression programs were upregulated during differentiation (Figures 2F, S3J-L). Altogether, these data showed that serial sampling of VLP-containing cell media can elucidate the transcriptional dynamics of individual genes and gene programs from live self-reporting cells.

Self-reported RNAs capture acute transcriptional dynamics from primary cell spheroids

We next asked whether self-reported RNA could capture acute transcriptional changes. Live-cell assays are uniquely powerful for repeated analysis of structured biological systems that require preservation of the three-dimensional physical architecture of cell populations. In light of our prior finding that primary cells can constitutively self-report their transcriptional states (Figure 1J), we elected to study the acute transcriptional dynamics of primary cell spheroids in response to environmental stimuli (Figure 3A).

Fig. 3. Live-cell transcriptional time-course of HUVEC spheroid response to TNFɑ stimulation.

(A) Experimental scheme to measure transcriptional dynamics from self-reporting HUVEC spheroids stimulated with TNFɑ over 14 hours. Time-point measurements reflect the cumulative VLP-exported transcripts since the prior media collection. The t = 0h time-point reflects spheroids incubated over 19 hours prior to stimulation.

(B) Representative microscopy of HUVECs grown in monolayer (top) or spheroid (bottom) formats. Cells were fixed and immunofluorescently stained for nuclei (DAPI), CD31, and MMLV Gag p30 protein.

(C) Gene expression log₂(fold-changes) between t = 14h and t = 0h time-points detected using cell media samples from HUVECs grown in monolayers (x-axis) and spheroids (y-axis). Leading edge genes from GSEA analysis using the Hallmark “TNFɑ signaling via NFκB” gene set (Fig. S4F) are highlighted.

(D) Expression of significantly up-regulated and down-regulated genes over time detected from cell media of Gag-expressing HUVEC spheroids (P adjusted < 0.05). Each point represents a replicate (n=3). TPM = transcripts per million.

(E) Expression of all statistically significant (P adjusted < 0.05) DEGs detected from cell media samples that are unique to spheroids compared to monolayer grown cells. Each column represents a replicate. Expression values are Z-score standardized across rows.

We engineered primary human umbilical vein endothelial cells (HUVECs) to express Gag via lentiviral transduction, and validated that Gag-expressing HUVECs faithfully self-reported their transcriptional states with comparable technical performance in spheroid and monolayer culture formats (Figures 3B, S4A-D). We then stimulated HUVECs with TNFɑ and sampled VLP-containing cell media, which captured substantial transcriptional changes within 14 hours of stimulation (Figure S4E). In particular, genes canonically regulated by NFκB in response to TNFɑ were highly up-regulated (Figure 3C) and to comparable degrees between spheroid and monolayer cultures (Figure S4F), as expected. More generally, we identified several temporal DEGs with varying effect sizes using VLP-containing cell media from self-reporting HUVEC spheroids (Figure 3D) and 113 DEGs were unique to spheroids compared to cells grown in 2D monolayers (Figure 3E). Several genes involved in cell adhesion and cytoskeleton (ICAM1, ACTG1, and VAPA), as well as transcriptional and epigenetic response to inflammation (IL8, TNFAIP2, KLF7, BRD3, TRIM56, BCL6B, and CCL5), significantly varied over time in VLP-containing media from spheroids, but not monolayer cultures, suggesting the relevance of the culture format and physical architecture of cell populations for the identification of biological effects in this system. These data collectively support the feasibility of self-reported temporal live-cell transcriptomics from three-dimensionally structured cell populations and to capture acute transcriptional dynamics.

Epitope-functionalized VLPs enable multiplexed and selective self-reporting

To further extend the capabilities of self-reporting beyond whole-population readouts, we sought to resolve cell-type-specific transcriptomes from heterogeneous cell populations. An advantage of VLPs is their amenability to pseudotyping with surface-displayed proteins. We hypothesized that pseudotyping VLPs with envelope proteins displaying unique epitope tags would enable multiplexed capture of VLPs produced by different cell populations (Figure 4A). To test this hypothesis, we utilized a common envelope protein derived from the vesicular stomatitis virus (VSV-G), inserted a FLAG tag near the N-terminus and transfected it alongside Gag into HEK293T cells to produce pseudotyped VLPs. FLAG immunoprecipitation (IP) of the cell media enabled epitope-dependent isolation of VLPs (Figure S5A), and comparable performance with HA-tagged VSV-G and HA IP of the cell media (Figure S5B). RNA-seq of the IP output revealed reads mapping to the human transcriptome with enrichment for exonic regions (Figures S5C-D).

Fig. 4. Deconvolution of cell-type-specific VLPs in cell co-cultures with epitope-tagged envelope proteins.

(A) Architecture of VLPs pseudotyped with engineered envelope proteins containing epitope-tag barcodes.

(B) Experimental design to isolate cell-type-specific VLPs from cell co-cultures via immunoprecipitation of epitope-tag barcodes. Gag-expressing HEK293T and HT1080 cells were lentivirally transduced with FLAG or HA-tagged VSV-G expression vectors (schematically depicted), respectively. Self-reporting HEK293T and HT1080 cells were pooled and cultured together for 48 hours to allow for VLP export into the cell media. The cell media from the co-cultures was then subjected to FLAG and HA IPs to isolate HEK293T and HT1080 VLPs, respectively.

(C) Western-blot validation of cell type-specific isolation of VLPs from cell line co-cultures via epitope-tag based IPs on the cell media.

(D) Quantification of the number of genes detected by RNA-seq from cell lysate and IP-purified VLPs from cell media.

(E) CellNet classification of samples for embryonic kidney and fibrosarcoma GRNs. Asterisks indicate an adjusted P value < 0.05, computed with a two-tailed Wilcoxon rank sum test and corrected for multiple hypothesis testing.

We then assessed the performance of simultaneously measuring live-cell transcriptomes from different cell types grown in co-culture. We stably introduced FLAG-tagged VSV-G into Gag-expressing HEK293T cells and HA-tagged VSV-G into Gag-expressing HT1080 cells (Figure 4B). We verified the epitope-specificity of each IP protocol through cross-over experiments, confirming that each protocol specifically isolated the desired epitope in the presence of other epitopes (Figure S5E). We then co-cultured the epitope-tagged cell lines, and following 48 hours of VLP production, subjected cell media samples to epitope-specific IPs (Figures 4C-D). The resulting RNAs isolated from FLAG and HA IPs reflected substantial enrichment of the HEK293T and HT1080 cell transcriptomes, respectively (Figure 4E), demonstrating multiplexed, cell type-resolved self-reporting from mixed cell populations.

Finally, we identified that human membrane proteins, such as the truncated nerve growth factor receptor (tNGFR), can also enable epitope-dependent isolation of VLPs from cell media when engineered with a FLAG tag in the topological domain (Figures S5F-G). While VSV-G is commonly utilized to pseudotype lentiviral particles – given its broad tropism – we observed no lentiviral transduction with FLAG-tNGFR pseudotyped particles, similar to negative controls with non-pseudotyped particles (Figures S5H-I). This finding suggests that the application of non-viral envelope proteins can limit cellular uptake of surface-engineered VLPs, and we recommend engineering VLPs that lack a capability for transduction in self-reporting applications.

Gag-RNA binding domain fusions modulate the repertoire of VLP-packaged RNA

We extended our VLP engineering efforts to explore whether and how the self-reported transcriptional profiles depend on the RNA binding properties of VLP components. We hypothesized that VLP-packaged RNA could be modulated by complementing the intrinsic RNA binding capacity of the MMLV Gag NC domain with C-terminal fusions to human RNA binding domains (RBDs) (Figure 5A).

Fig. 5. Engineering MMLV Gag with RNA binding proteins.

(A) Schematic of VLPs derived from Gag proteins engineered with C-terminal fusions to RNA binding proteins.

(B) AlphaFold2-predicted structure of engineered Gag-eIF4E protein, whereby the 217-amino acid human eIF4E protein (UniProt ID: P06730) was fused to the C-terminus of MMLV Gag via a 5x(GS) linker.

(C) AlphaFold3-predicted structure of wild-type eIF4E (left) or eIF4E containing W56L, W102L, and W166L mutations (right), each in complex with guanosine-5′-diphosphate (GDP).

(D) Scatterplots of normalized gene expression measured from cellular lysate and cell media. HEK293T cells were transfected with plasmid vectors to express Gag, Gag-eIF4E, or Gag-eIF4E triple mutant (TM) and cell media was collected after 48 hours of VLP production. Each point represents a gene detected from RNA-seq. Pearson correlation coefficients for each gene class are noted in each plot. TPM = transcripts per million.

(E) MA plots visualizing the degree of quantitative agreement between RNA-seq measurements on cell lysate and cell media for Gag, Gag-eIF4E, or Gag-eIF4E TM protein variants. Statistically significant DEGs (P adjusted < 0.05) are highlighted in blue. RNA count normalization was performed with DESeq2.

(F) Quantification of the number of statistically significant up and down-regulated genes between cell media and cell lysate from (E).

(G) CellNet classification scores for cell media samples from cells expressing H2B-GFP, wild-type Gag, Gag-eIF4E and Gag-eIF4E TM.

(H) ROC analysis for DEG detection from wild-type Gag or Gag-EIF4E cell media samples. HEK293T-versus-HeLa DEGs from cell media were used to predict the DEG fold change from matched cell lysate RNA-seq. AUC = area under the curve.

Prior reports demonstrated the utility of MS2 hairpin aptamers to enrich the packaging of specific RNA transcripts in VLPs.⁴⁰ We built upon this approach by modifying the popular CROP-seq sgRNA expression vector⁵⁰ to enable live-cell readouts of functional genomics experiments. The introduction of MS2 hairpins into the polyadenylated transcript produced by the CROP-seq system (Figure S6A) and the fusion of the MS2 coat protein (MCP) to Gag (Figure S6B) resulted in MS2- and MCP-dependent enrichment of CROP-seq transcripts in cell media (12-fold enrichment relative to cell lysate and 10-fold enrichment relative to wild-type Gag cell media, Figures S6C-F). We then delivered a pooled sgRNA CROP-seq library to cells and developed protocols to read out the abundances of sgRNAs from cell media analogous to a pooled genetic screen (Figures S6G-I, Table S5). Gag-MCP yielded sgRNA abundances that were better-correlated with the cellular library composition than wild-type Gag (Pearson r = 0.822 versus 0.477, Figure S6J), demonstrating how the self-reporting system can be tailored for accurate live-cell readout of specific RNA transcripts, including sgRNA libraries.

We then asked whether this Gag-RBP approach could generalize beyond targeted synthetic transcripts to broader classes of endogenous RNA. Cellular mRNA characteristically contain a singular 5′ cap modification with the enzymatic addition of a N⁷-methylguanosine (m⁷G). eIF4E is the principal and evolutionarily conserved protein that recognizes and binds the 5′ m⁷G cap to initiate translation.⁵¹ Thus, we fused eIF4E to the C-terminus of Gag (Figure 5B), and found this fusion protein to be amenable for producing VLPs in HEK293T cells (Figure S7A). Importantly, tryptophans 56, 102 and 166 within eIF4E are known to establish π stacking and Van der Waals interactions with m⁷G based on crystallographic structural studies.⁵² We modeled eIF4E bound to guanosine-5′-diphosphate (GDP) using AlphaFold3⁵³, and, guided by the structure and prior literature, mutated the three tryptophans to aliphatic leucines to ablate the aromatic interactions with m⁷G to produce a triple mutant (termed Gag-eIF4E TM) for use as a cap binding loss-of-function control (Figure 5C). Cell media samples from Gag-eIF4E expressing HEK293T and HeLa cells exhibited higher concordance of protein-coding transcripts to cell lysate than cell media from wild-type Gag-expressing cells across independent experiments and different cell types (Figures 5D, S7B). We did not observe this improvement with Gag-eIF4E TM cell media, suggesting that the higher transcriptional concordance observed in protein-coding transcripts from Gag-eIF4E VLPs could be attributed to 5′ m⁷G cap binding (Figure 5D). There were 1.47-fold more genes detected on average from Gag-eIF4E cell media compared to wild-type Gag cell media, and higher rates of NGS reads that mapped to exons (Figure S7C). This finding suggests different RNA packaging preferences for Gag-eIF4E compared to wild-type Gag. In support of this hypothesis, we found that protein-coding genes detected from Gag-eIF4E cell media exhibited higher concordance to cell lysate (r = 0.863) compared to non-coding genes (r = 0.658), and genes detected by Ribo-seq⁵⁴ were globally more abundant in Gag-eIF4E cell media compared to wild-type Gag cell media (Figure S7D). Collectively, these data are consistent with the role of eIF4E in translation initiation on mature, spliced mRNAs and self-reporting readouts focused on gene transcripts actively translated in the cytosol.

