Identifying cell types based on expression profiles is a pillar of single cell analysis. Existing machine-learning methods identify predictive features from annotated training data, which are often not available in early-stage studies. This can lead to overfitting and inferior performance when applied to new data. To address these challenges we present scROSHI, which utilizes previously obtained cell type-specific gene sets and does not require training or the existence of annotated data. By respecting the hierarchical nature of cell type relationships and assigning cells consecutively to more specialized identities, excellent prediction performance is achieved. In a benchmark based on publicly available PBMC data sets, scROSHI outperforms competing methods when training data are limited or the diversity between experiments is large.

Identifying cell types based on expression profiles is a pillar of single cell analysis. Existing machine-learning methods identify predictive features from annotated training data, which are often not available in early-stage studies. This can lead to overfitting and inferior performance when applied to new data. To address these challenges we present scROSHI, which utilizes previously obtained cell type-specific gene lists and does not require training or the existence of annotated data. By respecting the hierarchical nature of cell type relationships and assigning cells consecutively to more specialized identities, excellent prediction performance is achieved. In a benchmark based on publicly available PBMC data sets, scROSHI outperforms competing methods when training data are limited or the diversity between experiments is large.

Human tumors are highly heterogeneous in their cell composition; specifically, they exhibit heterogeneity in transcriptional states of malignant cells, as has been recently discovered through single-cell RNA sequencing (scRNA-seq). Distinct states of malignant cells have been linked to variability in tumorigenic properties and resistance to anti-cancer treatment. Despite the fact that scRNA-seq data contain all necessary information to uncover shared transcriptional states of malignant cells in tumors, jointly analyzing cells from multiple cancer patients comes with its set of challenges including batch correction and accounting for patient-specific genetic background driving differences between gene expression vectors. We propose CanSig, an easy-to-use approach designed to discover known and de novo shared signatures in cancer single cells. CanSig preprocesses, integrates and analyzes scRNA-seq data to provide new signatures of shared transcriptional states and links these states to known pathways. We show that CanSig successfully rediscovers ground truth pathways determining shared transcriptional states in two simulated and three experimental datasets; the latter spanning 135 patients and 72,000 cells. We then illustrate CanSig's investigative potential by discovering novel signatures in esophageal squamous cell carcinoma possibly linked to targeted patient treatment; we also point out a de novo signature in breast cancer predictive of patients' survival. In the cancer types studied, we juxtapose copy number variation with discovered shared transcriptional states and uncover a genetic component predisposing cancer cells to activation of specific transcriptional programs. In sum, CanSig, specifically developed to analyze shared transcriptional heterogeneity of malignant cells of different genetic backgrounds, can greatly facilitate the exploratory analysis of scRNA-seq cancer data and efficiently identify novel transcriptional signatures linked to known biological pathways.

Tumor heterogeneity is regarded as a significant obstacle to successful personalized cancer medicine. Specifically, multiple cancer types have been shown to exhibit heterogeneity in the transcriptional states of malignant cells even within the same tumor. Some of these specific transcriptional states of malignant cells have been linked to cancer relapse and resistance to treatment. However, today there is no universal and easy-to-use computational method to extract information about shared transcriptional states from single-cell RNA sequencing measurements (scRNA-seq). To reliably identify shared transcriptional states of cancer cells, we propose a novel computational tool, CanSig. CanSig automatically preprocesses, integrates, and analyzes cancer scRNA-seq data from multiple patients to provide novel signatures of shared transcriptional states; it also associates these states to known and potentially targetable biological pathways. CanSig jointly analyzes cells from multiple cancer patients while correcting for batch effects and differences in gene expressions caused by genetic heterogeneity. For this, CanSig automatically infers copy number variations in malignant cells and trains a deep learning architecture based on conditional variational autoencoders integrating scRNA-seq measurements. In our benchmarks, CanSig shows state-of-the-art performance in data integration and reliably re-discovers known transcriptional signatures on four previously published cancer scRNA-seq datasets. For instance, CanSig re-discovers four main cellular states of glioblastoma cells previously reported by Neftel et al., Cell, 2019. We further illustrate CanSig’s investigative potential by uncovering signatures of novel transcriptional states in several cancer types; some of the novel signatures are linked to such cancer hallmarks as cell migration and proliferation and are enriched in more advanced tumors. We also uncover signatures associated with specific genomic aberrations such as gene copy number gains. In conclusion, CanSig detects shared transcriptional states in tumors using as input scRNA-seq data for cells with different genetic backgrounds and possibly exhibiting strong batch effects. It can significantly facilitate the analysis of scRNA-seq cancer data and efficiently identify transcriptional signatures linked to known biological pathways. The CanSig method implemented in Python is available at https://github.com/BoevaLab/CanSig Citation Format: Josephine Yates, Florian Barkmann, Pawel Piotr Czyz, Agnieszka Kraft, Niko Beerenwinkel, Valentina Boeva. CanSig: De novo discovery of shared transcriptional states in cancer single cells [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 3166.