A precision oncology-focused deep learning framework for personalized selection of cancer therapy

Area of Science:

• Oncology
• Computational Biology
• Genomics

Background:

• Precision oncology relies on actionable molecular alterations for targeted therapies.
• Most tumors lack these alterations, limiting treatment to less optimized cytotoxic chemotherapies.
• Data-driven prioritization strategies for chemotherapy are scarce.

Purpose of the Study:

• To develop an integrated computational/experimental approach for predicting efficacious cancer treatments.
• To create a deep learning model (ScreenDL) applicable to both targeted agents and chemotherapies, irrespective of actionable alterations.
• To improve treatment selection beyond existing biomarkers and methods.

Main Methods:

• Generated functional drug response data from a large collection of patient-derived tumor models.
• Trained ScreenDL, a novel deep learning model, using combined tumor omic and functional drug screening data.
• Validated ScreenDL's predictive accuracy against diverse drug mechanisms and existing biomarkers.

Main Results:

• ScreenDL accurately predicts drug response across various mechanisms, outperforming current methods and biomarkers.
• The integrated approach demonstrated superior clinical benefit and objective response rates in preclinical breast cancer models (patient-derived xenografts).

Conclusions:

• The ScreenDL model offers a promising data-driven strategy for selecting effective cancer therapies, even in the absence of actionable alterations.
• This computational/experimental approach warrants further investigation in clinical trials for improved cancer patient outcomes.