Next, we characterized the performance of Gag-eIF4E for live-cell transcriptional profiling applications and first noted that the expression of Gag-eIF4E, similar to that of wild-type Gag, was minimally perturbative to cells (Figures S1L, S7E). We quantified transcriptional differences between cell media and cell lysate measurements across Gag protein variants (Figure 5E). Gag-eIF4E resulted in approximately half as many DEGs between cell media and cell lysate (557 genes significantly upregulated, and 476 significantly downregulated), compared to wild-type Gag (988 genes significantly upregulated, and 1,203 significantly downregulated) (Figure 5F), and corresponded with improved cell-type classification performance using cell media samples (P = 0.0122, Figures 5G, S7F). Additionally, DEGs from cell media samples with Gag-eIF4E were more concordant with DEGs detected from cell lysate measurements (Gag ROC AUC: 0.641, Gag-eIF4E ROC AUC: 0.860, Figure 5H). Taken together, these Gag protein engineering results show that Gag-RBP fusions can tune the representation of RNA molecules in VLPs for live-cell transcriptional profiling applications.

Live-cell transcriptomics from enclosed microphysiological systems identifies stroma-specific signatures of endothelial vasculogenesis

Finally, we leveraged our VLP engineering efforts to demonstrate the enabling capabilities of live-cell transcriptomics with self-reported RNAs. The ability to monitor transcriptional trajectories in real time is particularly beneficial for microphysiological systems (MPS), which are uniquely positioned to facilitate disease modeling and drug screening.⁵⁵ For example, MPSs offer superior physiological relevance for disease modeling since they can incorporate self-assembled vascular networks.⁵⁶ Endothelial and perivascular stromal cells subjected to biophysical cues in MPS devices can recapitulate the complex, multi-cellular architectures of human vasculature.⁵⁷ Transcriptomic profiling could aid in parsing mechanistic responses to interventions and benchmarking self-assembled microvasculature to in vivo data; however, the difficulty of retrieving cells from within the microfluidic device renders traditional destructive methods impractical for longitudinal study. To overcome this challenge, we leveraged our ability to garner self-reported transcriptomes from primary human endothelial cells (Figure 3) as a solution to monitor the molecular development of engineered vascular networks.

Here, we investigated how distinct perivascular environments influence endothelial vasculogenesis. We selected two contrasting stroma, normal human lung fibroblasts (NHLFs) that are commonly used to maintain a homeostatic environment, and patient-derived endometrial stromal cells (ESCs) that are derived from the remodeling-intensive environment of the endometrium. We engineered HUVECs to constitutively express Gag-eIF4E, seeded the cells with either stromal population into MPS devices containing a vascular-supportive synthetic hydrogel⁵⁷, and subjected the co-cultures to continuous fluid flow for 7 days to capture the transition from initial endothelial sprouting to the formation of lumenized networks (Figures 6A-B). We further verified that constitutive Gag-eIF4E expression did not impede endothelial network formation (Figures 6B, S8A). Daily sampling of the perfused media from three independent devices per stroma condition and sequencing of VLP-containing cell media yielded longitudinal transcriptomes with high signal-to-background (Figure S8B). While we detected over 10,000 genes per sample from day three onward, gene counts in the first 48 hours were approximately 20% lower, a possible consequence of physical factors influencing VLP recovery prior to substantial cellular remodeling of the MMP-cleavable hydrogel and the formation of a lumenized vascular network.

Fig. 6. Transcriptional dynamics of vasculogenesis in microphysiological systems.

(A) Schematic of microphysiological system (MPS) experiments. Primary human umbilical vein endothelial cells (HUVECs) were engineered to constitutively express Gag-eIF4E and mNeon as a translation reporter via a P2A ribosomal skipping peptide. HUVECs were mixed with either normal human lung fibroblasts (NHLFs) or endometrial stromal cells (ESCs) within a synthetic hydrogel and seeded into MPS devices. Perfused media contains VLPs produced by HUVECs.

(B) Live-cell transcriptomics time-course of vasculogenesis in MPS devices. Cells were seeded into MPS devices on day zero as in (A) and cultured for seven days under continuous fluid flow. VLP-containing perfused media was completely collected every 24 hours and RNA-seq libraries were prepared to read out transcriptional dynamics of HUVECs. Time-point measurements therefore reflect the cumulative VLP-exported transcripts since the prior media collection. Representative mNeon microscopy is presented to visualize vascular network formation from mNeon+ HUVECs co-cultured with NHLFs. The scale bar reflects 500 μm.

(C) PCA of cell media RNA-seq samples from MPS devices for both NHLF and ESC stroma conditions. PCA was performed using variance-stabilizing transformed RNA-seq counts on all samples, but visualized separately for each stroma condition. Lines connect the same MPS device across each time point.

(D) Heatmap of 1,472 genes with temporal dynamics shared between NHLF and ESC stroma conditions identified by a joint GAM. Rows represent significant genes (P adjusted (time) < 0.05 and P adjusted (time, condition) > 0.05) and columns reflect model-predicted, VST-normalized expression across a 100-point time grid. The mean trajectory across all genes within each cluster is depicted to the left of the heatmap and shaded ribbons reflect +/− one standard deviation.

(E) Hallmark gene sets with temporal dynamics shared between NHLF and ESC stroma conditions (P adjusted (time) < 0.05 and P adjusted (time, condition) > 0.05). Single-sample GSEA (ssGSEA) scores were computed for each sample and analyzed using a joint GAM to assess temporal significance. Fitted ssGSEA scores are shown across a 100-point time grid, with shaded ribbons indicating 95% confidence intervals.

(F) Temporal modules of genes exhibiting dynamics differing between NHLF and ESC stroma conditions. Left, Pearson correlation matrix of the difference in GAM-predicted trajectories (NHLF – ESC) for significant genes. The matrix is clustered to identify four modules. Middle, heatmaps of model-predicted expression across a 100-point time grid for NHLF and ESC stroma conditions. Genes (rows) are ordered by the Pearson correlation matrix. Right, mean trajectories across all genes in each module for each stroma condition. Shaded ribbons reflect +/− one standard deviation.

(G) Gene programs with temporal dynamics differing between NHLF and ESC stroma conditions (P adjusted (time, condition) < 0.1). ssGSEA scores and shaded ribbons were computed equivalently as in (E).

We first performed PCA to parse transcriptional information garnered from cell media, which captured variance associated with transcriptional changes over time and between stroma conditions (Figures 6C, S8C-D). Motivated by the temporal variation in the data, we then implemented generalized additive models (GAMs) to identify genes that significantly varied over time and exhibited stroma-dependent dynamics (Table S6). The non-destructive nature of self-reported transcriptional measurements enabled a repeated-measures statistical design that increased statistical power by controlling device-to-device variability, as evident from reduced pergene residual variance compared with an independent-measures statistical design (Figure S8E). Cellular self-reporting also reduced experimental resources to capture a transcriptional time-course, requiring only three MPS devices per stroma type, with repeated sampling of each of the six devices over the seven-day time course. Conventional destructive approaches would have required 42 independent MPS devices to capture the same time-course, which would be prohibitively resource intensive and difficult to scale with additional conditions.

The improved performance of the repeated-measures model enabled us to identify 1,472 significant genes exhibiting shared temporal dynamics between stroma conditions compared to 1,363 genes detected with an independent-measures model. The cohort of genes with shared dynamics reflected a core vasculogenic signature, including known regulators such as KLF2, FGF23, TCF4, FZD4, and EDN1 (Figures 6D, S8F). This analysis also nominated a poorly annotated gene, CCDC85B, that may exhibit yet-to-be-appreciated effects on endothelial vasculogenesis. Clustering the GAM-predicted trajectories revealed six gene modules characterized by distinct gene activation or repression kinetics (Figure 6D). Furthermore, temporally dynamic genes reflected known signaling pathways involved in vasculogenesis, such as Hedgehog, Wnt/β-catenin, PI3K/AKT/mTOR, and Notch, as well as transcriptional programs involved in endothelial functions, including migration, extracellular matrix assembly, and response to laminar fluid shear stress (Figures 6E, S8G).

We then extended the GAM framework to identify genes with differential dynamics between perivascular stroma environments, which amounted to 232 significant genes. Of these genes, we clustered the difference in GAM-predicted trajectories between stroma conditions to reveal four modules of rewired gene expression dynamics (Figure 6F). Several metabolic genes (COX6B1, ATP5L, ATP5G1, ATP5E) and cytoskeletal regulators (ARPC3, PFN1, TMSB4X, VIM) were characteristic of Module 2, which exhibited temporal attenuation in the NHLF condition yet increased over time in the ESC condition (Figures 6F, S8H-I). Genes with differentially phased dynamics across stroma conditions (Modules 1 and 3) included the angiogenesis-modulating transcription factors ZEB1 and ATOH8, and the soluble factors MIF and IGF2. These gene modules were further supported by the differential dynamics of gene programs associated with those genes, such as oxidative phosphorylation (OxPhos) and reactive oxygen species (ROS) pathways (Figures 6G, S8D).

The temporal dynamics of gene programs also pointed to functional differences between perivascular stromal environments. For example, the significant rise of a TGFβ signaling program in the NHLF condition, which was generally invariant in the ESC condition, combined with the significant increase in an endothelial barrier establishment gene program in the NHLF condition (Figure 6G), suggested that the lung fibroblast niche drove HUVECs from active sprouting to a more functionally mature transcriptional state. Conversely, the sustained temporal increase in MYC, ROS, and OxPhos programs in the ESC condition relative to the NHLF condition, and differential onset of vessel remodeling programs (Figure 6G), suggest a proliferative and metabolically active vascular network consistent with the high cellular turnover characteristic of the uterine microenvironment.

Collectively, these data and our analytical framework demonstrated that live-cell transcriptomics with cellular self-reporting enabled longitudinal measurements within otherwise inaccessible MPSs and the discovery of divergent temporal logic employed by primary endothelial cells navigating different physiological niches.

DISCUSSION

In this work, we developed cellular self-reporting for longitudinal, transcriptome-wide profiling of dynamic biological processes from living cells. Non-destructive live-cell self-reporting enables gene expression to be monitored by RNA-seq across timepoints and subsequent complementary studies on the same living biological sample. Self-reported readouts are particularly advantageous for analyzing post-mitotic cell types and tissues, and advanced biological models such as organoids and embryoid bodies. In these settings, high culture-to-culture variability often renders sub-sampling strategies difficult to implement or weakly powered. By comparing transcriptomes from the same sample, self-reporting enables a repeated-measures experimental design to bolster statistical inferences of gene expression dynamics across time points.⁵⁸ In addition to transcriptome-wide profiling of RNA native to the host cell, we showed that self-reporting enabled detection of specific RNA subsets, such as CROP-seq sgRNAs suitable for CRISPR screening with readout of dynamic transcriptional phenotypes. We expect this longitudinal approach to be highly advantageous for in vivo CRISPR screens, particularly when cells are inaccessible and effect timescales are unknown or variable. Future work will focus on pairing perturbation identities with the VLP-exported transcriptome.

The indexing, recording, and storage of transcriptional state information or transcripts themselves in cells has been an active area of technology development for more than a decade. These concepts enable new applications and insights not accessible with standard transcriptional readouts, in some cases showing useful performance at the single-cell level.^{26,30,31,59,60} Cellular self-reporting of mRNA transcriptome-wide, here realized using a VLP export approach, represents a distinct paradigm with overlapping and differentiated sets of use cases for the research community to explore. The large dynamic range in time, selectivity for cell populations of interest, compatibility with barcode export, availability of living cells after sampling exported transcriptomes, compatibility with genetic, biological, optical, or chemical control of self-reporting across time and space, and the ability to collect cell-free samples distant from the reporting cells are key features of self-reporting likely to support new types of gene expression profiling applications.

We demonstrated the versatility of the cellular self-reporting paradigm through engineering of VLP components and the applicability to structured biological systems. MPSs, in particular, have garnered increased attention to aid in drug discovery and development campaigns. The utilization of self-reporting in MPS studies can provide high-dimensional, temporally-resolved transcriptomic readouts to parse responses to drugs, and to facilitate further refinement of disease models. More broadly, we expect that cellular self-reporting will find wide application in biological research⁶¹, including potential applications to report on other analytes such as proteins and metabolites. Altogether, the work presented here establishes cellular self-reporting as an attractive tool for live-cell transcriptomics, enabling the study of dynamic biological processes across a spectrum of biological contexts and new experimental designs.

