VEIL: An Agentic Framework for Target Prioritisation from In-House Knowledge Bases

Companies and academic consortia have invested heavily in knowledge bases for drug target discovery, yet turning that evidence into rankings remains subjective, hard to reproduce, and prone to bias. We present VEIL (Veiled Evidence Interpretation Loop), an agent-based framework that organises evidence into separate parameters, each one line of evidence with a single hypothesis, interpreted independently by a specialised agent. The agents then reconcile their judgements, a reviewer checks the reasoning, and all decisions are fixed before scoring. Applied to a molecular model of Chronic Kidney Disease (CKD) with 921 genes and 51 evidence columns, VEIL recovered known drug targets at a rate above chance, a sign it picks up real biological signal rather than noise. Compared with ranking genes by literature attention, whether measured globally or specifically for CKD, VEIL prioritised substantially different, understudied candidates, relying less on literature-driven bias. Across three runs, agents read the evidence the same way and produced similar prioritisation logic, though the exact weighting varied. So today's agents seem more consistent in interpreting what the evidence means than at judging how much it should count.