Intratumor transcriptional heterogeneity (ITTH), defined as the coexistence of diverse cell states within a single tumor, complicates cancer treatment and contributes to variable therapeutic responses. Although single-cell RNA sequencing (scRNA-seq) can resolve this complexity, its cost and technical demands limit large-scale use. Bulk RNA sequencing (bulk RNA-seq) provides a scalable alternative but requires computational methods to deconvolve bulk transcriptomes into distinct cell states. Existing supervised approaches rely on accurate reference data, which are lacking for many cancer types, while unsupervised methods are not tailored to capture heterogeneity within the malignant compartment. To address these limitations, we developed CDState, an unsupervised deconvolution method based on nonnegative matrix factorization with a sum-to-one constraint and a cosine-similarity-based optimization, which infers malignant cell states using bulk RNA-seq data. CDState demonstrated robustness using pseudobulk scRNA-seq datasets from five cancer types, outperforming existing unsupervised methods in estimating both state-specific gene expression and cell proportions. Applied to 33 cancer types from The Cancer Genome Atlas, CDState revealed recurrent gene programs, including epithelial-mesenchymal transition, MYC targets, and oxidative phosphorylation, as major contributors to malignant ITTH. The malignant state proportions were linked to clinical features, including patient survival and therapeutic response. Finally, mutations and copy number alterations in genes such as TP53, KRAS, PIK3CA, SOX2, and SATB1 were identified as potential genetic drivers of malignant cell ITTH across cancer types. This study demonstrates the utility of CDState for characterization of malignant cell states from bulk RNA-seq data, establishing a framework for investigating malignant cell ITTH in large-scale cancer atlases.

Single-cell RNA sequencing (scRNA-seq) has become a cornerstone experimental technique in tissue biology, with gene expression data for over 100 million cells available in public repositories. The high dimensionality, sparsity, and technical noise inherent to scRNA-seq data have motivated the development of a broad spectrum of representation learning approaches. These methods learn compressed, lower-dimensional representations of single-cell transcriptomes that are meant to preserve essential variation while reducing noise, and can be used for clustering, visualization, trajectory inference, and other downstream tasks. Furthermore, methods have emerged that aim to integrate data from multiple experiments by learning a common latent representation. In this review, we frame factor models, autoencoders, contrastive learning approaches, and transformer-based foundation models as distinct instances of the representation learning paradigm for scRNA-seq. We provide a coherent taxonomy of these methods that articulates their conceptual foundations, shared assumptions, and key distinctions. We also discuss benchmarking and identify major challenges and open questions that will shape the future of the field.

Single-cell RNA sequencing facilitates the discovery of gene expression signatures that define cell states across patients, which could be used in patient stratification and precision oncology. However, the lack of standardization in computational methodologies used to analyze these data impedes the reproducibility of signature detection. To address this, we developed CanSig, a comprehensive benchmarking tool that evaluates methods for identifying transcriptional signatures in cancer. CanSig integrates metrics for batch correction and biological signal conservation with a transcriptional signature correlation metric to score methods according to signature rediscovery, cross-dataset reproducibility, and clinical relevance. CanSig was applied to 13 methods on 12 single-cell RNA sequencing datasets from five human cancer types-glioblastoma, breast cancer, lung adenocarcinoma, rhabdomyosarcoma, and cutaneous squamous cell carcinoma-representing 185 patients and 174,000 malignant cells. The signatures identified with these methods correlated with clinically relevant outcomes, including patient survival and lymph node metastasis. These results identified Harmony, BBKNN, and fastMNN as the highest-scoring integration methods for discovering shared transcriptional states in cancer. Overall, CanSig provides a standardized, reproducible framework for uncovering clinically relevant cancer cell states in single-cell transcriptomics. SIGNIFICANCE:The development of CanSig facilitates computational strategies for the reproducible discovery of shared cancer cell states, improving cross-study reliability and enabling the detection of clinically relevant gene signatures. This article is part of a special series: Driving Cancer Discoveries with Computational Research, Data Science, and Machine Learning/AI .

