Multimodal hierarchical classification of CITE-seq data delineates immune cell states across lineages and tissues

Summary

Single-cell RNA sequencing (scRNA-seq) is invaluable for profiling cellular heterogeneity and transcriptional states, but transcriptomic profiles do not always delineate subsets defined by surface proteins. Cellular indexing of transcriptomes and epitopes (CITE-seq) enables simultaneous profiling of single-cell transcriptomes and surface proteomes; however, accurate cell-type annotation requires a classifier that integrates multimodal data. Here, we describe multimodal classifier hierarchy (MMoCHi), a marker-based approach for accurate cell-type classification across multiple single-cell modalities that does not rely on reference atlases. We benchmark MMoCHi using sorted T lymphocyte subsets and annotate a cross-tissue human immune cell dataset. MMoCHi outperforms leading transcriptome-based classifiers and multimodal unsupervised clustering in its ability to identify immune cell subsets that are not readily resolved and to reveal subset markers. MMoCHi is designed for adaptability and can integrate annotation of cell types and developmental states across diverse lineages, samples, or modalities.

Graphical abstract

Highlights

• •MMoCHi is a supervised machine learning framework for annotating single-cell data
• •MMoCHi enables annotation of multimodal CITE-seq and spatial profiling data
• •Method outperforms competitors in joint benchmarking with FACS and CITE-seq
• •We demonstrate insights into T cell memory subsets from a cross-tissue immune atlas

Motivation

Many of the features that define cell types are not captured by single-cell RNA sequencing (scRNA-seq), motivating the development of multimodal single-cell technologies, including cellular indexing of transcriptome and epitopes by sequencing (CITE-seq), RNA expression and protein sequencing assay (REAP-seq), and antibody sequencing (Ab-seq), for simultaneous profiling of surface proteomes and transcriptomes. However, integrating these high-dimensional modalities to identify cell subsets, developmental states, and other cellular properties with high fidelity across disparate datasets remains a challenge. Furthermore, unsupervised methods like clustering often fail to delineate cell types with established markers. Thus, there is a need for computational tools that can handle batch effects that are unique to multimodal technologies and identify cell subsets based on known markers.

Accurate cell-type annotation often requires multimodal data, such as joint profiling of mRNA and protein from the same individual cells. Caron et al. develop MMoCHi, which uses supervised machine learning to enable accurate annotation of multimodal single-cell data, including CITE-seq and spatial profiling, without relying on reference atlases.

Introduction

New computational tools have greatly advanced the utility of single-cell genomics. It is now possible to infer trajectories and regulatory networks, identify biomarkers associated with clinical covariates, and perform clonal lineage tracing, all from the same measurements. However, applying these exciting methods usually requires high-quality classification of each cell in the dataset in terms of canonical subsets with known markers. While many analytical tools have been developed for the cellular annotation of single-cell RNA sequencing (scRNA-seq) data (reviewed in Cheng et al.¹ and Heumos et al.²), we have identified a trade-off between classification accuracy and the cell-type information supplied by the user to the classifier algorithm. Specifically, we find that providing protein-level information from cellular indexing of transcriptome and epitopes by sequencing (CITE-seq) and classifying according to a user-supplied hierarchy of cell types, each labeled with known markers, can lead to greatly enhanced classification accuracy compared to fully automated, reference-based methods or manual annotation of clusters. While such detailed information is not always available, it is increasingly common to profile specimens comprising known cell types with well-characterized markers to uncover phenotypic responses to drugs,³ alterations in disease or aging,^4,5 or interactions between known cell types.⁶ We describe a computational framework and user-friendly software that uses supervised machine learning to capitalize on such domain knowledge for highly accurate cell-type classification from multimodal, single-cell sequencing and spatial profiling data.

Unsupervised clustering is commonly used to divide cells into groups with similar expression profiles,^7,8 and manual annotation of clusters is likely the most common approach to classifying cells from single-cell sequencing data. Indeed, this approach has proven invaluable for characterizing cellular heterogeneity,⁹ has led to the discovery of new cell types,^10,11 and has been adapted for CITE-seq datasets.^12,13 However, the number, type, and identity of clusters can be difficult to compare across studies,^9,14 and there is no guarantee of concordance between clusters and canonical cell types that are known to be present in a given specimen.

On the other hand, supervised machine learning methods can address these drawbacks by incorporating reference data or marker-based definitions of cell types.¹⁵ Examples of such tools include CellTypist,¹⁶ ImmClassifier,¹⁷ HieRFIT,¹⁸ and Garnett.¹⁹ CellTypist uses logistic regression and stochastic gradient descent for automated annotation based on a reference atlas.¹⁶ To overcome the many, sometimes overlapping, annotations across atlases, it includes an option called “majority voting” (MV), where users can elect to perform conventional clustering and annotate all cells in each cluster with the most dominant classification. ImmClassifier addresses the reverse problem—existing reference datasets for cell types of interest (i.e., for immune cells) may not be pre-annotated with the desired granularity or may not include all the necessary cell subsets.¹⁷ Thus, it first annotates input data with multiple random forest classifiers trained on independent reference datasets and then generates a consensus classification using a deep neural network. To improve annotation accuracy of highly similar subsets, HieRFIT utilizes a hierarchical design, implementing multiple random forest classifiers to progressively annotate cell types.¹⁸ Together, these reference-based methods enable cross-study comparisons^15,20 but require reference atlases, which are not available for all tissues and contexts.^21,22,23 Reference-free annotation methods address these shortcomings, instead relying on knowledge-based marker definitions.^19,24 Garnett, for example, classifies scRNA-seq data using a hierarchy of cell types defined by user-supplied markers and a regularized elastic-net generalized linear model for training.¹⁹ Overall, existing tools have successfully leveraged a variety of machine learning algorithms, both reference-based and marker-based cell-type definitions, and hierarchical designs for cell-type classification, all with exclusively RNA-level features.

Here, we developed a supervised approach for cell-type annotation of CITE-seq data, designated multimodal classifier hierarchy (MMoCHi), that incorporates surface protein and transcript features for marker-based, reference-free classification. The development of reference-free, multimodal classifiers to analyze single-cell data is of particular importance for studying the immune system. Immune cells comprise multiple disparate lineages—each of which can be subdivided into functionally distinct, but closely related, subsets that represent various developmental and activation states.²⁵ Decades of research have delineated numerous immune cell subsets across the lymphocyte and myeloid lineages, often defined by surface phenotypes, which shape our biological questions^25,26,27,28 but are challenging to annotate by transcriptome alone.^{17,29,30,31,32} Thus, to benchmark this tool against other annotation methods, we sorted and profiled T cell subsets using CITE-seq and demonstrate improved performance by MMoCHi over the alternative methods described above, particularly in the annotation of subsets with highly similar expression profiles. We applied MMoCHi to produce integrated annotations for a cross-tissue CITE-seq atlas of diverse immune cell populations isolated from complex samples. Extracting important features from MMoCHi’s underlying random forests revealed highly interpretable learned representations of cell types, which we used to identify markers for distinguishing transcriptionally similar T cell subsets. Finally, we demonstrated the versatility of MMoCHi by extending it to applications beyond CITE-seq. This includes identification of transformed tumor cells from joint profiling of aneuploidy and gene expression (GEX) and annotation of in situ spatial profiling (10× Genomics Xenium) by joint classification with GEX and morphological features. Together, MMoCHi readily enables multimodal cell-type annotation based on marker genes and proteins, is designed for applicability to CITE-seq of any cell lineage or sample type, and shows promise for extension to forthcoming modalities such as in situ spatial profiling of complex tissues.

Results

Algorithm overview

The MMoCHi algorithm uses a hierarchy of random forest classifiers trained on GEX and antibody-derived tags (ADTs) for cell-type classification of CITE-seq data (Figure 1). Prior to classification, ADT expression is batch-corrected using landmark registration, as previously applied to flow cytometry³³ and CITE-seq³⁴ (see STAR Methods). We identify populations exhibiting negative (background) and positive ADT expression and apply warping functions to align their midpoints (landmarks) across batches to effectively integrate CITE-seq expression (Figure 1A). MMoCHi then classifies cell types based on a user-supplied hierarchy of cell subsets paired with marker-based definitions (Figure 1B). At each classification node, high-confidence members of each subset are identified using manual thresholds on user-provided gene and/or protein markers. A random forest is then trained on a representative set of these high-confidence cells and used to annotate all cells—including those not labeled by high-confidence thresholding (Figures 1B–1D). Once trained, classifiers can be interrogated for features important for cell classification or applied to extend cell-type annotation to other datasets.

Figure 1.

Schematic showing general workflow of MMoCHi

(A) Batch correction of antibody-derived tag (ADT) expression is performed by landmark registration on log-normalized counts (1). After detection of positive (white ticks) and negative (black ticks) peaks in each batch (2), a warping function is applied to align these landmarks across batches (3).

(B) MMoCHi hierarchy demonstrating the classification workflow. User supplies a hierarchy of cell subsets and marker definitions for each subset (1). Thresholding is performed to select high-confidence events for each subset (2). A portion of these high-confidence events are used to train a random forest (3). Finally, the trained random forest is used to predict subset identities of all events from the parent subset (4). This process is repeated for each classifier node within the hierarchy until all cells are classified to endpoint subsets.

(C) Prior to training, high-confidence events could be resampled to remove potentially mislabeled events (“noise”) and amplify potentially challenging-to-classify events (“in danger”). Nearest neighbors are calculated on high-confidence events (1). Events with no neighbors in agreement with their high confidence label are identified as noise (2). Events with only some neighbors in agreement or in poorly represented clusters are identified as in danger (3 and 4). Events labeled “noise” are removed from the training data. Events labeled “in danger” are duplicated five times in the training dataset (5).

(D) Training data are selected from a random sample of cleaned-up high-confidence events with rare subsets oversampled so all subsets are equal in the training data. In the case of insufficient training events in individual batches, high-confidence events can be spiked into the training data from other batches.

Superior annotation of closely related immune cell subsets by MMoCHi

MMoCHi is designed to improve the annotation of subsets with highly similar transcriptomic profiles compared to conventional manual annotation of clusters or methods that rely exclusively on mRNA profiles. We chose to test its performance using T cell subsets, which are well defined by surface-marker expression and functional readouts,^26,27 but challenging to annotate by scRNA-seq alone. For example, CD4⁺ and CD8⁺ T cells are often imperfectly resolved, partially due to low CD4 transcript expression.^17,18,30 Moreover, conventional human T cells are delineated into subsets based on differentiation state and migration capacity into naive T cells (CCR7⁺ CD45RA⁺ CD45RO⁻) and memory T cells, which comprise central memory (T_CM; CCR7⁺ CD45RA⁻ CD45RO⁺), effector memory (T_EM; CCR7⁻ CD45RA⁻ CD45RO⁺), and terminally differentiated (T_EMRA; CCR7⁻ CD45RA⁺ CD45RO⁻) subsets.^26,27 Distinguishing naive T cells from T_CM and T_EM from T_EMRA by transcriptome alone is challenging.^17,29,30,35 We sorted and performed CITE-seq on seven T cell subsets (CD4⁺ naive, CD4⁺ T_CM, CD4⁺ T_EM, CD8⁺ naive, CD8⁺ T_CM, CD8⁺ T_EM, and CD8⁺ T_EMRA) and monocytes (Figures 2A and S1A; Tables S1 and S2). CCR7 staining by CITE-seq is suboptimal,³⁶ so we used CD62L, which has high concordance in human blood.^26,27 To eliminate batch effects, we performed CITE-seq antibody staining before sorting and labeled sorted populations with hashtag antibodies prior to pooling all samples for library preparation and sequencing. Hashtagged populations then served as known references to evaluate classification.

