Mohammed Balkhair

Artificial Intelligence Specialization · Cohort 2025

Deep Learning Approaches to Improve Drug–Target Binding Affinity Prediction in Precision Oncology

This study presents a benchmarking effort to evaluate and improve drug–target binding affinity (DTA) prediction models for precision oncology. We focus on graph-based deep learning approaches that represent drugs and proteins, with particular emphasis on the protein modeling component. To ensure fair evaluation, we address issues of information leakage across datasets and assess model robustness under distribution shifts. Our experiments systematically incorporate various forms of protein dynamics, including contact maps, embeddings from pre-trained protein language models, and conformational ensembles generated by structure prediction tools such as AlphaFold2. We investigate the individual and combined effects of these dynamic features on model performance. By refining protein graph representations at both the node and edge levels, we aim to establish a strong baseline for DTA prediction tasks and provide insights into how structural dynamics enhance model generalization. Ultimately, our findings support the development of more reliable models for integration into personalized cancer therapy pipelines.

Mentor: Sushant Kumar

Read the final report (PDF)

See it in the program report (PDF)