Gene signature scoring is integral to single-cell RNA sequencing (scRNA-seq) data analysis, particularly for unsupervised cellular state annotation based on maximum signature score values. However, this application requires robust and comparable score distributions across diverse signatures and experimental conditions. Our systematic evaluation of established scoring methodologies-Seurat, SCANPY, UCell, and JASMINE-across nine healthy and cancer scRNA-seq data sets demonstrates their insufficiency in fulfilling this requirement. To address this limitation, we present Adjusted Neighborhood Scoring (ANS), a deterministic algorithm with enhanced control gene selection that significantly improves score stability and cross-signature comparability, achieving cell-state annotation accuracy comparable to supervised methods. We demonstrate the practical utility of ANS by developing and validating a gene signature to differentiate cancer-associated fibroblasts from malignant cells undergoing epithelial-to-mesenchymal transition. Overall, ANS provides a robust and reliable gene signature scoring framework, significantly improving the accuracy of score-based annotation of cell types and states in single-cell studies.

Intratumor transcriptional heterogeneity (ITTH), defined by the coexistence of diverse cell states within one tumor, complicates cancer treatment by contributing to variable therapeutic responses. Although single-cell RNA sequencing can resolve this complexity, its cost and technical demands limit its large-scale use. Bulk RNA-seq data provide a scalable alternative, but most deconvolution methods depend on predefined references, restricting their ability to detect novel malignant states. Unsupervised approaches avoid these constraints but are not tailored to capture heterogeneity within the malignant compartment. To address these limitations, we introduce CDState, an unsupervised method for inferring malignant cell subpopulations from bulk RNA-seq data. CDState utilizes non-negative matrix factorization improved with sum-to-one constraint and a cosine similarity-based optimization to deconvolve bulk gene expression into distinct cell state profiles. We demonstrate robustness of CDState on bulkified single-cell RNA-seq datasets from five cancer types, showing that it outperforms existing unsupervised deconvolution methods in the estimation of both cell state proportions and gene expression profiles. Applied to 33 cancer types from The Cancer Genome Atlas, CDState reveals recurrent gene programs, including epithelial-mesenchymal transition, MYC targets, and oxidative phosphorylation, as major contributors to malignant cell ITTH. We further link malignant states to patient clinical features, identifying states associated with poor prognosis. We propose an intratumor heterogeneity index and show its association with patient survival, clinical characteristics, and therapeutic response. Finally, we identify mutations and copy number alterations in genes such as TP53 , KRAS , PIK3CA, SOX2, and SATB1 as potential genetic drivers of malignant cell ITTH across cancer types.

Single-cell foundation models (scFMs) hold promise for applications in cell type annotation and data integration, but their internal mechanisms remain poorly understood. We investigate the structure of these models by training sparse autoencoders (SAEs) on the hidden representations of two widely used scFMs, scGPT and scFoundation. The learned features reveal diverse and complex biological and technical signals, which emerge even in pre-trained models. We also observe that the encoding of this information differs between scFMs with distinct training protocols and architectures. Further, we find that while many features capture the information about cell types across several studies, they often fall short of unifying it into a single generalized representation. Finally, we demonstrate that SAE-derived features are causally related to model behavior and can be intervened upon to reduce unwanted technical effects while steering model outputs to preserve the core biological signal. These findings provide a path toward more interpretable and controllable single-cell foundation models. ### Competing Interest Statement The authors have declared no competing interest. Swiss National Science Foundation, https://ror.org/00yjd3n13, 205321 207931

Self-supervised learning (SSL) has proven to be a powerful approach for extracting biologically meaningful representations from single-cell data. To advance our understanding of SSL methods applied to single-cell data, we present scSSL-Bench, a comprehensive benchmark that evaluates nineteen SSL methods. Our evaluation spans nine datasets and focuses on three common downstream tasks: batch correction, cell type annotation, and missing modality prediction. Furthermore, we systematically assess various data augmentation strategies. Our analysis reveals task-specific trade-offs: the specialized single-cell frameworks, scVI, CLAIRE, and the finetuned scGPT excel at uni-modal batch correction, while generic SSL methods, such as VICReg and SimCLR, demonstrate superior performance in cell typing and multi-modal data integration. Random masking emerges as the most effective augmentation technique across all tasks, surpassing domain-specific augmentations. Notably, our results indicate the need for a specialized single-cell multi-modal data integration framework. scSSL-Bench provides a standardized evaluation platform and concrete recommendations for applying SSL to single-cell analysis, advancing the convergence of deep learning and single-cell genomics.