Figure 2.

MMoCHi classification of pre-defined T cell subsets outperforms other annotation methods

(A) PBMCs were stained with fluorescently labeled antibodies and a cocktail of oligo-tagged antibodies for sequencing. Seven T cell subsets and monocytes were sorted, hashtagged, pooled, and sequenced.

(B) MMoCHi hierarchy defining subsets using the same markers used for sorting.

(C) UMAPs of totalVI latent space colored by sorted cell type (identified by hashtag oligo; HTO), MMoCHi classification, or manual cluster annotations.

(D) Row-normalized heatmap comparing MMoCHi classification to sorted cell type. Color represents proportion of cells in each MMoCHi classification from each sorted subset.

(E) Performance comparison using F1 scores, calculated for each cell subset using HTO-derived sorted cell-type labels as truth. F1 scores for each method were aggregated in box-and-whisker plots (right). Features of each method are labeled for context (left) using green checks: “Trained on GEX/ADT,” whether gene expression (GEX) or antibody-derived tag (ADT) expression data were used during model training; “HC-thresholds,” whether high-confidence thresholds were used for training data selection; “Hierarchical,” whether annotation was performed on multiple levels; “Cluster-based,” whether annotations were applied to unsupervised clusters. For scRNA-seq classifiers, the percentage of classifications that were relevant to this analysis, were alternative (but potentially valid) classifications, were unlabeled, or were entirely erroneous are shown using pie charts (far right). Only relevant classifications were included for calculation of F1 scores. Statistical significance was calculated using a Friedman rank-sum test, matched by subset, followed by multiple comparisons testing using paired Wilcoxon signed-rank tests and false discovery rate (FDR) correction. Significance groups for each method are shown by lettering, where methods not sharing a letter are significantly different (p < 0.05). Methods that could not label T cell memory subsets were excluded from significance testing, and the merged efficacy of CD4⁺ or CD8⁺ T cell classifications was displayed instead.

MV, majority voting; T_CM, central memory T cell; T_EM, effector memory T cell; T_EMRA, terminally differentiated effector memory T cell. Schematic in (A) was created with BioRender.com.

All sorts were of high purity (>91%; Table S3), and subsets reflected expected ADT marker expression (Figure S1B). We devised and applied a MMoCHi hierarchy with the same protein markers used for sorting (Figure 2B; Table S4). For comparison to unsupervised approaches, we used totalVI¹³ to calculate a multimodal latent space and then performed clustering and manual annotation (Figures 2C and S1C–S1F). The high concordance between sorted cell type and MMoCHi classification compared to manual annotation is visually evident from uniform manifold approximation and projection (UMAP) plots in Figure 2C. Indeed, MMoCHi classification had 92% agreement with sorted labels, while manual annotation had only 70% agreement (Figures 2C and 2D; Table S5). MMoCHi-classified populations showed expected ADT expression (Figure S1B). MMoCHi classification remained accurate when downsampling GEX and/or ADT reads or when subsampling the number of monocytes or CD8⁺ T_CM cells used for training, revealing robustness to lower data quality or fewer, imbalanced, training events (Figures S1G–S1P). Although this classification uses both ADT and GEX features, MMoCHi was likely insensitive to GEX coverage because both the reference dataset and the hierarchy for classification were defined by protein expression.

To compare performance between various methods, we calculated precision, recall, and F1 score for each subset as well as overall accuracy (Figure 2E; Table S5). MMoCHi accurately classified all subsets, with an average F1 score of 0.91 (Figure 2E, MMoCHi). We tested variations to validate MMoCHi’s design. Performance worsened when training random forests using only GEX and/or when classifying all subsets with only one random forest (instead of a hierarchy) (Figure 2E, MMoCHi GEX, MMoCHi flat, and MMoCHi GEX flat). We noted a modest decrease in performance when relying exclusively on automatic thresholds (Figures S2A and S2C; Table S4) and thus encourage users to manually evaluate and adjust automatic thresholds (similar to best practices for flow cytometry gating³⁷). For Figure 2, we applied MMoCHi using manually defined hyperparameters and with the in-danger-noise algorithm enabled (see STAR Methods), but we noted similar performance with MMoCHi’s built-in automatic hyperparameter optimization or in-danger-noise disabled (Figures S2B and S2C). Overall, these results support MMoCHi’s multimodal, hierarchical, and semi-automated design.

We next compared MMoCHi classification to manual annotation of unsupervised clusters derived from GEX, ADT expression, or the totalVI latent space (see STAR Methods; Figure 2E, GEX Leiden, ADT Leiden, and totalVI Leiden; Figures S1C–S1F). MMoCHi outperformed all manual annotation methods, particularly GEX Leiden, which failed to identify CD4⁺ T_CM or CD8⁺ naive T cells (Figure S1F). To ensure manual annotation was not hampered by a lack of clustering resolution, we also overclustered the data (see STAR Methods), but this resulted in only a slight improvement (Figure 2E, GEX Leiden OC, ADT Leiden OC, and totalVI Leiden OC). Furthermore, since probabilistic generative modeling on small datasets can be sensitive to feature numbers,³⁸ we also showed that manual annotation using totalVI was insensitive to the number of input genes (Figures S3A–S3D).

Last, we tested four supervised classification methods for scRNA-seq: ImmClassifier,¹⁷ CellTypist,¹⁶ HieRFIT,¹⁸ and Garnett.¹⁹ We applied ImmClassifier, CellTypist, and HieRFIT as intended, using their pre-designed reference-based training (“pre-trained models,” see STAR Methods). CellTypist was applied to annotate either individual events or by MV across clusters, as advised.¹⁶ Garnett was designed, similar to MMoCHi, for classification using marker-based hierarchies¹⁹ and was applied using both a published hierarchy and custom hierarchy manually selected markers (“provided markers” and “manual markers,” respectively; see STAR Methods; Table S6). These pre-trained models include classification of some non-terminal or other, potentially valid, subsets (e.g., T cells, mucosal-associated invariant T [MAIT] cells, regulatory T cells [T_regs], or exhausted T cells; “alternative”) and classification of populations excluded by our fluorescence-activated cell sorting (FACS) gating (e.g., neutrophils, B cell lineage, natural killer [NK] cells, or γδ T cells; “erroneous”). In addition, HieRFIT and Garnett avoid low-confidence classifications, leaving some events undetermined (“unlabeled”). For benchmarking, we examined only relevantly classified events (i.e., belonging to one of the sorted subsets), but notably, large portions of CellTypist and ImmClassifier classifications were alternative or incorrect, and most Garnett annotations were undetermined (Figure 2E). Focusing on relevant classifications, all scRNA-seq classifiers performed on par with GEX Leiden, significantly underperforming compared to MMoCHi classification (Figure 2E). HieRFIT pre-trained and Garnett provided markers were excluded from benchmarking analyses, as they could not annotate T cells at an appropriate granularity.

All of the supervised approaches described above effectively distinguished monocytes from T cells (Figure 2E), consistent with previously documented high performance of these tools.^16,17,18,19 Our evaluations emphasize the difficulty of segregating T cell subsets that share highly similar transcriptomic profiles without the robust training data and surface protein profiling that MMoCHi leverages. To isolate these two effects from performance of the underlying machine learning algorithms employed, we assessed additional (impractical) conditions. First, we trained these tools on orthogonally labeled, batch-free training data—training classifiers using the hashtag-derived sorted subset labels and evaluating performance on a 20% holdout with 5-fold cross-validation (“Sort-ref”; see STAR Methods). Although this sort-reference condition is unrealistic (it requires all subsets be FACS sorted within the same dataset, obviating the need for classification), it revealed differences in performance enabled by MMoCHi’s high-confidence thresholding. Notably, we identified only a modest improvement with idealized training data compared to standard MMoCHi classification, highlighting the efficacy of high-confidence thresholding (Figure S3D). Using sort-reference, classifiers designed for scRNA-seq performed on par with MMoCHi GEX, but worse than MMoCHi (Figure S3D, comparing within Sort-ref). Although Garnett Sort-ref outperformed MMoCHi GEX Sort-ref, it failed to annotate 10% of the cells (Figure S3D). We next coerced these scRNA-seq classifiers to accept multimodal data (“Multi”; see STAR Methods). Two variants (HieRFIT Multi Sort-ref and Garnett Multi Sort-ref) performed on par with MMoCHi Sort-ref, suggesting that MMoCHi’s underlying machine learning algorithms perform as well as these established tools. Importantly, however, MMoCHi trained using high-confidence thresholds was the only method that could approach this ultra-high accuracy without requiring access to the subset labels from FACS (i.e., “cheating”). Because these subsets were all exclusively defined by surface protein, MMoCHi could achieve similarly high accuracy when trained on ADT expression alone (Figure S3D), suggesting MMoCHi is highly effective at selecting the most informative features, regardless of which modalities are included. Together, these data provide strong support for MMoCHi’s high-confidence thresholding and multimodal training, which together enable effective annotation of these highly similar cell types.

Using these same 80% train and 20% test splits, we also compared MMoCHi’s performance to other relevant methods that lacked any pre-trained classifiers for this task, including two hierarchical scRNA-seq classifiers (CHETAH³⁹ and scClassify⁴⁰) and one multimodal multitask learning method that enables CITE-seq classification (Matilda⁴¹). As before, although these methods could effectively distinguish distinct cell subsets (monocytes and T cells), MMoCHi outperformed these tools at annotation of nuanced T memory subsets (Figure S3D).

As described above, MMoCHi also has a built-in batch-correction tool for CITE-seq (landmark registration), which we tested on publicly available datasets using diverse antibody panels, chemistries, and sequencing depth (Figure S4A). Landmark registration (run here with no manual input; see STAR Methods) integrated ADT expression more effectively than other ADT normalization techniques (CLR³⁶ and dsb⁴²), was robust to cell-type subsampling, and resulted in a modest improvement in classification performance, especially when the same high-confidence thresholds were applied across batches (Figures S4B–S4H; Table S4).

MMoCHi integrates classification across diverse human tissue immune cells

We next applied MMoCHi to total immune cells from lymphoid and mucosal tissue samples obtained from human organ donors¹⁶ (Figure 3A; Table S1). Immune cells were enriched from eight sites across two donors and included lung (LNG), bronchial alveolar lavage (BAL), lung-associated lymph node (LLN), spleen (SPL), jejunum epithelial layer (JEL), jejunum lamina propria (JLP), bone marrow (BOM), and blood (BLD), using methods optimized for each site.^16,30 We performed CITE-seq to profile over 270 surface markers (Table S2). In previous analysis of single-cell transcriptomes from this dataset, we detected all lineages of immune cells, including T cells, B cells, innate lymphocytes, and myeloid cells across multiple sites.¹⁶ However, refined immune cell subsets and differences across tissue sites were difficult to resolve in this complex dataset and required extensive manual annotation. Here, we integrated protein and transcriptome profiling using MMoCHi to determine whether transcriptionally similar immune subsets could be identified across diverse tissues.

