Gene dependency-informed inference of response to targeted cancer therapies

Area of Science:

• Computational biology
• Cancer genomics
• Pharmacogenomics

Background:

• Targeted cancer therapies rely on blocking essential proteins, but omics-based drug sensitivity models often lack mechanistic understanding.
• Developing predictive models for drug response is crucial for personalized cancer treatment.

Purpose of the Study:

• To introduce FORGE (Factorization Of Response and Gene Essentiality), a novel joint matrix factorization framework.
• To co-model drug response and target gene dependency for biologically informed treatment stratification.
• To derive a "Benefit Score" from gene expression for estimating therapeutic potential.

Main Methods:

• Implemented a joint matrix factorization framework (FORGE) to model drug response and gene essentiality simultaneously.
• Derived a Benefit Score based on basal gene expression to predict therapeutic potential.
• Validated FORGE's performance in predicting drug dependency and IC50 values for erlotinib in unseen cell lines.

Main Results:

• FORGE demonstrated high concordance for gene dependency (0.69) and IC50 (0.62) in erlotinib-treated cell lines.
• Stratification by Benefit Score revealed increased gene dependency and decreased IC50 across quartiles.
• Joint modeling outperformed single-task approaches, improving predictive performance and agreement between gene-level effects (p=0.039).
• Higher benefit scores correlated with tumor regression in patient-derived xenografts and predicted drug susceptibility in a large dataset.

Conclusions:

• FORGE provides a mechanistically grounded approach for predicting targeted therapy response.
• The Benefit Score effectively stratifies patients, identifying those likely to respond to treatment.
• FORGE enhances the predictive power of omics data for personalized cancer medicine.