GNN-MA: Soft Molecular Alignment with Cross-Graph Attention for Ligand-Based Virtual Screening

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Ligand-based virtual screening (LBVS) aims for high early enrichment in large libraries.
• Traditional LBVS often uses 1D/2D descriptors, neglecting 3D information due to computational cost and uncertainty.
• Accurate molecular alignment and conformer generation are significant challenges in 3D-based screening.

Purpose of the Study:

• To develop a novel retrieval-style pairwise scoring model for LBVS.
• To leverage molecular graphs for unified representation and avoid explicit 3D modeling.
• To enhance early enrichment and discrimination in virtual screening campaigns.

Main Methods:

• Proposed GNN-MA, a graph neural network model using intra-graph message passing and cross-graph attention.
• Implemented atom-level soft alignment for focusing on key substructures.
• Introduced a bond-to-atom semantic aggregation module for improved chemical bond cue utilization.
• Utilized 2D molecular graphs derived from SMILES, eliminating the need for 3D conformer generation.

Main Results:

• GNN-MA achieved competitive overall discrimination (ROC-AUC) on DUD-E and LIT-PCBA datasets.
• Demonstrated consistent gains in early enrichment metrics (EF@1-5%) on DUD-E compared to ablated variants.
• Showed target-dependent improvements on LIT-PCBA.
• Learned soft alignment provided qualitative interpretability in case studies.
• Throughput benchmarks indicate suitability as a re-ranking model.

Conclusions:

• GNN-MA offers an effective 2D graph-based approach for LBVS, enhancing early enrichment.
• The model provides a balance between performance and computational efficiency by avoiding 3D modeling.
• GNN-MA shows promise as a re-ranking tool in large-scale virtual screening workflows.