Figure 3.

MMoCHi classification applied to human immune cells from blood and diverse tissue sites

(A) Single-cell suspensions of immune cells were isolated from blood and the indicated tissue sites of two organ donors. CD45⁺ immune cells were selected for by magnetic enrichment, hashtagged by tissue site, pooled, and sequenced.

(B) UMAP of donor-integrated totalVI latent space, colored by tissue site.

(C) MMoCHi hierarchy used for classification (see Table S4 for full specification).

(D) Performance curves showing the F1 score for each endpoint subset when training random forests with various numbers of trees (estimators). F1 scores were calculated on held-out data using high-confidence thresholded events as truth. The red line indicates the median F1 score across all predicted cell types.

(E) Column-normalized heatmap depicting the distribution of classified cell types across tissue sites. The number of total events in each classification or tissue site is displayed.

LNG, lung; BAL, bronchoalveolar lavage; LLN, lung lymph node; SPL, spleen; JEL, jejunum epithelial layer; JLP, jejunum lamina propria; BOM, bone marrow; BLD, blood; T_CM, central memory T cell; T_reg, regulatory T cell; T_EM, effector memory T cell; T_RM, resident memory T cell; T_EMRA, terminally differentiated effector memory T cell; NK, natural killer cell; ILC, innate lymphoid cell; DC, dendritic cell; pDC, plasmacytoid dendritic cell; Mono, monocyte; Mem, memory; N/CM, naive/central memory; Res, residency; Eff, effector. Schematic in (A) was created with BioRender.com.

For visualization, we embedded a donor-integrated totalVI latent space with UMAP, revealing multiple groupings, some of which corresponded to different tissue sites (Figure 3B; see STAR Methods). We constructed and applied a MMoCHi hierarchy representing all expected cell types using a combination of transcript and surface protein markers (Figure 3C; Tables S4 and S7). Training and classifying these approximately 198,000 immune cell events into 26 subsets took less than 15 min (Figure S5A; see STAR Methods), demonstrating MMoCHi’s scalability. MMoCHi thresholds were manually defined, as automated thresholds (see STAR Methods) resulted in discrepancies, most of which could be resolved by manually tuning thresholds on as few as eight markers (Figure S2D; Table S4). The resultant model, using 200 estimators in each random forest, was well fit at every level of the hierarchy, as measured by the prediction accuracy for internally held-out subsets in the same classification layer and across the full hierarchy (Figures 3D and S5B–S5Q; see STAR Methods). MMoCHi classified cell types across tissue sites and consistently across donors (Figures 3E and S6).

We compared MMoCHi classification to manually annotated clusters of the multimodal totalVI latent space (Figures 4A, 4B, S7A, and S7B). The methods were broadly concordant, with 72% of events labeled identically; however, disagreements occurred between cell types with similar expression profiles. These disagreements were not resolved by increasing the number of clusters nor by altering the number of genes input to totalVI (Figures S7C and S7D). Notably, unsupervised clustering failed to resolve any CD8⁺ T_CM cells, γδ T cells, naive B cells, or dendritic cells (DCs). We investigated discrepancies using marker expression (Figures 4C–4E and S8). A substantial percentage (6%) of αβ T cells had conflicting CD4⁺ or CD8⁺ annotations, and by protein and transcript expression, MMoCHi classifications were more appropriate (Figure 4C). Cytotoxic lymphocytes are found across multiple lineages (NK cells, innate lymphoid cells [ILCs], CD8⁺ T_EMRA cells, and γδ T cells) and are difficult to resolve; however, MMoCHi correctly classified NK cells and ILCs by their lack of CD3 or TCRαβ surface expression, despite expression of CD3E and TRDC transcripts^23,31 and their co-clustering with CD8⁺ T_EMRA (Figures 4A–4D). MMoCHi also correctly distinguished CD8⁺ T_EMRA and γδ T cells where cluster-based manual annotation failed, marked by expression of either CD3 and TCRαβ or CD3, TCRγδ, and variable expression of TRDV1 and TCRVδ2, respectively (Figure 4D). MMoCHi also improved identification of T cell memory subsets by expression of CD62L, CCR7, CD45RA, and CD45RO (Figure 4E). Across these populations with conflicting annotations, we used differential ADT expression to further quantify that MMoCHi-classified populations better reflected expected marker expression, such as upregulation of CD4 on MMoCHi-annotated CD4⁺ T cells and upregulation of CD3, CD8, and CD45RA on MMoCHi-annotated CD8⁺ T_EMRA (Figures S8B–S8D; Table S8). Together, MMoCHi classifications improved identification of known immune cell subsets, particularly in cases of discordant mRNA and protein expression.

Figure 4.

MMoCHi-classified human immune cells better reflect expected marker expression patterns

(A) UMAPs of organ donor cells, colored by MMoCHi classification and manual cluster annotations. Dashed ellipses highlight areas of major disagreement between MMoCHi classification and manual annotation.

(B) Row-normalized heatmap comparing MMoCHi classification to manual annotation. Color represents proportion of cells in each MMoCHi-classified population that were manually annotated as each subset. Yellow columns indicate subsets that were not detected by manual annotation.

(C–E) Plots depicting expression of selected cell-type markers on cells grouped by their MMoCHi classification and manual annotation. Dot plots display gene expression (GEX). Dot size represents the percentage of cells in the group expressing a gene, and the dots are colored by the mean log-normalized GEX counts per 10 thousand. Violin plots display the distribution of antibody-derived tag (ADT) expression for each population. Violins are colored by the median log-normalized ADT counts per thousand. Behind each violin plot are gray guides displaying the range of ADT expression (light gray rectangle) and a threshold delineating positive from background expression (dashed line). The number of events in each grouping is displayed above the dot plots. Events are denoted by a “+” where MMoCHi and manual annotation agree and a “−” where they disagree.

T_CM, central memory T cell; T_reg, regulatory T cell; T_EM, effector memory T cell; T_RM, resident memory T cell; T_EMRA, terminally differentiated effector memory T cell; ILC, innate lymphoid cell; NK, natural killer cell; pDC, plasmacytoid dendritic cell; DC, dendritic cell; Mono, monocyte.

We additionally sought to evaluate MMoCHi’s performance when pre-trained classifiers were applied to new datasets. To this end, we applied MMoCHi classifiers trained using the FACS dataset (“FACS-trained”) to the αβ T cells and monocytes isolated from the peripheral blood mononuclear cells (PBMCs) of organ donors (Figures S9A and S9B; see STAR Methods). The FACS-trained MMoCHi model outperformed pre-trained or FACS-trained ImmClassifier, CellTypist, and Garnett models at replicating the correct annotation for each cell type (Figure S9; Table S5). Although we do not anticipate that this will be a primary use case of the algorithm, these data further support MMoCHi’s efficacy.

MMoCHi classifiers are highly interpretable and identify subset-associated features

Having demonstrated MMoCHi’s utility for cell-type annotation, we next determined which features (transcript and surface protein expression) were useful for subset delineation. Many single-cell classifiers are trained on dimensionally reduced data,²⁰ hampering feature-level interpretability, but random forests within MMoCHi are trained using all available protein-coding genes and surface proteins. During training, features are selected for their contribution to decreasing impurity, thus providing a natural ranking of features by their importance for classification^43,44 (see STAR Methods). We analyzed the important features at each level of the hierarchy, which included subset markers defined by the user (for high-confidence thresholding), and revealed other known subset markers, identified de novo, including IL1R1 and GATA2 for mast cells⁴⁵; APOE, ACP5, and C1QC for macrophages⁴⁶; CD79A, MARCKS, MEF2C, and CD32 for B cell-like^47,48,49,50; and TYROBP, GZMB, and CD123 for plasmacytoid DCs (pDCs)^51,52 (Figure S10; Tables S4, S8, and S9). Notably, important features included both genes and proteins, emphasizing the importance of using both modalities for this classification. This demonstrates that MMoCHi learns informative cell-type representations—an important pre-requisite for novel marker identification.

Next, we wondered if MMoCHi could be leveraged to improve transcriptome-based segregation of naive T cells and T_CM cells, which are defined by specific CD45 isoform surface expression but share similar transcriptomes.^17,27,29,35 First, we trained GEX classifiers using high-confidence CD4⁺ and CD8⁺ naive/T_CM cells selected by multimodal thresholding (Table S4). Despite ADTs being excluded from training, these classifiers were well fit and highly accurate, and resulting subsets had expected surface-marker expression (Figures 5A and 5B). To identify GEX markers of these subsets, we next interrogated impurity-based important features (Figures 5C and 5D; Tables S8 and S9). We applied robust filtering to highlight genes with clear differences in expression and reduce the potential for tissue-specific contamination (see STAR Methods) and identified a suite of memory-associated genes, including ITGB1, LMNA, and FAM129A,^26,29 as markers upregulated in CD4⁺ and CD8⁺ T_CM cells. Improved prediction accuracy by classifiers trained with only the top 1,000 important features confirmed that these features were useful for GEX classification (Figure 5E).

Figure 5.

MMoCHi reveals additional transcript markers distinguishing naive and central memory T cells

(A and B) Performance of gene expression (GEX)-only classification of naive/T_CM CD4⁺ (A) and CD8⁺ (B) T cells from organ donor samples. Column-normalized confusion matrices (left) comparing classification to held-out, high-confidence threshold labels. Violin plots (right) display the distribution of antibody-derived tag (ADT) expression for each population as classified by MMoCHi trained on multimodal data (ADT + GEX) or GEX only. Violins are colored by the median log-normalized ADT counts per thousand. Behind each violin plot are gray guides displaying the range of ADT expression (light gray rectangle) and a threshold delineating positive from background expression (dashed line).

(C and D) Dot plots displaying expression of important features for GEX-only classification of naive/T_CM CD4⁺ (C) or CD8⁺ (D) T cells. Dot size represents the percentage of cells in the group expressing a gene, and the dots are colored by the mean log-normalized GEX counts per 10 thousand. The top important features associated with a single subset (log₂(fold change) > 2 and greater than 10% change in dropout rate) are displayed. All features displayed were within the top 1,000 important features when training with GEX only on all tissues and with blood (BLD) only. The importance of each feature is shown in a bar chart (right). All displayed features were significantly differentially expressed (p < 0.05).

(E) Bar plots comparing the performance of naive/T_CM CD4⁺ and CD8⁺ T cell classifiers trained using multimodal data (ADT + GEX), GEX only, GEX with only BLD cells, or only the top 1,000 important GEX features.

(F and G) Volcano plots displaying differential gene expression between sorted blood naive and T_CM CD4⁺ (F) or CD8⁺ (G) T cells. Points represent genes and are colored by differential expression (|log₂(fold change)| > 1 and p < 0.05). Important features for classification that were differentially expressed are highlighted.