LIMITATIONS OF THE STUDY

While cellular self-reporting offers broad potential for dynamic transcriptional profiling, several limitations warrant careful consideration. The quantity of self-reported RNA available for library construction is dependent on several factors including the number of VLP-producing cells, the VLP production rate per cell, and sampling interval length. These factors collectively can limit sensitivity for lowly expressed genes (i.e. rare transcript detection) during short self-reporting periods (i.e. time resolution) and in small samples (i.e. single cells). While we demonstrated transcript detection from individual exporting cells, detecting lowly expressed transcripts with high time resolution in small samples will be challenging, and engineering high VLP production rates to improve sensitivity may affect the biological system under study. Most of the results in this study were produced from populations of more than 100,000 cells with VLP accrual times of at least 12 hours, and recovery of the entire VLP-containing culture medium.

While conventional RNA-seq relies on cell lysis to capture an instantaneous snapshot of the cellular transcriptome, VLP-exported RNA measures cytosolic-biased gene expression integrated over a sampling time interval. The time for the VLP export process sets limits on the achievable time resolution and sensitivity for transient or burst-like dynamic signals. Also, VLPs collected at one time point from culture media were produced by the cell within an extended time interval, increasing the dispersion in transcript age typically underlying RNA-seq data. Additionally, VLP collection comprises a media change which may alter the cellular states and properties observed at later times if not part of a standard protocol. We reason that this issue could be mitigated by fractional sampling of the culture media or replacing sampled media with conditioned culture media.

STAR METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Cell culture

Immortalized cell lines, such as HEK293T (ATCC cat. no. CRL-3216), HT1080 (ATCC cat. no. CCL-121) and HeLa cells (a gift from Iain Cheeseman at the Whitehead Institute) were cultured in a humidified incubator at 5% CO₂ at 37°C using high-glucose DMEM (Invitrogen cat. no. 10566016) supplemented with 10% FBS and 1% penicillin/streptomycin.

Human iPS cells were cultured on plasticware coated with hESC-qualified matrigel (Corning cat. no. 354277), according to the manufacturer's instructions. The 1157.2 human iPSC line (Boston Children’s Hospital Stem Cell Core) was maintained in StemFlex medium (Stem Cell Technologies cat. no. A3349401), whereas the iNGN iPSC line (a gift from George Church at Harvard Medical School) was cultured in mTeSR medium (Stem Cell Technologies cat. no. 85850). iPSCs were propagated as colonies using ReleSR (Stem Cell Technologies cat. no. 100-0483) upon reaching 70% confluency and cultured in a humidified incubator at 5% CO₂ and 37°C.

Stromal fibroblasts were isolated from proliferative phase eutopic endometrial biopsies and all participants provided informed consent in accordance with a protocol approved by the Partners Human Research Committee and the Massachusetts Institute of Technology Committee on the Use of Humans as Experimental Subjects. Primary isolated endometrial stromal fibroblasts were routinely cultured at 37°C and 5% CO₂ in phenol red-free DMEM/F12 media supplemented with 1% penicillin/streptomycin and 10% v/v dextran/charcoal treated fetal bovine serum (Atlanta Biologicals). Stromal fibroblasts were split 1:2 once they reached 70-80% confluency, and the medium was replaced every 2-3 days.

Primary human umbilical vein endothelial cells (HUVECs) for spheroid experiments in Fig. 3 were purchased from Lonza (cat. no. C2519A, Lot: 23TL223958) and propagated on 0.1% gelatin-coated tissue-culture plates (Stem Cell Technologies cat. no. 07903) in M199 medium (Sigma cat. no. M4530) supplemented with 10% FBS, 20 mM HEPES (Invitrogen cat. no. 15630080), GlutaMAX (Gibco cat. no. 35050061), 100 ug/mL heparin (Stem Cell Technologies cat. no. 07980), 10 ng/mL VEGF (R&D Systems cat. no. 293VE500), 10 ng/mL bFGF (Gibco cat. no. 13256029), 10 ng/mL EGF (R&D Systems cat. no. 236-EG-200), and 10 ng/mL IGF-I (PeproTech cat. no. 100-11). HUVECs for MPS experiments in Fig. 6 were purchased from Angio-Proteomie (cat. no. cAP-0001, Lot: 2021020301) and cultured in VascuLife medium (Lifeline Cell Technology cat. no. LL-0003) on tissue-culture plates coated with Quick Coating Solution (Angio-Proteomie cat. no. cAP-01). HUVECs were seeded at 2,500 cells/cm², propagated upon reaching ~70% confluency and cultured at 37°C and 5% CO₂.

METHOD DETAILS

Molecular cloning

All plasmid vectors were designed in Benchling and cloned via Gibson Assembly using NEB’s Gibson Assembly Master Mix (NEB cat. no. E2611L), according to the manufacturer’s protocol unless otherwise stated. The Moloney MLV Gag ORF (UniProt ID: P03332) was synthesized by IDT as a gBlock and cloned in frame with a P2A-GFP sequence into the pcDNA3 vector in order to generate the plasmid expression vector, pGag. As a negative control for VLP formation, we also created a Gag truncation mutant where the MA domain was deleted in order to abolish VLP maturation. We PCR dialed-out amino acids 216-530 from the MLV Gag ORF gBlock and equivalently cloned it in frame with a P2A-GFP sequence into the pcDNA3 vector in order to generate plasmid, pGagΔMA. We fused mCherry to the C-terminus of Gag via a 3x glycine-serine linker and cloned the ORF into pcDNA3 to create plasmid, pGag-mCherry. We utilized the pLenti vector to create all lentiviral expression plasmids. We PCR dialed out the Gag-P2A-GFP ORF from pGag and cloned it in frame with P2A-ZeoR into pLenti, downstream of a CAG promoter in order to create the tricistronic vector, pLV_CAG_Gag. We performed the comparable PCR dial out from pGagΔMA in order to create pLV_CAG_GagΔMA. Lentiviral vectors expressing Gag fusion proteins were generated by cloning ORFs encoding RNA binding proteins into pLV_CAG_Gag at the C-terminus of Gag with a 5x glycine-serine (GS) linker. The eIF4E ORF was PCR dialed out from a whole-transcriptome amplification library in order to create pLV_CAG_Gag-EIF4E. We created a triple mutant variant of eIF4E by designing Gibson Assembly primers to substitute the tryptophans 56, 102 and 166 for leucines then PCR amplified the eIF4E cDNA ORF with said primers and assembled the fragments via Gibson Assembly to generate pLV_CAG_Gag-EIF4E-TM. Epitope-tagged VSV-G constructs were generated utilizing the VSV-G ORF from pMD2.G (a gift from D. Trono, Addgene #12259) and inserting a FLAG (DYKDDDDK) or HA (YPYDVPDYA) sequence following the signal peptide at amino acid position 23. FLAG or HA tags were inserted into tNGFR after the signal peptide and the ORFs were synthesized by Twist Biosciences. The epitope-tagged ORFs were cloned into pcDNA3 or pLenti backbones with a 2A-mCherry-2A-PuroR cassette to derive plasmid or lentivirus vectors, respectively. Finally, piggyBac vectors for expression in iPS cells were generated with a piggyBac transposon backbone (System Biosciences) containing a CBX3 ubiquitous chromatin opening element (UCOE) upstream of an EF1ɑ promoter driving expression of Gag or GagΔMA in frame with a P2A ribosomal skipping peptide and mNeonGreen reporter.

All Gibson Assembly reactions were purified with a DNA Clean & Concentrator column (Zymo cat. no. D4033) prior to transformation into NEB Stable Competent E. coli (cat. no. C3040H). Transformed E. coli were grown overnight at 30°C on agar plates with 50 μg/mL carbenicillin selection. Individual colonies were picked for liquid culture in LB media supplemented with 100 μg/mL ampicillin and plasmid DNA was subsequently isolated using a QIAGEN Plasmid Plus Midi Kit with endotoxin removal. Plasmid sequences were fully verified by whole plasmid sequencing (GeneWiz or Primordium Labs, Inc). Plasmids have been deposited with Addgene.

Transient transfection of self-reporting constructs

Unless otherwise stated, HEK293T cells were seeded into tissue culture-treated 6-well plates at a density of 100,000 cells/cm² (1 million cells per well in 2 mL of media). The following day, the media was removed and replenished with 1 mL of fresh media. The cells were subsequently transfected with 3000 ng of total DNA using Lipofectamine 3000 (ThermoFisher cat. no. L3000015) per the manufacturer’s instructions. For non-pseudotyped VLP experiments, 3000 ng of H2B-GFP (Addgene #67928), GagΔMA, MMLV Gag (wild-type), or engineered MMLV Gag variants were transfected. For VLP pseudotyping experiments, 3000 ng of total plasmid consisting of a Gag expression plasmid, VSV-G/tNGFR expression plasmid and pUC19 plasmid (2:3:4 ratio by mass) was transfected. The media was changed 6 hours post-transfection and was collected for downstream processing 48 hours post-transfection.

VLP purification from cell media

VLP-containing cell media from transient transfection experiments or stable Gag-expressing cells was first processed by centrifugation at 1,000 g for 5 minutes at 4°C. Next, the media was further filtered through a 0.45 μm low protein binding filter (VWR cat. no. 28145-481) and subsequently concentrated with an Amicon 100 kDa MWCO centrifugal filter (Millipore Sigma cat. no. UFC810096) at 2500 g for 15 minutes at 4°C. The retentate within the filter was then either frozen at −80°C for storage or used immediately for subsequent assays.

As a technical aside, we have been able to successfully prepare RNA-seq libraries from concentrated cell media samples that have been stored at −80°C for several months. We also observed that low-serum media (such as VascuLife for endothelial cell culture) results in a greater degree of media concentration with Amicon filters compared to media containing 10% FBS, and is more conducive for SPRI-based purification of RNA directly from TCL-lysed media concentrate. We thus recommend, when possible, to utilize low-serum media especially in experiments with limited VLP-producing cells (~100,000) to maximize the RNA input for SMART-Seq2 library construction.

For FLAG-based immunoprecipitation of VLPs, Anti-FLAG M2 Magnetic Beads (Sigma cat. no. M8823) were first washed three times with TBS (50 mM Tris HCl, 150 mM NaCl) buffer. The beads were resuspended in 500 μL TBS + 1% Tween-20 and incubated with concentrated VLP-containing cell media on a rotisserie at 4°C overnight. The next day, the beads were washed three times with TBS + 1% Tween-20. VLPs were eluted from the beads eluted in 500 ng/uL 3x FLAG peptide (Sigma cat. no. F4799) at 1200 RPM for 30 minutes at 4°C. For HA-based immunoprecipitation of VLPs, Pierce Anti-HA Magnetic Beads (ThermoFisher Scientific cat. no. 88836) were washed three times with TBS + 0.05% Tween-20, resuspended in TBS + 1% Tween-20 and incubated with VLP-containing cell media on a rotisserie at 4°C overnight. The beads were then washed three times with TBS-T and elution occurred with 1 μg/μL of HA peptide (ThermoFisher Scientific cat. no. 26184) in 1X TBS at 1200 RPM for 20 minutes at 37°C. The eluted VLPs were used directly in downstream assays or stored at −80°C.

Microscopy for intracellular localization of MMLV Gag proteins

HEK293T cells were seeded into a 24-well glass-bottom culture plate at a density of 100,000 cells/cm². The following morning, each well was transfected with 165 ng of plasmid encoding a Gag-mCherry fusion protein and 330 ng of pUC19 using Lipofectamine 3000, per the manufacturer’s instructions. 48 hours following the start of the transfection, the cells were fixed with 4% PFA for 15 minutes, permeabilized with 0.2% Triton-X for 20 minutes, and blocked with 1% BSA for 30 minutes at room temperature. The samples were then stained with mouse anti-EEA1 antibody (BD Biosciences cat. no. 610457, RRID: AB_397830, 1:500 dilution) in 1% BSA for 2 hours at room temperature. The cells were then incubated with goat anti-mouse IgG Alexa Fluor 647 (ThermoFisher Scientific cat. no. A21235, RRID: AB_2535804, 1:500 dilution) and phalloidin Alexa Fluor 488 (ThermoFisher Scientific cat. no. A12379, 1:500 dilution) in 1% BSA for 30 minutes at room temperature. The cells were then stained with 100 ng/mL DAPI in 2X SSC and imaged at 40X magnification.

