Deep-Learning Tool ScVital Enables Species-Agnostic Integration of Cancer Cell States

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Genetically engineered mouse models (GEMMs) are crucial for studying human tumors.
• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity in cancer.
• Cross-species comparisons of scRNA-seq data are hindered by batch effects and species differences.

Purpose of the Study:

• To develop a computational tool, scVital, to integrate scRNA-seq data across species.
• To overcome batch effects and identify conserved cell states between mouse models and human tumors.
• To introduce a new metric, latent space similarity (LSS), for evaluating batch correction accuracy.

Main Methods:

• Developed scVital, a tool utilizing a variational autoencoder and discriminator for species-agnostic embedding of scRNA-seq data.
• Introduced the latent space similarity (LSS) score for batch correction evaluation.
• Applied scVital to scRNA-seq data from normal tissues, pancreatic ductal adenocarcinoma, lung adenocarcinoma, and undifferentiated pleomorphic sarcoma.

Main Results:

• scVital effectively integrated normal tissue scRNA-seq data across species with high fidelity.
• The tool accurately aligned biologically similar cell states in pancreatic and lung adenocarcinoma data.
• scVital identified a novel, chemotherapy-resistant cell state shared between mouse and human models of undifferentiated pleomorphic sarcoma.

Conclusions:

• scVital successfully overcomes batch effects and species differences in cross-species scRNA-seq analysis.
• The tool enhances the identification of conserved cell states, improving the translational value of GEMMs.
• scVital facilitates deeper biological insights by enabling robust cross-species comparisons.