MKFGO

Assigning Gene Ontology (GO) terms to an uncharacterized protein is one of the oldest problems in bioinformatics and one of the least well served by any single method. Sequence-homology transfer works when a close, well-annotated relative exists and fails when it does not; protein language models capture functional signal that homology search misses but blur fine distinctions between related terms; interaction networks and term-frequency priors each contribute information the others lack. MKFGO — Multi-source Knowledge Fusion for Gene Ontology prediction — is built around the premise that these sources are complementary and that the fusion should happen at the decision level rather than by concatenating features.

Developed by Yi-Heng Zhu and colleagues at Nanjing Agricultural University and Nanjing University of Science and Technology with collaborators at City University of Hong Kong, Nanjing Forestry University, and Yangzhou University, MKFGO was posted to bioRxiv in April 2025 and published in Briefings in Bioinformatics later that year.

It is explicitly an ensemble, not a single end-to-end network. Five pipelines run independently and their confidence scores are combined by a multilayer perceptron: two are deep-learning models built on pretrained language model embeddings, and three are established bioinformatics heuristics. Users supply a FASTA file and receive GO terms; the trained components are fixed, so no retraining is involved.

#Key Features

• HFRGO, handcrafted-feature deep learning: An LSTM-attention network over PSSM profiles, predicted secondary structure, and InterPro family-domain vectors, trained with a triplet loss so that feature-space similarity tracks functional similarity for guilt-by-association transfer.
• PLMGO, protein language model pipeline: ProtTrans embeddings are decoded by a fully connected network, contributing the evolutionary signal that handcrafted features miss.
• DLMGO, gene-sequence pipeline: Nucleotide Transformer embeddings of the protein-coding gene sequence provide a DNA-side view and, run standalone, predict function for non-coding genes from DNA alone.
• Two classical components: PPIGO infers terms through STRING interaction partners and homology, while NAIGO assigns each term its background frequency in the training set as a calibrated baseline.
• Decision-level fusion: The five score vectors are combined by an MLP and passed through hierarchical post-processing that enforces GO's true-path rule; the paper shows this beats feature-level fusion of the same inputs.

#Technical Details

Benchmarking follows the CAFA protocol, with a temporally separated test set of 1,522 non-redundant, manually reviewed proteins. Against 12 published predictors, MKFGO reached the best Fmax and Smin on all three GO aspects: Fmax of 0.710 (molecular function), 0.459 (biological process), and 0.639 (cellular component), with AUPRC of 0.716, 0.400, and 0.668 and Smin of 6.97, 23.08, and 7.38. Relative to the second-best method, ATGO+ (Fmax 0.693/0.430/0.607), that is an average gain of 4.5% in Fmax and 7.7% in Smin. Ablations show HFRGO and PLMGO carry most of the accuracy while the remaining three still add measurable value, and DLMGO contributes least to the protein-side ensemble. On a separate non-coding benchmark of 147 genes drawn from COXPRESdb and ATTED-II, DLMGO alone performed comparably to the composite TripletGO method.

#Applications

The system targets groups annotating newly sequenced or poorly characterized proteomes, where experimental characterization is out of reach and homology transfer is unreliable. A public webserver accepts sequences directly and is the fastest route for small jobs. Local installation is heavier: the pipeline calls BLAST, SPOT-1D-LM, InterProScan, ProtTrans, and Nucleotide Transformer, plus a roughly 14 GB reference library the authors host for download, so batch annotation of whole proteomes needs GPU hardware and setup effort. The DLMGO branch is the more unusual capability, extending GO prediction to non-coding genes that have no protein product to embed.

#Impact

MKFGO's contribution is less a new architecture than a careful demonstration of where complementary evidence actually lives and how best to merge it. The finding that decision-level fusion outperforms feature-level fusion of the same signals is directly reusable by other function-prediction systems, and the ablation table makes the contribution of each source legible rather than assumed. Caveats are worth stating plainly: the deep-learning components are probes on frozen pretrained embeddings rather than models trained end to end; the gains over ATGO+ are meaningful but incremental; the repository carries no license file, and the trained models arrive through an external archive rather than the code repository itself.