MODIFINDER: An Evidence-First Agent for Genetic Modifier Verification and Hypothesis Construction

Genetic modifiers can explain why the same pathogenic background produces differences in disease severity, penetrance, or treatment response, and may point to therapeutic targets. Yet identifying modifier relationships in the literature is difficult: relevant evidence is often distributed across sources, curated resources are incomplete, and gene co-mentions do not by themselves establish a modifier relationship. We present MODIFINDER, a two-layer evidence-first agent for literature-based genetic-modifier hypothesis construction. Layer1 starts from the abstract, retrieves additional source context only when needed, and verifies extracted direct relations against their supporting text. Verified direct records enter a source-grounded evidence graph. Separately, role resolution extracts typed (A)-(B) bridges from the same registered source set. Layer2 uses these intermediates to search independent literature for candidate genes and constructs inspectable hypothesis cards that preserve the evidence and provenance for both links. On the strict sensitivity subset, conditional source acquisition achieved reference-conditioned (F_1=.470), compared with .348 for full-document one-shot processing, while using 72.3\% fewer source characters. Conditional acquisition required additional model calls. Compared with abstract-only processing, recall increased from .294 to .395 with similar reference-conditioned precision. Across 50 query contexts, MODIFINDER constructed 71 source-valid indirect paths, and follow-up evidence was collected for 66 (93.0\%). Context-stratified analysis identified 21 paths with matched or partially matched context evidence, including 6 strict matches. A blinded pilot audit by a domain expert found that both evidence legs could be supported even when differences in biological context prevented modifier-level conclusions. By retaining the supporting literature for each leg, the cards make this gap explicit and identify what must be resolved in follow-up.