Computational tracking of cell origins using CellSexID from single-cell transcriptomes

Summary

Cell tracking in chimeric models is essential yet challenging in developmental biology, regenerative medicine, and transplantation research. Current methods like fluorescent labeling and genetic barcoding are technically demanding, costly, and often impractical for dynamic tissues. We present CellSexID, a computational framework that uses sex as a surrogate marker for cell-origin inference. By training machine-learning models on single-cell transcriptomic data, CellSexID accurately predicts individual cell sex, enabling in silico distinction between donor and recipient cells in sex-mismatched settings. The model identifies minimal sex-linked gene sets through ensemble feature selection and has been validated using public datasets and experimental flow sorting, confirming biological relevance. We demonstrate CellSexID’s applicability beyond chimeric models, including organ transplantation and sample demultiplexing. As a practical alternative to physical labeling, CellSexID facilitates precise cell tracking and supports diverse biomedical applications where mixed cellular populations need to be distinguished.

Graphical abstract

Highlights

• •In silico cell-origin tracking via sex prediction in scRNA-seq
• •Robust ensemble marker selection across tissues and species
• •Broadly validated in chimeric, transplant, and demultiplexing studies
• •Bypasses costly, labor-intensive experimental labeling approaches

Motivation

Accurately determining the origin of individual cells is essential for studies involving chimeric models, organ transplantation, and pooled single-cell experiments. Existing genetic labeling, barcoding, or chromosome-based assays are costly, technically demanding, and often lack scalability or single-cell resolution. CellSexID is a machine-learning framework that leverages sex-specific gene expression as a natural surrogate for origin in sex-mismatched settings. It identifies minimal, robust marker sets and infers per-cell origin directly from transcriptomes, eliminating the need for genetic engineering or physical labeling. This scalable, cost-effective approach is applicable across tissues, species, and diverse biomedical research scenarios.

Tai et al. introduce CellSexID, a machine-learning framework for in silico tracking of cell origin in sex-mismatched single-cell RNA-seq data. By leveraging sex prediction as a surrogate, it bypasses experimental labeling and enables accurate origin tracking across chimeric models, organ transplantation, and sample demultiplexing.

Introduction

Understanding how cellular ontogeny influences function requires accurate cell-origin tracking across developmental biology, regenerative medicine, immunology, and disease modeling.^1,2 In both experimental and clinical contexts, cell populations of mixed origin frequently arise—for example, in transplantation models, tissue regeneration, or pooled donor samples.³ Chimeric models, which contain cell populations derived from genetically distinct sources, are a critical tool for such investigations.⁴ These models allow researchers to trace cell lineages throughout an organism’s lifespan, offering insights into processes such as development, aging, immune reconstitution, and disease progression. For example, studies of macrophages—immune cells that arise from both prenatal progenitors and postnatal hematopoietic stem cells (HSCs)—have revealed distinct functional roles based on their developmental origin.⁵ Prenatal-origin macrophages are seeded into tissues early in development and can self-renew, while HSC-derived macrophages are continuously replenished by blood-derived monocytes originating from the bone marrow.⁶ Despite macrophages’ prevalence and functional diversity across all tissues, the specific contributions of these distinct populations of different origins to tissue maintenance, immune responses, and disease remain poorly understood.

Despite its importance, accurate identification of cellular origin at single-cell resolution remains technically challenging. For example, mouse chimeras generated through bone-marrow transplantation into recipient hosts of different origins are a frequently employed and powerful tool in biological experimentation. However, existing methods for distinguishing the donor from recipient cells face significant restraints. Traditional approaches rely on tracking genetically engineered artificial markers, such as distinct CD45 alleles or fluorescent reporters. Unfortunately, these markers are often not readily available in the mouse models under study, leading to the need for labor-intensive breeding programs and preparation steps.^7,8 Additionally, classical methods involving reporter systems designed for lower-throughput analyses often rely on invasive labeling procedures or genetic modifications that are impractical for large-scale single-cell studies or primary human samples. Another strategy that does not require genetic engineering employs sex-mismatched chimeras, using biological sex as a surrogate marker to determine cell origin by detecting either sex chromosomes^9,10 or their associated gene products.¹¹ Techniques like fluorescence in situ hybridization (FISH) can detect Y chromosomes in individual cells, providing single-cell resolution, but are only semi-quantitative and incompatible with high-throughput single-cell transcriptomic analysis.^9,10 On the other hand, PCR-based methods to detect sex-related genes allow for bulk-level quantification of chimerism but lack the single-cell resolution necessary for individual cell-origin identification.¹¹ The inherent limitations of these current techniques hinder our ability to fully characterize and quantify ontogenically distinct cell populations within tissues at the single-cell level and consequently restrict advances in fields such as immunology, oncology, developmental biology, and regenerative medicine.

To overcome these limitations, we present CellSexID, a machine-learning framework designed to leverage sex as a surrogate marker to identify cellular origin in sex-mismatched chimeric models at single-cell resolution. Unlike prior methods, CellSexID does not require genetic engineering or external labeling and integrates seamlessly into experimental designs by simply including male and female donors or recipients. One of the core functionalities of CellSexID is its ensemble feature-selection scheme, which rigorously identifies a minimal yet informative set of sex-linked marker genes that optimize classification performance from a training reference single-cell dataset with male and female cells. By training on our default gene panel, CellSexID achieved robust area under the precision-recall curve (AUPRC > 0.94) on a variety of datasets from different tissues and species. CellSexID is also a versatile tool, not limited to chimeric mouse models, justified by its accurate prediction in other scenarios like organ transplantation and demultiplexing. Notably, CellSexID reliably distinguished male donor from female recipient cells in our validation datasets. Furthermore, we demonstrate that CellSexID’s annotations can enable important biological insights by applying it to an independent dataset (chimeric mouse diaphragm experimental dataset). CellSexID enabled separation of HSC-derived donor (male) from non-HSC-derived host (female) macrophages. Among the nine macrophage clusters identified in healthy adult skeletal muscle by single-cell RNA sequencing (scRNA-seq),¹² only one was almost exclusively associated with a non-HSC origin and exhibited a unique anti-inflammatory gene-expression profile.

Our findings underscore CellSexID’s ability to uncover ontogeny-specific gene signatures and functional characteristics. This scalable and efficient approach enables high-resolution cell-origin tracking across diverse biological contexts. CellSexID advances studies of ontogeny across tissues, with applications in inflammation, cancer, and stem cell therapies.¹³ It is also applicable to diverse settings such as tracking donor and host cells in human organ transplants, distinguishing mixed-sex populations in organoids or pooled samples, and improving sex-aware quality control in large-scale single-cell datasets. Additionally, CellSexID facilitates retrospective sex stratification in public datasets and supports investigations of sex-biased diseases. CellSexID enables precise single-cell-origin analysis for developmental biology, immunology, and regenerative medicine, advancing understanding of cellular function, lineage-specific responses, and disease progression.

Results

Method overview

CellSexID, as illustrated in Figure 1, first identifies gene features that are predictive of sex using a committee of four machine-learning classifiers: XGBoost,^14,15,16 support vector machine,^17,18,19 random forest,^{20,21,22,23,24} and logistic regression.^25,26,27 The committee determines important features based on classifier consensus (Figure 1A). Using publicly available mouse adrenal gland scRNA-seq datasets containing samples from both sexes,²⁸ the committee identified 14 critical sex-linked markers that enable accurate sex prediction. Based on these selected features, the machine-learning classifiers are then trained and predict the sexes for single cells. Validations using public datasets from various tissues and species demonstrated near-perfect prediction performance, highlighting the robustness of the selected features and confirming the feasibility of using single-cell gene expression for sex prediction with machine learning. Please refer to STAR Methods and GitHub repository for machine-learning model details.

Figure 1.

Overview of the CellSexID method

(A) Identification of sex-specific gene features for predicting cell origin in single-cell RNA sequencing (scRNA-seq) data. Four classifiers (random forest, logistic regression, support vector machine [SVM], XGBoost) select sex-distinguishing genes via majority voting.

(B) Experimental validation of CellSexID using chimeric mouse models. Female mice are transplanted with male bone marrow, creating chimeric populations. Diaphragm macrophages are isolated, flow sorted by CD45 markers, barcoded with CMOs, and sequenced. The scRNA-seq data from chimeric mice are used to test classifiers trained on a public dataset, with predictions visualized in UMAP space and compared against flow-cytometry ground truth for validating CellSexID predictions.

(C) Evaluation of the model’s predictive performance showing predicted cell-sex percentages compared to flow-cytometry-derived ground truth, providing an assessment of reliability.

(D) Application of CellSexID for annotating cell origin in chimeric mice through a streamlined workflow enables a range of single-cell analyses, supporting studies of cellular dynamics and differences between recipient and donor cells in diverse research contexts.

See also Figure S1.

To evaluate the real-world applicability of CellSexID, we validated CellSexID using sex-mismatched chimeric mice with CD45 lineage tracing. Diaphragm macrophages from sex-mismatched chimeric mice were sorted with fluorescence-activated cell sorting (FACS), cell multiplexing oligonucleotide (CMO) labeled to encode male and female origin, and profiled by scRNA-seq (Figure 1B). This validation dataset, with flow-cytometry-confirmed ground truth, assessed model performance in distinguishing donor (male) from recipient (female) cells. Models trained on public adrenal gland datasets achieved high accuracy on this chimeric mouse diaphragm validation dataset, confirming CellSexID’s generalizability (Figure 1C). To demonstrate biological utility, we applied CellSexID to a separate chimeric mouse diaphragm experimental dataset (Figure 1D). Differential expression and pathway analyses revealed distinct donor-recipient profiles, highlighting CellSexID’s capacity to track origins and dissect cell-specific programs relevant to disease and therapy.

