Improving ΔΔG Predictions with a Multitask Convolutional Siamese Network

Abstract

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

Scoring Functions for Protein-Ligand Binding Affinity Prediction Using Structure-based Deep Learning: A Review

“…For example some methods trained to predict binding affinities performed poorly on the different task of predicting the differences in binding affinity upon protein mutation (Aldeghi et al, 2018b). DL methods specifically designed for ranking-computing relative binding affinities-have been developed (Jiménez-Luna et al, 2019) and are an active area of research (McNutt and Koes 2022).…”

Data-driven generation of perturbation networks for relative binding free energy calculations

“…This has been attempted before and offers additional information such as pose differences between input ligands which are highly inuential to the RBFE reliability. 27,28 However, the bottleneck in this scenario is that a large number of RBFE simulations must be run. Additionally robust experimental binding affinity must be available for each RBFE edge Indeed, during early investigations of this work attempts were made to create a training set that included original ligands, but the chemical space associated with training such a model appeared too large with respect to the data available.…”

“…perturbations have been proposed in the form of siamese neural networks. 27,28 The current work proposes a data-driven RBFE network generator as an alternative to expert-driven approaches. To accomplish this, a transfer learning ML framework was designed that allows predictions of statistical uncertainties for molecular perturbations typically handled in RBFE.…”

Data-driven Generation of Perturbation Networks for Relative Binding Free Energy Calculations

“…This has been attempted before and offers additional information such as pose differences between input ligands which are highly influential to the RBFE reliability. 27,28 However, the bottleneck in this scenario is that a large number of RBFE simulations must be run. Indeed, during early investigations of this work attempts were made to create a training set that included original ligands, but the chemical space associated with training such a model appeared too large with respect to the data available.…”

“…21 Additionally, novel machine learning (ML) techniques of describing RBFE perturbations have been proposed in the form of siamese neural networks. 27,28 The current work proposes a data-driven RBFE network generator as an alternative to expert-driven approaches. To accomplish this, a transfer learning ML framework was designed that allows predictions of SF for molecular perturbations typically handled in RBFE.…”