Improving ΔΔG predictions with a multi-task convolutional Siamese Network

• DOI: 10.26434/chemrxiv-2021-vcmzz
• Publication Date: 2021-12-14T08:21:56Z
• Authors: Andrew McNutt; David Koes
• Keywords: Regularization; Ranging

The lead optimization phase of drug discovery refines an initial hit molecule for desired properties, especially potency. Synthesis and experimental testing the small perturbations during this refinement can be quite costly time consuming. Relative binding free energy (RBFE, also referred to as ∆∆G) methods allow estimation changes after a ligand scaffold. Here we propose evaluate Convolutional Neural Network (CNN) Siamese network prediction RBFE between two bound ligands. We show that our multi-task loss is able improve on previous state-of-the-art via increased regularization latent space. architecture well suited in comparison standard CNN trained same data (Pearson’s R 0.553 0.5, respectively). When evaluated left-out protein family, shows variability its predictive performance depending family being ranging from-0.44 0.97). improved generalization by injecting only few examples (few-shot learning) from evaluation dataset model training.