Ensemble selection identifies minimal 14-gene set

First of all, we justify the necessity of using a machine-learning-based method for selecting multiple gene features and predicting sex by comparing with a rule-based method relying on a single sex chromosome gene. Figure S1D shows that our machine-learning models significantly outperform a rule-based classifier: if Uty > 0, then male; otherwise, female; this is based on the assumption that sex can be determined solely from Y chromosome gene expression. This is likely due to the challenges posed by high dropout rates and variability in the expression of sex-chromosome-linked genes in scRNA-seq data. Such limitations make the genes on sex chromosomes insufficiently reliable for accurate classification without additional features.

For our machine-learning classifiers, using a limited number of key gene features for sex-based cell tracking decreases both the complexity and cost of future experimental applications. Therefore, to enhance real-world applicability of our sex-based cell-origin model, we focused on identifying a minimal gene set capable of ensuring accurate sex determination. As shown in Figure S1A, our feature-selection strategy achieves optimal prediction performance compared to models without feature selection or those with varying numbers of top features. We compared models selecting top 10, 20, or 200 genes per committee member, plus a baseline using all genes, deriving the final panel from genes selected by ≥3 members. Our results demonstrate that selecting the top 20 genes from each model—resulting in a final panel of 14 genes—achieves the highest predictive performance. Additionally, optimizing feature selection can improve the computational efficiency and accuracy of our model. To identify the most predictive gene features, we employed an ensemble feature-selection approach,^29,30,31,32 combining four machine-learning algorithms: XGBoost, support vector machine, random forest, and logistic regression. Figure S1B shows that this ensemble approach outperforms individual models, resulting in superior prediction accuracy.

Our ensemble approach was then trained and tested on publicly available scRNA-seq datasets from the Gene Expression Omnibus (GEO), featuring mouse adrenal gland immune cell from female (GEO: GSM6153751; 3,963 cells) and male (GEO: GSM6153750; 2,685 cells) mice.²⁸ Figure 2A displays the top 20 genes with the highest feature-importance scores across each model. Figure 2B shows the same results when the feature selection is applied to a human acute myeloid leukemia (AML) bone-marrow dataset (GEO: GSE289435).³³ Certain genes consistently ranked as top predictors across different models. By analyzing the committee votes of each gene, we identified a robust set of genes (14 for mouse and nine for human) selected by at least C = 3 of the four models (Figures 2C and 2D; see Figure S1C for justification of this threshold). These sex-linked gene markers were used in the cell sex predictions in subsequent analyses due to parsimony and their consistently high performance across multiple algorithms.

Figure 2.

Committee-based machine-learning approach identifies key gene markers for sex prediction in mouse and human datasets

(A) Stacked bar plot showing scaled feature-importance contributions from each committee member—logistic regression (LR), support vector machine (SVM), XGBoost, and random forest (RF)—for mouse gene markers identified using the adrenal gland dataset (GEO: GSM6153751 and GEO: GSM6153750). The contribution scores are the average of five technical replicates generated using random seeds. In each run, 80% of the data are used for training the model and the remaining 20% are for testing and computing the score. Red boxes highlight genes with the highest consensus scores. Feature importance was calculated using min-max normalization within each model followed by summation across committee members.

(B) Gene markers identified from the human AML bone-marrow dataset (GEO: GSE289435, “single-cell transcriptional atlas of hematopoiesis”). Due to the large dataset size (>80,000 cells), one-fifth of the data were randomly sampled for feature selection in each technical replicates.

(C and D) Final gene-marker tables showing top-ranked consensus genes included in the final marker list for mouse (C) and human (D), including chromosomal locations, full gene names, and NCBI gene IDs for cross-species ortholog mapping.

See also Figure S2.

It is notable that, among the 14 selected genes, only five are found on sex chromosomes, whereas the rest are autosomal. The sex-chromosome-linked genes have a combination of X chromosome (Xist) and Y chromosome (Ddx3y, Eif2s3y, Uty, Kdm5d) locations. Xist plays a central role in X chromosome inactivation in females.³⁴ Ddx3y and Eif2s3y dictate male-specific functions such as spermatogenesis.^35,36 Uty³⁷ and Kdm5d,³⁸ located on the Y chromosome, are also recognized as reliable sex markers, with their expression limited to males and their roles linked to histone modifications and chromatin organization.³⁹ Among the nine autosomal genes identified, several are also known to exhibit sexual dimorphism in other cell types.^40,41 To further validate our findings, we compared the performance of our 14 selected genes against features limited to only the Y chromosome or the two sex chromosomes together. As shown in Figure S2A, the genes selected by our model, including those located on autosomal chromosomes, led to a superior predictive performance. Specifically, the "all chromosomes" condition includes all 14 selected gene features (selected using our feature-selection strategy from all genes on all chromosomes), while the "established sex markers from selected" condition uses only the sex-linked genes from the 14-gene set. This comparison between “all chromosomes” and "established sex markers from selected" provides a clear justification for the inclusion of autosomal genes. The results clearly show that models using the full 14-gene set outperform those restricted to sex-chromosome-linked genes. In this regard, we applied the same feature-selection procedure to sex-linked genes alone “X and Y chromosomes only” and Y chromosome genes alone “Y chromosome only,” and both subsets exhibited inferior performance compared to the first bar with our 14 features selected from all genes. This underscores the value of our end-to-end feature-selection approach, powered by the machine-learning model committee, in identifying robust sex markers beyond sex chromosomes. A more comprehensive list of identified features is available in Table S1.

In addition, to determine whether these genes have been previously implicated in sex determination or related biological processes, we performed literature searching and found that all the autosomal genes in the sex-linked marker list of human and mouse are associated with sex-relevant functions. In mice, sex relevance was supported for Gm42418,⁴² Uba52,⁴⁰ Wdr89,⁴³, Lars2,⁴⁴ and Rps27.⁴⁵ For example, Uba52 exhibits sex-specific regulation in the amygdala during fear memory formation,⁴⁰ and Wdr89 shows sex-biased expression in the postnatal pituitary gland, implicating its potential role in hormonal regulation.⁴³ Lars2 has been identified as a gene with sex-dimorphic expression across multiple cell types, as revealed by single-cell transcriptomic analyses.⁴⁴ In human, IFIT2^46,47 and IFIT3^47,48 are interferon-stimulated genes with higher expression in females and well-established roles in immune responses with known sex differences. In addition, the literature also suggests that it is common for autosomal genes to be relevant to sex. For example, Nr5a1 (on mouse chromosome 2) is known to play a critical role in sex determination and development.^49,50,51 Using the identified autosomal genes from the sex-linked marker lists, we performed enrichment analysis using ToppGene⁵² and PubMed gene sets (Enrichr, version 2023), as shown in Figure S2B. The significantly enriched terms were related to sex difference, including male germ cell-specific ribosomes and translation-related pathways, which are functionally linked to sex-biased expression and male fertility.

In summary, our ensemble approach identified a biologically relevant minimal gene set, demonstrating the power of combining multiple algorithms for robust feature selection. This consensus-based selection method was able to restrict the number of predictors to an easily manageable level without sacrificing performance. In addition, to provide more flexibility, in CellSexID released in our GitHub repository we also provide the functionalities allowing the users to directly provide their own gene list or perform feature selection on their own datasets.

Cross-tissue and cross-species validation

To assess the effectiveness and generalizability of our feature-selection approach, we validated its performance across a diverse range of datasets spanning multiple tissues and species. These validations included (1) testing the default set of 14 sex-linked markers on various datasets, (2) evaluating the robustness of our feature-selection pipeline by applying it to new datasets, and (3) demonstrating applicability across species using human data.

We first conducted a baseline cross-validation using the publicly available mouse adrenal gland dataset from Dolfi et al. (GEO: GSM6153751 for female mice and GEO: GSM6153750 for male mice). In this dataset, the mouse adrenal gland dataset (CD45⁺ hematopoietic cells) was isolated from adrenal glands, and scRNA-seq was performed using the 10x Genomics platform. The sex origin of the cells was definitively known based on the source animals. Using an 80/20 train-test split, we trained and evaluated our machine-learning models on this dataset (Figure 3A). All models demonstrated consistently high and comparable performance across multiple metrics with those 14 selected features. Our default classifier, random forest, achieved an area under the receiver operating characteristic curve (AUROC) of 0.9723, supporting both the validity of our marker selection and the strong baseline performance of CellSexID.

Figure 3.

CellSexID performance evaluation across tissues and species

(A) Baseline performance of four machine-learning models—logistic regression (LR), support vector machine (SVM), XGBoost, and random forest (RF)—trained on the public mouse adrenal gland dataset. Internal validation was performed using an 80/20 train-test split, with 14 sex-linked marker genes selected from the same dataset.

(B) Cross-tissue validation (kidney). Models trained on the adrenal gland dataset were tested on a mouse kidney dataset (GEO: GSE129798) using the adrenal-derived marker list.

(C) Cross-tissue validation (hypothalamus). Models trained on the adrenal gland dataset were tested on a mouse hypothalamus dataset (GEO: GSE201032) using the adrenal-derived marker list.

(D) Cross-tissue validation (colon). Models trained on the adrenal gland dataset were tested on a mouse-colon dataset (GEO: GSM2967048) using the adrenal-derived marker list.

