High Performance of Gradient Boosting in Binding Affinity Prediction

TL;DR

This paper demonstrates that combining protein-ligand interaction features with graph-level features in gradient-boosted decision trees significantly improves binding affinity prediction, offering a scalable and efficient alternative to graph neural networks.

Contribution

The paper introduces a novel approach that integrates interaction and graph features into GBDT for better binding affinity prediction, outperforming existing methods.

Findings

GBDT with combined features outperforms GNN-based models

The approach is computationally efficient and scalable

Results show improved prediction accuracy over prior methods

Abstract

Prediction of protein-ligand (PL) binding affinity remains the key to drug discovery. Popular approaches in recent years involve graph neural networks (GNNs), which are used to learn the topology and geometry of PL complexes. However, GNNs are computationally heavy and have poor scalability to graph sizes. On the other hand, traditional machine learning (ML) approaches, such as gradient-boosted decision trees (GBDTs), are lightweight yet extremely efficient for tabular data. We propose to use PL interaction features along with PL graph-level features in GBDT. We show that this combination outperforms the existing solutions.