Results

Data coverage and pipeline output characteristics

All five weighted and five uniform ancestry runs contained both long-format receptor-level data and drug-level risk files. The ancestry-specific TWAS directories contained between 17,620 and 18,057 genes. Each run included 70 drugs in the initial list, nine drugs dropped for missing defined daily dose, 58 matched ligands, and 1,285 drug-receptor pairs (Table 1).

Mode | Ancestry | Genes with TWAS data | Drugs in input list | Drugs dropped for no DDD | Drugs matched to ligand | Drug × receptor pairs | Ki values imputed | TWAS values imputed | Strong binders | TWAS-significant rows | Receptor genes without TWAS
Weighted | LDL_AFR | 18,016 | 70 | 9 | 58 | 1,285 | 0 | 210 | 247 | 111 | 9
Weighted | LDL_EAS | 17,620 | 70 | 9 | 58 | 1,285 | 0 | 239 | 247 | 97 | 10
Weighted | LDL_EUR | 18,057 | 70 | 9 | 58 | 1,285 | 0 | 210 | 247 | 150 | 9
Weighted | LDL_HIS | 18,017 | 70 | 9 | 58 | 1,285 | 0 | 210 | 247 | 147 | 9
Weighted | LDL_SAS | 17,936 | 70 | 9 | 58 | 1,285 | 0 | 210 | 247 | 175 | 9
Uniform | LDL_AFR | 18,016 | 70 | 9 | 58 | 1,285 | 0 | 210 | 247 | 111 | 9
Uniform | LDL_EAS | 17,620 | 70 | 9 | 58 | 1,285 | 0 | 239 | 247 | 97 | 10
Uniform | LDL_EUR | 18,057 | 70 | 9 | 58 | 1,285 | 0 | 210 | 247 | 150 | 9
Uniform | LDL_HIS | 18,017 | 70 | 9 | 58 | 1,285 | 0 | 210 | 247 | 147 | 9
Uniform | LDL_SAS | 17,936 | 70 | 9 | 58 | 1,285 | 0 | 210 | 247 | 175 | 9

No rows were removed because the affinity-dose score was zero, and no Ki values were imputed in the reported runs. TWAS values were imputed for 210 rows in the African, European, Hispanic, and South Asian datasets and for 239 rows in the East Asian dataset. The number of nominally significant drug-receptor rows varied from 97 in the East Asian dataset to 175 in the South Asian dataset. Between nine and ten receptor genes lacked direct TWAS data in each ancestry run (Table 1).

These figures indicate that the planned computational outputs were available across all ancestry and weighting conditions. They are workflow-coverage metrics and should not be interpreted as evidence of clinical representativeness or biological completeness. In particular, missing TWAS values, deterministic receptor-to-gene mapping, and the restricted receptor set remain relevant to the interpretation of contribution shares.

Receptor-class and gene-level reallocation

The clearest difference between the two models was a redistribution of relative reconstructed model contribution across receptor classes. In the literature-weighted model, the pooled reconstructed contribution share of serotonergic receptors was 40.75%, histaminergic receptors 25.46%, dopaminergic receptors 17.61%, adrenergic receptors 12.91%, muscarinic receptors 3.14%, transporters 0.07%, and sigma receptors 0.06%. Under uniform weighting, the reconstructed dopaminergic share increased to 44.02%, representing a gain of 26.41 percentage points. Histaminergic contribution fell to 8.70%, a loss of 16.76 percentage points. Serotonergic contribution decreased to 32.70%, a loss of 8.05 percentage points. Adrenergic contribution decreased by 2.62 percentage points to 10.29%. Muscarinic, transporter, and sigma contributions changed only modestly (Table 2).

