MFBind: a Multi-Fidelity Approach for Evaluating Drug Compounds in Practical Generative Modeling

• DOI: 10.48550/arxiv.2402.10387
• Publication Date: 2024-02-15
• Authors: Peter Eckmann; Dongxia Wu; Germano Heinzelmann; Michael K. Gilson; Rose Yu

Current generative models for drug discovery primarily use molecular docking to evaluate the quality of generated compounds. However, such are often not useful in practice because even compounds with high scores do consistently show experimental activity. More accurate methods activity prediction exist, as dynamics based binding free energy calculations, but they too computationally expensive a model. We propose multi-fidelity approach, Multi-Fidelity Bind (MFBind), achieve optimal trade-off between accuracy and computational cost. MFBind integrates simulators train deep surrogate model active learning. Our utilizes pretraining technique linear heads efficiently fit small amounts high-fidelity data. perform extensive experiments that (1) outperforms other state-of-the-art single baselines modeling, (2) boosts performance markedly higher