(E) Marker transfer in mouse: kidney test. A new marker list was selected using our feature-selection pipeline applied to the mouse-colon dataset (GEO: GSM2967048). Based on this colon-derived marker list, models were trained on the colon dataset and tested on the kidney dataset.

(F) Marker transfer in mouse: hypothalamus test. Using the colon-derived marker list, models trained on the colon dataset were tested on the hypothalamus dataset.

(G) Marker transfer in human: T cell leukemia test. Models were trained on the human AML bone-marrow dataset (GEO: GSE289435) using nine human sex-linked markers and tested on a human adult T cell leukemia/lymphoma (ATL) blood dataset (GEO: GSE294224).

(H) Marker transfer in human: thymic epithelial test. Models trained on the human AML bone-marrow dataset (GEO: GSE289435) using the same nine sex-linked markers were tested on a human thymic mimetic epithelial cell dataset (GEO: GSE262749).

To validate the robustness of our sex-linked markers and the flexibility of our method, we tested whether the machine-learning models trained on our default mouse adrenal gland dataset with the default 14 sex-linked markers can maintain accurate sex prediction on other very different datasets from other mouse tissues. In Figures 3B–3D, we present these cross-tissue validation results of training our machine-learning classifiers with the selected sex-linked gene markers on the mouse adrenal gland dataset and testing on a mouse kidney (GEO: GSE129798),⁵³ a mouse hypothalamus (GEO: GSE201032),⁵⁴ and a mouse-colon (GEO: GSM2967048)⁵⁵ dataset. The consistently high prediction performance in these new settings (most AUROC > 0.95) further validates the robustness and the generalizability of CellSexID and our selected sex-linked marker list.

To evaluate cross-tissue generalizability, we trained new models using the mouse-colon (GEO: GSM2967048)⁵⁵ dataset and applied the same methodology to identify 11 markers from this tissue. We then tested the model performance using these markers on scRNA-seq datasets from other mouse tissues: mouse kidney (GEO: GSE129798) and mouse hypothalamus (GEO: GSE201032). As shown in Figures 3E and 3F, CellSexID maintained high prediction accuracy across these datasets, with random forest AUROC values exceeding 0.96. These results further demonstrate the robustness of our marker set and the transferability of models trained on one tissue to other distinct tissues.

Next, to test generalizability across species, we applied our feature-selection and -prediction pipeline to human data. Feature selection was performed on human bone-marrow mononuclear cells (BMMCs) from AML dataset (GEO: GSE289435), and the resulting markers were used to train models that were then validated on independent human datasets, including human peripheral blood mononuclear cells (PBMCs) from T cell leukemia-lymphoma (ATL) patient (GEO: GSE294224)⁵⁶ and human thymus cell (GEO: GSE262749)⁵⁷ datasets. As shown in Figures 3G and 3H, the models achieved excellent predictive accuracy across these human datasets, further demonstrating the robustness and generalizability of CellSexID across species.

Together, these results show that (1) our default 14 sex-linked gene markers perform well across tissues, (2) feature selection is robust to dataset choice and consistently yields strong classifiers, and (3) CellSexID is broadly applicable across species, tissues, and experimental conditions. These validations collectively demonstrate that CellSexID is a reliable and broadly applicable tool for sex-based cell-origin inference, offering a scalable and label-free alternative for diverse single-cell transcriptomic studies.

Experimental validation using cell sorting

After demonstrating the strong sex-prediction performance of CellSexID on public datasets, we next evaluated its applicability for tracking cell origin in a real-world setting by performing “de novo validation” in sex-mismatched chimeric mice. The chimeric mice were created by transplanting bone marrow from male donor mice (CD45.1 allele) into female (CD45.2 allele) host recipient mice (Figure 4A). The resident macrophages present in the diaphragm muscle of female recipient mice were protected from the pre-transplant irradiation protocol through the use of lead shielding as we have recently described.⁵⁸ At 8 weeks post transplantation, macrophages were sorted and pooled from the diaphragms of two chimeric mice by FACS based on their CD45 allele status (CD45.1 or CD45.2) and then labeled with CMO barcodes. Using our scRNA-seq analysis pipeline, these distinct CMO barcodes served as ground truth for sex-based macrophage origin, providing a robust benchmark for experimentally evaluating CellSexID’s annotation performance in this chimeric mouse diaphragm validation dataset.

Figure 4.

Reliable performance of cell-origin classifiers on experimentally generated single-cell validation data from chimeric mouse with cell sorting as ground truth

(A) Generation of a chimeric mouse diaphragm validation dataset by creating chimeric mouse, collecting single-cell data, and using cell sorting as ground truth. A female CD45.2 mouse, irradiated with diaphragm shielding, was transplanted with male CD45.1 bone marrow. After 8 weeks, diaphragm macrophages were isolated, stained to identify live macrophages (CD11c−, SiglecF−, F4/80+, CD11b+), flow-sorted by CD45.1 and CD45.2 expression, barcoded (CD45.2, CMO306; CD45.1, CMO305), pooled, and sequenced using 10x Genomics.

(B) Confusion matrix of random forest predictions based on the 14 sex-linked markers, demonstrating high classification accuracy for cell origin.

(C) AUROC scores of all four classifiers (trained on the public dataset), confirming reliable performance.

(D) Precision-recall (PR) curves for the same classifiers, showing consistent predictive robustness.

(E and F) Comparative analysis of AUROC and AUPRC values for each classifier using bootstrap resampling. Error bars represent 95% confidence intervals (2.5th–97.5th percentile) from 100 bootstrap iterations with replacement applied to the training data. In each iteration, models were retrained on the resampled training set and evaluated on the fixed test set. All models achieved AUROC > 0.82 and AUPRC > 0.95.

(G) Bar plot comparing random forest-predicted cell-origin ratios (based on sex) with ground truth, illustrating the classifier’s high accuracy using the selected sex-linked markers.

See also Figures S3 and S4.

The validation results (Figures 4B–4G) demonstrate CellSexID’s high classification accuracy in a real-world setting. Figure 4B shows the confusion matrix for the default random forest model, while Figure 4C presents ROC curves for all four algorithms, with area under the curve (AUC) as the primary performance metric. Figure 4D displays PRC curves evaluating precision-recall trade-offs, and Figures 4E and 4F show bootstrapped AUROC and AUPRC performance confirming model stability across multiple samples. Figures S3A and S3D provide more comprehensive details for all classifiers. Figure 4G compares predicted versus actual male-to-female ratios in the chimeric mouse diaphragm validation dataset, confirming CellSexID’s accuracy in real-world population-composition estimation.

This comprehensive evaluation with the minimal 14-gene set supports CellSexID’s reliability across multiple scRNA-seq scenarios. Notably, models trained on public mouse adrenal gland data achieved high predictive accuracy when tested on the independent bone-marrow-transplantation-derived chimeric mouse diaphragm dataset, demonstrating strong cross-dataset generalizability. CellSexID operates without assumptions about tissue types, cellular contexts, or species, enabling broad applicability across diverse experimental systems.

To provide a more comprehensive view of the model’s performance, we calculated class-specific metrics in Figure S3E. In this supplementary figure, AUROC and AUPRC for male and female predictions provide a more detailed view of the model performance, showing decent performance in predicting both sexes. To better understand the performance and investigate the characteristics of the misclassified cells, we performed comparative analysis between the correctly classified and misclassified cells. We discovered that misclassified cells tend to be noisier and have statistically significant higher dropout rate than correctly classified cells (Figure S4A). A higher dropout rate (zero expression) of genes is a kind of typical technical noise in single-cell data, and it naturally leads to the loss of information of the sex-linked gene and thus might mislead the classification. In this analysis, we compared correctly classified cells versus misclassified cells within female and male groups, focusing on the 14 sex-linked gene markers used by our machine-learning classifiers (Figures S4B–S4F). In female cells, correctly classified cells showed significant upregulation of Xist, which is expected and consistent with the expected high expression of Xist in female cells due to X chromosome inactivation.⁵⁹ In contrast, among female cells that are misclassified as male cells, the loss of expression of Xist implying that they are classified as male cells might be due to the key female marker not being detected. In male cells, correctly classified cells showed significant upregulation of Lars2 and Uba52, two genes previously reported to be more highly expressed in male cells.^41,60 The low expression or dropout of these genes in male cells may lead to insufficient male-specific transcriptional signatures, thereby increasing the chance of being misclassified as female cells. Overall, these observations are consistent with a well-known property of scRNA-seq data: low mRNA capture efficiency often leads to frequent dropout zeros, which can affect the accurate classification of cell identity.⁶¹

Biological insights into macrophage ontogeny

While our validations have been focused on chimeric models, CellSexID’s capability of cell-origin tracking based on sex prediction enables a wide range of downstream applications without requiring genetic modification, physical barcodes, or antibody-based labeling. To demonstrate this, we explored its performance in two additional experimental contexts—organ transplantation and sample demultiplexing.

We tested CellSexID’s utility in distinguishing donor- and recipient-derived cells in organ-transplantation contexts using sex-mismatched models. Figure 5A illustrates results from human organ-transplantation experiments. Our default random forest classifier, trained on the human AML bone-marrow dataset (GEO: GSE289435) using nine human sex-linked markers, accurately differentiated donor from recipient cells in a human kidney transplant dataset (GEO: GSE151671)⁶² based solely on sex-specific gene-expression patterns.

Figure 5.