Level | Receptor/gene | Class | Weighted contribution (%) | Uniform contribution (%) | Delta or mean delta (percentage points) | Consistency
Class | Dopaminergic | Dopaminergic | 17.61 | 44.02 | +26.41 | —
Class | Histaminergic | Histaminergic | 25.46 | 8.70 | −16.76 | —
Class | Serotonergic | Serotonergic | 40.75 | 32.70 | −8.05 | —
Class | Adrenergic | Adrenergic | 12.91 | 10.29 | −2.62 | —
Class | Muscarinic | Muscarinic | 3.14 | 3.46 | +0.32 | —
Class | Transporter | Transporter | 0.07 | 0.44 | +0.37 | —
Class | Sigma | Sigma | 0.06 | 0.40 | +0.34 | —
Gene | DRD2 | Dopaminergic | — | — | +13.326 | 1.00
Gene | DRD4 | Dopaminergic | — | — | +5.894 | 1.00
Gene | DRD3 | Dopaminergic | — | — | +5.838 | 1.00
Gene | HTR2B | Serotonergic | — | — | +3.754 | 1.00
Gene | HTR1A | Serotonergic | — | — | +1.598 | 1.00
Gene | HTR7 | Serotonergic | — | — | +1.124 | 1.00
Gene | DRD1 | Dopaminergic | — | — | +1.026 | 1.00
Gene | ADRA1A | Adrenergic | — | — | −2.048 | 1.00
Gene | HRH1 | Histaminergic | — | — | −17.175 | 1.00
Gene | HTR2A | Serotonergic | — | — | −9.265 | 1.00
Gene | HTR2C | Serotonergic | — | — | −6.599 | 1.00
Gene | HTR6 | Serotonergic | — | — | −0.802 | 1.00
Gene | CHRM3 | Muscarinic | — | — | −0.724 | 1.00

The gene-level analysis showed that this was not only a class-level effect. At the model level, DRD2 gained 13.326 percentage points of relative contribution, while DRD4 and DRD3 gained 5.894 and 5.838 points, respectively. Other consistently gaining genes included HTR2B, HTR1A, HTR7, DRD1, HTR1D, HTR1B, HTR3A, CHRM2, ADRA1D, HRH2, SIGMAR1, and SLC6A2. The largest losses were observed for HRH1, which declined by 17.175 points, HTR2A, which declined by 9.265 points, and HTR2C, which declined by 6.599 points. ADRA1A, ADRA1B, HTR6, and CHRM3 also lost relative contribution (Table 2).

These percentages describe the allocation of the model’s reconstructed signal. They should not be interpreted as estimates of the percentage of clinical LDL risk mediated by a receptor or receptor class. The reallocation is mathematically expected to some extent because the weighted model gives much greater initial importance to HRH1 and HTR2C than to DRD2, DRD3, or DRD4. It is nevertheless informative because it identifies which model-based biological explanations become prominent when those priors are removed.

Transancestral directional consistency and dominant model contributors

All 39 genes represented in the exported gene-direction consistency file showed the same direction of model contribution change across the five ancestry-defined LDL datasets. DRD2, DRD3, and DRD4 gained relative contribution in every ancestry. HRH1, HTR2A, and HTR2C lost relative contribution in every ancestry. The same directional consistency was observed for the smaller changes affecting adrenergic, muscarinic, serotonergic, histaminergic, sigma, and transporter genes (Table 2).

Under uniform weighting, DRD2 was the leading gene by reconstructed contribution in all five ancestry datasets. DRD3 occupied second place in the African and European datasets, whereas HTR2A occupied second place in the East Asian, Hispanic, and South Asian datasets. HTR2A and DRD3 exchanged positions near the top of the ranking, but DRD2 remained the number-one contributor in every dataset. DRD4, HRH1, HTR7, ADRA1A, and HTR2B or HTR2C also appeared among the leading model contributors.

The perfect directional consistency and universal DRD2 leadership indicate that the broad reallocation pattern was not driven by one ancestry dataset. However, the five analyses shared the same Ki matrix, defined daily dose values, receptor-to-gene map, and prior-weight comparison. The consistency therefore reflects the combined influence of ancestry-specific TWAS information and a common pharmacological structure. It demonstrates reproducibility of model sensitivity rather than independent confirmation of causal receptor biology. These findings also do not constitute evidence of ancestry-specific clinical LDL risk.

