Conclusions

This Stage 3 analysis evaluated a computational prioritization index, not a validated patient-level clinical risk predictor, by examining the sensitivity of an integrated Ki-DDD-TWAS antipsychotic metabolic-risk framework to literature-derived receptor priors in ancestry-defined LDL datasets. The main finding was a clear separation between the stability of drug ordering and the stability of mechanistic attribution. Drug ordering remained highly stable across ancestry-defined datasets in both models, although uniform weighting reduced the stability of membership in the highest-ranked set. Clozapine remained the highest-ranked drug across all ancestry-defined datasets, and olanzapine remained in the upper tier of the weighted model. At the receptor level, removing the literature priors produced a substantial and directionally consistent redistribution of contribution away from histaminergic and serotonergic systems and toward dopaminergic receptors. DRD2 remained the leading contributor under uniform weighting across all ancestry-defined datasets. HRH1 was the most ranking-sensitive gene in the weighted leave-one-gene-out analysis.

These findings support the interpretation that receptor priors stabilize the clinical interpretability of the model while also shaping its mechanistic conclusions. The uniform model is valuable for sensitivity analysis and hypothesis generation, but its clinically discordant movements should prevent its use as a standalone treatment-ranking tool. ADRB1 is a plausible candidate for further research because of its strong LDL TWAS signal and low literature weight, but it lacked support across the ancestry datasets and should not be treated as a confirmed causal mechanism. Because the analyses shared pharmacological inputs across ancestries and lacked patient-level clinical validation, the findings describe model behavior rather than clinical risk. We recommend dual reporting: the literature-weighted model as the clinically informed reference and the uniform model as a sensitivity analysis of receptor-prior assumptions. The findings demonstrate sensitivity of mechanistic attribution to receptor-prior assumptions while drug ordering remains relatively stable; they do not establish clinical risk, causality, or treatment-ranking validity.