Biologically-informed integration of drug representations for breast cancer treatment using deep learning.

Abstract

Approximately 50% of breast cancer patients receiving neoadjuvant therapy do not achieve pathological complete response (pCR). Identifying personalized optimal therapeutic strategy is an unmet major challenge. Here, based on 4371 eligible patients from 31 datasets, we develop GDnet, an interpretable deep learning model integrating drug representations and tumor transcriptome profiles to predict the responses to neoadjuvant therapies and aid in selecting optimal treatment strategy. We demonstrate that GDnet significantly outperforms transcriptome-only model in predicting treatment response. Then we conduct two series of in-silico simulated clinical trials based on I-SPY2 trial and test datasets respectively and show that GDnet can significantly increase the pCR rate in all trials. Moreover, the odds ratios for pCR increase from 1.6 to 2.5 linearly as optimization intensifies. Overall, GDnet can function as a digital drug-testing surrogate to optimize treatment decision-making. This approach may have broader applications across various treatment settings and cancer types.