Transancestral stability of drug rankings

Drug ordering was highly correlated across ancestry datasets in both weighting modes. The weighted model had a mean off-diagonal Spearman correlation of 0.996. Pairwise correlations ranged from 0.991 to 0.998. The uniform model had a mean off-diagonal Spearman correlation of 0.995, with pairwise values ranging from 0.993 to 0.997 (Table 3).

Mode | Mean off-diagonal Spearman rho | Pairwise rho range | Mean pairwise top-10 Jaccard | Jaccard range
Weighted | 0.996 | 0.991–0.998 | 1.000 | 1.000–1.000
Uniform | 0.995 | 0.993–0.997 | 0.758 | 0.667–0.818

The weighted model also showed complete top-10 set identity across every ancestry pair. Every pairwise weighted top-10 Jaccard value was 1.000. The uniform model retained substantial overlap but was less stable, with a mean pairwise top-10 Jaccard value of 0.758. Individual uniform-model Jaccard values ranged from 0.667 to 0.818. The reduction in top-10 overlap indicates that the literature priors stabilized membership of the highest-ranked set, even though the overall rank correlations remained very high without them.

The weighted model produced perfect rank stability across the five datasets for several drugs, including clozapine, olanzapine, aripiprazole, brexpiprazole, iloperidone, cariprazine, chlorpromazine, melperone, levosulpride, triflupromazine, sultopride, and thioridazine. Sulpiride and amisulpride were among the least stable weighted-model drugs, each with a rank standard deviation of 2.17. Under uniform weighting, clozapine, aripiprazole, fluphenazine, perphenazine, levosulpiride, zotepine, and thioridazine had standard deviations of zero. Sulpiride had the largest uniform-model standard deviation, at 2.74, followed by chlorprothixene at 2.51 (Table 4).

Drug | Weighted mean rank | Weighted rank SD | Uniform mean rank | Uniform rank SD
Clozapine | 1.00 | 0.00 | 1.00 | 0.00
Aripiprazole | 20.00 | 0.00 | 14.00 | 0.00
Fluphenazine | 13.80 | 0.84 | 15.00 | 0.00
Perphenazine | 13.80 | 1.30 | 16.00 | 0.00
Zotepine | 4.40 | 1.14 | 6.00 | 0.00
Thioridazine | 8.00 | 0.00 | 7.00 | 0.00
Ziprasidone | 5.60 | 0.55 | 2.40 | 0.55
Olanzapine | 7.00 | 0.00 | 12.00 | 1.41
Quetiapine | 14.80 | 0.84 | 25.20 | 1.48
Sulpiride | 31.20 | 2.17 | 22.00 | 2.74
Chlorprothixene | 9.20 | 0.45 | 10.40 | 2.51

The delta-rank analysis provided a related measure of transancestral consistency. Across 52 common drugs, pairwise Spearman correlations of the uniform-minus-weighted rank-change profiles ranged from 0.931 to 0.976. All pairwise comparisons were associated with very small p-values. Thus, ancestry-specific LDL inputs produced not only similar absolute rankings but also similar responses to removal of the literature priors.

Clinical context and rank movements under prior removal

The weighted rankings were qualitatively concordant with selected clinical patterns at the highest-liability end, but this was not an external clinical validation. Clozapine was ranked first in all five ancestry datasets, with a mean rank of 1.00 and a rank standard deviation of zero. Olanzapine had a mean weighted rank of 7.00 and also showed no ancestry-related rank variation. Aripiprazole had a mean rank of 20.00 and a standard deviation of zero. Ziprasidone had a mean weighted rank of 5.60 and a standard deviation of 0.55. The model therefore preserved the prominence of clozapine and placed olanzapine in the upper tier, while aripiprazole was lower in the predicted-risk ordering. Ziprasidone was more intermediate than its generally favorable clinical and metabolic profile might suggest.