Scanning electron microscopy

VLPs were produced from HEK293T cells via plasmid transfection of MLV Gag expression plasmids, as described in “Transient transfection of self-reporting constructs.” After 48 hours-post transfection, the cell media was centrifuged at 2000 g for 15 minutes to remove debris followed by filtration through 0.45 μm cellulose-acetate filters (VWR cat. no. 28145-481). The cleared media was then ultracentrifuged at 100,000 g for 2 hours using a 20% sucrose cushion. Carbon-coated copper mesh grids were coated in 1 μg/mL Alcian Blue solution. Grids were washed with ultrapure water, dried and then incubated at room temperature with VLP-containing supernatants. SEM micrographs were acquired on a Zeiss Crossbeam 540 scanning electron microscope.

Temporal stability of RNA in VLPs

We assessed the temporal stability of RNA within VLPs by first producing VLPs via transient transfection of MMLV Gag into HEK293T cells with n=3 transfection replicates, as described above. VLP-containing cell media was first centrifugated at 1000 g for 5 minutes and then filtered through 0.45 μm cellulose acetate filters. The media was then incubated at 37°C for 4 days and sampled every 24 hours to quantify GAPDH expression via RT-qPCR. RNA was extracted from cell media samples using a Qiagen QIAamp Viral RNA Mini Kit and then treated with DNase I (NEB cat. no. M0303S) for 10 minutes at 37°C. The DNase I reaction was inactivated with 5 mM EDTA and then subsequently cleaned up with 1X RNAClean XP SPRI beads (Beckman Coulter cat. no. A63987). 500 ng of RNA per sample was reverse transcribed using an iScript cDNA Synthesis kit (Bio-Rad cat. no. 1708890). GAPDH expression was quantified with a sensitive TaqMan-based qPCR assay using a commercial primer/probe set from ThermoFisher Scientific (Hs02758991_g1). Briefly, we first made a PCR standard by gel extracting the 93-bp PCR amplicon produced from the GAPDH primer/probe set using a 2% agarose gel and DNA Gel Extraction kit (Zymo cat. no. D4007). The concentration of gel-extracted GAPDH amplicon was quantified with a QuBit high sensitivity dsDNA assay and then diluted in nuclease-free water to create 10-fold serial dilutions ranging from 1e9 to 1e2 molecules/uL. Standards, cDNA samples and no-template controls were assayed by RT-qPCR in technical triplicate using JumpStart Taq DNA polymerase ReadyMix (Sigma cat. no. P2893), with Rox as a background control dye and the following thermal program: 94°C for 2 minutes, and 35 cycles of 94°C for 30 seconds, 60°C for 30 seconds and 72°C for 30 seconds. The PCR standard curve exhibited high technical reproducibility and a coefficient of determination of R² = 0.9962. GAPDH copy number expression for each technical replicate was determined from the PCR standard curve and expression for each sample is the arithmetic mean across all technical replicates. To determine temporal stability, the copy number expression of GAPDH at each time-point was normalized to the starting Day 0 copy number baseline expression.

Nuclease treatment of VLP-containing cell media

HEK293T cells were transfected in 6-well plates with 3000 ng of plasmid encoding Gag, as described in “Transient transfection of self-reporting constructs.” The cells were incubated for 48 hours to allow for VLP production, after which the cell media was collected in 2 mL nuclease-free microcentrifuge tubes, centrifuged at 1000g for 5 minutes, then 90% of the media was filtered through a low-protein binding 0.45 μm filter (Avantor cat. no. 28143-312) into new 2 mL nuclease-free microcentrifuge tubes and frozen at −80°C until ready for subsequent processing. Amicon 100 kDa filters (Millipore Sigma cat. no. UFC810024) were washed with 4 mL of sterile PBS at 1500g for 5 minutes to remove trace amounts of glycerine on the filter surface. The cell media was thawed from −80°C storage, topped off to 4 mL total volume with sterile PBS and concentrated at 1500g in 5 minute increments until ~500 μL of retentate volume remained. Each sample was normalized to 840 μL total volume with sterile PBS. We utilized 140 uL of sample to treat with an RNase cocktail (ThermoFisher Scientific cat. no. AM2286, 1 uL per reaction) or PBS at 37°C for 15 minutes with three replicates per treatment condition. Total RNA was isolated from 1 million HEK293T cells using an RNeasy Plus Mini kit (Qiagen cat. no. 74134) to use as a “free RNA” control for RNase activity and equivalently treated +/−RNase as the cell media samples. Free RNA samples were purified with RNAClean XP SPRI beads and cell media samples were purified with a QIAamp Viral RNA Mini Kit (Qiagen cat. no. 52904). We performed RT-qPCR for GAPDH as described in “Temporal stability of RNA in VLPs.” GAPDH expression was then normalized to the mean of the PBS treatment samples for free RNA and cell media conditions.

Protein structure predictions

We utilized ColabFold, a Google cloud implementation of AlphaFold2 ⁶² to predict the structure of engineered variants of MLV Gag and epitope-tagged envelope proteins. For all predictions, amino acid sequences were queried as monomers in AlphaFold2_mmseqs2 (v1.5.5) with default model/MSA settings and num_recycles set to 6. To model the interaction of the 5′ mRNA cap with either wild-type eIF4E or eIF4E exhibiting W56L, W102L, W166L mutations, we utilized AlphaFold3 ⁵³, implemented through AlphaFold Server (alphafoldserver.com), with the amino acid sequence of the proteins and guanosine-5′-diphosphate as a ligand. The resulting PDB file for each predicted protein structure was visualized with PyMOL (v2.5.7).

Lentivirus production

HEK293FT cells were seeded at 1 million cells/well in 6-well plates. The following day, cells were transfected when 90-95% confluent with pMD2.G (Addgene #12259), psPAX2 (Addgene #12260), and a lentiviral transfer plasmid (2:3:4 ratio by mass, 3 μg total plasmid) using Lipofectamine 3000 (ThermoFisher Scientific cat. no. L3000015). Media was exchanged after 6 hours and viral supernatant was harvested 48 hours after transfection and filtered through 0.45 μm filters (VWR cat. no. 28145-481). The purified lentivirus was frozen at −80°C until needed for infection.

Generation of constitutive MMLV Gag expressing cells

Constitutively self-reporting HEK293T cells, HT1080 cells and primary stromal fibroblasts were generated by lentiviral transduction of MMLV gag. For some experiments, cells were then subsequently infected with lentivirus to introduce epitope-tagged envelope proteins. Lentiviruses were produced using plasmids pLV_CAG_GagΔMA or pLV_CAG_Gag (as described above). Cells were then transduced with lentiviruses in the presence of 8 μg/mL polybrene in order to achieve ~30% fluorescent reporter-positive cells. The cells were incubated in virus for 24 hours followed by antibiotic selection 48 hours post-transduction at the following concentrations: 1 μg/mL puromycin (ThermoFisher Scientific cat. no. A1113803), and 200 μg/mL zeocin (ThermoFisher Scientific cat. no. R25001). Cells were under antibiotic selection until uninfected control cells died (~5 days for puromycin and ~9 days for zeocin).

Stable Gag-expressing iPS cells were generated through piggyBac transposition to mitigate transgene silencing, which is commonly observed with lentiviral vectors. iPS cells were dissociated into single cells with TrypLE reagent (ThermoFisher Scientific cat. no. 12604013) and 800,000 cells were nucleofected with 2.5 μg of Super piggyBac Transposase (SBI cat. no. PB200A-1) and 10 μg of transposon plasmid using Lonza Cell Line Nucleofector Kit V (Lonza cat. no. VCA-1003) on an Amaxa IIb Nucleofector with program B-016. iPS cells post- nucleofection were cultured with ROCK inhibitor, RevitaCell (ThermoFisher Scientific cat. no. A2644501) for 24 hours and sorted one week later using a Sony MA900 FACS to enrich for mNeon+ cells. Cells were submitted for karyotypic analyses (Cell Line Genetics), where 20 metaphase cells were analyzed using G-band karyotyping.

CellTiter-Glo time-course assay

We utilized CellTiter-Glo assays to determine cell growth kinetics from constitutively self-reporting cells. First, 5,000 cells were seeded per well in Nuncleon white polystyrene 96-well assay plates (ThermoFisher Scientific cat. no. 136101), avoiding edge wells. Each condition was prepared in technical triplicate, and replicate plates were prepared at the start of the experiment for each timepoint in the time-course. The CellTiter-Glo assay (Promega cat. no. G7572) was performed according to the manufacturer’s protocol on a BioTek Synergy Neo microplate reader. The raw luminescence signal for each timepoint was normalized to the Day 0 baseline signal.

RNA-sequencing with SMART-Seq2

RNA from cell media or IP-purified VLPs was extracted using 2X TCL Buffer (Qiagen cat. no. 1070498), whereas RNA from cells was extracted with 1X TCL Buffer (Qiagen cat. no. 1031576). Lysed samples were frozen at −80°C until ready for RNA-seq library construction. RNA-seq libraries were prepared using the SMART-Seq2 protocol, as previously published ⁴⁸ with some modifications. Briefly, lysed samples were cleaned with 2.2X RNAClean XP SPRI beads (Beckman Coulter cat. no. A63987). RNA was eluted from SPRI beads directly into a master mix, comprising the SMART-Seq2 reverse transcription primer, dNTPs (NEB cat. no. N0447L) and RNase inhibitor (ThermoFisher Scientific cat. no. AM2694). The eluted RNA was incubated at 72°C for 3 minutes to denature secondary structure and immediately followed by incubation on ice for 2 minutes. A reverse transcription master mix was then added, comprising Maxima RNase H-minus RT (ThermoFisher Scientific cat. no. EP0751), Betaine (Sigma cat. no. B0300-1VL), a template switching oligo (Exiqon), RNase inhibitor (ThermoFisher Scientific cat. no. AM2694) and MgCl2. Reverse transcription proceeded with the following thermal program: 50°C for 90 minutes then 85°C for 5 minutes. The cDNA was then directly whole-transcriptome amplified with KAPA HiFi HotStart ReadyMix (Roche cat. no. 07958935001) using an ISPCR primer according to the following thermal program: 98°C for 3 minutes, 21 cycles of 98°C for 15 seconds, 67°C for 20 seconds, and 72°C for 6 minutes, and a final extension step of 72°C for 5 minutes. The amplified cDNA was cleaned with 0.8X AMPure XP SPRI beads (Beckman Coulter cat. no. A63881). Ten nanograms of DNA was tagmented at 58°C for 10 minutes in a 10 μL reaction containing 2 μL of 5X tagmentation buffer (50 mM Tris-HCl, 25 mM MgCl2 pH 8.0), 2 μL of Tris Buffer (10 mM Tris-HCl, 1% Tween-20 pH 8), and 4 μL Nextera (Illumina). The reaction was stopped with 1% SDS and incubated at 72°C for 10 minutes, then 4°C for 3 minutes. The tagmented library was cleaned with 1X AMPure XP SPRI beads followed by 12 cycles of PCR with NEBNext High Fidelity polymerase (NEB cat. no. M0541L) to incorporate Illumina indices and flow cell handles. The libraries were assessed using a High Sensitivity DNA kit (Agilent cat. no. 5067-4626) on an Agilent Bioanalyzer, then pooled, diluted and denatured for sequencing on either Illumina or Aviti sequencers. Paired-end reads were aligned to the hg19 human reference genome, unless otherwise stated. We assessed the concordance of gene expression measurements with various alignment tools, specifically STAR (version 2.6.0c) ⁶³ and kallisto (version 0.46.0) ⁶⁴. For kallisto-based analyses, the hg19 cDNA fasta reference from the UCSC Genome Browser was appended with transgene sequences to create a custom kallisto index via the “kallisto index” command. Experiment-specific analyses are detailed in experiment-specific method sections.

Gradient boosting framework to model VLP measurement biases

To understand gene- and sequence-level features that contribute to differences between VLP cell media and bulk cell lysate RNA-seq measurements, we first generated VLPs from HEK293T cells (as described in “Transient transfection of self-reporting constructs” and “VLP purification from cell media”) then prepared RNA-seq libraries on cell lysate and cell media samples (as described in “RNA-sequencing with SMART-Seq2”).

Raw sequencing reads were aligned to the human reference genome (GRCh38, GENCODE release 47) using STAR (v2.7.10a). Gene-level quantification was performed using the “--quantMode GeneCounts” command-line option during STAR alignment. We assembled a gene-level expression matrix for all samples using unstranded counts, which is appropriate for SMART-Seq2 libraries. We then performed differential gene expression analysis using DESeq2 (v1.50.2) ⁶⁵ in R by contrasting the cell media samples against the cell lysate samples. Log₂(fold-changes) (LFCs) were stabilized using the apeglm shrinkage estimator (v1.32.0) to account for the inherent noise in low-abundance transcripts. Genes with a shrunken ∣LFC∣ > 2 and a Benjamini-Hochberg-adjusted P-value < 0.05 were retained for downstream machine learning.

