From Proteins to Ligands: Decoding Deep Learning Methods for Binding Affinity Prediction

Abstract

Accurate in silico prediction of protein−ligand binding affinity is important in the early stages of drug discovery. Deep learning-based methods exist but have yet to overtake more conventional methods such as giga-docking largely due to their lack of generalizability. To improve generalizability, we need to understand what these models learn from input protein and ligand data. We systematically investigated a sequence-based deep learning framework to assess the impact of protein and ligand encodings on predicting binding affinities for commonly used kinase data sets. The role of proteins is studied using convolutional neural network-based encodings obtained from sequences and graph neural network-based encodings enriched with structural information from contact maps. Ligand-based encodings are generated from graph-neural networks. We test different ligand perturbations by randomizing node and edge properties. For proteins, we make use of 3 different protein contact generation methods (AlphaFold2, Pconsc4, and ESM-1b) and compare these with a random control. Our investigation shows that protein encodings do not substantially impact the binding predictions, with no statistically significant difference in binding affinity for KIBA in the investigated metrics (concordance index, Pearson's R Spearman's Rank, and RMSE). Significant differences are seen for ligand encodings with random ligands and random ligand node properties, suggesting a much bigger reliance on ligand data for the learning tasks. Using different ways to combine protein and ligand encodings did not show a significant change in performance.

Do molecular fingerprints identify diverse active drugs in large-scale virtual screening? (no)

“…In recent years, deep learning strategies have entered the drug discovery toolkit (e.g. Gentile et al (2022); Gorantla, Kubincova, Weiße, & Mey (2023); Yang et al (2021)), but these have not yet solved the problem of rapid virtual screening. Though the path forward is not clear, we suggest that it is vital that molecules be represented in such a way that the potential context of the molecule (i.e.…”

Benchmarking Active Learning Protocols for Ligand-Binding Affinity Prediction

“…The main focus is on the AL design and not on the method used for labeling data. Different strategies for labeling are possible, such as docking, AFE methods (relative or absolute methods), experimental measurements, or an ML property prediction model . How to choose best labeling strategies and how to mix different ones will not be evaluated here.…”

“…Computational potency prediction methods have evolved over the last four decades from traditional docking, , alchemical free energy (AFE) techniques , to more recently ML approaches. − However, the use of AL applications together with computational potency estimation such as virtual screening − or relative binding free energy (RBFE) calculations using molecular dynamics simulations − only emerged in the past 8 years, driven by the increase in automation and throughput of computational tools for drug discovery. In these cases, 100s to 1000 compounds are selected out of pools containing up to 100,000 samples.…”

Guided Docking as a Data Generation Approach Facilitates Structure-Based Machine Learning on Kinases

“…As of December 2023, this includes 19,443 QM and 16,972 MD simulations. Other studies investigate augmenting the structure space using AlphaFold or homology modeling, with two recent examples from the kinase field. , The KinCo dataset focuses on kinase inhibitors and provides docked poses for 137,778 kinase-ligand complexes, annotated with bioactivity measurements. The complexes are generated using blind docking into kinase homology models covering the complete kinome tree.…”