Predicting Single-Cell Drug Sensitivity Utilizing Adaptive Weighted Features for Multi-Source Domain Adaptation.

Area of Science:

• Genomics
• Computational Biology
• Pharmacology

Background:

• Single-cell sequencing generates vast transcriptional data, enabling identification of drug-resistant tumor subpopulations.
• Current single-cell drug response prediction models have suboptimal performance.
• Developing accurate single-cell drug response prediction is crucial for personalized cancer therapy.

Purpose of the Study:

• To propose scAdaDrug, a novel multi-source domain adaptation model for predicting individual cell drug response.
• To enhance drug response prediction accuracy at the single-cell level.
• To leverage diverse datasets for robust model training.

Main Methods:

• Developed scAdaDrug, a multi-source domain adaptation model utilizing adaptive importance-aware representation learning.
• Employed a shared encoder with adversarial domain adaptation to extract domain-invariant features.
• Introduced a plug-and-play module for adaptive modulation of latent representations using importance-aware weights.

Main Results:

• scAdaDrug achieved state-of-the-art performance in drug response prediction across multiple independent datasets.
• The model demonstrated effectiveness on single-cell data from cell lines, patient-derived xenografts (PDX), and clinical tumor cohorts.
• Ablation experiments confirmed the model's ability to capture underlying drug response patterns.

Conclusions:

• scAdaDrug offers a significant advancement in single-cell drug response prediction.
• The model's adaptive importance-aware representation learning effectively integrates information from multiple data sources.
• scAdaDrug holds promise for improving personalized treatment strategies in oncology.