Graph neural networks
Can graph networks on 3D protein shapes predict drug binding?
Open access · cc by · source: Europe PMC
Representing both drugs and AlphaFold-predicted proteins as graphs, and letting the two interact through an attention-like module, predicted drug-target binding better than four earlier deep-learning methods.
Study at a glance
- Design
- Computational / modelling — New GNN architecture (iNGNN-DTI) compared with DeepDTA, ML-DTI, MolTrans and DGraphDTA on Davis, KIBA and BIOSNAP, with five runs per dataset, cold-start splits and an ablation.
- N
- Three benchmark datasets of drug-target pairs; sizes are given only in a table not included in the text, so no single N is stated.
- Population
- Drug-target pairs from the Davis, KIBA and BIOSNAP benchmarks (kinase-focused for Davis and KIBA)
- Outcome
- AUROC and AUPRC (plus sensitivity and specificity) for predicting whether a drug binds a target
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
The model had the highest AUROC and AUPRC on all three datasets; the largest gain was in AUPRC on Davis, rising from 0.382 to 0.473, while gains on KIBA were small. On unseen drugs and unseen proteins it also beat the baselines, though absolute AUPRC stayed low. Removing either the interaction module or the pretrained features hurt performance. Highlighted drug atoms sometimes matched real hydrogen-bond partners, but highlighted protein residues did not reliably match true binding sites.
Methodology
The authors turned each drug's chemical string into an atom graph and each protein's AlphaFold2-predicted structure into a residue contact graph. A nested GNN encoded each graph, a cross attention-free transformer module let drug and protein nodes exchange information, and features from pretrained Chemformer and ESM models were added before a final classifier. They compared this model with four baselines on three public benchmarks, tested it on held-out unseen drugs and proteins, ran an ablation, and inspected which atoms and residues the model focused on.
Limitations
Improvements on KIBA and on AUROC generally are small, and results are averages of only five runs with no formal significance tests reported. Davis and KIBA cover only kinase proteins, and negatives in BIOSNAP were randomly sampled pairs that may include true but untested interactions. The interpretability and COVID-19 drug-repurposing analyses rely on virtual docking and literature, not laboratory binding experiments, and the authors admit protein attention weights do not consistently align with real binding residues.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Network context adds information beyond node features.
In drug-target and synthetic-lethality prediction, using graph structure helped over raw features (iGRLDTI), and combining a knowledge graph with the gene-pair graph beat either alone (KG4SL, AUC 0.947).
Evidence for the claim as stated.
Gains over strong baselines were often small (about one to four points for iGRLDTI; small on KIBA for iNGNN-DTI) and came from benchmarks with randomly sampled negatives, single datasets or unequal evaluation protocols, while the protein-structure GNNs showed larger margins.
Evidence for the claim as stated.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
Gains over strong baselines were often small (about one to four points for iGRLDTI; small on KIBA for iNGNN-DTI) and came from benchmarks with randomly sampled negatives, single datasets or unequal evaluation protocols, while the protein-structure GNNs showed larger margins.
Related papers in this topic
Same topic cluster — not a recommendation engine.
- Can spreading signals through a protein network find disease genes?
- Can a known gene network make microarray classifiers interpretable?
- Can a graph neural network sort unknown phage DNA into families?
- Can self-supervised learning predict how mutations change binding?
- Does letting each node choose its own depth fix GNN over-smoothing?