Concept
Graph neural networks
8 studiesEvidence last moved Sep 27, 2026
Graph neural networks learn from data arranged as nodes and edges, passing information between neighbours, for example atoms in a protein or drugs, targets and genes in a biological network. The evidence here is from computational benchmarks on protein model quality, protein interaction sites, drug-target interaction, synthetic lethality and drug response, plus an early gene-network smoothing method.
Many biological objects are naturally graphs. These studies show what building that structure into a model buys, and why reported gains need careful reading because of benchmark design.
Studies
8
Findings
4
7 supporting · 0 challenging · 1 qualifying citations
Open tensions
2
Latest change
Concept page published
Graph neural networks
Currently
What we know
- Building 3D symmetry into the network helped on protein structure tasks.
- The graph structure itself carried much of the signal.
- Network context adds information beyond node features.
- Graph priors can aid interpretation even when accuracy does not change.
Largest unresolved question
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.
Common misconceptions
An 'unknown' drug-target or gene pair is a true negative.
Several benchmarks sampled negatives from unrecorded pairs, some of which may be real but untested interactions, which can distort metrics and make the task easier than discovery.
Attention weights in a GNN show where the drug binds.
In iNGNN-DTI, highlighted drug atoms sometimes matched real hydrogen-bond partners, but highlighted protein residues did not reliably match true binding sites.
Related
Claim ledger
What the evidence shows
Drawn from 8 studies in this library. Mix labels say which citation roles are present; they are not a strength score. Supports means evidence for a finding; Challenges means evidence against a stated position; Qualifies marks scope.
Building 3D symmetry into the network helped on protein structure tasks.
E(3)-equivariant GNNs, which respect 3D rotations and translations, outperformed alternatives: EquiPPIS was the only method above 0.8 ROC-AUC for interaction sites, and its non-equivariant version and plain GCN/GAT networks did worse.
- Can an equivariant graph network judge protein models better?— Did not isolate equivariance from its extra AlphaFold2 and language-model features; removing AlphaFold2's confidence input caused the largest drop.
Study Role Design N Population Outcome Does building 3D symmetry into a GNN improve protein site prediction? Supports Computational / modellingEquiPPIS (E(3)-equivariant GNN) trained on 335 proteins, tested on 60, compared with nine prior predictors plus invariant/no-attention/GCN/GAT ablations; also tested on unbound and AlphaFold2 structures N=60 · 60 test proteins (Test_60); 335 training proteins; 31-protein unbound test subset; 42-protein validation set Protein chains from three public PPI benchmark datasets Per-residue interaction-site classification: ROC-AUC, PR-AUC, F1, MCC Can an equivariant graph network judge protein models better? Supports Computational / modellingNew equivariant graph neural network (EnQA) trained on CASP and AlphaFold2 structural models and compared with existing quality-assessment methods on held-out test sets No single N: separate test sets of 9501 CASP14 models (64 targets), 945 CAMEO models (38 targets) and 56 AlphaFold2 test targets Predicted 3D protein structural models from CASP, CAMEO and AlphaFold2 Agreement (MSE, MAE, Pearson correlation) between predicted and true lDDT quality scores at residue and model level The graph structure itself carried much of the signal.
GraphQA clearly outperformed ProQ4, which uses the same input features with a 1D-CNN, for ranking protein models on CASP13, and a variant using only amino-acid identity was almost as good.
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).
- Does letting each node choose its own depth fix GNN over-smoothing?
- Does background biological knowledge help predict deadly gene pairs?
- Can graph networks on 3D protein shapes predict drug binding?
Study Role Design N Population Outcome Does letting each node choose its own depth fix GNN over-smoothing? Supports Computational / modellingGraph of drugs and protein targets with learned node features; node-dependent smoothing depth, then a gradient-boosted tree classifier; 10-fold cross-validation against five published methods. N=1923 · 1923 known drug-target interactions among 549 drugs and 424 targets, plus an equal number of randomly paired negatives (a 1:10 imbalanced setting was also run). Drugs and protein targets from DrugBank 3.0 and HPRD AUC, AUPR and F1 for predicting drug-target interactions; mean absolute distance between node embeddings as an over-smoothing measure Does background biological knowledge help predict deadly gene pairs? Supports Computational / modellingLink prediction on SynLethDB gene pairs; KG4SL (knowledge-graph message passing on gene-specific subgraphs) versus 10 baselines using AUC, AUPR and F1, plus sensitivity, convergence and knowledge-graph ablation analyses. N=72804 · Balanced dataset of gene pairs: known synthetic-lethal pairs from SynLethDB plus an equal number of randomly drawn unknown pairs. Human gene pairs from SynLethDB and the SynLethKG knowledge graph AUC, AUPR and F1 for predicting synthetic-lethal gene pairs Can graph networks on 3D protein shapes predict drug binding? Supports Computational / modellingNew 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. 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. Drug-target pairs from the Davis, KIBA and BIOSNAP benchmarks (kinase-focused for Davis and KIBA) AUROC and AUPRC (plus sensitivity and specificity) for predicting whether a drug binds a target Graph priors can aid interpretation even when accuracy does not change.