CellSexID enables accurate cell-origin tracking in organ transplantation and demultiplexing scenarios

(A) Performance evaluation for transplantation prediction using AUROC and AUPRC metrics. Random forest classifiers were trained on human AML BMMCs (GEO: GSE289435) using different gene-marker sets and tested on human kidney-transplantation data to distinguish donor from recipient cells. Four marker sets were compared: baseline four genes (XIST, RPS4Y, EIF1AY, DDX3Y highlighted by original dataset providers), full CellSexID gene set, sex chromosome genes within CellSexID, and Y chromosome genes only to ensure robustness; results are reported as mean ± SD across 15 independent runs with different random seeds. p values from the two one-sided tests (TOSTs) indicated that the CellSexID marker sets achieve comparable AUROC, and a one-tailed test demonstrates significantly superior AUPRC compared to the baseline method, suggesting overall better performance.

(B and C) Demultiplexing performance on synthetic datasets with skewed sex ratios of 4:1 male:female (B) and 1:4 male:female (C), created by sampling 12,000 cells from human thymic medullary epithelial cells dataset (GEO: GSE262749). ROC curves (top) and precision-recall curves (bottom) demonstrate near-perfect classification performance across all machine-learning models (logistic regression [LR], support vector machine [SVM], XGBoost [XGB], random forest [RF]) with AUC values >0.98, indicating robust performance even under challenging class imbalance conditions.

This enabled reliable cellular origin tracking without genetic modifications or invasive labeling, achieving >0.9 AUROC in complex in vivo transplantation scenarios. While the original study identified recipient-derived fibroblasts as key contributors to graft fibrosis, CellSexID provided more accurate, complete, and scalable cell-origin classification, reducing ambiguity in sex-mismatched transplant samples with superior performance compared to the original markers (Figure 5A). These results demonstrate CellSexID’s effectiveness in addressing critical transplantation biology challenges—including immune reconstitution studies, organ-engraftment analyses, and in vivo lineage-dynamics monitoring—with direct applicability to clinical human transplantation where donor-recipient sex mismatches frequently occur. This extends CellSexID’s utility beyond animal chimeric studies, establishing it as a broadly relevant tool for both basic biomedical research and clinical translation. Additionally, we tested CellSexID’s ability to perform single-cell demultiplexing using sex prediction as a surrogate target in sex-mismatched multiplexed samples. By leveraging endogenous gene expression rather than physical barcodes or antibody-based panels, CellSexID eliminates additional reagents and protocol modifications, significantly reducing experimental complexity, costs, and hands-on time compared to traditional approaches like antibody hashing.

We validated CellSexID’s ability to discriminate cell origins in mixed male/female samples using synthetic datasets with extreme imbalances: 1:4 (2,400 male:9,600 female) and 4:1 (9,600 male:2,400 female) ratios across five compositions of 12,000 cells each. These datasets were generated by systematic sampling from human thymic medullary epithelial cells (GEO: GSE262749) to simulate challenging scenarios where sex-mismatched cells from multiple individuals are processed in single sequencing runs.

The model was trained on the human AML bone-marrow dataset (GEO: GSE289435) containing 87,171 cells and tested on these synthetic imbalanced datasets using a panel of nine sex-linked marker genes (RPS4Y1, EIF1AY, XIST, DDX3Y, UTY, KDM5D, IFIT3, IFIT2, and RPS4X). CellSexID successfully distinguished cells solely based on their sex-linked expression patterns across all synthetic ratios (Figures 5B and 5C), achieving near-perfect performance with AUROC scores exceeding 0.996 and accuracy rates above 96% even in the most extreme imbalanced scenarios, demonstrating robust demultiplexing capabilities without relying on traditional physical barcodes or antibody-based multiplexing approaches. This approach validates that sex-linked gene expression provides a reliable biological signature for cell-origin determination across diverse mixing ratios and population compositions, and it is therefore applicable to real-world demultiplexing scenarios.

Notably, our previously described validation experiment (Figure 4)—in which male and female cells from two mice were pooled and sequenced together—serves as a real-world demonstration of barcode-free demultiplexing. In this setting, externally added CMOs provided the ground truth, allowing us to directly test whether CellSexID could accurately assign cell identities based solely on sex-linked gene expression. All four machine-learning models achieved strong annotation performance, with AUROC values exceeding 0.82 and AUPRC values above 0.95. These results validate CellSexID’s cell-origin-tracking performance and demonstrate its practical utility as a cost-effective and robust alternative to traditional barcode-based multiplexing approaches.

These examples demonstrate CellSexID’s broad applicability beyond conventional chimeric models. By using sex-linked gene expression, CellSexID reliably differentiates cell origins while requiring less effort and cost than traditional methods. This versatility enables diverse applications, including single-cell quality control, retrospective dataset analysis, and sex-associated disease studies, positioning CellSexID as a valuable, cost-effective tool for biological, clinical, and bioinformatic research.

CellSexID provides biological insights into macrophage ontogeny

Adult skeletal muscle macrophages originate from two sources: prenatal embryonic progenitors that self-renew locally (bone marrow independent) and postnatal monocytes requiring continuous bone-marrow replenishment.⁵ Accordingly, we next sought to determine whether CellSexID could ascertain the relative contributions of these two populations to macrophage subsets and their associated transcriptomic profiles in the main respiratory muscle, the diaphragm—where the role of macrophage ontogeny remains underexplored. To address this question, we generated additional chimeric mice in which CD45.1 male bone marrow was transplanted into CD45.2 female recipients, generating another chimeric mouse diaphragm experimental dataset (Figure 6A). Using our trained random forest model to predict cell origins from scRNA-seq analysis of sorted diaphragm macrophages at 8 weeks post transplantation, the predicted percentages of donor (bone marrow-dependent) and recipient (bone-marrow-independent) macrophages were 74% and 26%, respectively. This closely matched the percentages obtained from flow-cytometry staining of CD45.1+ and CD45.2+ cells (70% and 30%, respectively) (Figure 6B). This confirms that CellSexID can be used to determine cellular origin in chimeric mice without the need for conventional tracking methods such as the CD45.1 allele or fluorescent marker genes, which are not readily available in most genetically modified mouse model strains. In addition to these label predictions, Figure S4G visualizes the probability assigned to each cell by our classifiers. These visualizations provide an intuitive understanding of prediction uncertainty, showing how confidence varies across different regions of the embedding.

Figure 6.

CellSexID enables origin-specific analysis of macrophage subpopulations, revealing unique gene-expression profiles and functional pathways

(A) Schematic overview of the experimental design.

(B) Proportional distribution of macrophages from recipient and donor origins in each mouse, determined by flow-cytometry sorting and scRNA-seq predictions using our default classifier. Data represent eight biological replicates (n = 8; mean ± SEM).

(C) Unsupervised clustering analysis identifies nine distinct macrophage groups, segregating cells from recipient (female) and donor (male) origins based on unique gene-expression profiles.

(D) UMAP visualization showing the distribution of recipient- and donor-derived macrophages. Cell origins were annotated using our default classifier model and validated against the monocyte-derived macrophage marker Ccr2, supporting the accuracy of the predictions.

(E) Proportion of recipient- and donor-derived macrophages (annotated using our default classifier) within each identified cluster, highlighting origin-specific predominance in certain clusters (n = 8; mean ± SEM).

(F) Volcano plot displaying the top 10 upregulated and downregulated genes in macrophage cluster C4 compared to all other clusters, highlighting significant gene-expression differences.

(G) Pathway enrichment analysis of cluster C4 (top 20 pathways), indicating enhanced cell communication and proliferation alongside reduced immune and inflammatory responses compared to other clusters.

(H) Volcano plot highlighting genes with significant upregulation or downregulation in recipient- versus donor-derived macrophages within macrophage cluster C3.

(I) Pathway analysis of macrophage cluster C3 (top 20 pathways), revealing significantly downregulated immune-related pathways in recipient-derived macrophages compared to donor-derived macrophages.

See also Figures S5 and S6.

We applied a standard scRNA-seq data-analysis pipeline for our preprocessed data according to common practice, which includes dimensionality reduction, clustering, and downstream analysis such as differentially expressed genes (DEG) and enrichment.⁶³ Visualizations of uniform manifold approximation and projection (UMAP) were generated after dimensionality reduction. Our analysis identified nine distinct macrophage clusters in the diaphragm (Figure 6C). The rationale for selecting clustering resolution is provided in Figure S6D, while Figures S5, S6F, and S6G demonstrate the robustness of the clustering, showing consistent downstream results across different resolutions. The sex origin of each macrophage cluster in the muscle was determined using our default classifier, showing strong concordance between male-derived cells and Ccr2 expression (Figures 6D and S6E), a marker of bone-marrow-derived macrophages.^64,65 All of the clusters were predominantly derived from donor (male) bone-marrow monocytes with the exception of cluster C4, in which 89% of the macrophages were of recipient (female) origin (Figure 6E). To further evaluate transcriptional consistency within the female-enriched cluster, we calculated a label score of 0.8041 based on local neighborhood sex-label agreement, indicating strong homogeneity of sex identity within this cluster.

