Summary
A bioRxiv preprint describes PRIORITI, a framework that combines language-model synthesis of human genetic evidence with supervised learning to forecast whether a proposed drug target–disease pairing will succeed. The authors report results on historical and post-cutoff benchmark sets.
A bioRxiv preprint from researchers at the University of Texas MD Anderson Cancer Center and collaborating institutions describes PRIORITI, a framework for estimating whether a proposed drug target–disease pairing may translate into clinical success. The authors report that it performed well on both a held-out historical test set and a separate benchmark of target–indication pairs from after the training period.
The framework treats the proposed target gene and disease indication as the starting point for a forecast. Its reported results are retrospective model evaluations, rather than outcomes from a prospective test in drug-development programmes.
How PRIORITI builds its forecasts
PRIORITI—short for Prospective Rationale-Informed Outcome Reasoning for Integrated Target-indication Intelligence—uses a domain-instructed large language model to synthesise drug-agnostic human genetic evidence across prespecified channels. In other words, the evidence synthesis concerns the gene–disease relationship, not a particular drug designed to act on the target.
The system converts that evidence into an embedded representation and combines it with static information drawn from gene summaries and disease definitions. A supervised model then uses this representation to predict target–indication outcomes. The authors describe the separation between evidence synthesis and outcome prediction as a way to make the forecasts leakage-resistant.
A separate, label-blinded language-model explainer turns locked predictions into prioritisation rationales. These rationales are intended to make the basis for ranking a proposed target–disease pairing easier to inspect; the probability estimate and its explanation are produced in distinct steps.
Results on historical and later benchmarks
The supervised model was trained on 6,784 temporally curated historical target–indication outcomes. On a held-out historical test set of 1,696 pairs, the authors report a ROC-AUC of 0.883 and an expected calibration error (ECE) of 0.023. ROC-AUC measures how well a model ranks positive cases above negative ones, while ECE assesses how closely its predicted probabilities correspond to observed frequencies. The authors say PRIORITI outperformed a baseline using static target and indication context, as well as a human-curated genetic-evidence reference.
The researchers also evaluated the framework on 765 target–indication pairs from strictly after the historical cutoff. On this out-of-time benchmark, they report a ROC-AUC of 0.817, a precision–recall AUC of 0.357 and an ECE of 0.044. They say this performance exceeded that of the static-context baseline and direct zero-shot language-model forecasting. The authors describe the 765-pair evaluation as the largest out-of-time cohort reported for this task.
These results suggest that structured synthesis of genetic evidence can help rank target–disease hypotheses across historical and later datasets. The evaluation is nevertheless retrospective: it does not show whether using PRIORITI changes real-world target-selection decisions or improves the success of drug-development programmes. The abstract also does not specify how the historical outcome labels define clinical success, a detail that matters when interpreting what the forecast predicts.