GOBoost

Sort the Gene Ontology terms in a standard protein annotation benchmark by how often they appear and the curve falls off a cliff: a few dozen generic terms sit on tens of thousands of proteins, while most of the 2,752 terms are attached to a handful each. A multi-label classifier trained on that distribution can post a respectable score by learning the head and effectively ignoring everything else — and the terms it ignores are the ones a biologist actually wants, such as which reaction an enzyme catalyses.

GOBoost treats that skew as the modelling problem rather than as a nuisance. Three base models share one architecture but see different slices of the label space — one trained on the high-frequency head, one on the medium- and low-frequency tail, one on all terms — and their predictions are averaged so that a term's estimate comes from a model that saw enough of it. Inside each base model sit two further pieces: a global-local label graph that learns GO term co-occurrence both corpus-wide and per protein, and a multi-granularity focal loss that decouples the focusing parameter along two axes at once, positive versus negative labels and head versus tail labels.

The method comes from Renzhi Cao's group at Pacific Lutheran University with collaborators at Anhui University, Hamilton College, Saint Louis University and the University of Washington Bothell. Code is on GitHub under an MIT license and the trained checkpoints are archived on Zenodo.

#Key Features

• Head/Tail/All ensemble: Three checkpoints trained on different label partitions are averaged, so rare terms are scored by a model that was not competing against the head.
• Global-local label graph: A corpus-wide co-occurrence matrix captures the relationships among frequent terms; a second, per-protein matrix learned from the query's own features supplies what the shared matrix is too coarse to hold.
• Multi-granularity focal loss: The focal parameter is split into positive/negative and head/tail components, the tail ratio set from the training label distribution, pushing gradient toward under-represented terms.
• Structure plus sequence: Residue features combine a frozen ESM-1b embedding with a learned amino-acid-type embedding, then propagate over a contact graph built from the query's 3D coordinates.
• Annotation from a single structure: The released predictor takes one PDB file and a probability threshold and returns GO terms with confidences, with no fitting on user data.

#Technical Details

Each residue is represented by the 1280-dimensional output of ESM-1b's 33rd layer — sequences are truncated at 1,000 residues — to which an amino-acid-type encoding, mapped to those same 1280 dimensions, is linearly added before projection to 512 hidden dimensions. The protein becomes an undirected graph whose nodes are residues and whose edges connect Cα atoms within 10 Å, over which a graph convolutional network aggregates, a weighted sum across the deeper layers forming the final graph feature. Class activation mapping converts residue-level features into per-term label embeddings, which pass through the label graph module; a graph-pooling classifier and a label-feature classifier are then averaged into the final probability.

Training uses the DeepFRI PDB split — an 8:1:1 division yielding 29,893 training proteins, with 489 MF, 1,943 BP and 320 CC terms — augmented with the HEAL AF2 set's 38,185 AlphaFold 2 training structures, themselves clustered at 25% sequence identity with MMseqs, for a combined pool of roughly 68,000 proteins. The head model covers the top 100, 300 and 50 terms for MF, BP and CC. Training ran on a single RTX 3080 with batch size 32 and Adam at 1e-4. On the PDB test set GOBoost reaches AUPR of 0.765, 0.458 and 0.573 and Fmax of 0.787, 0.659 and 0.745 for MF, BP and CC, improving AUPR over HEAL by 10.71%, 35.91% and 22.71%. Stratified by information content, the gains are largest on the most specific terms (IC ≥ 10), where BP AUPR rises 52.48%.

#Applications

The practical use is annotating proteins that homology transfer handles badly — orphan sequences, structural genomics targets, engineered or metagenomic proteins with no well-annotated relative. Because the input is a structure, an AlphaFold 2 model works in place of an experimental one, which puts most of the proteome in reach. Groups needing a first-pass functional hypothesis before committing to assays can run the predictor over a set of structures and filter on the returned probabilities.

#Impact

GOBoost is a concrete demonstration that the long-tail structure of the Gene Ontology is worth modelling explicitly, and its largest gains land where that argument predicts — on deep, specific, low-frequency terms rather than on aggregate averages. The costs are real: a 3D structure is required for every query, ESM-1b's 1,000-residue truncation discards the tails of long proteins, the annotations come from a 2019 GO snapshot, and the full method loads three checkpoints per sub-ontology from a 1.3 GB archive. The work was published in Bioinformatics in June 2025, though the comparisons are the authors' own rather than an independent assessment such as CAFA.