AI-Based Functional Annotation of Unknown Genes

Abstract

Despite decades of genome sequencing, approximately 30 to 40 percent of genes in most sequenced organisms lack experimentally validated functional annotations, and AI-based methods that predict gene function from sequence, structure, expression, and network features now offer scalable solutions for closing this annotation gap, yet their adoption varies across methodological categories. We evaluated 216 AI-based functional annotation programmes across centres in Spain, Italy, and Austria between 2017 and 2021, spanning five method categories: homology-based transfer with deep sequence embeddings, protein language model function prediction, guilt-by-association from co-expression and interaction networks, structure-to-function inference from predicted folds, and multi-evidence integration frameworks. A Functional Annotation AI Index (FAAI) was constructed from five sub-scores -- Gene Ontology term prediction accuracy, annotation specificity beyond generic terms, cross-organism generalisation, experimental validation rate, and proteome-scale coverage -- with weights from regression against sustained adoption into genome annotation pipelines. FAAI correlated with adoption at r = +0.84 and discriminated adopted from non-adopted methods with an AUC of 0.884. Multi-evidence integration scored highest (mean FAAI 0.826), while structure-to-function inference trailed at 0.600. Only 35.6 percent exceeded the 0.75 threshold. GO prediction accuracy carried the largest weight (beta = +0.280), followed by annotation specificity (beta = +0.228).