A disentangled transformer-based transfer learning framework to predict patient drug response from tumor single-cell...

Area of Science:

• Computational biology
• Genomics
• Pharmacogenomics

Background:

• Intratumoral cellular heterogeneity complicates cancer treatment efficacy.
• Single-cell transcriptomics provides high-resolution data but faces challenges in clinical drug response prediction.
• Existing transfer learning methods often mask crucial cellular heterogeneity by operating at the bulk level.

Purpose of the Study:

• To develop a disentangled transfer learning framework, scTAPE, for predicting patient drug response using tumor single-cell transcriptomics.
• To address the limitations of bulk-level transfer learning by preserving cellular heterogeneity information.
• To enable accurate cross-domain generalization from pre-clinical data to clinical cohorts.

Main Methods:

• scTAPE employs a pre-training and fine-tuning paradigm.
• Disentangled learning extracts pharmacological signals from matched bulk and single-cell expression profiles.
• A supervised drug response model is fine-tuned on labeled cell-line data for cross-domain generalization.

Main Results:

• scTAPE accurately predicts drug response across cell-line datasets and two independent clinical cohorts.
• The framework outperforms existing state-of-the-art single-cell-based predictors.
• scTAPE identifies potential therapeutic agents targeting drug-resistant subpopulations and predicts response to single and combination treatments.

Conclusions:

• scTAPE offers a powerful approach for predicting cancer patient drug response by leveraging single-cell transcriptomics.
• The framework's ability to analyze cellular subpopulations provides insights into treatment strategies and drug resistance.
• scTAPE advances the clinical translation of single-cell data for personalized cancer therapy.