(H and I) Column-normalized heatmaps displaying performance of GEX-only classifiers trained on organ donor BLD applied to sorted naive/T_CM CD4⁺ (H) and CD8⁺ (I) T cells.

Statistical significance was calculated using a two-sided Wilcoxon with tie correction, followed by a Benjamini-Hochberg adjustment for multiple comparisons. T_CM, central memory T cell.

We sought to validate these findings using our scRNA-seq of sorted CD4⁺ and CD8⁺ naive and T_CM populations. Most markers identified for naive and T_CM were also differentially expressed within the sorted dataset (Figures 5F and 5G; Table S4). Random forests trained on organ donor PBMCs were well fit, effectively recapitulating the sorted labels (Figures 5H and 5I). Overall, these findings demonstrate the capacity of MMoCHi to leverage CITE-seq to identify GEX markers and train classifiers that can be effectively applied to scRNA-seq datasets.

MMoCHi effectively annotates other multimodal datasets and cell types

Last, we explored whether MMoCHi could be easily extended beyond CITE-seq. First, we applied MMoCHi to publicly available paired transcriptomes and surface proteomes of sorted T and NK cells generated using the Ab-seq platform.^53,54 We obtained high concordance with cluster-based annotations and expected marker expression (Figures 6A, 6B, S11A, S12A, and S12D; Tables S4 and S9). As in the above examples, using exclusively automated thresholds (see STAR Methods) resulted in reduced performance, which could be largely rescued by manually tuning thresholds for only two markers (Figure S2E; Table S4). Next, we applied MMoCHi to scRNA-seq of a high-grade glioma biopsy.⁵⁵ In these data, malignant cells were challenging to identify based on marker genes, because they are highly heterogeneous and express many of the same genes as non-neoplastic cells in the brain microenvironment. However, gliomas often harbor clonal aneuploidies that are detectable by averaging GEX across whole chromosomes.⁵⁵ Using a joint feature set comprising transcriptome and computed whole-chromosome expression (see STAR Methods), MMoCHi effectively segregated tumor cells from other non-neoplastic subsets, which MMoCHi also annotated accurately (>98% agreement overall; Figures 6C, 6D, S11B, S12B, and S12E; Tables S4 and S9). Notably, chromosomal expression was both useful as a marker for high-confidence thresholding and important internally for classification (Figure S11B; Table S9).

Figure 6.

MMoCHi applied to other multimodal datasets

MMoCHi was applied to Ab-seq data of sorted T and NK cells (A and B), scRNA-seq of a high-grade glioma biopsy (C and D), and spatial transcriptomics of a human lymph node using the 10× Xenium platform (E–H). (A, C, and E) MMoCHi hierarchies used for classification (see Table S4 for full specification) and schematics for extraction of multimodal data for classification (see STAR Methods). (B, D, and F) Row-normalized heatmaps comparing MMoCHi classification to manual annotation. Color represents proportion of cells in each MMoCHi-classified population that were manually annotated as each subset. Yellow columns indicate subsets that were not detected by manual annotation. (G) Images displaying DAPI staining intensity on a human lymph node, overlaid with detected cell boundaries (see STAR Methods). Cell boundaries are colored by MMoCHi classification or manual annotation. (H) Magnified sections of the images in (G), as outlined by the red boxes.

We further explored whether MMoCHi could be adapted to spatial profiling, using a human lymph node dataset generated by the Xenium platform (10× Genomics), which enables multiplexed profiling of hundreds of RNAs in situ with subcellular resolution. By pairing the 377 profiled genes with physical attributes of each cell extracted from the image and segmentation (see STAR Methods), MMoCHi effectively classified 17 subsets, revealing underlying spatial structure, including B cell follicles and lymphatics (Figures 6E–6H; Table S4). MMoCHi outperformed manual annotation of provided cell clusters (see STAR Methods), which failed to segregate CD4⁺/CD8⁺ T cells and various B cell lineage subsets (Figures 6F, S11C, S12C, and S12F). In areas of disagreement, MMoCHi-classified subsets better reflected expected marker expression (Figure S11C). For example, MMoCHi-classified memory B cells featured upregulation of CD27 GEX and downregulation of TCL1A compared to MMoCHi-classified naive B cells, and MMoCHi-classified T_reg cells featured upregulation of IL2RA and downregulation of IL7R, CCR7, and SELL compared to CD4⁺ naive/T_CM cells (Figures S13A–S13G; Table S8). As before, the added modality was useful for MMoCHi classification, where physical attributes (particularly nuclear circularity and DAPI intensity) ranked in the top 5% of features used (Figure S11C; Table S9).

Finally, as the comparator manual annotation for the above example was performed internally, we sought to evaluate MMoCHi’s high performance using an independently annotated Xenium dataset. For this, we used a published dataset of an ER⁺/HER2⁺/PR⁻ breast tumor biopsy profiled using a panel of 313 genes.⁵⁶ Following cell segmentation, the authors manually annotated clusters and subclusters using average marker gene expression. Using a hierarchy designed for this sample (Table S4) and the same classification technique (combining GEX and physical cell attributes into MMoCHi), we obtained similarly high performance. Overall, we identified ∼70% agreement in annotation, but in areas of disagreement (e.g., CD4⁺ and CD8⁺ T cells, B cell memory subsets, monocytes and macrophages, or normal epithelial vs. tumor cells), MMoCHi’s annotation better reflected expected marker expression (Figures S13H–S13O; Tables S4 and S8). For example, MMoCHi-classified cancer-associated fibroblasts featured higher POSTDN and PDGFRB GEX and lower MYH11 expression compared to MMoCHi-classified perivascular-like cells⁵⁷ (Figures S13I and S13N), and MMoCHi-classified CD8⁺ T cells featured upregulation of CD8A and downregulation of CD4 compared to CD4⁺ T cells (Figures S13M and S13O). MMoCHi-annotated CD4⁺ T cells also featured higher CCL5 expression, likely attributable to disagreement between CD4⁺ T_EM subtypes and CD8⁺ T cells (Figure S13O; Table S8). Together, these results suggest that MMoCHi classification can effectively annotate populations within Xenium spatial transcriptomics based on canonical, marker-based definitions. Overall, MMoCHi’s success in annotating these datasets of varying subset compositions and modalities demonstrates the algorithm’s high adaptability to classification of diverse multimodal single-cell datasets.

Discussion

The advent of multimodal single-cell technologies has enabled high-dimensional profiling of many systems, organs, diseases, and species. However, the development of analytical tools to identify cell states and their features consistently across multimodal studies is lagging behind these data acquisition technologies. Here, we present MMoCHi, a multimodal, hierarchical classification approach for cell-type annotation of CITE-seq data, which integrates both gene and protein expression, does not require reference datasets, performs landmark registration for harmonization across large datasets, and accurately classifies both closely related subsets within a single lineage and diverse subsets from across lineages. We devised an experimental validation approach that allowed us to directly compare MMoCHi and several established annotation modalities directly with FACS-based subsetting for the exact same sample of cells. This enabled an extensive benchmarking study across conventional clustering-based methods, supervised machine learning algorithms, and both reference-based and marker-based classifiers with a uniform and independent reference annotation. Applied to immune cell CITE-seq datasets, MMoCHi outperforms current annotation algorithms and identifies new markers for subset delineation. Altogether, MMoCHi provides an adaptable approach for applying marker-based annotation to multimodal datasets.

Currently available single-cell classifiers are primarily designed for scRNA-seq; however, CITE-seq can be leveraged to improve annotation.^36,58 In contrast with other marker-based strategies, MMoCHi does not perform k-nearest-neighbor smoothing⁵⁹ and does not assume bimodal distributions of marker expression,⁶⁰ allowing for robust classification of closely related subsets with discordant transcriptome and surface proteome profiles. Here, we show that MMoCHi effectively distinguishes T cell subsets defined by surface expression of CD45 isoforms and homing receptors but that share overlapping transcriptomes,^17,27,29,35 namely naive T cells from T_CM and T_EM from T_EMRA. By classifying hierarchically, MMoCHi was also able to correctly annotate diverse immune lineages without compromising its performance at segregating functionally and transcriptionally similar cell types, including cytotoxic NK cells, CD8⁺ T_EMRA cells, and γδ T cells.^23,31,32,61 MMoCHi can accurately classify cell types despite other sources of variation, as shown by integrated classification of multiple immune cell lineages across blood and eight disparate tissue sites of two organ donors.

By providing a platform for using high-confidence, marker-based thresholding for selection of training data, MMoCHi enables the flexible design of new classifiers based on prior knowledge without a requirement for carefully curated reference atlases. In particular, applying high-confidence thresholding to CITE-seq protein expression serves as a natural extension of the vast knowledge base developed of flow cytometry markers and gating.^25,37 Similar to other marker-based strategies, these subset definitions are highly interpretable and can be robustly applied across studies and specific sequencing conditions to train new classifiers.^19,24 To reduce redundancy in efforts to design MMoCHi hierarchies, the hierarchies defined in this article, along with other community contributions, have been made available online (https://mmochi.readthedocs.io). Here, we also apply MMoCHi’s multimodal training data selection to improve transcriptome-based annotation of T cell subsets across datasets, demonstrating potential for MMoCHi to advance annotation of scRNA-seq data as well.

In contrast with efforts to automate cell-type annotation,^15,16 MMoCHi’s thresholding schemas require careful marker curation and domain expertise, however; by leveraging this expertise, MMoCHi classifications may better reflect canonically defined and biologically relevant designations. In addition, MMoCHi is not optimized for novel subset identification, requiring the user to define all cell types within a sample for classification. Thus, MMoCHi is complemented by multimodal unsupervised exploration of the dataset either before or after classification to annotate or identify markers of unexpected cell types and states or to identify novel subsets of canonical cell types.

Beyond cell-type classification, MMoCHi learns key features of protein expression and GEX, which can be used to identify new cell-type markers and derive biological insights. MMoCHi random forests do not require prior dimensionality reduction, enabling evaluation of individual surface proteins and protein-coding genes as subset markers. In classifying diverse and highly similar subsets, we show that MMoCHi extracts relevant markers, relying on both features used for training data selection and other known gene and protein markers of each subset. This learned feature-importance ranking serves as a built-in complement to conventional feature extraction methods (i.e., differential GEX) to highlight important, interpretable subset differences. In using MMoCHi to identify improved transcriptional markers of naive T cells and T_CM cells, we discovered that circulating T_CM actually upregulate markers that had previously been associated with T cells in tissue (including ANXA2, ITGB1, LMNA, and FAM129A)^29,30 and T_EM cells (including FAS, PRDM1, and KLRB1).^26,29,62 This finding raises the possibility that early stages of tissue programming occur during T_CM differentiation and indicate a gradual transition between T cell memory subsets. Importantly, we did not previously identify these expression patterns of circulating T_CM in earlier studies of T cells in tissues and blood,^29,30 likely due to the lack of CITE-seq and multimodal classification.

