MENDELSEEK algorithm outperforms AI methods in identifying disease-causing genes

A patient walks into a genetic clinic with a rare disease no one can quite name. Sequencing reveals hundreds of mutated genes. Which one caused the illness? For more than half of all known Mendelian diseases—single-gene disorders that follow predictable inheritance patterns—medicine still cannot answer that question. The mutations are there. The diagnosis remains elusive.

Researchers at multiple institutions have now built a tool called MENDELSEEK that cuts through this fog with unusual precision. The algorithm integrates genetic variation scores with information about protein interactions, cellular pathways, and evolutionary history to identify which genes among hundreds of candidates actually drive disease. When tested against established methods, including DeepMind's AlphaMissense, MENDELSEEK achieved an area under the precision-recall curve of 0.737—a 17.7 percent improvement over the next-best approach and more than double the performance of AlphaMissense, which scored 0.324 despite its state-of-the-art artificial intelligence foundation.

The work began with a practical frustration. Existing variation-based methods—tools designed to flag harmful mutations—cast too wide a net. When researchers evaluated ENTPRISE, SIFT, and PolyPhen2 on ten patient samples, ENTPRISE predicted roughly 100 disease-causing genes per patient, while SIFT and PolyPhen2 predicted between 400 and 500. A clinician faced with such lists cannot act. MENDELSEEK was built to filter those false positives by asking a different question: not which mutations are harmful, but which genes themselves have the intrinsic properties of disease drivers.

The algorithm works by assembling a portrait of each gene from multiple angles. It calculates how damaging mutations would be across the entire protein using ENTPRISE scores for missense variants and ENTPRISE-X for frameshifts and stop codons. It maps which of 2,363 Reactome pathways the gene participates in and which of 12,535 Gene Ontology biological processes it engages. It incorporates a protein language model—a deep learning representation of amino acid sequences trained on vast protein databases. These features combine into a 15,926-dimensional vector per gene, which the algorithm then feeds into an Extreme Gradient Boosting machine learning model trained on 4,823 known Mendelian genes from the OMIM database.

When the researchers tested MENDELSEEK on a harder validation set—genes that excluded training data from competing methods to avoid bias—the algorithm still dominated. Its precision-recall score reached 0.489, compared to 0.334 for ENTPRISE and 0.225 for REVEL, a meta-predictor combining 13 different methods. The enrichment factor for the top 180 ranked predictions was 4.31, meaning those genes were 4.31 times more likely to be true Mendelian genes than random selection would predict. AlphaMissense, despite its recent development and deep learning sophistication, ranked fourth.

The analysis revealed something unexpected about Mendelian genes themselves. Genes that cause single-gene diseases tend to have more protein-protein interactions than genes associated with complex, polygenic disorders. They are also evolutionarily ancient—genes that emerged near the origin of life itself. The top Gene Ontology processes associated with Mendelian genes involve gene expression and transcription, processes essential to all living cells. The top pathways are metabolism, signal transduction, and immune function. When these ancient, central processes malfunction, the consequences are severe and visible.

Using MENDELSEEK, the researchers predicted 1,277 novel candidate Mendelian genes not yet catalogued in OMIM. To validate these predictions, they deployed Valsci, a literature-mining tool that searches scholarly databases for evidence linking genes to disease. Of the 1,277 predictions, 108 genes had strong literature support. Cross-referencing with the DECIPHER database of developmental disease genes added another 465 overlaps, bringing the validated set to 530 genes—41.5 percent of predictions. Examples include ITGB1, which encodes integrin beta 1 and is involved in kidney disease; ND6, linked to mitochondrial disease and Leber's hereditary optic neuropathy; and RIMS1, documented as causing cone-rod dystrophies. The remaining predictions remain unvalidated but catalogued for future investigation.

The work also illuminated a spectrum of genetic disease. Some genes drive only Mendelian disorders. Others drive both Mendelian and complex diseases. Still others contribute only to polygenic conditions. Genes that bridge Mendelian and complex disease tend to have more protein interactions, participate in more pathways, and express at higher levels than purely Mendelian genes. This suggests they evolved later and gained complexity over time. For clinicians and researchers, MENDELSEEK offers immediate utility: a way to rank candidates and prioritize investigation. For the field, it offers a map of what makes a gene a disease driver—knowledge that may eventually help predict which genes matter most when the next patient arrives with hundreds of mutations and no diagnosis.