To identify the determinants of RNA packaging bias, we constructed a high-dimensional feature matrix using the GenomicFeatures (v1.62.0) ⁶⁶ and Biostrings (v2.78.0) packages in R. For each gene, we prioritized the Ensembl Canonical transcript as defined in the GENCODE release 47 GTF. In cases where a canonical transcript was not designated, features were calculated as the mean across all annotated transcripts. We extracted structural metrics, including cDNA (sum of all exons), genomic span (TSS to TES), total intron counts and splice ratio (ratio of cDNA length to genomic span). We also extracted compositional metrics, including global cDNA GC content, and GC content of annotated 5’ and 3’ untranslated regions (UTRs). For genes that lacked annotated UTRs, we used the global cDNA GC values instead. For sequence motif features, we calculated the frequencies of all possible 64 trinucleotides (3-mers) in the spliced transcript sequence. Finally, we integrated subcellular RNA localization data by analyzing cytoplasmic and nuclear fractions from publicly available HEK293T RNA-seq datasets (GSE195493) ⁶⁷. These data were processed identically to our cell lysate and cell media samples, yielding a shrunken LFC representing the cytoplasmic-to-nuclear partitioning coefficient for each gene. The compiled set of features are documented in Table S2A.

We implemented a gradient boosting framework using XGBoost (v3.1.3.1) ⁶⁸ in R and the assembled feature matrix. The model was trained to predict the shrunken LFC between cell media and cell lysate as a continuous outcome. The following parameters were utilized for model training with the xgb.train() function: objective = “reg:squarederror”, eta = 0.02, max_depth = 10, gamma = 0.01, subsample = 0.98, colsample_bytree = 0.86. To assess model generalization, 10-fold cross-validation was performed. The dataset was partitioned into 10 non-overlapping subsets using the caret package (v7.0-1) in R; in each iteration, the model was trained on 9 subsets and validated on the 10th. Model performance was quantified by calculating the Pearson correlation coefficient between the experimentally derived LFC and the values predicted by the model across all validation folds (mean r = 0.67, standard deviation = 0.025). The model was then trained on all genes in the features matrix and the contribution of each feature to the model’s predictions was interpreted using Shapley additive explanations (SHAP). SHAP values were calculated for all features across all genes using the SHAPforxgboost package (v0.2.0) to determine the global importance (mean absolute SHAP) and the directionality of the impact (Table S2B). Finally, feature importance was visualized using beeswarm summary plots for gene-level and sequence-motif features.

Cell type classification and gene regulatory network inference with CellNet

We utilized CellNet ⁴⁹ to derive active gene regulatory networks (GRNs) from RNA-seq samples of cell lysate and VLPs, and to assess the ability of VLP-derived RNAs to reflect the GRNs in the lysate of the host cell. We retrained the CellNet classifier to incorporate GRNs reflective of HEK293T, HT1080, HeLa or primary stromal fibroblast cells. First, we quantified transcript abundances in cell lysate samples with cn_salmon() using a pre-prepared Salmon transcript index, salmon.index.human.122116.tgz available from https://github.com/pcahan1/CellNet. We constructed new cell-type specific GRNs with cn_make_grn() using samples from the June 20, 2017 edition of the human CellNet Processor (https://github.com/pcahan1/CellNet) and the cell lysate samples. We assessed the random forest classifier using cn_splitMakeAssess() and generated a new CellNet processor object with cn_make_processor(). We then applied CellNet with this retrained classifier using default settings to cell lysate and cell media RNA-seq samples, and plotted the sample classification scores in R.

VLP sampling interval experiments

We assessed VLP-derived transcriptomes as a function of the amount of time allowed for VLP production. First, 1.2 million cells were seeded into T75 flasks per replicate. For these experiments, we utilized HT1080 cells engineered to constitutively express MMLV Gag and HA-VSVG via lentiviral transduction or uninfected control cells. We assessed VLP production intervals of 1, 3, 6, 12, 24 and 48 hours. To account for cell division during the VLP production window, we staggered the start of the collection windows so that the terminal time-points were the same across all intervals. Starting at a minimum of 24 hours post-seeding (e.g. in the case of the 48-hour production interval), we performed a 100% media change. At the terminal time-point, 100% of the cell media was collected, processed, subjected to HA immunoprecipitation and lysed with 2X TCL buffer, as described in “VLP purification from cell media.” Cells were trypsinized at the terminal time-point for each production interval condition and lysed with 1X TCL buffer. SMART-Seq2 libraries were generated for cell lysate and HA immunoprecipitated cell media samples, as described in “RNA-sequencing with SMART-Seq2.” RNA-seq paired-end reads were aligned to the hg19 human reference genome using STAR aligner (version 2.6.0c) ⁶³. BAM alignment files were produced using STAR with command line options –outSAMtype BAM SortedByCoordinate –twopassMode Basic and –outSAMstrandField intronMotif. RNA-SeQC ⁶⁹ was run using STAR-produced BAM files and an hg19 transcriptome GTF file retrieved from the UCSC Genome Browser to determine sample quality and alignment metrics. The number of detected transcripts and genes per sample were plotted from RNA-SeQC quantifications. CellNet was re-trained to include fibrosarcoma cell lysate samples and applied to all samples. The fibrosarcoma GRN scores were then plotted as a function of the VLP production interval time.

Single cell export rate measurements

HT1080 cells were engineered to constitutively express MMLV Gag via lentiviral transduction with vector pLV_CAG_Gag. We sorted single cells and 100-cell minibulk populations from the top quartile of the GFP expression distribution into 384-well plates (Corning cat. no. 8794BC) containing 10 μL of media per well using a Sony SH800Z cell sorter. The plate was centrifuged at 200 g for 1 minute at room temperature after FACS to pellet cells and then transferred to a 37°C incubator. After 24 hours of culture, brightfield and fluorescence images were acquired using a Nikon Ti2 microscope for each single-cell sorted well to verify the presence of a viable GFP+ cell. Wells with no cells or with more than one GFP+ cell were excluded from downstream sample processing. Cell media was then harvested from each well, lysed with 2X TCL buffer (Qiagen cat. no. 1070498), and imaged again to determine if cells were inadvertently aspirated as a result of cell media collection. The corresponding cell lysates were prepared by directly lysing cells within the well using 10 μL of 1X TCL buffer (Qiagen cat. no. 1031576). RNA-seq libraries were generated using SMART-Seq2 and uniquely indexed, as previously described. For single-cell samples, we only retained coupled cell lysate and media RNA-seq libraries for wells where a single cell was identified in both rounds of imaging during sample collection. The final set of libraries were pooled and sequenced on an Illumina sequencer.

Paired-end sequencing reads were pseudoaligned with kallisto (v0.43.1) ⁶⁴ to a custom reference generated by concatenating the MMLV Gag ORF to the human hg19 reference transcriptome. Single cell and 100-cell minibulk samples for which the cell media measurements contained over 5% of the mitochondrial gene counts detected in the matched cell lysate, and samples for which the cell lysate had fewer than 1,000 unique genes detected were excluded from the analysis. This conservative filtering removed samples which may have had cell lysate contamination in the collected cell media. To compare samples and account for varying sequencing depths, total transcript counts after pseudoalignment were then subsampled to 20,000. Transcripts detected in cell media were counted only if they were also present in the corresponding cell lysate. Transcript-level counts were then aggregated to gene-level counts for downstream analyses and plotting.

Immunoblotting for VLPs

Cell media was collected and processed as described in the section “VLP purification from cell media.” The cells were then washed once with PBS and incubated with TrypLE (Gibco) in order to create a single cell suspension. An aliquot of cells was resuspended in a protein extraction solution (150 mM NaCl, 1% Triton X-100, 50 mM Tris-HCl pH 7.5) supplemented with a protease inhibitor (Sigma cat. no. 4693159001) for 30 minutes at 4°C on a rotisserie. The cell lysate was centrifuged at 10,000 g for 5 minutes at 4C, after which 90% of the supernatant was transferred to a new microcentrifuge tube and frozen at −80°C or used directly for immunoblotting. Cell protein lysate, concentrated cell media, and/or IP-purified VLPs were reduced using 2X Tris-Glycine SDS Sample Buffer (Life Technologies cat. no. LC2676) and 10X NuPAGE Reducing Agent (Life Technologies cat. no. NP0009) at 98°C for 5 minutes. The reduced samples were then run on a Novex WedgeWell 8-16% Tris-Glycine Gel (Life Technologies cat. no. XP08165BOX) for 50 minutes at 225V and 400 mA. The gel was then transferred to a PVDF membrane (Life Technologies cat. no. IB24002) using an iBlot 2 device (Life Technologies). The PVDF membrane was blocked in 5% milk (Bio-Rad cat. no. 1706404XTU) for 1 hour at room temperature with gentle shaking, followed by an overnight incubation at 4°C with gentle shaking in primary antibodies against Gag at 1:2000 dilution (Abcam cat. no. ab100970, RRID: AB_10696216) and Actin at 1:5000 dilution (Abcam cat. no. ab179467, RRID: AB_2737344). The next day the membrane was washed three times with 5% milk and incubated for 4 hours at 4°C in an anti-rabbit secondary antibody (ThermoFisher Scientific cat. no. A32733, RRID: AB_2633282, 1:10,000 dilution). In experiments where we utilized GagΔMA-expressing cells as a VLP control, we stained for actin with a primary mouse antibody (ThermoFisher Scientific cat. no. MA5-11869, RRID: AB_11004139, 1:400 dilution) and anti-mouse secondary (ThermoFisher Scientific cat. no. A11001, RRID: AB_2534069, 1:10,000 dilution) to distinguish the similarly sized proteins. The membranes were imaged on a Azure Biosystems C600 Imaging System and analyzed in Fiji (v2.14.0/1.54f) ⁷⁰.

Cell-type-specific VLP purification from co-cultures using epitope-tagged envelope proteins

HEK293T and HT1080 cells were engineered to constitutively express either MMLV Gag or GagΔMA via lentiviral transduction. After selection and expansion, HEK293T cells were transduced with pLV_FLAG-VSV-G, whereas HT1080 cells were transduced with pLV_HA-VSV-G. VLP production and the specific purification of FLAG- and HA-tagged VLPs were determined in each cell line via Western-blotting for MMLV Gag p30 protein. HEK293T cells expressing MMLV Gag and FLAG-VSV-G were then co-cultured at a 1:1 ratio with HT1080 cells expressing MMLV Gag and HA-VSV-G in T75 flasks. After 72 hours of co-culture, cell lysates and cell media were collected. The cell media was divided into aliquots and subjected to either FLAG- or HA-based immunoprecipitation, as described in “VLP purification from cell media”, to specifically purify VLPs from HEK293T or HT1080 cells, respectively. Western-blotting and RNA-seq were performed as previously documented. RNA-seq reads were aligned with STAR (version 2.6.0c) ⁶³ to the hg19 reference genome using command line options –outSAMtype BAM SortedByCoordinate – twopassMode Basic and –outSAMstrandField intronMotif. Gene and transcript alignment metrics were determined with RNA-SeQC ⁶⁹ using STAR-produced BAM files and an hg19 transcriptome GTF file retrieved from the UCSC Genome Browser. CellNet was re-trained to include embryonic kidney and fibrosarcoma cell lysate references as described above and then applied to all samples.

Cellular transduction with pseudotyped VLPs

We assessed the ability of VLPs pseudotyped with various envelope proteins to transduce wild-type HEK293T cells. First, 1 million HEK293T cells were plated per well in a 6-well plate and the following day were transfected with plasmids via Lipofectamine-3000, per the manufacturer’s instructions. In total, 3000 ng of DNA was transfected per well in a 2:3:4 ratio by mass of envelope plasmid, Gag plasmid (psPAX2, Addgene #12260) and H2B-GFP transfer plasmid (pHR_dSV40-H2B-GFP, Addgene #67928), respectively. As a negative control, we transfected 3,000 ng of H2B-GFP only. As a non-pseudotyped VLP control, we substituted the envelope plasmid for pUC19 within the three-plasmid cocktail. We used vectors pFLAG-VSVG and pPB_EF1a_FLAG-tNGFR_ZeoR to express FLAG-VSV-G and FLAG-tNGFR, respectively. The cell media was collected 48 hours post-transfection, centrifuged at 1,000 g for 5 minutes then 0.45 um filtered with cellulose acetate filters. We quantified the titer of each condition using a Lenti-X GoStix Plus assay (Takara cat. no. 631280). For each condition, the processed media was titrated by volume (10, 50, 100, and 500 uL) and administered to 20,000 wild-type HEK293T cells per well in a 24-well plate. We performed three independent transductions per VLP dose. Flow cytometry was performed for GFP 48 hours later and wild-type HEK293T cells were utilized as a factor-minus-one control to set a positive gate.