While we have developed MMoCHi with CITE-seq applications in mind, the algorithm is designed for easy extension to other modalities frequently paired with single-cell transcriptomes, such as profiling chromatin accessibility, T and B cell receptors, mutational landscapes, intracellular proteins, or a combination of these modalities.^{63,64,65,66,67,68} We also anticipate applications to emerging technologies for multimodal, single-cell spatial profiling.^69,70 We demonstrated the potential generality of MMoCHi through extension to joint aneuploidy/GEX profiles from a solid tumor and joint classification of in situ spatial profiling data using morphological and GEX features measured using the 10× Genomics Xenium platform. While we mainly focused on immunology, cell-type classification is a ubiquitous problem in single-cell genomics. Thus, we expect broad utility for MMoCHi in diverse biological applications, including identification of developmental states, building atlases of complex tissues and tumors, profiling model organisms, and analyzing clinical specimens.

Limitations of the study

Here, we developed MMoCHi as a marker-based approach for cell-type classification of multimodal single-cell datasets. Accurate comparison of annotation performance between MMoCHi and other classifiers required multimodal single-cell data with orthogonally validated cell-type labels, which has not been performed on publicly available datasets. To overcome this, we performed CITE-seq on PBMCs whose subset identity had been orthogonally classified using FACS. Although we compared MMoCHi’s performance using other published annotations, extensive comparison across tools was limited to this one PBMC dataset. One caveat of MMoCHi’s marker-based approach is that it is best suited to well-defined populations. In addition, MMoCHi assumes the user has identified within the hierarchy all major cell types present within a dataset. For novel subset or marker discovery, as well as data exploration, we recommend using MMoCHi in tandem with unsupervised clustering and differential expression. Although MMoCHi can perform minor integration across batches during its training, we show that, in cases with extreme batch artifacts (such as with ADT expression), MMoCHi’s accuracy benefits from additional batch correction to the input data (such as landmark registration of ADT expression). Last, although we demonstrate that MMoCHi is generalizable to a variety of single-cell modalities, including GEX, antibody staining, predicted aneuploidy status, and physical attributes, dimensionality reduction or other transformations may be required to extend MMoCHi to other modalities, especially those that suffer from higher sparsity.

Resource availability

Lead contact

Requests for further information and reagents should be directed to and will be fulfilled by the lead contact, Peter A. Sims (pas2182@cumc.columbia.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All CITE-seq data generated in this study have been deposited at NCBI GEO: GSE229791. Accession numbers are listed in the key resources table. This paper also analyzes existing, publicly available data available from 10× Genomics, Open Science Framework, and NCBI GEO. DOIs or accession numbers for these datasets are listed in the key resources table.
• •Code for MMoCHi is available at https://github.com/donnafarberlab/MMoCHi (https://doi.org/10.5281/zenodo.14060888). Code for demultiplexing CITE-seq and hashtag barcodes is available at https://github.com/simslab/DropSeqPipeline8 (https://doi.org/10.5281/zenodo.5534457).
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We thank Joshua I. Gray and Rory E. Morrison-Colvin for helpful discussions and members of the Farber laboratory for help with tissue processing. This work was supported by a Seed Networks for the Human Cell Atlas grant from the Chan Zuckerberg Initiative (CZF2019-002452) and NIH grants AI128949 and AI106697 awarded to P.A. Sims and D.L.F. D.P.C. was supported by the Columbia University Graduate Training Program in Microbiology and Immunology (T32AI106711). P.A. Szabo was supported by a Canadian Institutes of Health Research Fellowship. Research reported here was performed in the Columbia Stem Cell Initiative Flow Cytometry Core, the Sulzberger Columbia Genome Center, and the Columbia Single Cell Analysis Core (supported by grant P30CA013696).

The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. We wish to thank the donor families for their generosity and are grateful for the exceptional efforts of the transplant coordinators and staff of LiveOnNY for making this study possible.

Author contributions

D.P.C. designed and performed experiments, analyzed data, made figures, and wrote the manuscript. W.L.S. performed experiments, analyzed data, made figures, and edited the manuscript. D.C., S.B.W., P.A. Szabo, and I.J.J. prepared samples for single-cell sequencing. P.A. Sims and D.L.F. designed experiments, analyzed data, and wrote and edited the manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
CD45 APC Clone: MB4-6D6	Miltenyi	Cat#130-123-787; RRID:AB_2819522
CD3 BUV496 Clone: OKT3	BD Biosciences	Cat#750969; RRID:AB_2875038
CD8 BUV395 Clone: RPA-T8	BD Biosciences	Cat#753756
CD4 BV605 Clone: OKT4	BioLegend	Cat#317438; RRID:AB_11218995
CD62L BV711 Clone: SK11	BD Biosciences	Cat#565039; RRID:AB_2744438
CD45RA FITC Clone: JS-83	Thermo Fisher	Cat#11-9979-42; RRID:AB_11151708
TCRGD PE-Vio770 Clone: REA591	Miltenyi	Cat#130-113-513; RRID:AB_2751122
CD14 BV510 Clone: 63D3	BioLegend	Cat#367124; RRID:AB_2716229
CD19 BV421 Clone: HIB19	BioLegend	Cat#302234; RRID:AB_11142678
CD33 PE Clone: WM53	BioLegend	Cat#303404; RRID:AB_314348
TotalSeq-A Custom Human Panel	BioLegend	Cat#99786
TotalSeq-A HTO 1	BioLegend	Cat#394601; RRID:AB_2750015
TotalSeq-A HTO 2	BioLegend	Cat#394603; RRID:AB_2750016
TotalSeq-A HTO 3	BioLegend	Cat#394605; RRID:AB_2750017
TotalSeq-A HTO 4	BioLegend	Cat#394607; RRID:AB_2750018
TotalSeq-A HTO 5	BioLegend	Cat#394609; RRID:AB_2750019
TotalSeq-A HTO 6	BioLegend	Cat#394611; RRID:AB_2750020
TotalSeq-A HTO 7	BioLegend	Cat#394613; RRID:AB_2750021
TotalSeq-A HTO 8	BioLegend	Cat#394615; RRID:AB_2750022
Biological samples
Human adult deceased organ donor samples	LiveOnNY	https://liveonny.org/
Human Subjects Blood Samples	Columbia University Irving Medical Center (CUIMC)	CUIMC IRB AAAP8763
Chemicals, peptides, and recombinant proteins
Brilliant Stain Buffer	BD Biosciences	Cat#566349
Collagenase	Millipore Sigma	Cat#11088882001
DAPI	Chemometec	Cat#910-3013
DNAse	Millipore Sigma	Cat#DN25-5G
DPBS	Corning	Cat#20-030-CV
EDTA	Corning	Cat#46-034-CI
Fetal Bovine Serum	GeminiBio	Cat#100-106
Ficoll-Paque PLUS	GE Healthcare	Cat#17-1440-03
Zombine NIR Fixable L/D	BioLegend	Cat#423105
Trustain FcX	BioLegend	Cat#422302
True-Stain Monocyte Blocker	BioLegend	Cat#426103
Critical commercial assays
10x Chromium Next GEM Single Cell 3’ Kit v3.1	10x Genomics	Cat#PN-1000121
NovaSeq 6000 S4 Reagent Kit v1.5 (200 cycles)	Illumina	Cat#20028313
Nextseq 500/550 Hi Output kit v2.5	Illumina	Cat#20024907
Deposited data
CITE-seq of sorted T cell subsets and organ donor MNCs	This paper	NCBI GEO: GSE229791
10k 3’ CITE-seq of PBMCs	10x Genomics	https://www.10xgenomics.com/resources/datasets
10k 5’ CITE-seq of PBMCs	10x Genomics	https://www.10xgenomics.com/resources/datasets
5k 3’ CITE-seq of PBMCs	10x Genomics	https://www.10xgenomics.com/resources/datasets
16-plex CITE-seq of PBMCs using fixed RNA	10x Genomics	https://www.10xgenomics.com/resources/datasets
Ab-seq of sorted T and NK cells	Trzupek et al.54	OSF: https://doi.org/10.17605/OSF.IO/EDCTN
scRNA-seq of a high-grade glioma biopsy	Levitin et al.55	NCBI GEO: GSE116621
10x Xenium of a human lymph node	10x Genomics	https://www.10xgenomics.com/resources/datasets
10x Xenium of human breast cancer biopsy	Janesick et al.56	NCBI GEO: GSE243280
Software and algorithms
MMoCHi v0.3.1	This paper	https://github.com/donnafarberlab/MMoCHi
DropSeqPipeline8	Yuan and Sims71	https://github.com/simslab/DropSeqPipeline8
scipy v1.10.1	Virtanen et al.72	https://github.com/scipy/scipy
scikit-fda v0.8.1	Ramos-Carreno et al.73	https://github.com/GAA-UAM/scikit-fda
sklearn v1.3.0	Pedregosa et al.74	https://scikit-learn.org/stable/
scanpy v1.9.3	Wolf et al.75	https://github.com/scverse/scanpy
kallisto v0.46.2	Bray et al.76; Melsted et al.77	https://github.com/pachterlab/kallisto https://github.com/pachterlab/kb_python
bustools v0.40.0	Melsted et al.78	https://github.com/pachterlab/kallistobustools
scvi-tools v1.0.2	Gayoso et al.13	https://github.com/scverse/scvi-tools
ImmClassifier v1.0	Liu et al.17	https://github.com/xliu-uth/ImmClassifier
CellTypist v1.6.2	Dominguez Conde et al.16	https://github.com/Teichlab/celltypist
HieRFIT v0.1.0	Kaymaz et al.18	https://github.com/yasinkaymaz/HieRFIT
Garnett v0.2.20	Pliner et al.19	https://github.com/cole-trapnell-lab/garnett
CHETAH v1.16.0	de Kanter et al.39	https://github.com/jdekanter/CHETAH
scClassify v1.5.1	Lin et al.40	https://github.com/SydneyBioX/scClassify
Matilda	Liu et al.41	https://github.com/PYangLab/Matilda
Muon v0.1.5	Bredikhin et al.79	https://github.com/scverse/muon
Scrublet	Wolock et al.80	https://github.com/swolock/scrublet
harmonypy	Korsunsky et al.81	https://github.com/slowkow/harmonypy
Xenium Onboard Analysis pipeline v1.5.0	10x Genomics	https://www.10xgenomics.com/support/software/xenium-onboard-analysis/latest
Shapely v2.0.2	Sean Gillies	https://github.com/shapely/shapely
Rasterio v1.3.9	Mapbox	https://github.com/rasterio/rasterio
R stats-package v4.3.1	R-core	https://www.r-project.org/

Experimental model and study participant details

Peripheral blood from a consenting healthy volunteer (34-year-old male) was obtained by venous puncture, through a protocol approved by the Columbia University IRB and complying with relevant ethical regulations for work with human participants. Human tissues were obtained from deceased organ donors as previously described.^16,30,82,83 The use of tissues from organ donors is not considered human subjects research as confirmed by the Columbia University IRB because the donors are deceased. Demographic information for all human samples can be found in Table S1.

