medRxivSource publication:
Loading content…
Rare-variant studies now yield more significant gene-phenotype associations than teams can place against prior evidence, and that placement has become the tighter bottleneck. PESTO combines large language model literature curation with structured Open Targets genetic evidence in one agentic pipeline that assigns Novel, Hypothesized, Existing or Established with supporting citations. Across 940 unique associations from two rare-variant studies, 7% were called Novel and 21% Hypothesized; 86% of 50 randomly paired negative controls fell into those two tiers.
Recently reported gene-phenotype associations can now be graded by how much prior evidence stands behind them, rather than only retrieved or not retrieved. Expert-curated resources such as ClinGen, OMIM and Open Targets cover only part of the association space, and absence from them does not establish novelty; synonyms, renamed disease nosology and related but distinct traits can all hide evidence. Across 940 unique gene-phenotype pairs from two rare-variant studies, the combined branches left 68 (7%) Novel and 194 (21%) Hypothesized, while 43 of 50 (86%) randomly recombined negative controls fell into those two tiers (Fisher's exact test, P = 1.3 × 10⁻¹⁶).
Verdicts come from two independent branches, one reading the literature and one querying a structured database, with the stronger tier retained. The literature branch expands gene and phenotype terms, searches PubMed, sorts records into relevance tiers, and has a model allocate confidence weights across the four tiers; the Open Targets branch retrieves traits linked to the gene, shortlists the 20 nearest to the queried phenotype by semantic similarity, and grades each. The two branches agree exactly on 41% of associations and fall within one tier on 83%; against GenCC expert curation, 96 of 100 Definitive pairs were called Established and 83 of 100 Absent controls were called Novel.
The assessment covers the state of prior evidence, not the validity of a newly reported association, so associations called Novel may still be false positives that need independent statistical and biological evaluation. The intended setting is significant gene-phenotype pairs from rare-variant and biobank studies, and each output carries a verdict, confidence weights and supporting PMIDs for pair-by-pair manual review.
The method does not currently assess concordance in mechanism of action or inheritance pattern between prior and new findings, and it does not integrate ClinGen, OMIM or evidence reported only in supplementary tables, which an optional full-text mode partly addresses. Uneven research attention across genes and diseases means evidence landscapes may reflect historical ascertainment and publication bias as well as disease biology.