Live-cell transcriptomics of iPS-to-neuron differentiation

We employed the human iNGN iPSC line (courtesy of Dr. George Church, Harvard Medical School) expressing the transcription factors Neurogenin-1 and Neurogenin-2 under the control of a doxycycline-inducible promoter ⁷¹ to study the transcriptional dynamics of neurogenesis. iNGN iPSCs were cultured on tissue culture plates coated with hESC-qualified matrigel (Corning cat. no. 354277) and grown in mTeSR media (Stem Cell Technologies, cat. no. 85850) with daily media changes. Cells were propagated as colonies with ReLeSR (Stem Cell Technologies, cat. no. 100-0483) upon reaching 70% confluency. iNGN iPSCs were engineered to constitutively express MMLV Gag via piggyBac transposition using transposon plasmid pPB_EF1a_Gag, as described in “Generation of constitutive MMLV Gag expressing cells”.

iNGN iPSCs were differentiated to bipolar-like neurons by first seeding 45 colonies/cm² in a 15-cm dish on Day -1 in 20 mL of media. We equivalently plated cells in multiple wells of a 24-well plate in order to destructively sample a well at each time-point of the differentiation for cell lysate control measurements. The following day (Day 0), the media was collected, the cells were washed with PBS to remove residual VLPs from the prior time-point, and then 20 mL of mTeSR media containing 0.5 μg/mL doxycycline (Sigma, cat. no. D9891-5G) was administered to initiate differentiation. Cells were destructively sampled for cell lysate control measurements by first washing the cells once with PBS, followed by singularization with TrypLE and lysis with TCL buffer (Qiagen). This process was repeated every 24 hours over 4 days of differentiation. Cell surface expression of NCAM1 was assayed via flow cytometry at the end of the differentiation (BioLegend cat. no. 362546, RRID: AB_2565964, 1:100 dilution, staining for 20 minutes at room temperature). Two independent differentiation experiments were performed. The cell media was processed by first incubating the crude media with 100 U/mL benzonase at 37°C for 60 minutes with periodic mixing every 15 minutes. The media was centrifuged at 1000 g for 5 minutes, filtered through a 0.45 μm filter and then concentrated with a 100 kDa Amicon filter (Milipore Sigma, cat. no. UFC910096) at 2500g for 20 minutes at 4°C. The retentate was lysed with 2X TCL buffer (Qiagen) and immediately frozen at −80°C. RNA-seq libraries were prepared with for cell lysate and cell media samples (as described in “RNA-sequencing with SMART-Seq2”) and sequenced on an Illumina NextSeq in read configuration 75-8-8-75.

The hg19 UCSC reference transcriptome was appended with the coding sequences of mouse Neurog1 and Neurog2, and MMLV Gag to generate a custom kallisto index using the “kallisto index” command. RNA-seq paired-end reads were pseudoaligned to the custom index with kallisto (v.0.46.0) ⁶⁴. Transcript-level estimates were imported into R and summed to gene-level estimates using the tximport package (version 1.28.0) ⁷². Differential gene expression analysis was performed using DESeq2 (version 1.44.0) ⁶⁵ on the gene-summarized estimated count matrix from kallisto, filtering for genes with at least 10 reads in 2 or more samples. Statistically significant genes between Day 0 and subsequent time-points were identified using an FDR based P value adjustment method and alpha = 0.05. We defined genes as significantly differentially expressed if FDR < 0.05 and log₂(fold-change) > 1 or log₂(fold-change) < −1. Gene set enrichment analysis ⁷³ was performed using GSEApy ⁷⁴ with genes ranked by the DESeq2 log₂(fold change). CellNet was applied to RNA-seq data from all time-points using the June 20, 2017 edition of the human CellNet Processor (available from https://github.com/pcahan1/CellNet) and without any further modifications. Embryonic stem cell GRN scores were then visualized as a function of time for cell lysate and cell media RNA-seq samples. All other downstream analysis was performed with custom Python scripts.

TNFɑ stimulation of self-reporting HUVEC spheroids

HUVECs were first engineered to produce VLPs via lentiviral transduction of MMLV Gag using vector pLV_CAG_Gag. HUVECs were cultured for one passage following the initial thaw then seeded into 10-cm, gelatin-coated dishes at 5,000 cells/cm² in the presence of 5 ug/mL polybrene (Santa Cruz Biotech cat. no. 28728-55-4) and lentivirus. We infected cells at high MOI to avoid the need for selection and minimize the number of cell culture passages. Lentivirus was removed 24 hours later, and transduction rates were assayed by flow cytometry for GFP 72 hours following the initiation of transduction. Each condition exhibited >88% GFP+ positive cells. We also verified endothelial identity by flow cytometry using antibodies targeting CD31 (BioLegend cat. no. 303123, RRID:AB_2562179, 1:100 dilution) and CD144 (BD Biosciences cat. no. 561567, RRID:AB_10712766, 1:100 dilution).

We formed MMLV Gag expressing HUVECs into spheroids based on established protocols ⁷⁵ with some modifications. First, three days prior to the initiation of spheroid formation the cells were transitioned to complete EGM-2 medium (Lonza cat. no. CC-3162). Cytodex-3 beads (Cytiva cat. no. 17048501) were reconstituted in PBS for 3 hours at room temperature and then autoclaved for 20 minutes at 125°C. On the day of spheroid formation, the beads were washed four times in EGM-2 medium. Next, 1 million cells were co-incubated with 2,500 beads in 1.5 mL of complete EGM-2 media at 37°C for 4 hours with gentle agitation every 20 minutes. The spheroids were then transferred to a 0.1% gelatin coated 6-well plate and cultured in complete EGM-2 media. In parallel, cells were equivalently grown in 2D monolayer format on 0.1% gelatin coated 6-well plates. After 19 hours (t = 0h time-point), the media was collected and processed, as previously described in “VLP purification from cell media”. The spheroids and monolayer cells were then stimulated with 30 ng/mL human TNFα (PeproTech cat. no. 300-01A-50UG). After another 14 hours (t = 14h time-point), the media was collected again and equivalently processed. RNA-seq libraries were prepared from processed cell media samples, as described in “RNA-sequencing with SMART-Seq2” and paired-end sequenced in a 76-8-8-76 read configuration on an Illumina MiniSeq.

Paired-end reads were pseudoaligned with kallisto (version 0.46.0) ⁶⁴ and gene-summarized count estimates were imported using the tximport package (v1.28.0) ⁷² in R (version 4.3.1). PCA was performed using the rlog() function on genes with at least 1 count across all samples. Differentially expressed genes between t = 14h and t = 0h time-points were identified with DESeq2 (v1.40.2) ⁶⁵ on the estimated counts matrix from kallisto. Statistically significant genes were defined as FDR-adjusted P value < 0.05 and log₂(fold-change) > 1 or log₂(fold-change) < −1. Gene set enrichment analysis for the Hallmark TNFa Signaling via NFkB gene set was performed using the clusterProfiler package (v4.8.2) and unfiltered gene lists ranked by descending DESeq2 log₂(fold-change).

Immunofluorescence microscopy of HUVEC spheroids

HUVECs cultured in monolayer format were stained in 24-well glass-bottom plates. HUVEC spheroids were cultured as described above and gently transferred to microcentrifuge tubes before being stained in suspension. All samples were fixed with 4% PFA for 30 minutes and permeabilized with 0.1% Triton-X for 15 minutes. Samples were stained with rabbit anti-Gag antibody (Abcam cat. no. ab100970, RRID: AB_10696216, 1:1,000 dilution) and Alexa Fluor 594 mouse anti-human CD31 antibody (BioLegend cat. no. 303126, RRID: AB_2563303, 1:200 dilution) in 3% BSA overnight at 4°C. Samples were then incubated with goat anti-rabbit IgG Alexa Fluor 647 (ThermoFisher Scientific cat. no. A32733, RRID: AB_2633282, 1:500 dilution) and donkey anti-mouse IgG Alexa Fluor 594 (ThermoFisher Scientific, RRID: AB_2556543, 1:1,000 dilution). Samples were stained with 100 ng/mL DAPI in 2X SSC and imaged at 20X and 40X magnification.

Evaluation of MMLV Gag fusion proteins

HEK293T and HeLa cells were transfected with MMLV Gag, Gag-eIF4E or Gag-eIF4E-TM vectors as described in “Transient transfection of self-reporting constructs.” Cell lysate and media samples were collected and processed 48 hours post-transfection, as described in “VLP purification from cell media” and RNA-seq libraries were prepared, as described in “RNA-sequencing with SMART-Seq2.” Paired-end reads were pseudoaligned with kallisto (v0.46.0) and gene-summarized count estimates were imported using the tximport package (v1.28.0). DEGs between cell lysate and media samples for each Gag construct were identified with DESeq2 (v1.40.2) and significance was defined if FDR-adjusted P value < 0.05 and ∣LFC∣ > 1. Sample quality and alignment metrics were also evaluated with RNA-SeQC on STAR-aligned BAM files. CellNet was re-trained to contain either HEK293T or HeLa cell lysate samples and then applied to cell lysate and cell media samples. For ROC analysis of MMLV Gag or Gag-eIF4E constructs, we first identified DEGs between HEK293T and HeLa cells for both cell lysate and cell media samples. Genes with FDR-adjusted P value < 0.05 and log₂(fold change) > 1 were defined as significantly up-regulated, and genes with FDR-adjusted P value < 0.05 and log₂(fold change) < −1 were defined as significantly down-regulated. HEK293T-versus-HeLa DEGs identified from cell lysate samples were considered the ground-truth. ROC analysis was then performed on the HEK293T-versus-HeLa log₂(fold changes) of genes from cell media samples using the evalmod() function in the precrec package (v0.14.5) in R.

Live-cell transcriptional time-course of vasculogenesis in microphysiological systems

HUVECs (Angio-Proteomie cat. no. cAP-0001, Lot: 2021020301) for MPS experiments were maintained in VascuLife medium (Lifeline Cell Technology cat. no. LL-0003), released with TrypLE and propagated on tissue culture plastic-ware coated with Quick Coating Solution (Angio-Proteomie cat. no. cAP-01). HUVECs were engineered to constitutively express Gag-eIF4E by lentiviral gene transfer using pLV-Gag-EIF4E-2A-mNeon-2A-PuroR. In brief, 1.5 million HUVECs were seeded into a 15-cm dish and incubated with lentivirus for 48 hours, after which cells were selected with 1 μg/mL puromycin (ThermoFisher Scientific cat. no. A1113803) for 5 days and then expanded for an additional passage.

To set up MPS experiments, Gag-eIF4E expressing HUVECs (Passage 7 or lower) were mixed with primary stromal cells in a 7:1 ratio within v-CS-ECM synthetic hydrogel and seeded into microfluidic MPS devices, as previously described ⁵⁷. We set up 3 device replicates per stroma condition and loaded 10 μL of cell-hydrogel mixture into each device, which amounted to 180,000 HUVECs and 25,714 NHLFs (Lonza cat. no. CC-2512, Lot: 20TL356516) or endometrial stromal cells per device. Devices were incubated for 30 minutes to allow gelation of the synthetic hydrogel before being connected to continuous flow, as previously described ⁵⁷. MPS devices were incubated in a 37°C, 5% CO₂ incubator under continuous fluid flow for 7 days and imaged daily using a 4X objective on a Keyence BZ-X Microscope. DNA low-bind microcentrifuge tubes were used as media reservoirs and perfused cell media (~1 mL) was completely collected and replenished every 24 hours. The cell media was centrifuged at 1,000g for 5 minutes and then 95% of the media was filtered through a low-protein binding 0.45 μm filter (Avantor cat. no. 28143-312). The cleared media was topped-off to 4 mL with sterile PBS, concentrated with 100 kDa Amicon filters (Millipore Sigma cat. no. UFC810096) at 2,500g to a final volume of ~25 μL and then frozen in nuclease-free microcentrifuge tubes at −80°C.

Following the completion of the time-course experiment, concentrated cell media samples were thawed on ice, then 10 μL of cell media was lysed with 10 μL of 2X TCL buffer and used immediately as input for SMART-Seq2 (as described in “RNA-sequencing with SMART-Seq2”). RNA-seq libraries were assessed using an Agilent High Sensitivity DNA Bioanalyzer, pooled and sequenced on an Aviti Sequencer using a 2x150 cycle High Output CloudBreak kit.

