MSNGO

Most proteins in most genomes have never been characterized experimentally, and Gene Ontology annotation is the standard way of assigning them a provisional function. Single-species predictors have improved steadily, helped enormously by the availability of predicted structures, but the species that need annotation most are the ones with the fewest labels to learn from. Methods that work across species have generally fallen back on sequence features plus a protein-protein interaction network, leaving the structural signal on the table.

MSNGO, from the School of Computer Science and Technology at Harbin Institute of Technology (Shenzhen) with the Center for Bioinformatics at Harbin Institute of Technology, closes that gap. It learns a structural representation directly from residue contact maps derived from AlphaFold 2 models, concatenates it with frozen ESM-2 sequence embeddings, and then propagates the combined representation through a heterogeneous network that spans thirteen species. Labels can therefore travel from a well-annotated organism to a sparsely annotated one along both interaction and homology edges. The work was published in Bioinformatics in 2025 and sits alongside GO-annotation predictors such as DeepGO, differing in its explicit multi-species propagation step.

#Key Features

• Structure as a first-class input: A graph convolution pooling module runs over the residue contact map of each protein and pools amino acid representations into a protein-level structural vector, rather than treating structure as an auxiliary feature.
• Frozen sequence encoder: ESM-2 supplies per-residue embeddings that are averaged into a 1280-dimensional protein vector, so the trained parameters are spent on structure and propagation instead of re-learning sequence statistics.
• Attention-weighted cross-species propagation: A graph attention mechanism sets the propagation strength on each edge of a heterogeneous network built from interaction and homology relationships, letting distantly related species contribute where the evidence supports it.
• All three sub-ontologies: Separate models are trained for molecular function, biological process, and cellular component, covering 6,445, 19,323, and 2,645 GO terms respectively.
• FASTA-in prediction: A released checkpoint annotates arbitrary new sequences through a single predict.py call, with training and inference cleanly separated in the codebase.

#Technical Details

The structural encoder uses two graph convolution pooling modules, each containing three graph convolution layers, with a graph pooling rate of 0.75; contact maps come from AlphaFold DB, with ESMFold predictions filling in proteins the database lacks. Sequence features come from the 650M-parameter ESM-2 checkpoint. The heterogeneous network combines STRING v11.0b interaction edges with homology similarity edges across thirteen species spanning mammals, ectotherms, plants, bacteria, and fungi. Evaluation follows a temporal split: proteins annotated before January 2021 train the model, those first annotated between January 2021 and July 2022 form the validation set, and those first annotated between August 2022 and August 2023 form the test set.

MSNGO reaches an Fmax of 0.7332 on biological process, 0.8102 on molecular function, and 0.7920 on cellular component, with AUPR values of 0.7485, 0.8100, and 0.8640 — improvements of 10.43%, 5.04%, and 1.14% over the second-best method in each branch. The nearest competitor with matched ESM-2 features, PSPGO, scores 0.6289, 0.6395, and 0.6758; SPROF-GO is closest on cellular component at 0.7806. An ablation on molecular function attributes most of the margin to the two novel components: dropping structural features takes Fmax to 0.6395, and dropping the network propagation layer takes it to 0.5449.

#Applications

MSNGO is aimed at annotating proteomes where experimental characterization is thin — non-model plants, agricultural pathogens, environmental bacteria and fungi — by borrowing evidence from better-studied relatives. Groups that have sequenced a new organism can run predicted GO terms over its full proteome to prioritize which proteins to characterize at the bench, and the per-branch models let a project focus on molecular function alone when that is what the downstream analysis needs.

#Impact

The result that matters here is the size of the structural contribution: on molecular function, removing structure costs more than any other component except propagation itself, which is a concrete argument that structure-aware features belong in multi-species function prediction rather than sequence-plus-network alone. Practical adoption faces friction. The repository has no license file, the pretrained checkpoint is distributed through a Google Drive link rather than a model hub, and the thirteen-species training network bounds how far predictions can be trusted for lineages outside it.