Method details

MMoCHi

MMoCHi is designed to standardize classification of cell subsets using multimodal (e.g. CITE-seq) data (Figure 1). To train a new MMoCHi model, the user specifies a hierarchy of cell subsets, along with marker definitions (based on thresholds of either gene or protein expression) for each. Similar to other hierarchical classifiers,^18,19 hierarchies progressively segregate subsets by similarity of expression profiles, which may or may not reflect a cell type’s ontogeny. Importantly, classifiers in MMoCHi hierarchies can handle any number of subsets, allowing support for non-binary divisions of cell types at a single level. MMoCHi iterates through each level of the hierarchy, selecting high-confidence members of each subset, training a random forest classifier on normalized gene and protein expression, and classifying cell types. Once trained, MMoCHi classifiers can be interrogated for feature importances or applied to extend cell type annotation to other datasets.

Feature selection, normalization, and batch correction

As input to MMoCHi, gene expression (GEX) and antibody derived tag (ADT) expression count matrices were log-normalized, while other modalities were provided directly as calculated. For modalities with batch artifacts, MMoCHi supports batch-corrected expression, but this was not necessary for GEX in this study. For many markers, ADT expression distributions matched the expected bimodal, trimodal, or gradients of expression observed by flow cytometry. To account for variation in antibody staining across samples, ADT expression in multi-sample datasets was batch-corrected using a landmark registration algorithm (Figure 1A), similar to previously described techniques for flow cytometry and CITE-seq.^33,34 Landmark registration was applied to expression of all ADTs, except for isotype controls, which were excluded from downstream analyses. First, landmarks (peaks) were identified in the distribution of expression for each ADT in each sample as automatically detected local maxima (scipy.signal.find_peaks, using the Scipy package v1.10.1) on kernel-density-smoothed (scipy.stats.gaussian_kde) ADT expression, or manually identified. Curve registration and warping functions were applied to align these landmarks across samples (skfda.preprocessing.registration.landmark_elastic_registration_warping, using the scikit-fda package v0.8.1). For each ADT in each sample, the lowest detected peak (assumed to represent “negative” expression) and the uppermost peak (assumed to represent “positive” expression) were aligned to center on fixed values: 1 and 3, respectively. In samples with only one peak detected, the peak was aligned to the “negative” location, although this could be adjusted manually to align batches with only positive-expression for an ADT. As a batch-correction technique, landmark registration assumes that changes across samples in the degree of marker positivity on marker-positive populations is due to technical artifacts, not biological differences. This is a reliable assumption for annotation purposes, especially when samples are from similar sources or contain the same cell types at varying frequencies.³³ Once batch-corrected, thresholds delineating positive and negative populations for ADT markers could be applied across batches.

Selection of training data

At each level of the hierarchy, we selected high-confidence events for each subset using a combination of automatically-derived and manually-adjusted thresholds on the user-supplied marker definitions (Figure 1B; Table S4). MMoCHi supports logical operators (e.g., “and,” “or,” “any of the following”) in cell type marker definitions, allowing for effective training data selection for subsets that lack binary, exclusively expressed markers by using combinations of expression. ADT markers were fitted with 1- or 2- component Gaussian mixture models (sklearn.mixture.GaussianMixture, using the Sklearn package v1.3.0), and thresholds were automatically defined as 1 to 4 standard deviations above or below the means of each Gaussian. ADT thresholds were then manually refined using established flow cytometry gating principles (e.g., Ref. 37). This manual refinement was especially important for non-bimodal ADT markers, many of which are essential for many cell type classifications, but for which there are not generalizable automatic thresholding algorithms. GEX thresholds were drawn similar to Garnett,¹⁹ primarily capturing events with any marker expression, or occasionally (in genes with high background expression; Table S4) capturing only the highest expressing events using the automatic thresholding and manual refinement pipeline described for ADT markers. A portion (20%) of these high-confidence events were held out for testing and validation, and the remaining 80% could be used for training. We then resampled this training dataset (Figure 1C) to remove events likely to be mistakenly labeled due to imperfect marker thresholding (noise) and to overrepresent events likely to be misclassified (in danger). To identify “noise” and “in danger” events, principal component analysis (PCA; scanpy.pp.pca, using the Scanpy package v1.9.3) was run on the scaled expression of the top 5000 highly variable genes (scanpy.pp.highly_variable_genes) and all ADTs. The 5 nearest neighbors were calculated for each event (sklearn.neighbors.NearestNeighbors). Events were considered “noise” and removed if all neighbors disagreed with their high-confidence label. Events were considered “in danger” if less than half the neighbors agreed with their high-confidence label. Training events were also clustered using the Leiden algorithm⁸⁴ (scanpy.tl.leiden) to identify “in danger” clusters—clusters representing less than 5% of a subset’s training events. All events considered “in danger” were oversampled 5 times to increase representation during training. At classification nodes where subsets were not expected to segregate by unsupervised approaches—including Lymphocytes and TCR—these selection steps were skipped. To account for class imbalance in the training dataset, events for all subsets were oversampled to equal numbers. Finally, training events were subsampled without replacement to a maximum of 20,000 events for computational performance. The selection of training data occurred separately for each batch. If a minimum of 100 high-confidence events were not identified in a batch, events would be spiked into that batch from other batches (Figure 1D).

Training and calibration of random forest classifiers

At each level of the hierarchy, we trained a random forest (sklearn.ensemble. RandomForestClassifier) with a default of 100 trees, each with a max depth of 25 to reduce overfitting. The random forests are trained considering each feature from each modality separately, which accounts for inconsistencies in marker expression across modalities. In datasets with multiple batches, 100 trees were trained separately for each batch and added to the forest. Once trained, the forest would predict the subset identity of events and provide the proportion of trees in agreement with the classification. To convert these proportions of trees in agreement to probabilities, they were transformed using an isotonic regression (sklearn.calibration.CalibratedClassifierCV) trained on a subset of the held-out data.⁸⁵ Classification performance was then evaluated on the remaining hold-out data.

Hyperparameters were tuned in cases of poor fit, either automatically or manually. Hyperparameter tuning on the underlying random forest classifiers focused on optimizing values for the number of trees in the forest (n_estimators), the maximum depth of each tree (max_depth), and the maximum number of features to consider when looking for the best split during random forest training (max_features). For each hyperparameter tested, fit was determined by calculating the balanced accuracy score (sklearn.metrics.balanced_accuracy_score) between predicted and high-confidence labels on a subset of the held-out data (distinct from the subset used for performance evaluation). For automatic optimization, n_estimators were tested at 50, 100 (MMoCHi default), 200, 400, 800, and 1200; max_depth of 10, 25 (MMoCHi default), or unlimited; and max_features of log₂(total), square-root of total (MMoCHi default), 5% of total, or 10% of total features. Optimal hyperparameter values were selected as either the maximum score, or as the lowest value before this maximum score where score did not improve by at least 0.004. To increase fit, the max_features hyperparameter was manually set to 10% of total features for CD4⁺ and CD8⁺ memory classification for sorted T cells, as well as Lymphocyte, TCR, and CD4/8 classification for organ donor cells.

The trained classifier was then used to predict subset identity for all events at that level. This process of training data selection, classifier training, and prediction were repeated until all cells had been labeled to terminal subsets.

CITE-seq profiling of FACS sorted subsets

Peripheral blood mononuclear cells (PBMCs) were isolated from a fresh sample of peripheral blood using RosetteSep Granulocyte Depletion Cocktail (StemCell Technologies), following manufacturer’s protocols for density gradient centrifugation. Briefly, samples were incubated with the cocktail at 20°C for 10 min, diluted 1:1 with FACS buffer (DPBS 10% FBS 2mM EDTA), then layered over Ficoll-Plaque in SepMate PBMC isolation tubes (StemCell Technologies). Samples were centrifuged at 1200 x g for 10 min at 20°C, and PBMC layers were isolated according to instructions. Samples were washed (400 x g for 10 min) with FACS buffer. For further erythrocyte removal, pellets were resuspended in ACK lysis buffer (Gibco) incubated for 2 min at 37°C and washed with FACS buffer.

PBMCs were stained with Zombie NIR Fixable Viability dye (BioLegend) for 30 min. Samples were kept at 4°C in the dark. Cells were washed thrice with FACS buffer, resuspended in TrueStain FcX and TrueStain Monocyte Blocker (BioLegend), and incubated for 10 min. We designed a FACS-sort antibody cocktail, prioritizing antibody clones with discrete epitopes from the TotalSeq-A Universal Human Panel (BioLegend) to reduce steric hinderance during CITE-seq staining (Table S2). Cells were incubated with the FACS-sort antibody cocktail for 30 min, then washed thrice with FACS buffer. Cells were incubated in TrueStain FcX (BioLegend) for 10 min, then stained using a custom TotalSeq-A Universal Human Panel (BioLegend) for 30 min, according to manufacturer instructions. Samples were washed thrice with FACS buffer. Cells were sorted using a FACS Aria II (BD Biosciences; Figure S1A). Seven T cell memory populations (CD4⁺ Naive, CD4⁺ T_CM, CD4⁺ T_EM, CD8⁺ Naive, CD8⁺ T_CM, CD8⁺ T_EM, CD8⁺ T_EMRA), and monocytes were sorted into sterile, heat inactivated FBS. Sort purity was calculated as the number of events falling within a subset’s gates divided by the total number of singlet events times 100, using the same gating strategy as the sort (Figure S1A). All sorts were high purity, with mean purity 94.4% (Table S3). Sorted populations were washed in FACS buffer and resuspended in TrueStain FcX (BioLegend) for 10 min. Samples were then stained with TotalSeq-A hashtag antibodies (BioLegend). In this experiment, the hashtag-oligos (HTOs) correspond to the sorted immune cell subsets (HTO1: CD4⁺ Naive, HTO2: CD4⁺ T_CM, HTO3: CD4⁺ T_EM, HTO4: CD8⁺ Naive, HTO5: CD8⁺ T_CM, HTO6: CD8⁺ T_EM, HTO7: CD8⁺ T_EMRA, HTO8: Monocytes). Samples were then washed thrice with FACS buffer, and pooled.

CITE-seq profiling of human tissue samples

Mononuclear cells (MNCs) were isolated from tissue sites of two organ donors (D496 and D503; Table S1), as described.¹⁶ Approximately 1 million MNCs per tissue site were washed in FACS buffer and resuspended in TrueStain FcX (BioLegend) for 10 min. Samples were then stained with TotalSeq-A hashtag antibodies (BioLegend) for each tissue. For each donor, the hash-tagged MNCs from each tissue site were pooled, washed with FACS buffer, and stained with the TotalSeq-A Human Universal Cocktail panel according to the manufacturer’s instructions (BioLegend).

Library preparation, sequencing, and alignment

The sorted immune cells were counted, diluted to an appropriate volume, and loaded across two lanes of a 10X Genomics Chromium instrument targeting 6,000 cells each. Samples from each organ donor were loaded across 16 lanes of a 10X Genomics Chromium instrument targeting 10,000 cells each. cDNA synthesis, amplification and sequencing libraries were generated using the Next GEM Single Cell 3′ Kit v3.1 (10X Genomics) with the recommended modifications for compatibility with the TotalSeq-A cell hashing and CITE-seq reagents (BioLegend). Organ donor GEX libraries were sequenced on a NovaSeq 6000 (Illumina) with 100 cycles for reads 1 and 2. Organ donor ADT and HTO libraries and all FACS sorted subset libraries were sequenced on a NextSeq 500 (Illumina) with 28 cycles for read 1 and 44 cycles for read 2.