To gain deeper insights into the differential functions of bone-marrow-dependent versus bone-marrow-independent macrophages, we performed DEG analysis. In macrophage cluster C4, which is primarily composed of bone-marrow-independent macrophages, 522 genes were significantly upregulated, and 266 genes were downregulated compared to all other clusters (Figure 6F). The top 10 upregulated genes included Cfh, Colec12, Timd4, Mir99ahg, Selenop, C4b, Cpne8, Tanc2, Stab1, and Mctp1, while the top 10 downregulated genes were Cd74, H2-Eb1, H2-Ab1, H2-Aa, Ccr2, H2-DMa, Cd52, H2-DMb1, Mpeg1, and Tmsb10. Pathway-enrichment analysis revealed that the upregulated genes in cluster C4 were significantly associated with pathways such as endocytosis, adherens junction, Ras signaling pathway, and Rap1 signaling pathway (Figure 6G). These pathways suggest enhanced cellular uptake processes, cell adhesion, and growth factor signaling, likely contributing to increased cell communication and proliferation. In contrast, the downregulated pathways included hematopoietic cell lineage, tumor necrosis factor (TNF) signaling pathway, and nuclear factor (NF)-κB signaling pathway, indicating reduced pro-inflammatory immune responses in this bone-marrow-independent macrophage population. To further assess whether these transcriptional differences reflected biological divergence by origin rather than intrinsic sex effects, we compared gene expression between male- and female-classified cells across the dataset. We found a significant overlap (899 genes, p = 5.35 × 10⁻⁷⁹⁵) between the DEGs from the male vs. female comparison and those from the C4 vs. rest comparison (Figure S6A), suggesting that CellSexID’s sex-based labeling faithfully captured origin-related identity. Since both donor and recipient macrophages share the same tissue microenvironment and experience identical local cues, the observed transcriptional divergence likely reflects ontogeny (donor vs. recipient origin) rather than sex-specific regulation. Consistent with this interpretation, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis revealed that genes downregulated in recipient-derived (female) cells were enriched in immune-related pathways—such as antigen presentation, Th1/Th2 differentiation, and IBD—paralleling those identified in the cluster-level analysis (Figures 6G and S6B). Together, these findings support the conclusion that CellSexID’s sex-based predictions reflect meaningful biological variation rooted in cell origin.

We applied CellSexID to mixed-origin cluster C3 (68% donor, 32% recipient macrophages) to examine whether similar differences existed within a single cluster. CellSexID enabled differential expression analysis between donor-derived and recipient-derived macrophages within C3—impossible without origin identification due to mixed populations. This analysis identified 60 upregulated and 53 downregulated genes in recipient-origin macrophages (Figure 6H). Pathway enrichment revealed that downregulated pathways in recipient-origin macrophages mirrored those in cluster C4, including antigen presentation, T cell differentiation, and other pro-inflammatory immune responses (Figure 6I), reinforcing distinctions between bone-marrow-dependent and -independent macrophages and demonstrating CellSexID’s utility for origin-specific analyses within mixed populations. DEG overlap analysis between predicted-origin groups within C3 and the global male vs. female comparison revealed 192 shared genes (p = 7.53 × 10⁻²⁴², hypergeometric test) (Figure S6C). These results demonstrate that predicted sex serves as a surrogate for studying cell-origin-associated transcriptional programs, enabling insights difficult to obtain in mixed-origin clusters without CellSexID.

Discussion

This study addressed the critical challenge of tracking cell origins in various biological contexts (e.g., chimeric models) by developing CellSexID, a machine-learning-based tool that uses sex-specific gene expression in scRNA-seq data as a natural surrogate marker. Applied to sex-mismatched chimeras, CellSexID is a simple, low-cost, precise method for donor-recipient distinction. Through careful feature selection, we identified 14 essential genes that include but are not limited to those located on sex chromosomes, which allow for robust sex prediction and thereby enable reliable cell-origin tracking. CellSexID simplifies the cell-tracking process by eliminating the need for biased labeling techniques dependent on exogenous marker genes, which commonly have variable levels of gene expression under different experimental conditions and can also lack cellular specificity (“leaky expression”).^66,67 In addition, expensive and labor-intensive cross-breeding strategies are often required to incorporate exogenous cell-origin markers into existing animal models.^7,8 By taking advantage of the endogenous gene-expression profile associated with biological sex, CellSexID is well positioned as an efficient, scalable solution in experimental as well as clinical settings.

There are four major contributions of CellSexID. First, the primary innovation of this study lies in its experimental design, which transforms the challenging task of cell-origin tracking into a more manageable problem—sex prediction based on transcriptomic data. By utilizing sex-mismatched samples, machine-learning models can efficiently identify informative gene features and classify cell origins using sex as a surrogate target, significantly reducing the costs and labor associated with traditional labeling methods. Second, our consensus-based committee feature-selection pipeline robustly identifies minimal yet highly informative gene-marker sets across diverse datasets. Notably, this pipeline consistently selects reliable markers, demonstrating high predictive performance across datasets derived from various tissues and species. The default marker list includes both widely recognized sex-linked genes, such as Xist, Ddx3y, Eif2s3y, Uty, and Kdm5d, and, surprisingly, numerous highly predictive autosomal genes. The identification of autosomal markers, including Uba52 (encodes a ubiquitin-ribosomal fusion protein that has been reported to play an important role in sex-related behavioral differences mediated by the amygdala⁴⁰), Lars2 (leucine-tRNA-synthase 2, was recently found to be highly enriched in endothelial cells of male mice⁴¹), and various ribosomal proteins (Rpl and Rps), challenges the conventional assumption that sex-chromosome-linked genes predominantly determine sex prediction. Including autosomal markers led to improved predictive metrics, such as AUROC and AUPRC. These findings underscore the strength of CellSexID’s integrative machine-learning framework in uncovering robust sex markers beyond traditional assumptions. Additionally, CellSexID offers flexibility by supporting feature selection on user-specified datasets and enabling users to input custom gene lists directly. Third, the general applicability of CellSexID extends beyond traditional animal chimeric models, as shown by successful validation in human datasets and clinically relevant contexts, including organ transplantation and sample demultiplexing. This broad applicability highlights the method’s potential utility across diverse experimental and clinical scenarios, where sex prediction serves effectively as a surrogate marker for cell-origin tracking. Finally, using sex as a surrogate target for cell-origin tracking enabled the discovery of biological insights into ontogeny-specific gene signatures and functional characteristics. Analyses performed on our chimeric mouse diaphragm experimental dataset illustrate these insights clearly. Thus, CellSexID provides a versatile and robust framework for generating biological discoveries across various tissues, species, and research contexts.

To demonstrate CellSexID’s real-world applicability, we validated its performance in chimeric mice with sex-mismatched bone-marrow transplantation. In tracking the sex origin of macrophages in skeletal muscle using our tool, we were able to show high concordance with ground-truth flow-cytometry results, confirming its accuracy in distinguishing bone-marrow-derived from bone-marrow-independent cells. Furthermore, here we show the presence of important transcriptional differences in skeletal muscle macrophages based on cell origin (i.e., between bone-marrow HSC-derived and embryonic-derived macrophages), thus revealing how ontogeny helps to shape immune responses. Notably, we identified a unique subset of macrophages in the diaphragm that are overwhelmingly bone marrow independent and display an anti-inflammatory and pro-regenerative gene-expression profile. This discovery highlights CellSexID’s capacity to reveal ontogeny-driven functional differences within tissues where traditional methods have struggled to provide sufficient resolution. The characteristics of embryonic-derived tissue-resident macrophages are particularly relevant to various pathological states in which these cells are often lost from the diseased tissue,^6,68 and CellSexID provides a powerful tool for tracking their behavior at different stages of pathology.

In conclusion, CellSexID provides a scalable, cost-effective tool for tracking cell origins without exogenous labeling. Applications include organ transplantation, dataset demultiplexing, quality control, and sex-associated disease studies. Our analysis on the chimeric mouse-diaphragm experimental dataset demonstrates CellSexID’s ability to identify functionally distinct cell populations, such as embryonic-derived macrophages with pro-regenerative properties, and underscores its utility in uncovering fundamental biological insights.

Limitations of the study

A limitation of CellSexID is its reliance on sex differences between donor and recipient animals, which could potentially influence cell states or functions in certain contexts. While this dependency does not affect the main goals of this study, it may introduce bias in specific cell types where sex differences play an important role in determining phenotype. To address this limitation, future studies could implement bi-directional chimeric experiments with both male and female donors and recipients. This approach not only strengthens the robustness of our method but could also promote a greater understanding of the more nuanced aspects of sex differences in biological systems.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Jun Ding (jun.ding@mcgill.ca).

Materials availability

This study did not generate new unique chemical or biological reagents. Detailed experimental protocols are available from the lead contact without restriction.

Data and code availability

• •Public scRNA-seq datasets used in this study are available from the GEO and are publicly available as of the date of publication. The primary training dataset consists of adrenal gland CD45⁺ immune cells from GEO: GSM6153751 (female mice) and GEO: GSM6153750 (male mice). Additional mouse datasets include GEO: GSE129798 (kidney), GEO: GSE201032 (hypothalamus), and GEO: GSM2967048 (colon). Human datasets include GEO: GSE294224 (ATL blood), GEO: GSE289435 (AML bone-marrow), GEO: GSE262749 (thymus), and GEO: GSE151671 (kidney transplant). The chimeric mouse diaphragm datasets generated in this study are deposited under GEO series GEO: GSE305521. This includes the validation dataset with gene expression (GEO: GSM9178708) and multiplexing barcodes (GEO: GSM9178709), and the experimental dataset with day 1 gene expression (GEO: GSM9178710), day 1 multiplexing barcodes (GEO: GSM9178711), day 2 gene expression (GEO: GSM9178712), and day 2 multiplexing barcodes (GEO: GSM9178713).
• •All original code has been deposited at GitHub and is publicly available at https://github.com/mcgilldinglab/CellSexID, under https://doi.org/10.5281/zenodo.16797276.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was funded by grants awarded to J.D. and B.J.P. We gratefully acknowledge the support from the Canadian Institutes of Health Research (CIHR) under grant nos. PJT-180505 and 180644; the Fonds de recherche du Québec - Santé (FRQS) under grant nos. 295298, 295299, and 366764; the Natural Sciences and Engineering Research Council of Canada (NSERC) under grant no. RGPIN2022-04399; and the Meakins-Christie Chair in Respiratory Research.