Paired-end sequencing fastq files were pseudoaligned with kallisto (v0.46.0) ⁶⁴ using a custom index consisting of the hg19 UCSC cDNA reference and the Gag-EIF4E-2A-mNeon-2A-PuroR transgene sequence. Transcript-level abundance estimates were aggregated to gene-level counts using the tximport package (v1.38.2) ⁷² in R. Gene-level counts were normalized and variance-stabilized using the vst() function from the DESeq2 package (v1.50.2) ⁶⁵. The VST-normalized expression matrix was subset for genes that have cumulative sums > 1 across all samples and then used as input for PCA with the plotPCA() function. Genes were ranked by PC1 or PC2 loadings for GSEA analysis with Hallmark gene sets using the fgsea package (v1.36.0) ⁷⁶ in R.

We employed Generalized Additive Models (GAMs) to identify genes with significant temporal expression profiles and differences between HUVECs grown in various stroma conditions (NHLF versus ESC). First, we utilized VST-normalized gene expression to account for the heteroskedastic nature of RNA-seq data and to satisfy the assumption of Gaussian error distributions for temporal modeling. We implemented a GAM framework with a difference-smooth specification to explicitly test for genes that vary differently between stroma conditions. For each gene i, a model was fit to relate the VST-normalized expression (y) to time (t), condition (C) and subject (S, e.g. MPS device replicate):

$$y_i={\beta }_0+{\beta }_{Ci}+f_{shared}(t_i)+f_{Ci}(t_i)+b_{Si}+{ϵ}_i$$

Where, β₀ is the global intercept, β_Ci is a parametric fixed effect representing the baseline difference in expression between conditions independent of time, f_shared(t_i) is a global smooth function representing the shared temporal program common to both conditions, f_Ci(t_i) is a factor-smooth interaction representing the condition-specific deviation from the global trend, b_Si is a random effect smooth accounting for the correlation of repeated-measures within the same biological subject, and ϵ_i is the residual error. Model parameters were estimated using Restricted Maximum Likelihood (REML). We implemented the model using the mgcv package (v1.9-3) in R as gam(expr ~ condition + s(time, k=5) + s(time, by=condition, k=5) + s(subject, bs=”re”), data, method=”REML”). P values for the parametric and smooth terms were extracted for each gene and adjusted for multiple hypothesis testing using the Benjamini-Hochberg procedure. Genes were classified as differentially dynamic (e.g. genes varying over time, but differently between stroma conditions) if P-adjusted_s(time) < 0.05 and P-adjusted_{s(time,by=condition)} < 0.05. Genes with condition-specific dynamics (e.g. genes that vary over time in one condition, but not the other) were defined as P-adjusted_s(time) ≥ 0.05 and P-adjusted_{s(time,by=condition)} < 0.05. Predicted gene trajectories were generated across a 100-point time grid from Days 1-7 using the model components, excluding subject-specific random effects (predict(fit, df.pred, exclude=”s(subject)”, type=”response”)). We then computed the difference in predicted gene trajectories between NHLF and ESC conditions to determine the magnitude and timing of divergence for differentially dynamic and condition-specific genes. The time of peak difference was defined as the time at which ∣f_ESC(t) - f_NHLF(t)∣ was maximized and we categorized the response timing as Early (≤Day 2), Intermediate (Day 2-6) or Late (≥Day 6). We then identified modules of genes with coordinated “rewiring” by calculating a symmetric Pearson correlation matrix for differentially dynamic and condition-specific genes using the difference in predicted gene trajectories between NHLF and ESC conditions. Clustering of the Pearson correlation matrix was performed using Ward’s D2 linkage and the stability of the clusters was assessed via bootstrapping (B = 100) using the fpc package (version 2.2-14). Heatmaps were plotted using ComplexHeatmap (v2.26.0) in R. We equivalently corroborated GAM results with a factor-smooth interaction implementation using gam(expr ~ condition + s(time, k=5) + s(time, condition, bs=”fs”, k=5) + s(subject, bs=”re”), data, method=”REML”). P values for smooth terms were adjusted for multiple hypothesis testing using the Benjamini-Hochberg procedure and shared dynamic genes between stroma conditions were defined as P-adjusted_s(time) < 0.05 and P-adjusted_{s(time,condition)} ≥ 0.05. To visualize trajectories of genes with varying magnitudes, predicted values were scaled to a [0,1] range based on absolute maxima and minima of the trajectory vector. Hierarchical clustering was performed using Ward’s D2 method on the first three principal components of the scaled NHLF trajectory matrix. This gene ordering was then applied to the scaled ESC trajectory matrix to generate heatmap visualizations of shared dynamic genes as in Fig. 6D. Six robust clusters were identified with 100 random bootstraps each with an average Jaccard similarity coefficient >0.7 across bootstraps.

Finally, we quantified the activity of Hallmark and GO BP gene sets in each sample over time using single-sample GSEA (ssGSEA). We utilized an R implementation of ssGSEA modeled from GSEApy (v1.1.9) and calculated ssGSEA scores for each gene set in each sample using the VST-normalized gene expression matrix. To determine the statistical significance of gene program dynamics, ssGSEA scores were modeled as a function of time and stroma condition using a factor-smooth interaction GAM framework, as previously described. Predicted gene program trajectories and 95% confidence intervals were generated from the GAM fits by excluding subject-specific random effects. All P values for time and time-condition interaction terms were corrected for multiple hypothesis testing using the Benjamini-Hochberg procedure.

Directed export of CROP-seq sgRNA transcripts in engineered VLPs

To facilitate export of specific RNA transcripts, we employed the MS2 hairpin aptamer system derived from the MS2 bacteriophage. First, the MS2 coat protein (MCP) (N55K variant) ORF was PCR amplified from MS2-P65-HSF1_GFP (Addgene #61423) ⁷⁷ and inserted with a serine-glycine linker via Gibson Assembly into pLV_CAG_Gag at the C-terminus of MMLV Gag in order to generate the plasmid pLV_CAG_Gag-MCP. We modified the CROPseq-Puro vector (Addgene #127458) ⁵⁰ by inserting a 24x tandem array of MS2 hairpins derived from phage-cmv-cfp-24xms2 (Addgene #40651) ⁷⁸ via restriction cloning in between the PuroR ORF and WPRE in order to generate the entry vector CROPseq-Puro-24xMS2.

We generated a 2,219 sgRNA library by subsampling the Broad Institute’s Dolcetto sgRNA library ⁷⁹. The library was synthesized as an oligo array (Agilent), minimally amplified with 12 rounds of PCR using KAPA HiFi HotStart ReadyMix (Roche cat. no. 07958935001) and purified with a DNA Clean & Concentrator column (Zymo cat. no. D4033). The amplified oligo pool was cloned into the CROPseq-Puro-24xMS2 entry vector or the unmodified CROPseq-Puro entry vector (termed CROPseq-Puro-0xMS2 in these experiments) at a 1:100 (vector:insert) molar ratio via Golden Gate Assembly with Esp3I (ThermoFisher Scientific cat. no. FD0454) and T7 ligase (Enzymatics cat. no. L6020L) with the following thermal program: 15 cycles of 37°C for 5 minutes then 20°C for 5 minutes. The Golden Gate reactions were purified via DNA precipitation with isopropanol and 500 ng of each library transformed into Endura ElectroCompetent Cells (Lucigen cat. no. 60242-2) following the manufacturer’s instructions. After transformation, Endura cells recovered for 1 hour at 37°C and then were grown for 16 hours at 30°C with 100 μg/mL ampicillin selection. Plasmid DNA was isolated using a Plasmid Plus Midi Kit (Qiagen cat. no. 12945).

To produce lentivirus, 22 million HEK293FT cells were seeded per 15-cm dish, and the following morning the cells were transfected with 12.1 μg pMD2.G (Addgene #12259), 18.15 μg psPAX2 (Addgene #12260) and 24.2 μg of the CROPseq plasmid library using Lipofectamine 3000 transfection reagent (ThermoFisher Scientific cat. no. L3000075) per the manufacturer’s instructions. Following 48 hours, the virus-containing cell media was collected, centrifuged at 1000g for 5 minutes, filtered through 0.45 μm filters (Avantor cat. no. 28143-312) and concentrated with a 100 kDa Amicon centrifugal filter (Millipore cat. no. UFC910024) at 2000g for 20 minutes at 4°C. Lentiviral particles were subsequently aliquoted and frozen at −80°C.

The CROP-seq libraries were introduced into HEK293T cells via reverse infection. First, HEK293T cells were singularized with TrypLE (Gibco cat. no.12604013) and 2 million cells were seeded into a T75 flask with the CROP-seq lentivirus pool (MOI = 0.25), supplemented with 8 μg/mL polybrene (Milipore Sigma cat. no. TR-1003-G). The cells were incubated with the lentivirus for 24 hours, after which 1 μg/mL of puromycin (ThermoFisher Scientific cat. no. A1113803) was administered for 5 days to select for infected cells. The puromycin-resistant cells were expanded and genomic DNA was extracted from a subset of cells using a custom extraction buffer (1 mM CaCl2, 3 mM MgCl2, 1 mM EDTA, 10 mM Tris-HCl, 1% Triton X-100, 0.2 mg/ml Proteinase K) and the lysate was subjected to the following thermal program: 65°C for 10 minutes followed by 95°C for 15 minutes. sgRNA barcode sequences were amplified from genomic DNA using NEBNext High-Fidelity 2X PCR Master Mix (NEB cat. no. M0541L) with 28 cycles of PCR and primers oDF-344_U6_NGS_P5 (5′-ACACGACGCTCTTCCGATCTTCTTGTGGAAAGGACGAAAC) and oDF-112_scaffold_NGS_P7 (5′-CTGGAGTTCAGACGTGTGCTCTTCCGATCAAGCACCGACTCGGTGCCAC). Illumina indices and flow cell handles were introduced into each amplicon library with 12 cycles of PCR and then sequenced on an Illumina MiniSeq (read configuration: 75-8-8-0).

We then assessed detection of CROP-seq sgRNAs within VLP-containing cell media. The puromycin-selected cell populations were seeded into 6-well plates (1 million cells per well) and transfected the following day with 3000 ng of expression vectors encoding either H2B-GFP, wild-type Gag, or Gag-MCP with Lipofectamine-3000, according to the manufacturer’s protocol. The cell media was collected and concentrated (as described in “VLP purification from cell media”) and RNA-seq libraries were prepared (as described in “RNA-sequencing with SMART-Seq2”) 48 hours post-transfection. The CROP-seq transcript was then PCR dialed-out from each sample’s SMART-Seq2 whole-transcriptome amplification library using NEBNext High-Fidelity 2X PCR Master Mix with 28 cycles of PCR and the same dialout primers listed above. Barcode amplicon libraries and RNA-seq libraries were sequenced on an Illumina MiniSeq (read configuration: 76-8-8-76).

A custom kallisto index was generated by appending the CROP-seq transcript to the hg19 UCSC reference transcriptome. To prevent misquantification of reads from conserved sequences in Gag-MCP and CROP-seq expression vectors, we took a conservative approach and only utilized the hU6 promoter sequence directly upstream of the sgRNA spacer sequence to append to the UCSC reference transcriptome. RNA-seq reads were pseudoaligned with kallisto (v.0.46.0)⁶⁴ to the custom index and the fold change in CROP-seq transcript expression between conditions was calculated using the normalized transcripts-per-million expression matrix. We quantified the abundance of barcodes from plasmid, genomic DNA and RNA-seq dialout libraries using custom scripts written in R. In brief, we scanned fastq reads for the “CACCG” motif upstream of the sgRNA spacer sequence and identified exact matches to sgRNA sequences in the library. Raw sgRNA barcode counts were normalized to the total read depth in each sample using the cpm() function in the edgeR package ⁸⁰. The representation of barcodes within each sample was determined by calculating a Gini AUC using the ineq package (v0.2-13) in R.

QUANTIFICATION AND STATISTICAL ANALYSIS

Quantification and statistical analyses were performed using R or Python. Details of statistical analyses are documented in the experiment-specific methods sections and figure legends. Experiments and analyses were performed unblinded to sample identities.