Reads were analyzed by pseudoalignment using kallisto v0.46.2 (GRCh38 with Gencode v24 annotation) and bustools v0.40.0.^76,77,78 CITE-seq and hashtag barcodes were demultiplexed and extracted using DropSeqPipeline8, as previously described.⁷¹ Hashtags were demultiplexed by CLR normalization, k-means clustering, and statistical identification of singlets by fitting a negative binomial model as described.³⁶ For both datasets, the GEX matrices were normalized to log(counts∗10,000/total_counts+1), and ADT expression matrix were normalized to log(counts∗1,000/total_counts+1).

Analysis and benchmarking using sorted T cells

Cells from the T cell sort were filtered to remove events with fewer than 1000 unique counts, fewer than 200 genes detected, or over 10% mitochondrial counts. A MMoCHi hierarchy was developed (Figure 2B) and classification performed using the algorithm above. In flat classification variations (MMoCHi Flat, MMoCHi GEX Flat), high-confidence thresholding and classification were performed for all subsets in a single classification node. In GEX variations (e.g., MMoCHi GEX, MMoCHi GEX Flat, etc.), ADT expression data was excluded from the detection of in-danger noise events and random forest training. In “Sort-ref” variants (e.g., MMoCHi Sort-ref, MMoCHi GEX Sort-ref, etc.), training events were selected using the sort labels instead of high-confidence thresholding. For benchmarking tests, to mirror intra-dataset performance testing of other reference-based tools^16,18 and avoid leaky pre-processing bias⁸⁶ a portion of the dataset (20%) was set aside to only be used for performance evaluation (completely held-out from the delineation of high-confidence thresholds, training data selection, random forest training, random forest calibration, and any manual or automatic hyperparameter optimization). To verify the benchmarking represented consistent trends when comparing MMoCHi to other tools, we applied 5-fold cross validation and reported the mean performance. To test MMoCHi’s robustness, performance was also tested by subsampling training event numbers and downsampling reads. Random subsampling was performed on the number of monocyte or CD8⁺ T_CM events included in the training dataset, down to 5% of the total of monocyte/CD8⁺ T_CM events profiled. Aligned reads for GEX and ADT libraries were downsampled prior to generation of raw-count expression matrices. Downsampling was performed down to 1% of the total aligned reads, and the effect of this downsampling on total counts is displayed.

We manually annotated unsupervised clusters by average expression of gene and protein markers. Leiden clusters of gene expression (GEX Leiden) were computed on a PCA of the highly variable genes with Scanpy⁷⁵ defaults. Leiden clusters of ADT expression (ADT Leiden) were computed on a PCA of all ADTs except for isotype controls. totalVI latent space was computed on all ADTs except for isotype controls, and highly variable genes selected using top 4000 genes as defined by the Seurat v3 method,⁸⁷ as recommended¹³ (scvi.model.TOTALVI.train and scvi.model.TOTALVI.get_latent_representation using scvi-tools v1.0.2). To ensure results of totalVI were not sensitive to the number of features used, we also computed totalVI latent spaces using the top 1000, 2000, and 8000 highly variable genes. Leiden clustering was performed on the 10 nearest neighbors of the top 40 principal components or the entire totalVI latent space. Over-clustering was also computed where highly variable gene selection, dimensionality reduction, and Leiden clustering were repeated to sub-cluster each cluster (excluding clusters of fewer than 40 events) resulting in 72 GEX clusters (GEX Leiden OC), 128 ADT clusters (ADT Leiden OC), and 116 totalVI clusters (totalVI Leiden OC). For visualization, UMAP (scanpy.tl.umap) embeddings of each of these feature spaces were calculated using defaults.

ImmClassifier,¹⁷ CellTypist,¹⁶ and HieRFIT¹⁸ were applied as recommended, using pre-trained models downloaded from online. ImmClassifier was run using docker with the built-in model, trained by inputting a log₂(counts∗10,000/total_counts+1) transformed GEX matrix, as recommended. CellTypist v1.6.2 was run by applying the “Healthy_COVID19_PBMC” model (trained on PBMCs from healthy individuals and individuals with COVID-19⁸⁸) with and without majority voting enabled (celltypist.annotate; CellTypist MV and CellTypist, respectively). A HieRFIT v0.1.0 model trained on 10x PBMC data was applied (HieRFIT::HieRFIT) with defaults settings.

Garnett¹⁹ models were trained using high-confidence thresholding with the provided online hierarchy (Provided Markers) and on manually and automatically (garnett::top_markers) selected transcript markers which were evaluated using built-in functions (garnett::check_markers; Manual Markers; Table S6). Garnett v0.2.20 models were trained (garnett::train_cell_classifier) with marker propagation disabled and prediction (garnett::classify_cells) was performed with cluster extension disabled.

Additionally, CellTypist, HieRFIT, Garnett, CHETAH, scClassify, and Matilda were trained using the HTO-derived sort labels as a reference dataset and evaluated using a 20% hold-out with five-fold cross validation (“Sort-ref”). A CellTypist model was trained (celltypist.train) with two-pass training enabled for feature selection. This model was applied to held-out data with and without majority voting enabled (CellTypist MV Sort-ref and CellTypist Sort-ref, respectively). HieRFIT and Garnett models were both trained (HieRFIT::CreateHieR; garnett::train_cell_classifier) using a hierarchy structured identically to the MMoCHi hierarchy (Figure 2B) and applied to held-out data (HieRFIT Sort-ref and Garnett Sort-ref, respectively). CHETAH and scClassify hierarchies were generated by the tools’ respective built-in algorithms using default settings during training on the HTO-derived sort labels. The scClassify model was trained using an ensemble method of feature selection methods: differential expression, χ²-test, Bartlett’s test, Kolmogorov-Smirnov test, and the bimodality index (selectFeatures=c(“chisq”, “limma”, “DV”, “DD”, “BI”)), and classification was performed using an ensemble of similarity metrics: Pearson correlation, Spearman correlation, Cosine similarity, Jaccard index, and Kendall-rank correlation coefficient (similarity = c(“pearson”, “spearman”, “cosine”, “jaccard”, “kendall”, “weighted rank”)). These models were applied to and evaluated on the held-out data as described (CHETAH Sort-ref; scClassify Ensemble Sort-ref). Additionally, scClassify was evaluated using the reported best-performing single feature-selection and similarity metric combination (scClassify Pearson DE Sort-ref). Matilda was trained on raw gene expression counts or on both raw gene and protein expression counts using the “main_matilda_rna_train.py” and “main_matilda_train.py” scripts, respectively. The models were then applied to held-out data using the “main_matilda_rna_task.py” and “main_matilda_task.py” scripts with the “query” flag set to “True” and evaluated as described (Matilda GEX Sort-ref; Matilda Sort-ref). CellTypist, HieRFIT, and Garnett were each also retrained using the combined multimodal matrix of GEX and ADT expression (“Multi”). To coerce these tools to accept the multimodal matrix, the check_expression parameter in CellTypist functions and the db parameter in Garnett functions were disabled.

Precision, recall, F1 scores and overall accuracy were calculated using sort labels as truth (sklearn.metrics.precision_recall_fscore_support; sklearn.metrics.accuracy_score). F1 score was calculated as the harmonic mean of precision and recall. “Agreement” between annotations was calculated as the total percentage of events that received identical labels between two methods. Garnett and HieRFIT can provide unknown and intermediate cell type labels, which were quantified and excluded from performance metrics calculations. Pre-trained classifier models could also provide erroneous (e.g., not included in the FACS sort) or alternative (i.e., potentially valid annotations but not using the subsetting-axis of interest) cell type labels, which were also quantified, then excluded from performance calculations.

Analysis of ADT batch-correction with published CITE-seq data

To test landmark registration against other, published CITE-seq normalization techniques, we used four public datasets of CITE-seq performed on healthy PBMCs available from 10x Genomics (Figure S4A). Each dataset was treated as a separate batch. For each batch, log(counts∗1,000/total_counts+1) and landmark registration were performed as described above. For these data, no manual modifications were made to the built-in automatic peak detection. CLR and dsb normalizations were performed using the Muon package⁷⁹ v0.1.5 (muon.prot.pp.clr with axis set to 1; muon.prot.pp.dsb). The dsb algorithm uses empty droplets to identify protein-specific noise originating from unbound antibodies.⁴² Empty droplets were defined as events that contained between 2 and 3 log₁₀(ADT-counts), or (for 5k_pbmc_3p) between 1.5 and 2.8 log₁₀(ADT-counts) in the provided unfiltered (“raw”) matrix and did not appear in the filtered matrix. Isotype controls (IgG1, IgG2a and IgG2b) were then used to remove cell-specific noise, as described.⁴² Methods were evaluated using UMAPs and Leiden clustering computed as described above, using normalized expression of the 13 proteins profiled in all batches (CD3, CD4, CD8a, CD14, CD16, CD19, CD25, CD45RA, CD45RO, CD56, CD127, PD-1, and TIGIT). Nearest-neighbors were computed using “spearmanr” as the metric, as ADT expression was not scaled (to best-preserve each integration method). Batch integration was quantified with adjusted rand index (ARI) (sklearn.metrics.adjusted_rand_score). To test whether landmark registration was robust to changing cell proportions, landmark registration was also performed after removing 95% of the events with positive expression of CD14 (defined as above 2 log(CP1k+1)). The effect of landmark registration on MMoCHi classification was tested by applying MMoCHi to classify subsets across batches to log(CP1k+1)-transformed or landmark registered protein expression, with thresholds manually defined for all batches together or each batch separately (Table S4).

Analysis of organ donor sequencing

CITE-seq data obtained from tissue immune cells was filtered separately. Although hashtag demultiplexing removes inter-sample multiplets, this method cannot detect multiplets between cells from the same tissue site. Thus, multiplets were detected by Scrublet⁸⁰ (scrublet.Scrublet.scrub_doublets), using default settings, an expected doublet rate of 0.015, and applying separately to each library-tissue combination of over 100 events. Cells were then filtered to remove events with fewer than 1000 unique counts, fewer than 600 genes detected, or over 10% hemoglobin counts. We used a cluster-based method to remove events either identified as doublets by Scrublet or with high mitochondrial counts, similar to the previously described percolation method.¹⁶ Briefly, a PCA on highly variable genes was calculated, integrated using Harmony⁸¹ (scanpy.external.pp.harmony_integrate), and used for nearest neighbor calculation and Leiden clustering with a resolution of 20. Clusters with significantly higher Scrublet scores (above 0.1) or percent mitochondrial counts (above 15%) were removed. Individual events not captured by these clusters that were identified as Scrublet doublets or had over 25% mitochondrial counts were also removed.

