A Transfer Learning Framework for Predicting and Interpreting Drug Responses via Single-Cell RNA-Seq Data.

Area of Science:

• Computational Biology
• Genomics
• Cancer Research

Background:

• Chemotherapy resistance is a major obstacle in cancer treatment, driven by tumor heterogeneity and complex cellular interactions.
• Existing predictive models for drug response face limitations in performance and interpretability due to data challenges.
• Single-cell RNA sequencing (scRNA-seq) offers potential for understanding resistance mechanisms, but data utilization is limited.

Purpose of the Study:

• To develop a novel transfer learning framework for aligning bulk and single-cell RNA sequencing data for drug response prediction.
• To enhance the interpretability of predictive models by mapping latent features to biological pathways.
• To leverage publicly available scRNA-seq drug response datasets for improved cancer therapy insights.

Main Methods:

• Collected and curated scRNA-seq drug response datasets from public repositories.
• Designed a shared encoder to project bulk and single-cell data into a unified latent space.
• Implemented a sparse decoder guided by prior biological knowledge for pathway-based interpretability.

Main Results:

• The proposed transfer learning model achieved superior performance in drug response prediction across five scRNA-seq datasets.
• Integrated gradient analysis revealed biologically meaningful insights into drug resistance mechanisms.
• Demonstrated the model's ability to effectively align and utilize diverse sequencing data types.

Conclusions:

• Deep learning, particularly transfer learning, holds significant potential for advancing drug response prediction in oncology.
• scRNA-seq data, when effectively utilized, provides valuable insights for understanding and overcoming cancer drug resistance.
• The developed framework offers a promising approach for personalized cancer treatment strategies.