V2G agent attaches an evidence trail to every GWAS locus, yet stripping the evidence lifts accuracy to 0.864

V2G turns variant-to-gene assignment from a single compressed score into an evidence trail that can be checked locus by locus: fine-mapping, expression QTLs, colocalization, Mendelian randomization and population allele frequencies are each issued as a separate recorded query, and a language model then ranks the assembled evidence and states what it relied on. Integrative scores compress regulatory and molecular evidence into one number, so a user sees a fixed model's feature contributions rather than a choice that can be checked against retrieved evidence, and the score does not say where the method declined to answer. On 135 held-out loci where retrieval succeeded naturally and Open Targets also predicts, the five-seed GPT-5.2 consensus reaches Precision@1 of 0.793 (107/135), above the evidence ordering it receives at 0.711 (11 loci to 0, P = 9.8 × 10⁻⁴) and nearest-gene assignment at 0.615 (35 to 11, P = 5.4 × 10⁻⁴).

The model rarely overrides the given order, its overrides cluster at a few loci, and most of them land on the reference gene. Across 1,480 locus-seed decisions (296 loci, five seeds each) the model keeps the given top candidate at 86%; on loci where the reference gene was retrieved naturally, 120 of its 142 departures select a reference gene, covering 31 of 37 loci. No departure displaced a top candidate that was already correct; across the 117 loci whose initial top candidate was correct, zero events are compatible with a displacement rate of up to 2.5% per locus (one-sided 95% bound).

Withholding the evidence and showing only the trait and shuffled gene symbols makes the same model more accurate, so it is drawing on recall of gene-trait associations. The full prompt scores 0.778–0.807 across three seeds, the names-only prompt 0.864 (0.859–0.867), and on the 97 primary loci with no gene-EFO overlap with training the two score 0.832 and 0.794. On the 114 loci where the reference gene was never retrieved and had to be force-added, the names-only prompt selects it at 0.594 against 0.076 for the full prompt; moving each gene symbol onto another gene's evidence row makes the model follow the top-scored first row and drops primary Precision@1 to 0.284.

The demonstrated scope is 296 held-out loci from a benchmark of 844 fine-mapped loci, with the primary analysis restricted to the 135 natural loci where candidate retrieval succeeded and Open Targets returns a prediction; the pipeline is deployed as a container image inside the user's institution, summary statistics never leave it, the demonstration image carries reference-data slices for two loci, and a full deployment requires roughly 100 GB of reference data.

The train-test split only partly isolates associations: 87 test loci share a reference gene and exact trait name with training, the overlap-free subset is a post hoc sensitivity analysis, and all 88 replicate-cohort loci share a reference gene-EFO pair with training. Two language-model backbones were characterized and the result depends on which is used; no comparison against general biomedical agents was made, and the worked example is a familiar association. Sensitivity to the fine-mapping method and to colocalization inputs was not tested, colocalization was available at only 17 test loci in the frozen benchmark and was computed from incomplete regional inputs, and the phenotype comparison stops at 18 organ-system categories drawn from literature a model may have seen. The authors report no independent reproduction.