Constraining a classifier to be smooth over a metabolic gene network did not improve accuracy on a 17-sample microarray dataset but produced more interpretable pathway-level weights.
Debates
Tensions and limits
Some items are genuine disagreements on the same question. Others mark 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.
Evidence for
PaperFren reads this as a limit on how far one study travels — different assays, populations, or outcomes — not a forced fight between papers.
GNN-based drug-response models did well on random splits but much worse on held-out cell lines (SWnet MSE rose from 0.94 to 2.28).
GNN-based drug-response models did well on random splits but much worse on held-out cell lines (SWnet MSE rose from 0.94 to 2.28).
PaperFren reads this as a limit on how far one study travels — different assays, populations, or outcomes — not a forced fight between papers.
Timeline
How understanding moved
Study years are when the paper was published. Evidence edits are dated changes to this page's claims. Explanations are when PaperFren added a Discovery — not a claim that the science happened that day.
2026
Concept page published
Graph neural networks
Change log
What changed
Dated edits to this page's evidence: studies added or removed from a claim, claims added or withdrawn, and new explanations tagged here. Rewordings are not listed.
- Concept page published
Papers
8 studies in this library bear on Graph neural networks, ordered by citations.
- Can a known gene network make microarray classifiers interpretable?
Forcing a classifier's gene weights to vary smoothly across a known metabolic network kept accuracy unchanged while making the classifier readable in terms of pathways.
- Does letting each node choose its own depth fix GNN over-smoothing?
Letting each drug or protein node decide how far to gather information from its neighbours kept graph features distinct and slightly improved drug-target prediction.
- Can a graph neural network judge predicted protein shapes?
Treating a protein as a graph of residues and contacts let a fairly simple neural network rank predicted protein structures about as well as the best existing single-model tools.
- Does background biological knowledge help predict deadly gene pairs?
Feeding a graph neural network a knowledge graph of pathways, diseases and drugs improved prediction of synthetic-lethal gene pairs compared with methods that only saw known gene pairs.
- Can a neural network predict which cancer drugs work on which cells?
Combining a graph network for drug structure with gene features predicts drug sensitivity well on familiar cell lines, but error more than doubles on cell lines never seen in training.
- Can an equivariant graph network judge protein models better?
A graph network built to respect 3D rotations and translations estimated the quality of predicted protein structures more accurately than earlier methods, and even better than AlphaFold2's own confidence scores.
- Can graph networks on 3D protein shapes predict drug binding?
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.
- Does building 3D symmetry into a GNN improve protein site prediction?
A graph neural network whose outputs transform correctly under rotations, reflections and translations predicted protein interaction sites better than earlier methods, even from AlphaFold2-predicted structures.
Compare studies
Select 2–10 studies. Design and N are labels, not a ranking.
Nothing selected yet.
Questions
What is still open
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.
GNN-based drug-response models did well on random splits but much worse on held-out cell lines (SWnet MSE rose from 0.94 to 2.28).
Ask PaperFren about Graph neural networks
Study this conceptflashcards and short-answer questions
What does 'equivariance' mean for a protein GNN, and what evidence suggests it matters?
An equivariant network's outputs transform consistently when the input structure is rotated or translated, so it does not need to learn every orientation. EquiPPIS outperformed its own invariant version and plain GCN/GAT networks for interaction-site prediction, reaching above 0.8 ROC-AUC. The test set was small (60 proteins), so margins rest on few examples.
Why should a reader be cautious about GNN drug-target benchmark gains?
Gains over the best baselines were often one to a few points, on old or single datasets, with randomly sampled negatives that may be true interactions. Some comparisons used different evaluation protocols (cross-validation for baselines vs a single split). Performance on unseen drugs or proteins remained low in absolute terms.
Flashcards
0 of 8 answers reviewed