GNN2Pfam

Pfam annotations are produced by profile hidden Markov models (HMMs) built from hand-curated seed alignments — one HMM per family, each trained in isolation. That design has been remarkably durable, but it inherits the limits of alignment: proteins with no detectable similarity to a characterized sequence stay unannotated, closely related families are easily confused, and nothing learned about one family transfers to any other.

GNN2Pfam, developed at sinc(i) — the signals, systems and computational intelligence institute of Universidad Nacional del Litoral and CONICET in Santa Fe, Argentina, with a co-author at the La Jolla Institute for Immunology — takes the opposite approach. It trains a single graph neural network across all species and all families at once, so evidence is shared rather than partitioned. Each protein is represented as a graph over its own predicted 3D structure: nodes are residues carrying per-residue embeddings from ESM-2, and edges encode the geometry of residue contacts taken from the AlphaFold2 structure database. The model emits a per-residue score for every Pfam family plus an explicit "no domain" class, so the output is a positional trace along the sequence rather than a single whole-protein label.

The work was posted as a bioRxiv preprint in September 2025 and published in the Journal of Structural Biology in 2026. The authors present it as the first Pfam domain predictor to combine predicted 3D structure with protein language model representations in one end-to-end architecture.

#Key Features

• One model for all families: A single shared checkpoint covers every family in the benchmark, replacing the per-family HMM regime so patterns learned from well-populated families inform sparse ones.
• Structure-aware residue graphs: Contacts are defined by the 5 Å CONS criterion — two residues connect when the minimum distance between any of their carbon, oxygen, nitrogen, or sulfur atoms falls under 5 Å — chosen over 10 Å Cβ and 4 Å centroid alternatives for capturing close-range physical contacts without spurious edges.
• Rich edge features: Every edge carries a 4×4 matrix of pairwise backbone atom distances, giving the network explicit local geometry rather than a binary contact map.
• CRF output layer: A conditional random field over the per-residue predictions enforces label consistency along the chain; ablating it roughly doubles the test error.
• Explicit no-domain class: Because scores are produced per residue against a competing "no domain" curve, domain boundaries emerge from where the curves cross, and low-confidence AlphaFold2 regions show up as score dropouts.

#Technical Details

The architecture is compact: two graph attention network (GAT) layers with ReLU and dropout (0.1), followed by a CRF layer, trained with Adam under cross-entropy loss at a learning rate of 1e-05 with an edge embedding size of 16. Node features are ESM-2 per-residue embeddings of dimension 1,280, with the language model used as a frozen feature extractor. Structures come from the AlphaFold Protein Structure Database, and proteins absent from it are dropped.

Training and evaluation use a subset of the clustered Pfam v32.0 seed split introduced by Bileschi et al., with single-linkage clustering at 25% identity within each family to force a remote-homology test regime. The benchmark comprises 6,169 training and 1,122 test proteins across 58 Pfam families — 7,291 structures in total — using full-length proteins rather than seed domains. The HMM baseline was retrained from scratch on the same training sequences with MUSCLE 3.8.31 alignments and HMMER 3.4. Both methods reach a median recall of 1.0 over the full test partition, but on the subset of families where at least one method errs, GNN2Pfam holds a median recall of 1.0 against 0.75 for the HMMs, a difference significant under a Mann-Whitney U test. It reaches perfect recall on 56% of those hard families and recall above 0.5 on 90% of them.

#Applications

The intended use is annotating proteins that alignment-based search leaves in the dark — sequences with no close characterized relative, and families that HMMs conflate. The per-residue output also supports domain localization within a chain and flags candidate domains in regions Pfam leaves blank; several such predictions the authors initially scored as false positives against Pfam v32.0 were later confirmed by subsequent Pfam and InterPro releases, including a reannotation from PF03364 to PF10604 that the model had already called.

#Impact

GNN2Pfam shows that structure-plus-language-model representations can beat the HMM machinery Pfam has relied on for decades, and it arrives as structure-informed function prediction consolidates around AlphaFold-derived graphs. Its scope is currently modest: the benchmark covers 58 families rather than the full Pfam catalog, and the authors name scaling to newer and larger Pfam releases plus automatic domain boundary segmentation as future work. Practical adoption is also limited by distribution — the public repository provides the data-preparation, training, and evaluation pipeline, but carries no license file and no downloadable trained checkpoint, so reusing the model means retraining it.