Foundation models have become a powerful tool in single-cell transcriptomics, enabling broad generalization across tasks such as cell type annotation, data integration, and drug response prediction. Yet, most current models are trained predominantly on healthy cells, with a strong bias toward peripheral blood mononuclear cells. This raises an important question: how well do these models generalize to cancer-specific contexts? In this talk, I will explore whether training a foundation model exclusively on malignant cells—across diverse cancer types—can improve performance on tasks relevant to cancer biology and treatment. I will introduce CancerFoundation, a single-cell foundation model trained on malignant cells from over 40 tumor types. The model incorporates strategies for addressing tissue imbalance and technical variation, including domain-invariant training and tailored sampling. Through this work, I aim to address whether disease-specific pretraining can better capture the molecular features of cancer and improve the utility of foundation models in oncology applications such as batch integration and drug response prediction. Alexander Theus, Florian Barkmann, David Wissel, Tobias Scheithauer, Maria Brbic, Valentina Boeva. Cancer-specific foundation models: Friend or foe in healthcare AI? [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr IA02.

We propose a novel method, scTree, for single-cell Tree Variational Autoencoders, extending a hierarchical clustering approach to single-cell RNA sequencing data. scTree corrects for batch effects while simultaneously learning a tree-structured data representation. This VAE-based method allows for a more in-depth understanding of complex cellular landscapes independently of the biasing effects of batches. We show empirically on seven datasets that scTree discovers the underlying clusters of the data and the hierarchical relations between them, as well as outperforms established baseline methods across these datasets. Additionally, we analyze the learned hierarchy to understand its biological relevance, thus underpinning the importance of integrating batch correction directly into the clustering procedure.

Abstract The concepts of transcriptional intratumor heterogeneity and phenotypic plasticity have gained significant attention in relation to the development of tumor resistance to cancer treatments [1]. In several cancer types, recent studies have linked certain somatic genomic aberrations to the presence of specific transcriptional states in malignant cells [2,3]. However, our general understanding of drivers of epigenetic and transcriptional plasticity stays somewhat fragmented. In theory, this gap could be filled through the generation and analysis of single-cell multi-omics datasets of human tumors coupled with clinical information. In practice, such datasets are pretty scarce, and existing data certainly do not allow for a global analysis of the genetic drivers of cell plasticity and heterogeneity across cancer types. On the other hand, the scientific community has already generated rich bulk transcriptomics datasets coupled with genetic and clinical information (e.g., TCGA) that could be used to answer the question about the relationship between genetic and epigenetic landscapes if proportions of cells in different malignant states could be extracted via a computational deconvolution of the transcriptomic signal. Unfortunately, available methods for accurate bulk data deconvolution require reference single-cell transcriptomics datasets, which limits the applicability of such approaches to study large pan-cancer datasets. We provide a solution to this problem by creating and benchmarking an algorithmic approach that allows extracting the transcriptional heterogeneity information from bulk tumor profiles without the need for a single-cell reference when a matched DNA-sequencing experiment is available. The latter is true for thousands of tumor samples, including all TCGA datasets. We present results of a pan-cancer analysis linking genetic aberrations with proportions of malignant cells in specific transcriptional states that shed light on the potential genetic drivers underlying malignant cell heterogeneity. References 1. Boeva, V. et al. Heterogeneity of neuroblastoma cell identity defined by transcriptional circuitries. Nat. Genet. 49, 1408–1413 (2017). 2. Neftel, C. et al. An Integrative Model of Cellular States, Plasticity, and Genetics for Glioblastoma. Cell 178, 835-849.e21 (2019). 3. Moorman, A. R. et al. Progressive plasticity during colorectal cancer metastasis. BioRxiv Prepr. Serv. Biol. 2023.08.18.553925 (2023) doi:10.1101/2023.08.18.553925. Citation Format: Agnieszka Kraft, Josephine Yates, Florian Barkmann, Valentina Boeva. Unraveling potential genetic drivers of intratumor transcriptional heterogeneity and phenotypic plasticity in malignant cells [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Translating Cancer Evolution and Data Science: The Next Frontier; 2023 Dec 3-6; Boston, Massachusetts. Philadelphia (PA): AACR; Cancer Res 2024;84(3 Suppl_2):Abstract nr IA024.