Accurately predicting protein-ligand binding affinities is the holy grail of structure-based drug design. In drug design process, different methods for calculation of relative binding free energies are commonly deployed. If accurately predicted, it can help us in estimating the potency increase or decrease when designing analogous compounds before synthesizing them.<br/>We performed retrospective free energy binding calculations for over 170 compounds from Biogen’s in-house collection spread over 4 different protein targets. We compared multiple methods ranging from fast end approaches (e.g., docking/scoring, MM-GBSA) to rigorous state-of-the-art free energy methods such as FEP+ and Amber-TI.  Apart from using these physics -based methods, we also evaluated machine learning based methods such as K<sub>DEEP</sub> and DeltaDelta neural networks. Overall, we find that FEP+ and MMGBSA are good at accounting for the solvation/desolvation penalty but they do not perform as well when accounting for huge conformational sampling such as loop motions. Performance of Machine Learning based methods depends highly on the training dataset.

Machine learning binding affinity predictions as a cost-effective alternative to physics based approaches

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.