When Literature Priors Are Removed: Clinical Rank Shifts and Transancestral Stability in Antipsychotic Low-Density...

Antipsychotic-associated metabolic abnormalities vary considerably between drugs, but the receptor mechanisms contributing to lipid-related liability remain incompletely resolved. The integrated Ki-defined daily dose-transcriptome-wide association study (Ki-DDD-TWAS) score is a computational prioritization index derived from receptor-binding affinities, defined daily doses, and ancestry-stratified transcriptome-wide association study (TWAS) summary statistics; it is not a validated patient-level clinical risk predictor. We evaluated the sensitivity of an integrated Ki-DDD-TWAS prioritization framework to the removal of literature-derived receptor weights in ancestry-stratified low-density lipoprotein cholesterol (LDL) analyses. Saved outputs were available for 52 antipsychotics and five ancestry-defined LDL datasets: African, East Asian, European, Hispanic, and South Asian groups. Two models were compared: a literature-weighted model incorporating predefined metabolic receptor priors and a uniform model assigning a weight of 1.0 to all recognized receptor genes. Stage 3 analyses reconstructed receptor-level contributions and examined gene-direction consistency, ancestry-specific drivers, drug-rank correlations, rank stability, top-10 Jaccard similarity, leave-one-gene-out sensitivity, high-TWAS/low-weight candidates, and weighted-versus-uniform mean-rank differences. Removing literature priors produced a large shift in relative reconstructed model contribution from histaminergic and serotonergic systems toward dopaminergic receptors. Dopaminergic contribution increased by 26.41 percentage points, whereas histaminergic contribution decreased by 16.76 percentage points. All 39 genes represented in the exported direction-consistency analysis changed in the same direction across all five ancestry datasets. Dopamine Receptor D2 (DRD2) was the leading gene-level contributor under the uniform model in every ancestry. Cross-ancestry drug-rank correlations were very high in both models, with mean off-diagonal Spearman correlations of 0.996 for the weighted model and 0.995 for the uniform model. The weighted model had complete top-10 set identity across ancestry pairs, whereas the mean top-10 Jaccard similarity under uniform weighting was 0.758. ADRB1 was the strongest high-TWAS/low-weight candidate, with a maximum absolute z-score of 6.079, but it met the candidate threshold in only one ancestry. These findings indicate that literature-derived priors strongly shape model-based mechanistic attribution while exerting a smaller influence on broad drug ordering. The weighted model is best regarded as a clinically informed reference model, whereas the uniform model is a sensitivity and hypothesis-generation analysis. Because the five ancestry runs shared the same Ki matrix, defined daily dose (DDD) values, receptor map, and prior dictionary and lacked patient-level LDL validation, these findings describe computational model behavior rather than ancestry-specific clinical risk. Neither model is a validated clinical risk predictor, nor does either provide evidence of causal receptor mechanisms.