Utilizing 2d And 3d Convolutional Neural Networks For Predicting Protein-Ligand Binding Affinity
Loading...
Files
Date
2026-05-26
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
he accurate prediction of protein-ligand binding affinity remains a cornerstone challenge in computational drug
discovery, directly influencing hit identification, lead optimization, and compound prioritization. Traditional
experimental methods such as isothermal titration calorimetry and surface plasmon resonance, while accurate,
are hindered by high costs and low throughput. This paper presents a comprehensive investigation of 2D and 3D
Convolutional Neural Network (CNN) architectures for protein-ligand binding affinity prediction. We
systematically evaluate multiple CNN-based approaches, including 2D CNNs operating on molecular graphs
and ligand images, 3D CNNs processing voxelized protein-ligand complexes, and hybrid architectures
combining both paradigms. Using the PDBbind v2020 dataset comprising 19,443 protein-ligand complexes and
the CASF-2016 core set for benchmarking, we demonstrate that 3D-CNN models achieve superior performance
with Pearson correlation coefficients of 0.82-0.86 and RMSE values of 1.27-1.00 on the CASF-2016
benchmark. Hybrid attention-based architectures such as HAC-Net and CGDeepAff further improve
performance, achieving Pearson's R of 0.846-0.855 and Spearman's ρ of 0.843-0.861. Our results reveal that 3D
spatial representations capture critical geometric complementarity features that 2D approaches miss, while 2D
methods offer superior computational efficiency for high-throughput screening. We also identify key challenges
including data quality limitations, model interpretability concerns, and generalization to novel protein targets.
This paper concludes by outlining future research directions, emphasizing the potential of multi-modal
architectures, attention mechanisms, and geometric deep learning for advancing binding affinity prediction
toward clinical implementation