Author contributions

B.J.P. and J.D. conceptualized the study. H.T., Q.L., and J.W. developed the methodology. H.T., Q.L., J.W., J.T., B.Z., and R.L. conducted the investigation. H.T., Q.L., and J.W. wrote the original draft, and all authors reviewed and edited the manuscript. Resources were provided by B.J.P. and J.D. B.J.P. and J.D. acquired funding and supervised the study.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-CD45.1 (clone A20)	BD Biosciences	Cat#553776; RRID:AB_395044
Anti-CD45.2 (clone 104)	BD Biosciences	Cat#562129; RRID:AB_10897142
Anti-SiglecF (clone E50-2440)	BD Biosciences	Cat#562757; RRID:AB_2687994
Anti-CD11c (clone HL3)	BD Biosciences	Cat#564986; RRID:AB_2739034
Anti-CD11b (clone M1/70)	Biolegend	Cat#101226; RRID:AB_830642
Anti-F4/80 (clone BM8)	Biolegend	Cat#123114; RRID:AB_893478
Anti-CD16/CD32 (clone Ab93)	BD Biosciences	Cat#567021; RRID:AB_2870011
Biological samples
Mouse adrenal gland scRNA-seq (female)	GEO28	GSM6153751
Mouse adrenal gland scRNA-seq (male)	GEO28	GSM6153750
Mouse kidney scRNA-seq	GEO53	GSE129798
Mouse hypothalamus scRNA-seq	GEO54	GSE201032
Mouse colon scRNA-seq	GEO55	GSM2967048
Human AML bone marrow scRNA-seq	GEO33	GSE289435
Human ATL blood scRNA-seq	GEO56	GSE294224
Human thymus scRNA-seq	GEO57	GSE262749
Human kidney transplant scRNA-seq	GEO62	GSE151671
Chemicals, peptides, and recombinant proteins
Busulfan	Sigma	Cat#B2635
Phosphate-buffered saline (PBS)	Wisent	Cat#811-425
Collagenase B	Roche	Cat#11088815001
Fetal bovine serum (FBS)	Wisent	Cat#080150
Chromium Next GEM Single Cell 3′ Kit v3.1	10X Genomics	Cat#1000269
Chromium Next GEM Chip G Single Cell Kit	10X Genomics	Cat#1000127
3′ Feature Barcode Kit, 16 rxns	10X Genomics	Cat#1000262
3′ CellPlex Kit Set A, 48 rxns	10X Genomics	Cat#1000261
Cell Multiplexing Oligo (CMO)	10X Genomics	Cat#1000261
Critical commercial assays
Illumina sequencing	McGill Genome Center
Deposited data
Chimeric mouse diaphragm validation dataset	GEO	GSM9178708, GSM9178709
Chimeric mouse diaphragm experimental dataset	GEO	GSM9178710-GSM9178713
Experimental models: Organisms/strains
C57BL/6 mice (CD45.1)	The Jackson Laboratory	RRID:IMSR_JAX:002014
C57BL/6 mice (CD45.2)	The Jackson Laboratory	RRID:IMSR_JAX:000664
Software and algorithms
SCANPY	Wolf et al., 201863	Version 1.9.5
Cell Ranger	10x Genomics69	Version 7.1.0
Scikit-learn	Pedregosa et al., 201170	Version 1.5.1
XGBoost	Chen and Guestrin, 201614,15,16	Version 2.1.0
GSEApy	Fang et al., 202371	Version 1.1.0
Scrublet	Wolock et al., 201972	Version 0.2.3
Harmony	Korsunsky et al., 201973	Version 0.0.9
CellSexID	This paper	https://doi.org/10.5281/zenodo.16797276
Other
X-RAD SmART irradiator	Precision X-ray	N/A

Experimental model and study participant details

Mouse strains

Experiments utilized Mus musculus mice on a C57BL/6J background, including C57BL/6J-Ptprc^a (CD45.1, Ly5.1) and C57BL/6J-Ptprc^b (CD45.2, Ly5.2) strains (The Jackson Laboratory, Bar Harbor, ME, USA). Both male and female mice, approximately 6 weeks old and weighing 20–23 g, were randomly assigned for use. A total of 15 mice were included: for the chimeric mouse diaphragm validation dataset, two female CD45.2 mice served as recipients; for the chimeric mouse diaphragm experimental dataset, eight CD45.2 female mice served as recipients; the remaining five CD45.1 male mice were used as donors. Animals were housed in individually ventilated cages (3–5 mice per cage) under specific pathogen-free (SPF) conditions, with a 12:12 h light/dark cycle, ambient temperature of 22 ± 1°C, and relative humidity of 50–60%. Standard rodent chow and autoclaved tap water were provided ad libitum. All procedures were approved by the McGill University Animal Care Committee (protocol number 3480) and conducted in accordance with the Canadian Council on Animal Care guidelines. Animals underwent a 7-day acclimation period before experiments, and their health and welfare were monitored daily by trained personnel.

Method details

Bone marrow transplantation

Chimeric mice were created following our previously outlined protocol.⁵⁸ In brief, 6-week-old female CD45.2 mice received whole-body X-ray irradiation using the X-RAD SmART irradiator (Precision X-ray, USA). Lead shielding was applied to the diaphragm area to protect resident macrophages. The irradiation was fractionated into two doses of 6 Gy, delivered 4 h apart, with settings of 225kV, 13mA, and 1.0265 Gy/min. Six hours after the final irradiation, an intraperitoneal injection of 30 mg/kg busulfan (Sigma, USA) was administered to eliminate any remaining bone marrow protected by diaphragm shielding. Twenty-four hours post-irradiation, CD45.2 mice received an intravenous injection of 2x10⁷ bone marrow cells from age-matched CD45.1 male donors. To reduce infection risk, the recipient mice were provided with enrofloxacin in their drinking water for 7 days following irradiation. An 8-week recovery period was then allowed to ensure complete bone marrow reconstitution. Two female mice were used as recipients for the chimeric mouse diaphragm validation dataset and eight female mice were used as recipients for the chimeric mouse diaphragm experimental dataset. For experimental validation of CellSexID, we established a ground-truth system using these sex-mismatched chimeric mice, enabling lineage tracing via CD45 allele differences. Macrophages were FACS-sorted, labelled with cell multiplexing oligonucleotides to distinguish male and female origins, and subjected to scRNA-seq following isolation from diaphragm muscle. Flow cytometry was performed on this dataset to establish ground truth labels, ensuring rigorous evaluation of model predictions.

Flow cytometry

To quantify and identify macrophages in diaphragm, flow cytometry was performed as described in prior studies.^58,74 Mice were anesthetized using isoflurane before euthanasia, followed by heart perfusion with 20 mL of phosphate-buffered saline (PBS) (Wisent, Cat#811-425), with an additional 20 mL administered after cutting the dorsal aorta. The diaphragm tissue was dissected, minced, and digested in PBS containing 0.2% collagenase B (Roche, Cat#11088815001) for 1.5 h to create a cell suspension. Diaphragm cells were stained with a viability dye to distinguish live from dead cells. For surface marker staining, cells were incubated at 4°C in the dark with primary antibodies, including anti-CD45.1 (clone A20, Cat#553776), anti-CD45.2 (clone 104, Cat#562129), anti-SiglecF (clone E50-2440, Cat#562757), and anti-CD11c (clone HL3, Cat#564986) (all from BD Biosciences, NJ, USA); and anti-CD11b (clone M1/70, Cat#101226) and anti-F4/80 (clone BM8, Cat#123114) (both from Biolegend, CA, USA). To prevent non-specific binding, cells were treated with Fc blocking solution (anti-CD16/CD32, clone Ab93, BD Biosciences, Cat#567021). To identify diaphragm macrophages, we defined them as cells positive for CD45 (either CD45.1 or CD45.2), CD11b and F4/80, and negative for SiglecF and CD11c. For the model validation experiment, we specifically added CMOs (10X Genomics, Cat#1000261) to sorted populations of CD45.1 and CD45.2 macrophages to differentiate between donor and recipient mouse macrophages. In experiments addressing our research question using the model, CMOs were added to the unstained muscle cell suspension before sorting to distinguish cells originating from different mice. All macrophages were sorted in a PBS solution containing 20% fetal bovine serum (FBS) (Wisent, Cat#080150) to maintain cell viability and integrity throughout the process.

Chimeric mouse diaphragm validation dataset

For the validation experiment, we utilized a ratio of 80% CMO-labelled CD45.1 macrophages to 20% CMO-labelled CD45.2 macrophages. This mixture was loaded onto a 10x Genomics Chip G (Cat#1000127) and processed using the 10x Chromium Controller, following the manufacturer’s instructions. The Single Cell 3′ mRNA and CMO barcode library preparation was then conducted in accordance with 10x Genomics protocol CG000388.

Chimeric mouse diaphragm experimental dataset

For the experiments aimed at addressing the research question, a mix of CD45.1 and CD45.2 positive macrophages was added to the chip without distinguishing between the two types. FACS sorting and cell preparation were performed on four mice per day over two consecutive days (Day 1 and Day 2) Single Cell 3′ mRNA library preparation was following the manufacturer’s instructions. All libraries were sequenced on an Illumina NovaSeq6000 S4 (v1.5) platform at the McGill Genome Center.