We devised a MMoCHi hierarchy (Figure 3D; Table S4) and performed classification using the algorithm above. Due to insufficient numbers, stem cell-like memory T cells (T_SCM) were not included in the hierarchy (Table S7). Classification fit across a varying number of random forest estimators (trees) was evaluated for prediction accuracy. Overall prediction accuracy was measured using internally held-out events (events that were high-confidence thresholded but excluded from the training of random forest classifiers). Performance metrics were calculated as above for individual classification nodes using precision, recall, and F1 scores for each child subset, weighted by the number of internally held-out, high-confidence events (average=’weighted’). Gini impurity-based feature importances were automatically calculated during random forest training by scikit-learn. Manual annotation and UMAP calculation were performed on the totalVI latent space, as described above. The clustered dendrogram of classified subsets in each donor was calculated (scanpy.tl.dendrogram) on the totalVI latent space with optimal ordering enabled. In select cases of disagreement between MMoCHi and manual annotations, differential expression was performed between the disagreeing events (e.g., MMoCHi-classified CD8⁺ T cells that were manually-annotated as CD4⁺ T cells against manually-annotated CD8⁺ T cells that were MMoCHi-classified as CD4⁺ T cells) on the ADT expression or GEX matrices (Figures S12B–S12D and S17E–S17G, and Table S8).

Pre-trained classifiers were applied to organ donor PBMC events that were classified as one of: CD4⁺ Naive, CD4⁺ T_CM, CD4⁺ T_EM, CD8⁺ Naive, CD8⁺ T_CM, CD8⁺ T_EM, CD8⁺ T_EMRA, or classical monocytes. MMoCHi or Garnett models trained on the sorted FACS dataset (FACS-trained MMoCHi, FACS-trained MMoCHi GEX, and FACS-trained Garnett) or pre-trained ImmClassifier or CellTypist models were applied as described above. For ImmClassifier, CellTypist, and Garnett, Alternative, Unknown, or Erroneous classifications were quantified as described above.

To display of features associated with each subset (Figure S10), the top important features with a log₂(fold-enrichment) > 2 for that subset. For identification of transcriptomic markers for Naive T cells and T_CM (Figure 5), MMoCHi models were retrained at the CD4⁺ and CD8⁺ Naive/T_CM level, given the same multimodal high-confidence thresholding, but trained on only the GEX matrix. To ensure these GEX markers were not learned through tissue-specific contamination, additional models were trained using only T cells from organ-donor blood, and features displayed had to be in the top 1% of important features in both models. To improve robustness, they were additionally filtered to only those with a greater than 10% change in dropout rate.

Application of MMoCHi to other modalities

To apply MMoCHi to other modalities beyond CITE-seq, we used three public datasets: Ab-seq of sorted T and NK cells,⁵⁴ scRNA-seq of a glioma biopsy,⁵⁵ and Xenium of a human lymph node and breast cancer biopsy⁵⁶ obtained from 10x Genomics. For the Ab-seq dataset, only resting cells were used. The gene expression matrix was normalized to log(counts∗1,000/total_counts+1), and the protein expression matrix was normalized to log(counts∗100/total_counts+1). Clustering was performed on a totalVI latent space, calculated as described above. Manual annotations of these clusters were highly congruent with published annotations.⁵⁴ In the glioma dataset, the gene expression matrix was normalized to log(counts∗10,000/total_counts+1). The ratio of the average expression of genes on chromosome 7 to chromosome 10 was used as an additional marker, effectively segregating malignant cells, as described.⁵⁵ Manual annotations and UMAP coordinates were obtained from the published clustering analysis. In the Xenium datasets, the gene expression matrix was normalized to log(counts∗100/total_counts+1). Gene-expression probes marked as “negative control” or “blank” were removed. The datasets were filtered to remove any events with fewer than 10 total counts and 5 unique genes expressed. The remaining events had sufficient gene expression for application of high-confidence thresholding, following the methods above. Morphological features were calculated using the nucleus and cell boundaries, as determined by the Xenium Onboard Analysis pipeline v1.5.0. Nuclear or cellular area was calculated in μm² by converting boundary vertices to Polygon objects using the Shapely package v2.0.2 (shapely.geometry.Polygon; shapely.area). Nuclear or cellular roundness, or circularity, was calculated as the ratio of a shape’s area to the area of the minimum bounding circle of that shape (shapely.minimum_bounding_circle). Nuclear DAPI intensity was calculated by obtaining average pixel-intensity of the provided 2D maximum intensity projection (“MIP") image. To obtain the pixels overlapping with the nucleus, we used the Rasterio package v1.3.9 (rasterio.features.rasterize) using a transformation (affine.Affine.scale) of 0.2125μm per pixel, as defined in image metadata. These features were supplied to MMoCHi without additional scaling. Manual annotation of clusters was performed on the provided graph-based clusters and displayed on the provided graph-based UMAP projection. MMoCHi was applied to all four datasets using hierarchies crafted using known marker genes, proteins, or chromosome-level expression patterns (Table S4). Where possible, markers were selected to match the cell-type definitions used for the original manual annotation. For figure generation, Xenium cell boundaries were plotted on top of the provided 2D autofocus projection image of DAPI staining using the same scaling as above. Cell boundaries were shrunk slightly for display to reduce drawing overlap (using the shapely.geometry.Polygon.buffer method, with join_style “mitre”). Feature importances were obtained as described above. Performance comparisons of MMoCHi versus cluster-based or expert Xenium annotations were displayed for known marker genes. Comparisons with <50 cells (breast cancer biopsy) or <200 cells (lymph node) were excluded from visualization.

Estimation of computational performance

We measured the computational resources required to train and apply MMoCHi classifiers using a predefined hierarchy and thresholds (Figure 3C; Table S4). Comparisons were performed with multiprocessing enabled for random forest training, using 3^rd generation Intel Xeon Scalable processors (3.5 GHz) with 32 vCPUs and 32 GiB of RAM (AWS/EC2 c6i.8xlarge instance). Tests at varying event counts were performed by randomly subsampling the dataset prior to classification.

Quantification and statistical analysis

Statistical comparison of classification performance was calculated using a Friedman rank-sum test (stats::friedman.test using the stats library v4.3.1), matched by cell-subset, on F1 scores, followed by multiple comparisons using a paired two-sided Wilcoxon signed-rank test (stats::pairwise.wilcox.test) with FDR correction. Significance groups for each method are displayed by lettering, where methods not sharing a letter are significantly different (p < 0.05).

Prior to differential expression, the two groups were subsampled to the same number of events and their expression matrices were downsampled to equal total counts. Differential expression was performed (scanpy.tl.rank_genes_groups) on log-normalized expression, with Wilcoxon mode, tie correction, and FDR correction enabled. Genes or proteins with |log₂(fold change)| > 2 and p < 0.05 were identified as significantly differentially expressed. Statistical details for each experiment can be found in figure legends for each respective analysis.

Published: January 14, 2025

Supplemental information

Table S1. Donor metadata, related to Figures 2 and 3

Metadata for the organ donors (D496 and D503) and the blood donor (Blood Donor A) used in this study.

Table S2. Antibodies, related to STAR Methods

A list of antibody clones and sources used in the FACS panel, in the custom TotalSeq-A CITE-seq panel, and for hashtag-oligo (HTO) staining. Where applicable, the conjugated fluorophore or barcode sequence used for alignment is included.

Table S3. Sort purity, related to Figure 2

Sort purity of T cell subsets and monocytes after FACS. Purity is represented as the percentage of singlet events falling within the sorted subset’s gates using the same gating strategy as the sort (Figure S1A).

Table S4. MMoCHi hierarchies, related to Figures 2, 3, and 6

Subset definitions and thresholds used to construct MMoCHi hierarchies. Each sheet represents a separate MMoCHi hierarchy or associated high-confidence thresholds. Hierarchical relationships are defined in the column “Parent” (subsets with no parent represent the first classification level). High-confidence events for each subset are selected as events matching the listed marker expression (above the positive threshold for all “Positive” markers and below the negative threshold for all “Negative” markers). Gene expression markers are suffixed with “_gex”. Thresholds for immune CITE-seq protein markers are defined using log-normalized expression (log(CP1k + 1) for the FACS datasets) or landmark-registered expression (for the organ donor dataset). Thresholds for FACS and organ donor CITE-seq gene expression markers are defined using log(CP10k + 1). Thresholds for gene expression markers are in log(CP100 + 1) for the Ab-seq and Xenium datasets and are in log(CP10k + 1) for the Glioma dataset. Thresholds for other modalities are in varied units, as described (see STAR Methods). The abbreviation “OD” refers to the organ donor dataset (Figure 3), “BC” refers to the Xenium breast cancer dataset (Figure S18), and “LN” refers to the Xenium lymph node dataset (Figure 6). Sheets labeled with “Automatic” thresholds refer to thresholds determined by MMoCHi’s automatic thresholding without manual tuning (see STAR Methods). Gene thresholds defined by Gaussian mixture modeling are indicated by ∗.

Table S5. Annotation performance metrics, related to Figures 2 and 4

F1 score, precision, recall, and accuracy of MMoCHi and various annotation methods compared to sorted population labels (Figure 2), as defined by hashtag expression (see STAR Methods). For tools that produced non-relevant classifications, the total proportions of alternative, unlabeled, or erroneous classifications are quantified (see STAR Methods). Cross-dataset sheet refers to the performance of pre-trained MMoCHi and other pre-trained classifiers on organ donor PBMCs compared to MMoCHi classification labels (Figure 4), as defined by hashtag expression (see STAR Methods).

Table S6. Garnett definitions, related to Figure 2

Garnett marker file providing the hierarchy and gene expression marker-based definitions used for classification. See STAR Methods for details on marker selection.

Table S7. Insufficient availability of T_SCM in organ donor dataset for classification, related to Figure 3

Total number of cells across D496 and D503 that could be defined as high-confidence CD4⁺ or CD8⁺ T cells by protein gates, and percentage of those cells that could be identified as T_SCM. Thresholds for each marker are defined in Table S4.

Table S8. Differential expression, related to Figures 3, 5, and 6

Differential expression of gene expression (GEX) or antibody-derived tag (ADT) expression between each MMoCHi-classified subset listed and all other subsets at the same classification level or between two specified subsets is shown. For discordant comparisons, differential ADT expression between MMoCHi/manual or supplied annotation groups with disagreeing annotation is shown. The abbreviation “OD” refers to the organ donor dataset (Figure 3), “BC” refers to the Xenium breast cancer dataset (Figure S18), and “LN” refers to the Xenium lymph node dataset (Figure 6).

Table S9. Important features for classification, related to Figure 5

Gini impurity-based importance for each feature used for classification at each level of the organ donor classification. Feature names for each level are suffixed with “_mod_GEX” for gene expression, “_mod_landmark_protein” for landmark registered protein expression, “_mod_protein” for Ab-seq protein expression, “_mod_aneuploidy” for averaged chromosome expression in the Glioma dataset, and “_mod_physical_attributes” for morphological information extracted from Xenium imaging. For naive/T_CM analysis, classification was performed on T cells from all donor tissue sites (cd4_ncm, cd8_ncm) and on only T cells from organ donor blood (cd4_ncm_bld, cd8_ncm_bld).