Geometric Graph Neural Networks for Drug Discovery: 3D Structure‑Based Binding Affinity Prediction and Molecular Property Learning

The biopharma industry faces diminishing returns due to exhaustive target and small‑molecule exploration; accelerating virtual screening with machine learning can reduce cost and time.

Baidu’s PaddleHelix platform, built on PaddlePaddle, provides open‑source tools for drug screening, ADMET prediction, molecule generation, protein structure prediction, and more.

Biological data consists of three main types—small‑molecule compounds, DNA/RNA sequences, and proteins—each requiring geometric representation because spatial conformation strongly influences properties.

Standard graph convolutions ignore geometric invariance, leading to poor modeling of molecular structures. Two solutions are highlighted: Equivariant Neural Networks, which enforce commutativity between convolution and geometric transforms, and Geometric‑Encoded Message Passing, which explicitly encodes spatial information.

For protein‑ligand binding affinity prediction, the authors review virtual screening, describe four mainstream approaches (1D‑CNN, feature‑based models, 3D‑CNN, and GNN), and introduce their Structure‑aware Interactive Graph Neural Network (SIGN) that incorporates polar‑coordinate‑inspired graph attention, distance/angle discretization, and node‑edge interaction.

Experimental evaluation on the PDBbind and CSAR‑HiQ benchmarks demonstrates that SIGN outperforms baseline methods, and ablation studies confirm the importance of spatial and interactive factors.

In molecular property prediction, a dual‑channel framework processes both 2D and 3D views of molecules. Geometric contrastive learning aligns representations from these views while preserving chemical rules, using tasks such as atom masking, edge distance/angle prediction, and global distance masking.

Results on multiple benchmarks show that the proposed GeomGCL method surpasses traditional message‑passing GNNs, geometry‑based GNNs, and existing contrastive learning approaches, with notable gains from incorporating both views and self‑supervised objectives.

The authors also discuss a large‑scale self‑supervised pre‑training strategy (GEM) that leverages geometric information and multiple pre‑text tasks, achieving significant improvements on 14 benchmark datasets.

All code is released on GitHub (https://github.com/PaddlePaddle/PaddleHelix) and the online platform (paddlehelix.baidu.com) allows users to upload sequences or structures for immediate prediction.