Single-cell data acquisition

For our generated single-cell data (chimeric mouse diaphragm validation and chimeric mouse diaphragm experimental datasets), at 6 weeks post-transplantation, we used the same bone marrow transplantation protocol as described for generating chimeric mice. Wild-type (WT) female recipient mice were transplanted with bone marrow cells from WT male donors. Following the transplantation protocol, the mice were allowed to recover for 8 weeks to ensure complete bone marrow reconstitution. Subsequently, macrophages from the diaphragms of these mice were isolated via cell sorting, specifically targeting populations expressing CD45⁺ and F4/80+ markers to ensure high purity. Single-cell RNA sequencing was then performed on these isolated macrophages. Sequencing reads were processed using 10x Genomics Cell Ranger 7.1.0. Each library was processed using the Cell Ranger⁶⁹ multi pipeline. Specifically, we used the reference genome refdata-gex-mm10-2020-A with a minimum assignment confidence of 0.7. The libraries included gene expression and multiplexing capture data. The samples analyzed were WT_to_WT samples. Using Cell Ranger multi, each library was aligned to an indexed mm10 genome, generating an aggregated matrix of WT cells. This aggregation normalized the number of confidently mapped reads per cell in each library, preparing the data for downstream analysis. The chimeric mouse diaphragm validation dataset was used to assess the predictive performance of models trained on public datasets in determining cell sex origin. A separate chimeric mouse diaphragm experimental dataset was generated to demonstrate CellSexID’s utility for biological discovery through differential gene expression and pathway enrichment analyses of donor and recipient macrophages.

Single-cell data preprocessing

The SCANPY⁶³ (version 1.9.5) was applied to preprocess single-cell raw count matrices for all datasets. For GEO: GSM6153751 (female mice cells) and GEO: GSM6153750 (male mice cells) from adrenal gland dataset, we applied minimum cell expression threshold of 3, followed by log transformation and normalization. Since this served as the major gene selection dataset, we retained all genes without gene filtering to preserve comprehensive gene expression information. For the chimeric mouse diaphragm validation dataset, we used the same processing pipeline but applied a gene threshold of 100 and doublet removal using Scrublet to exclude potential multiplets. As this was a relatively small dataset, this approach removed only 3 cells, resulting in 1649 cells. For all cross-tissue validation datasets including mouse kidney (GEO: GSE129798), mouse hypothalamus (GEO: GSE201032), mouse colon (GEO: GSM2967048), and human ATL blood (GEO: GSE294224), human AML bone marrow (GEO: GSE289435), and human thymus (GEO: GSE262749) datasets, we applied stringent quality control parameters: minimum cell expression threshold of 3, removing cells expressing fewer than 200 genes, and mitochondrial gene percentage threshold of 5%, followed by log transformation and normalization to a total count of 10,000 per cell. For the kidney transplant scRNA-seq data (GEO: GSE151671), we applied the same parameters except for the mitochondrial gene percentage threshold, as applying the 5% threshold would result in loss of nearly all cells due to the inherently high mitochondrial gene expression in transplant tissue samples, likely reflecting cellular stress and metabolic alterations associated with transplantation procedures. For the chimeric mouse diaphragm experimental dataset, the Day 1 and Day 2 samples were combined for downstream analysis. We first performed quality control filtering genes not expressed in at least three cells. We filtered cells expressing more than 6200 genes or fewer than 200 genes. Cells with more than 20% mitochondrial reads were regarded as dead cells and also filtered. Immunocytes were specifically filtered by selecting cells expressing Ptprc. We also performed doublet removal using Scrublet⁷² to exclude potential multiplets. Subsequently, the single-cell data matrix was normalized to a total count of 10,000 per cell and log1p-transformed. Then, for cell origin classification, only the sex markers' expression was used for model input. For clustering analysis, highly variable genes were detected using SCANPY’s tool 'pp.highly_variable_genes' with default parameters, resulting in 1895 genes. The effects of total counts and mitochondrial gene percentage were regressed out. The data was then scaled, capping the values at 10.

Classifier committee

We employed a classifier committee to predict sex based on gene expression data. The committee consists of four distinct machine learning classifiers, each contributing unique perspectives to the task. By working together, the committee helps to identify the most relevant features for maximizing prediction accuracy. While the best-performing classifier is ultimately chosen to make the final prediction on the Chimeric mouse diaphragm test dataset, the combined efforts of the committee enhance the overall robustness and effectiveness of the feature selection process.

The choices of the models are based on their capabilities in feature selection and dealing with our classification tasks. XGBoost Classifier is a highly scalable model grounded in gradient boosting frameworks and can capture complex nonlinear patterns.^75,76 It evaluates feature importance through a combination of metrics that describe how feature enhances accuracy and decision-making, which includes Gain – the improvement in accuracy a feature provides to the decision points it influences, Cover – the proportion of data points a feature impacts, and Frequency – how often a feature is utilized across the ensemble of decision trees. Random Forest Classifier⁷⁷ leverages an ensemble of decision trees and quantifies feature importance through the mean decrease in impurity, often expressed in terms of Gini impurity or entropy. This approach calculates the extent to which each gene reduces uncertainty across trees, with a higher reduction indicating a stronger influence on model predictions. This method is particularly effective in discerning the genes that consistently refine the model’s outputs by creating clearer, more defined paths in the decision process. For the Support Vector Machines (SVM)^78,79,80 with a linear kernel used in our method, feature importance is derived directly from the model’s coefficients. These coefficients are integral to the hyperplane that separates the classes in the dataset. Each coefficient quantifies the influence of a corresponding gene on the classification boundary, with its magnitude indicating the strength of the impact and its sign denoting the direction (positive or negative). This linear approach provides a clear and quantifiable measure of each gene’s contribution to the model’s decision-making process, facilitating the identification of key genetic features crucial for predictive accuracy. Linear SVMs are preferred for their ability to perform well in high-dimensional spaces with clear margins of separation, and their simplicity and effectiveness in binary classification scenarios. Lastly, Logistic Regression ^81,82 determines the importance of genes through its coefficients within a linear context, where the magnitude of each coefficient indicates the extent to which a gene affects the model’s output. Larger absolute values suggest a stronger influence, making this approach straightforward for gauging the impact of individual genes on the prediction of the target variable. Logistic Regression is favored for its interpretability and the provision of probability scores for outcomes, which can be crucial for understanding risk factors in medical or biological studies. These methods, grounded in robust statistical and mathematical principles, allow for a detailed exploration of gene significance in predictive modeling. In our implementation, Logistic Regression, Random Forest and Support Vector Machine are implemented in Python packages Scikit-learn⁷⁰ and XGBoost is implemented using package ‘XGBoost’.¹⁴

Sex-relevant feature selection

We used the public mouse adrenal gland dataset to select the most significant gene features for predicting the sex of individual cells. First, each model in the committee is trained on the public adrenal gland dataset using all gene features and employed the built-in feature selection functions provided by their respective libraries to compute feature importance scores. The top $K$ (default 20, see justification of this hyperparameter in Figure S1) important features are selected. Then features selected by at least $C$ committee members are used as final selected features. We set $C$ to be 3 out of 4 (75%) to claim a majority and Figure S1C justifies our choice. For any gene feature $g$, it will be selected in the final feature set ${\mathit{G}}_{\mathbf{\text{final}}}$ if and only if it is selected by at least three models.

$${g\in \mathit{G}}_{\mathbf{\text{final}}}\iff \sum _{model\in \left\{RF,SVM,XGB,LR\right\}}{1}_{\text{model}}\left(g\right)\ge C$$ (Equation 1)

Where each $1_{\text{model}}\left(g\right)$ represents an indicator function indicating if the gene feature is selected by a specific model (SVM: Support Vector Machine; LR: Logistic Regression; RF: Random Forest; XGB: XGBoost).

$$\begin{gathered} \{\begin{array}{c}{1}_{\text{model}}\left(g\right)=1,ifg\in \arg {\mathit{max}}_{\mathit{G}}\left(\sum _{i=1}^K\left|{\mathit{w}}_{ \\ \text{model},{\mathit{G}}_i}\right|\right)\\ 1_{\text{model}}\left(g\right)=0,otherwise\end{array} \end{gathered}$$ (Equation 2)

Different $\mathit{w}$ corresponds to different models’ feature importance score returned by the Python library after training using all gene features on the public mouse adrenal gland dataset. $\left|{\mathit{w}}_{SVM}\right|$ represents the absolute value of the trained SVM’s weights. $\left|{\mathit{w}}_{LR}\right|$ denotes the absolute value of the trained Logistic Regression classifier’s coefficients. ${\mathit{w}}_{RF}$ is the importance of the trained Random Forest derived from the mean decrease in impurity across all trees in the forest. ${\mathit{w}}_{XGB}$ is the importance of the score of trained XGBoost derived from the Gain, Cover, and Frequency as mentioned in the previous paragraph. The subscription ${\mathit{G}}_i$ specifies the gene that the weight corresponds to.

The model committee also computes a scaled feature importance score for each gene feature, designed to reflect consensus among the committee members. This score is determined as follows: first, each committee member’s score is normalized to a range between 0 and 1. The normalized scores are then summed across all committee members. To further refine the score, a bonus of 0.25 is added for each committee member that selects the feature, while a penalty of 0.25 is subtracted for each committee member that does not select the feature. As a result, the final bonus/penalty component of the score will range between −1 (if none of the four committee members select the feature) and +1 (if all four committee members select the feature), providing a balanced measure of agreement across the committee. All the gene features will be ranked based on their feature selection scores. Once the feature selection process is completed, we use Random Forest as the classifier to annotate cell origin by default.