Supplementary Material

2

Table S1: Amino acid sequences of self-reporting constructs

Table S2A: Gene and sequence features for a gradient boosting regression model to understand VLP measurement biases

Table S2B: SHAP feature importance values from gradient boosting regression model

Table S3: Differential gene expression results between cell media and cell lysate measurements from constitutive MMLV Gag-expressing human cells

Table S4A: Differential gene expression analysis during iNGN differentiation detected from cell lysate: Day 4 versus Day 0

Table S4B: Differential gene expression analysis during iNGN differentiation detected from VLPs: Day 4 versus Day 0

Table S5: Counts matrix of CROP-seq sgRNAs detected from plasmid, genomic DNA and VLP libraries

Table S6A: Shared dynamic genes across stroma conditions from vasculogenesis time-course experiments in microphysiological systems

Table S6B: Differentially dynamic genes across stroma conditions from vasculogenesis time-course experiments in microphysiological systems

Table S6C: Gene program dynamics from vasculogenesis time-course experiments in microphysiological systems

Figures S1–S8

Tables S1–S6

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit Anti-Gag	Abcam	Cat# ab100970; RRID: AB_10696216
Rabbit Anti-Actin	Abcam	Cat# ab179467; RRID: AB_2737344
Goat anti-Rabbit IgG AF647	ThermoFisher Scientific	Cat# A32733; RRID: AB_2633282
Mouse Anti-Actin	ThermoFisher Scientific	Cat# MA5-11869; RRID: AB_11004139
Goat anti-Mouse IgG AF488	ThermoFisher Scientific	Cat# A11001; RRID: AB_2534069
Mouse anti-EEA1	BD Biosciences	Cat# 610457; RRID:AB_397830
Goat anti-Mouse IgG AF647	ThermoFisher Scientific	Cat# A21235; RRID: AB_2535804
Anti-Human CD56 FITC	BioLegend	Cat# 362546; RRID: AB_2565964
Anti-Human CD31 BV421	BioLegend	Cat# 303123; RRID:AB_2562179
Mouse Anti-Human CD144 AF647	BD Biosciences	Cat# 561567, RRID:AB_10712766
Mouse anti-human CD31 AF594	BioLegend	Cat# 303126; RRID: AB_2563303
Bacterial and virus strains
NEB Stable Competent E. coli	NEB	Cat# C3040H
Endura Electrocompetent E. coli	Lucigen	Cat# 60242-2
Biological samples
Human umbilical cord vein endothelial cells (HUVEC)	Lonza	Cat# C2519A
Human umbilical cord vein endothelial cells (HUVEC)	Angio-Proteomie	Cat# cAP-0001
Normal human lung fibroblasts (NHLF)	Lonza	Cat# CC-2512
Endometrial stromal fibroblasts (ESC)	Massachusetts Institute of Technology	N/A
Stromal fibroblasts	Massachusetts Institute of Technology	N/A
Chemicals, peptides, and recombinant proteins
DMEM	Invitrogen	Cat# 10566016
StemFlex Medium	Stem Cell Technologies	Cat# A3349401
mTeSR Medium	Stem Cell Technologies	Cat# 85850
VascuLife Medium	Lifeline Cell Technology	Cat# LL-0003
hESC-Qualified Matrigel	Corning	Cat# 354277
ReleSR	Stem Cell Technologies	Cat# 100-0483
RevitaCell	ThermoFisher Scientific	Cat# A2644501
EGM-2 Medium	Lonza	Cat# CC-3162
M199 Medium	Sigma-Aldrich	Cat# M4530
Heparin	Stem Cell Technologies	Cat# 07980
Recombinant Human VEGF	R&D Systems	Cat# 293-VE-500
Recombinant Human bFGF	Gibco	Cat# 13256029
Recombinant Human EGF	R&D Systems	Cat# 236-EG-200
Recombinant Human IGF-I	PeproTech	Cat# 100-11
Recombinant Human TNFα	PeproTech	Cat# 300-01A-50UG
Lipofectamine 3000 Transfection Reagent	ThermoFisher Scientific	Cat# L3000015
Phalloidin AF488	ThermoFisher Scientific	Cat# A12379
Paraformaldehyde (PFA, 16% w/v)	Electron Microscopy Sciences	Cat# 15710
Polybrene	Santa Cruz Biotech	Cat# 28728-55-4
Quick Coating Solution	Angio-Proteomie	Cat# cAP-01
Puromycin	ThermoFisher Scientific	Cat# A1113803
Zeocin	ThermoFisher Scientific	Cat# R25001
Esp3I	ThermoFisher Scientific	Cat# FD0454
T7 Ligase	Enzymatics	Cat# L6020L
Anti-FLAG M2 Magnetic Beads	Sigma-Aldrich	Cat# M8823
3xFLAG Peptide	Sigma-Aldrich	Cat# F4799
Pierce Anti-HA Magnetic Beads	ThermoFisher Scientific	Cat# 88836
HA Peptide	ThermoFisher Scientific	Cat# 26184
1X TCL Buffer	Qiagen	Cat# 1031576
2X TCL Buffer	Qiagen	Cat# 1070498
Deoxynucleotide (dNTP) Solution Mix	NEB	Cat# N0447L
SUPERase·In RNase Inhibitor	ThermoFisher Scientific	Cat# AM2694
Betaine Solution	Sigma-Aldrich	Cat# B0300-1VL
Maxima H Minus Reverse Transcriptase	ThermoFisher Scientific	Cat# EP0751
RNase Cocktail Enzyme Mix	ThermoFisher Scientific	Cat# AM2286
EDTA-free Protease Inhibitor Cocktail	Sigma-Aldrich	Cat# 4693159001
2X Tris-Glycine SDS Sample Buffer	Life Technologies	Cat# LC2676
10X NuPAGE Reducing Agent	Life Technologies	Cat# NP0009
Novex WedgeWell 8-16% Tris-Glycine Gel	Life Technologies	Cat# XP08165BOX
PVDF iBlot 2 Transfer Stack	Life Technologies	Cat# IB24002
Blotting Grade Blocker Non Fat Dry Milk	Bio-Rad	Cat# 1706404XTU
Doxycycline Hyclate	Sigma-Aldrich	Cat# D9891-5G
Cytodex-3 beads	Cytiva	Cat# 17048501
Critical commercial assays
Plasmid Plus Midi Kit	Qiagen	Cat# 12945
QIAamp Viral RNA Mini Kit	Qiagen	Cat# 52904
RNeasy Plus Mini kit	Qiagen	Cat# 74134
DNA Clean & Concentrator columns	Zymo Research	Cat# D4033
Low protein binding 0.45 μm filters	VWR	Cat# 28145-481
Amicon 100 kDa MWCO centrifugal filter	Millipore Sigma	Cat# UFC810096
Agilent High Sensitivity DNA Kit	Agilent	Cat# 5067-4626
Gibson Assembly Master Mix	NEB	Cat# E2611L
RNAClean XP SPRI beads	Beckman Coulter	Cat# A63987
AMPure XP SPRI beads	Beckman Coulter	Cat# A63881
KAPA HiFi HotStart ReadyMix	Roche	Cat# 07958935001
NEBNext High-Fidelity 2X PCR Master Mix	NEB	Cat# M0541L
JumpStart Taq DNA polymerase ReadyMix	Sigma-Aldrich	Cat# P2893
iScript cDNA Synthesis kit	Bio-Rad	Cat# 1708890
CellTiter-Glo	Promega	Cat# G7572
Nuncleon white polystyrene 96-well assay plates	ThermoFisher Scientific	Cat# 136101
384-well TC-treated microplates	Corning	Cat# 3764BC
Human GAPDH qPCR primer/probe set	ThermoFisher Scientific	Cat# Hs02758991_g1
Lonza Cell Line Nucleofector Kit V	Lonza	Cat# VCA-1003
AVITI 2x150 Sequencing Kit Cloudbreak FS High Output	Element Biosciences	Cat# 860-00013
Deposited data
Subcellular RNA localization data	Nicholson-Shaw et al.67	GSE195493
Raw sequencing reads	This paper	GSE304897
Experimental models: Cell lines
HEK293T	ATCC	Cat# CRL-3216
HEK293FT	ThermoFisher Scientific	Cat# R70007
HT1080	ATCC	CCL-121
HeLa	Whitehead Institute (Iain Cheeseman Lab)	N/A
1157.2 hiPSCs	Boston Children’s Hospital Stem Cell Core	N/A
iNGN hiPSCs	Harvard Medical School (George Church Lab; Busskamp et al.71)	N/A
Oligonucleotides
SMART-Seq2 polyT RT primer	Picelli et al.48	AAGCAGTGGTATCA ACGCAGAGTACT30 VN
SMART-Seq2 template switching oligo	Picelli et al.48	AAGCAGTGGTATCA ACGCAGAGTACATr GrG+G
SMART-Seq2 ISPCR primer	Picelli et al.48	AAGCAGTGGTATCA ACGCAGAGT
oDF-344_U6_NGS_P5	This paper	ACACGACGCTCTTC CGATCTTCTTGTGG AAAGGACGAAAC
oDF-112_scaffold_NGS_P7	This paper	CTGGAGTTCAGACG TGTGCTCTTCCGAT CAAGCACCGACTCG GTGCCAC
Recombinant DNA
pMD2.G	Didier Trono	Addgene Plasmid #12259
psPAX2	Didier Trono	Addgene Plasmid #12260
pHR_dSV40-H2B-GFP	Ron Vale	Addgene Plasmid #67928
Super piggyBac Transposase	SBI	Cat# PB200A-1
MS2-P65-HSF1_GFP	Konermann et al.77	Addgene Plasmid #61423
phage-cmv-cfp-24xms2	Wu et al.78	Addgene Plasmid #40651
CROPseq-Puro	Datlinger et al.50	Addgene Plasmid #127458
CROPseq-Puro-24xMS2	This paper	Addgene Plasmid #212629
pGag-mCherry	This paper	Addgene Plasmid #212631
pGag	This paper	Addgene Plasmid #212630
pGagΔMA	This paper	Addgene Plasmid #212632
pLV_CAG_GagΔMA	This paper	Addgene Plasmid #212633
pLV_CAG_Gag	This paper	Addgene Plasmid #212634
pLV_CAG_Gag-EIF4E	This paper	Addgene Plasmid #212635
pLV_CAG_Gag-EIF4E-TM	This paper	Addgene Plasmid #212636
pLV_CAG_Gag-MCP	This paper	Addgene Plasmid #212637
pLV_CAG_Gag-EIF4E-2A-mNeon-2A-PuroR	This paper	Addgene Plasmid #253299
pPB_EF1a_Gag	This paper	Addgene Plasmid #212655
pFLAG-VSVG	This paper	Addgene Plasmid #212641
pLV_CMV_FLAG-VSVG	This paper	Addgene Plasmid #212642
pLV_CMV_HA-VSVG	This paper	Addgene Plasmid #212645
pPB_EF1a_FLAG-tNGFR_ZeoR	This paper	Addgene Plasmid #212649
Software and algorithms
Fiji (v2.14.0/1.54f)	Schindelin et al.70	https://imagej.net/software/fiji/
FlowJo (v10.10.0)	BD Biosciences	https://www.flowjo.com/
PyMOL (v2.5.7)	Schrodinger	https://www.schrodinger.com/products/pymol
ColabFold	Mirdita et al.62	https://github.com/sokrypton/ColabFold
AlphaFold3	Google DeepMind	https://alphafoldserver.com/
R (v4.3.1)	The R Foundation	https://www.r-project.org/
Python (v3.9)	Python Software Foundation	https://www.python.org/
STAR (v2.6.0c, v2.7.10a)	Dobin et al.63	https://github.com/alexdobin/STAR
Kallisto (v0.43.1, v0.46.0)	Bray et al.64	https://github.com/pachterlab/kallisto
RNA-SeQC	DeLuca et al.69	https://github.com/getzlab/rnaseqc
tximport (v1.28.0)	Soneson et al.72	https://bioconductor.org/packages/tximport
DESeq2 (v1.40.2, v1.50.2)	Love et al.65	https://bioconductor.org/packages/DESeq2
edgeR	Robinson et al.80	https://bioconductor.org/packages/release/bioc/html/edgeR.html
GenomicFeatures (v1.62.0)	Lawrence et al.66	https://bioconductor.org/packages/release/bioc/html/GenomicFeatures.html
XGBoost (v3.1.3.1)	Chen & Guestrin68	https://github.com/dmlc/xgboost
SHAPforxgboost (v0.2.0)	Liu & Just	https://github.com/liuyanguu/SHAPforxgboost
mgcv (v1.9-3)	Simon Wood	https://cran.r-project.org/web/packages/mgcv/index.html
fgsea (v1.36.0)	Korotkevich et al.76	https://github.com/alserglab/fgsea
GSEApy (v1.1.9)	Fang et al.74	https://github.com/zqfang/GSEApy
CellNet	Cahan et al.49	https://github.com/pcahan1/CellNet
Analysis code	This study	https://github.com/mnajia/Live-Cell-Transcriptomics-VLP-Manuscript

HIGHLIGHTS.

• MMLV Gag packages and exports cellular RNAs into VLPs for various cell types

• Repeated VLP sampling from cell media captures transcriptional dynamics

• VLP engineering enables multiplexed readouts and tuned RNA packaging preferences

• In situ microphysiological system profiling reveals dynamic processes in vasculogenesis