Because rank 1 denotes the highest predicted risk, the sign of the rank change requires careful interpretation. A negative value for uniform rank minus weighted rank means that the drug moved toward a numerically smaller rank and therefore toward a higher predicted-risk position under uniform weighting. A positive value means that the drug moved toward a numerically larger rank and therefore toward a lower predicted-risk position.

Several drugs with relatively favorable clinical metabolic profiles moved toward numerically smaller ranks under uniform weighting. Ziprasidone changed from a mean weighted rank of 5.60 to a mean uniform rank of 2.40. Aripiprazole changed from 20.00 to 14.00, and haloperidol changed from 17.80 to 9.60. Pimozide, cariprazine, amisulpride, and several dopamine-focused older agents also moved toward higher predicted-risk positions under the uniform model. These movements should not be described as clinical improvement; they represent increased model-predicted risk after the priors were removed (Table 5).

Movement group | Drug | Mean delta rank, uniform − weighted
Biggest risers | Levomepromazine | +20.00
Biggest risers | Acepromazine | +16.80
Biggest risers | Promazine | +16.80
Biggest risers | Loxapine | +11.80
Biggest risers | Quetiapine | +10.40
Biggest risers | Periciazine | +8.40
Biggest risers | Clotiapine | +7.40
Biggest risers | Paliperidone | +7.00
Biggest risers | Pipamperone | +6.40
Biggest risers | Mesoridazine | +6.00
Biggest fallers | Amisulpride | −9.40
Biggest fallers | Cariprazine | −9.40
Biggest fallers | Thioproperazine | −9.20
Biggest fallers | Trifluperidol | −9.20
Biggest fallers | Sulpiride | −9.20
Biggest fallers | Benperidol | −9.00
Biggest fallers | Haloperidol | −8.20
Biggest fallers | Moperone | −8.00
Biggest fallers | Pipotiazine | −8.00
Biggest fallers | Sultopride | −6.80

Other drugs moved toward numerically larger ranks under uniform weighting. Olanzapine changed from a mean weighted rank of 7.00 to 12.00, quetiapine from 14.80 to 25.20, levomepromazine from 18.00 to 38.00, loxapine from 18.20 to 30.00, promazine from 23.00 to 39.80, and acepromazine from 25.40 to 42.20. The largest positive mean shifts were observed for levomepromazine, acepromazine, promazine, loxapine, quetiapine, periciazine, clotiapine, paliperidone, pipamperone, and mesoridazine. The largest negative shifts were observed for amisulpride, cariprazine, thioproperazine, trifluperidol, sulpiride, benperidol, haloperidol, moperone, pipotiazine, and sultopride (Table 5).

These movements suggest that the weighted model’s greater emphasis on HRH1 and HTR2C helps preserve the relative prominence of drugs with greater reported average adverse effects on weight, glucose, or selected lipid measures, although the model is not a full clinical ranking. Conversely, uniform weighting makes dopaminergic binding more influential and produces several movements that are discordant with the commonly described clinical metabolic profiles of aripiprazole, ziprasidone, and haloperidol. This discordance is informative as a sensitivity result but does not establish that the uniform model is biologically more accurate.

Leave-one-gene-out sensitivity

Removing individual genes from the reconstructed weighted contributions identified HRH1 as the most ranking-sensitive component in the reconstructed weighted model. The mean absolute rank shift across drugs and ancestry datasets was 4.58 positions for HRH1. The corresponding values were 3.32 for HTR2A, 2.14 for DRD2, 1.76 for HTR7, 1.55 for HTR2C, 1.44 for ADRA1A, 1.31 for HTR6, and 1.17 for CHRM3 (Table 6).