Sex classifier evaluation

To evaluate the real-world applicability of CellSexID, we validated the method across multiple public scRNA-seq datasets spanning diverse tissues and species. Cross-tissue validation included mouse kidney (GEO: GSE129798), hypothalamus (GEO: GSE201032), and colon (GEO: GSM2967048) datasets. Cross-species validation utilized human datasets including AML bone marrow (GEO: GSE289435), T cellT-cell leukemia blood (GEO: GSE294224), and thymus (GEO: GSE262749). Using the selected 14 final gene features, we tested the same models (Logistic Regression, Random Forest, Support Vector Machine, XGBoost) both on the mouse adrenal datasets and the chimeric mouse diaphragm validation dataset. Based on the selected sex-linked markers, classifiers output the probability of belonging to each class, which is converted to class binary prediction by setting a threshold, a common practice implemented in Scikit-learn. The predicted labels are subsequently used for evaluations. For testing on the mouse adrenal dataset, we used 5-fold cross validation to compute Accuracy and F1 score. Standard deviations are shown as error bars. For calculating Area Under the Receiver Operating Characteristic curve (AUROC) and the Area Under the Precision-Recall curve (AUPRC) we used functions in scikit-learn. For testing the model on the chimeric mouse diaphragm macrophage dataset, the full public dataset was used for training the model. For generating the bar plots in Figures 4E and 4F, AUPRC and AUROC are calculated using bootstrapping with 100 iterations. In each iteration, the adrenal gland training dataset is randomly resampled with replacement to get the same number of samples as the original training set, but the test set (chimeric mouse diaphragm validation dataset) is unchanged, ensuring a consistent basis for performance comparison.

These metrics, along with their 95% confidence intervals derived from the bootstrap samples, provided comprehensive insights into the models' predictive capabilities. While all models demonstrated excellent performance, we selected Random Forest as our default classifier for CellSexID. However, all four classifiers achieved comparable results and can be used interchangeably. The Random Forest classifier was subsequently used for annotating the cells in the single-cell dataset from our experiments.

UMAP and clustering analysis

Principal Component Analysis (PCA)^83,84 was performed with 50 components. For the chimeric mouse diaphragm experimental dataset, Harmony^73,82,85 was applied to remove batch effects across 8 samples and adjust the PCA basis. The neighborhood graph of cells was computed using the PCA representation using the SCANPY’s tool 'pp.neighbors' by considering 10 nearest neighbors and using the first 40 principal components, followed by UMAP⁸⁶ embedding using the 'tl.umap' tool with default parameters and Leiden clustering using SCANPY’s 'tl.leiden' tool with a resolution of 0.6. After further filtering out clusters with fewer than 200 cells, 9 macrophage clusters were identified. The choice of clustering resolution is based on Silhouette scores. In Figure S6D we show Silhouette scores⁸⁷ for different resolutions. Based on the plot, we select the resolution 0.6, which is before significant score drop. In addition, to justify the robustness of our analysis, we performed the same set of analysis as in Figure 6 using resolution 0.4 in Figure S5, which shows very similar results. In Figure S5, in the C3 vs. rest comparison, shared genes such as Timd4 and Tmsb10, as well as pathways including NF-κB signaling and TNF signaling pathway, show strong similarity to the results at resolution 0.6. In the C1 female vs. male comparison, consistent genes like Tmsb10, H2-DMb1, and Timd4, together with pathways such as antigen processing and presentation and Th1 and Th2 cell differentiation, also indicate overlap with the resolution 0.6 analysis. Furthermore, we also compared the DEGs between corresponding clusters at different resolutions, shown in Figures S6F and S6G. These similarities justify that our analysis results in Figure 6 are robust to the choice of clustering resolution.

Enrichment analysis

DEGs were identified using SCANPY’s tool ‘tl.rank_genes_groups’ with the Wilcoxon rank-sum test method.⁸⁸ Sex-related genes were removed from the analysis. Batch correction was not applied because our analysis shows it removes biological signal and does not affect our analysis results much. To evaluate the impact of batch correction on downstream analysis, we applied Scanorama to correct the gene expression matrix of the chimeric mouse diaphragm validation dataset. However, this substantially reduced sex-based signals critical to our study. For instance, the AUROC of the Random Forest classifier dropped from 0.837 (uncorrected) to 0.607 (corrected) (Figures S3B and S3C), indicating that batch correction attenuated meaningful biological variation. This effect has been noted in prior studies (Xiong et al.,⁸⁹ and others have reported that batch correction may inadvertently remove genuine biological variation—especially when the variable of interest (e.g., sex) is partially confounded with batch structure.⁹⁰ To preserve sex-specific signals, batch correction was not applied to the gene expression matrix in our final pipeline. In the DEG analysis, up-regulated genes were selected with Benjamini-Hochberg adjusted p-values less than 0.05 and log2 fold changes greater than 0.6, while down-regulated genes were selected with Benjamini-Hochberg adjusted p-values less than 0.05 and log2 fold changes smaller than −0.6. Top regulated genes were visualized through volcano plots, generated using the custom scatterplot function implemented in Python tool Matplotlib. If the number of top regulated genes exceeded 200, the top 200 genes were used; if fewer than 200 top regulated genes were identified, all top genes were used. The top up- and down-regulated genes were then subjected to KEGG 2019 Mouse^91,92,93,94 enrichment analyses using the GSEApy⁷¹ tool ‘enrich’ with KEGG for enrichment analysis.

Quantification and statistical analysis

All statistical analyses were performed using Python 3.8+ with the following packages: Scikit-learn version 1.5.1, SCANPY version 1.9.5, XGBoost version 2.1.0, GSEApy version 1.1.0, and NumPy/SciPy for statistical computations. Statistical details for each analysis are reported in the corresponding figure legends and results sections.

Error reporting

Different error bar types are used appropriately for different data types: ±SD for technical replicates to show measurement variability, ±SEM for biological replicates to show precision of population estimates, and 95% confidence intervals for statistical model performance to show uncertainty ranges.

Machine learning model evaluation

For classifier performance assessment, we computed Area Under the Receiver Operating Characteristic curve (AUROC) and Area Under the Precision-Recall curve (AUPRC) using scikit-learn implementations. Internal validation on the mouse adrenal gland dataset used 5-fold cross-validation with accuracy and F1 scores reported as mean ± standard deviation across folds. For the chimeric mouse diaphragm validation dataset, bootstrap resampling was performed with 100 iterations to generate 95% confidence intervals (2.5th–97.5th percentile) for AUROC and AUPRC values. In each bootstrap iteration, the training set was resampled with replacement while maintaining the original sample size, and models were retrained and evaluated on the fixed test set.

Sample sizes and statistical power

Sample sizes (n) represent individual cells for single-cell analyses and individual mice for experimental validation. For the mouse adrenal gland training dataset: n = 6,648 cells (3,963 female, 2,685 male). For the chimeric mouse diaphragm validation dataset: n = 1,649 cells from 2 mice. For the chimeric mouse diaphragm experimental dataset: n = 8 mice with cell counts varying by cluster analysis. Cross-tissue validation datasets ranged from thousands to tens of thousands of cells as specified in individual analyses. No formal power analysis was conducted for machine learning model training, as performance was evaluated through cross-validation and independent test sets.

Differential gene expression analysis

DEGs were identified using the Wilcoxon rank-sum test implemented in SCANPY’s 'tl.rank_genes_groups' function. Multiple testing correction was applied using the Benjamini-Hochberg false discovery rate method. Significance thresholds were set at adjusted pp-value <0.05 with log2 fold change >0.6 for upregulated genes and log2 fold change < −0.6 for downregulated genes. For pathway enrichment analysis, the top 200 regulated genes (or all genes if fewer than 200) were used for KEGG pathway analysis using GSEApy.

Gene overlap and enrichment analysis

Hypergeometric tests were used to assess statistical significance of gene overlap between different comparisons, with p-values reported for overlapping gene sets. Pathway enrichment analysis was performed using GSEApy with KEGG 2019 Mouse database, applying Benjamini-Hochberg correction for multiple testing.

Clustering and dimensionality reduction

Leiden clustering was performed using SCANPY with resolution parameter 0.6, selected based on Silhouette score analysis across multiple resolutions (0.2–1.0). Principal Component Analysis used 50 components, and UMAP embedding was computed using the first 40 principal components with 10 nearest neighbors. Silhouette scores were calculated to evaluate clustering quality and optimize resolution parameters.

Data assumptions and validation

Single-cell RNA-seq data were log1p-transformed and normalized to 10,000 counts per cell to meet assumptions of downstream analyses. For machine learning models, no specific distributional assumptions were required as tree-based methods (Random Forest, XGBoost) are non-parametric. Feature importance scores were normalized between 0 and 1 within each model before consensus scoring. Bootstrap confidence intervals were used to assess model stability without assuming normal distributions.

Missing data and quality control

Cells with fewer than 200 expressed genes, more than 6,200 expressed genes (potential doublets), or greater than 20% mitochondrial gene expression were excluded from analysis. Genes expressed in fewer than 3 cells were filtered out. Doublet detection was performed using Scrublet version 0.2.3 with default parameters.

Reproducibility

Feature selection was performed using 5 technical replicates with different random seeds, and consensus features were selected based on agreement across replicates. Model training and evaluation procedures were repeated with multiple random initializations to assess stability.

Statistical significance

Unless otherwise specified, statistical significance was defined as p < 0.05 after appropriate multiple testing correction. For machine learning performance metrics, statistical significance was assessed through confidence intervals generated by bootstrap resampling. Exact p-values and confidence intervals are reported in figure legends and results sections for each specific analysis.

Published: September 17, 2025