Few-Shot Generalization to Novel Compounds in Single-Cell Drug Response via Graph-Infused Meta-Pretraining

Area of Science:

• Computational Biology
• Pharmacogenomics
• Drug Discovery

Background:

• Single-cell drug response analysis is vital for biomarker identification and understanding resistance.
• Current models often neglect drug structure-function relationships, limiting generalization to novel drugs.
• Genomic data alone is insufficient for comprehensive single-cell drug response prediction.

Purpose of the Study:

• To develop a novel framework integrating drug structural information with genomic data for single-cell drug response prediction.
• To enhance the generalization capability of predictive models to novel drugs.
• To identify key genes and pathways involved in drug resistance at the single-cell level.

Main Methods:

• Developed a graph-aware Transformer to integrate atomic features from drug structures with genomic profiles.
• Employed meta-pretraining and few-shot transfer learning using bulk RNA datasets to address data scarcity.
• Introduced a position-based feature extraction network and gene gradient attribution for pathway and gene identification.

Main Results:

• Achieved a ~5% improvement in accuracy for predicting responses to known drugs.
• Demonstrated a ~20% increase in generalization performance for predicting responses to unseen drugs.
• Successfully identified key drug resistance genes and elucidated drug action pathways.

Conclusions:

• The proposed framework effectively integrates drug structure and genomic data for improved single-cell drug response prediction.
• This approach offers a powerful tool for studying drug resistance mechanisms, especially for novel chemical compounds.
• The model enhances predictive accuracy and generalization, paving the way for more effective drug development and personalized medicine.