Gene | Mean absolute rank change
HRH1 | 4.58
HTR2A | 3.32
DRD2 | 2.14
HTR7 | 1.76
HTR2C | 1.55
ADRA1A | 1.44
HTR6 | 1.31
CHRM3 | 1.17

In the European dataset, removing HRH1 moved clotiapine and paliperidone eight positions toward numerically smaller or higher predicted-risk ranks. Moperone and lurasidone also shifted by eight positions. Periciazine, benperidol, trifluoperazine, and pipamperone shifted by seven positions, while pimozide and pipotiazine shifted by six and five positions, respectively.

These results identify dependence of the reconstructed ranking on particular receptor components. HRH1 was the strongest ranking-sensitive component in this analysis, which is consistent with the high weight assigned to HRH1 and with the established pharmacological association between H1 receptor affinity and antipsychotic-associated weight gain [9]. The leave-one-gene-out analysis should nevertheless be interpreted as a model-dependence analysis, not as evidence that HRH1 independently causes LDL abnormalities. In addition, the reported reconstruction removed receptor-level contributions but did not fully recompute every drug-level multiplier used in the saved final score. The results are therefore best described as sensitivity of the reconstructed contribution-based ranking.

High-TWAS/low-weight candidates and dual reporting

Under the strict downstream screen of an absolute z-score of at least 3.0 and literature weight no greater than 0.15, ADRB1 and SLC6A3 were identified. These were exploratory, single-ancestry signals rather than replicated candidate genes. ADRB1 had a maximum absolute z-score of 6.079, a literature weight of 0.08, and appeared in 20 drug-level records. SLC6A3 had a maximum absolute z-score of 3.384, a literature weight of 0.04, and appeared in 29 drug-level records (Table 7).

Gene | Screen | Maximum absolute TWAS z | Literature weight | Number of drugs | Number of ancestries
ADRB1 | Strict and exploratory | 6.079 | 0.08 | 20 | 1
SLC6A3 | Strict and exploratory | 3.384 | 0.04 | 29 | 1
SLC6A4 | Exploratory | 2.902 | 0.06 | 9 | 1
DRD4 | Exploratory | 2.698 | 0.10 | 39 | 1
HTR2B | Exploratory | 2.696 | 0.00 | 17 | 1

The more permissive Stage 3 threshold also identified SLC6A4, DRD4, and HTR2B. Each candidate was supported in only one ancestry dataset. HTR2B had no positive value in the supplied literature-weight dictionary, so its uniform-model contribution reflects restoration of a gene that was not active in the weighted model. The absence of a candidate supported across multiple ancestry datasets contrasts with the strong transancestral consistency of the broad receptor-reallocation pattern.

The dual-reporting analysis showed that clozapine remained ranked first under both models. Other drugs displayed more substantial changes. Ziprasidone moved from a weighted mean rank of 5.60 to 2.40 under uniform weighting, whereas olanzapine moved from 7.00 to 12.00 and quetiapine from 14.80 to 25.20. The dual-reporting result therefore separates two uses of the framework. The weighted model offers the more clinically interpretable reference ordering because it preserves established receptor knowledge, although this represents qualitative concordance rather than external validation. The uniform model reveals how the ranking and mechanistic explanation change when those priors are removed (Table 8).

Drug | Weighted mean rank | Uniform mean rank | Rank shift, uniform − weighted
Clozapine | 1.00 | 1.00 | 0.00
Ziprasidone | 5.60 | 2.40 | −3.20
Olanzapine | 7.00 | 12.00 | +5.00
Quetiapine | 14.80 | 25.20 | +10.40
Aripiprazole | 20.00 | 14.00 | −6.00
Haloperidol | 17.80 | 9.60 | −8.20
Levomepromazine | 18.00 | 38.00 | +20.00

Additional detailed outputs, including full ancestry-specific drug ranks, complete receptor- and gene-level contribution tables, pairwise rank-correlation and Jaccard matrices, leave-one-gene-out results, and the full high-TWAS/low-weight candidate listings, are provided in Appendix 1.