A multi-task domain-adapted model to predict chemotherapy response from mutations in recurrently altered cancer genes

Area of Science:

• Oncology
• Genomics
• Computational Biology

Background:

• Next-generation sequencing (NGS) is vital in oncology, but most patients do not benefit from targeted therapies.
• Existing drug response prediction (DRP) models often rely on comprehensive genomic data, which is not always clinically available.
• There is a need for DRP models that utilize clinically feasible data, such as restricted gene sets from clinical-grade NGS (cNGS) panels, for chemotherapy repurposing.

Purpose of the Study:

• To develop a DRP model capable of predicting drug response using only data from cNGS panels.
• To address challenges of data sparsity and limited patient drug response information in DRP model development.
• To evaluate the performance of DRP models using whole-exome versus cNGS data.

Main Methods:

• Demonstrated that existing DRP models perform comparably with whole-exome and cNGS data.
• Introduced Drug IDentifier (DruID), a novel DRP model designed for restricted gene sets.
• Employed transfer learning, variant annotations, domain-invariant representation learning, and multi-task learning within DruID.

Main Results:

• Existing DRP models showed similar performance using whole-exome sequencing and cNGS data.
• DruID significantly outperformed state-of-the-art DRP methods on pan-cancer datasets.
• DruID demonstrated robust drug response classification on two independent real-world clinical datasets.

Conclusions:

• DRP models can be effectively developed using limited gene sets from cNGS panels.
• DruID represents a significant advancement toward a clinically applicable DRP tool for chemotherapy repurposing.
• This approach enhances the potential for personalized treatment strategies for a broader patient population in oncology.