An interpretable deep learning framework for genome-informed precision oncology

Area of Science:

• Computational biology
• Genomics
• Pharmacology

Background:

• Cancers arise from altered cellular signaling due to somatic genome alterations (SGAs).
• Precision oncology necessitates understanding cancer cell states to guide targeted therapy selection.
• Current genomic approaches often lack the granularity to predict drug response effectively.

Approach:

• Developed a two-component computational framework: representation learning and drug-response prediction.
• The representation-learning component uses a deep learning model to capture cellular signaling states perturbed by SGAs.
• The drug-response-prediction component leverages these learned cell states to forecast drug efficacy.

Key Points:

• The cell-state-oriented framework significantly improves the accuracy of predicting drug responses compared to models using SGAs directly.
• This approach enhances genome-informed precision oncology by moving beyond solely molecularly targeted drugs.
• The framework successfully predicts responses to chemotherapy agents based on SGAs.

Conclusions:

• This novel computational framework offers a more accurate method for predicting cancer drug responses.
• It advances precision oncology by integrating cellular state information derived from genomic alterations.
• The framework's ability to predict chemotherapy response expands its utility in personalized cancer treatment.