Dmitri Kireev, Bridging Physics and Machine Learning in Drug Discovery: A Hybrid Approach to Bias, Scale, and Scarcity

par Dmitri Kireev

mardi 17 juin 2025, 15:00 → 16:00 Europe/Paris

S3 351 (Sciences 3)

Despite recent advances, applying machine learning to drug discovery remains challenging due to limited and biased datasets, complex structure–activity relationships, and the sheer scale of chemical space. In this talk, I will describe our hybrid discovery framework (FRASE-bot) that strategically combines machine learning (ML) with physics-based methods such as docking and alchemical binding free energy calculations. I will discuss challenges in generalization and data representation, introduce our Hit-Triage Pretrained Transformer (Hit-TPT) trained as a binary classifier, and explain how ML is used opportunistically – where data are sufficient – while physics-based models are employed to ensure robustness and interpretability. The talk will include insights from our participation in community benchmarking efforts such as CACHE.