Look mom, no experimental data! Learning to score protein-ligand interactions from simulations

TL;DR

This paper introduces a hybrid deep learning approach trained on molecular dynamics simulations to predict protein-ligand binding affinities, achieving high accuracy and efficiency in virtual screening for novel targets.

Contribution

It presents a target-specific neural network trained with force matching on MD simulations, bridging physics-based accuracy and deep learning efficiency.

Findings

Achieves competitive virtual screening performance with limited MD simulation time

Attains state-of-the-art early enrichment using true ligand poses

Demonstrates potential of physics-informed learning for novel protein targets

Abstract

Despite recent advances in protein-ligand structure prediction, deep learning methods remain limited in their ability to accurately predict binding affinities, particularly for novel protein targets dissimilar from the training set. In contrast, physics-based binding free energy calculations offer high accuracy across chemical space but are computationally prohibitive for large-scale screening. We propose a hybrid approach that approximates the accuracy of physics-based methods by training target-specific neural networks on molecular dynamics simulations of the protein in complex with random small molecules. Our method uses force matching to learn an implicit free energy landscape of ligand binding for each target. Evaluated on six proteins, our approach achieves competitive virtual screening performance using 100-500 $\mu$s of MD simulations per target. Notably, this approach achieves state-of-the-art early enrichment when using the true pose for active compounds. These results highlight the potential of physics-informed learning for virtual screening on novel targets. We publicly release the code for this paper at https://github.com/molecularmodelinglab/lfm under the MIT license.