1. Introduction

Bipolar disorder (BD) is a severe and recurrent psychiatric disorder characterized by episodes of mania, hypomania, and depression. Its clinical presentation is heterogeneous, and early differentiation from major depressive disorder (MDD) remains a persistent diagnostic challenge. Accurate phenotyping is therefore essential for genomic and pharmacogenomic studies because diagnostic misclassification can dilute association signals and distort treatment-response interpretation. Recent work on electronic health records, natural language processing, and clinical decision support systems has emphasized the need for validated diagnostic labels when studying BD and its differentiation from MDD

[1]

Genetic studies have consistently demonstrated that BD is highly heritable and polygenic. Early genome-wide association studies implicated loci such as ANK3, CACNA1C, and ODZ4/TENM4, supporting the involvement of neuronal excitability, calcium-channel biology, and neurodevelopmental mechanisms

[2][3][4][5][6][7][8]

In parallel, rare-variant studies have added another layer of biological interpretation. Exome sequencing has identified AKAP11 as a risk gene shared between BD and schizophrenia, with functional relevance to GSK3B, a major signaling node implicated in lithium biology

[9][10][11]

Despite this progress, the translation of genomic findings into psychopharmacological mechanisms remains limited. Current BD treatments include lithium, anticonvulsant mood stabilizers, atypical antipsychotics, and antidepressants in selected clinical contexts

[12][13]

The aim was to compare evidence-stratified BD locus-to-gene mappings with drug-associated genes for medications used in BD. We evaluated exact gene-level overlaps, predefined mechanistic bridges, and pathway patterns while separating direct pharmacodynamic targets, pharmacokinetic genes, pharmacogenomic markers, and intracellular pathway nodes.

2. Research methods and principles

This integrative study used genomic data and curated drug-associated evidence. Following methodological review, the primary workflow was redefined to separate heterogeneous locus-to-gene mappings and pharmacological evidence types. The initial workflow and the rationale for full reanalysis are described in Section 2.6.1. Primary analyses used the strict BD set and quantitative Tier 1 targets; broader mappings and other drug-gene relationships served as sensitivity layers.

BD-associated genes were curated from major GWAS, trans-ancestry GWAS, and exome sequencing studies. Principal GWAS sources were Stahl et al. (2019), Mullins et al. (2021), O'Connell et al. (2025), and Zhang et al. (2026)

[5][6][7][8][9][2][3][4][14][15][16]

Each source-level locus-to-gene record was assigned one or more explicit mapping methods: rare-variant exome evidence, a fine-mapped credible variant linked by VEP or ABC, formal TWAS/eQTL colocalization, significant TWAS, FOCUS transcriptome fine-mapping, isoTWAS, SMR with HEIDI support, MAGMA gene-based association, or proximity/manual locus assignment. All supporting methods were retained in an evidence matrix rather than reduced to the nearest gene alone. The complete mapping rules, evidence matrix, set membership, symbol-resolution checks, and overlap sensitivity results are provided in Supplementary Table S20.

Three nested BD gene sets were defined:

· strict mapping set: exome evidence, formal colocalization, or an O'Connell fine-mapped credible variant linked by VEP/ABC with at least three supporting evidence items (n = 80 symbols);

· functionally expanded set: the strict set plus significant TWAS, FOCUS posterior inclusion probability ≥ 0.8, isoTWAS, significant SMR with HEIDI support, MAGMA, VEP, or ABC evidence (n = 307);

· broad-locus set: the functionally expanded set plus proximity-to-credible-variant and manually retained historical or current locus assignments (n = 331).

Before method-specific mapping was introduced, BD genes were assigned to four interpretive categories: 39 core genes, 56 high-confidence functional genes, 3 extended historical genes, and one predefined bridge node. The 39 core and 56 high-confidence genes formed the 95-gene comparator; the three extended genes and GSK3B bridge were retained for interpretation rather than included in that set. Because this source-level classification combined different strengths and mechanisms of evidence without a uniform gene-level mapping rule, it was replaced for primary inference by the explicit method-specific sets defined above. The complete initial workflow and rationale for reanalysis are described in Section 2.6.1. Predefined pharmacological bridge nodes were not added to the revised BD sets. Official HGNC symbols were used where available; unmapped or retired transcript identifiers were retained in quality-control records but excluded from term-level interpretation when not recognized by g:Profiler.

The primary drug list was based on medications used in BD treatment according to major clinical guidance

[12][17][13][18][19]

Drug-associated information was initially extracted from DrugBank 6.0 / DrugBank Online records retrieved and archived locally by the authors on 10 June 2026

[20]

Database annotations were interpreted with reference to ChEMBL, DGIdb, the IUPHAR/BPS Guide to Pharmacology, and the Therapeutic Target Database

[21][22][23][24][25][26][22][24][27][28][18][29][30]

For each drug-gene relationship, the extracted fields included drug, HGNC symbol, interaction type and direction, evidence source, quantitative affinity or potency, assay context, CNS penetration, therapeutic exposure, and human target-engagement evidence when available. Tier 1 required a direct gene-resolved interaction with a quantitative Ki, Kd, IC50, or EC50 value. Tier 2 comprised experimentally supported mechanistic nodes without a qualifying quantitative direct-interaction record. Tier 3 comprised pharmacokinetic enzymes and transporters. PGx association markers were kept separate and were not treated as targets. Quantitative potency and CNS engagement were recorded as contextual evidence and did not by themselves determine tier membership. Contested, broad protein-group, or insufficiently gene-specific relationships were excluded from the primary target set. The complete adjudicated evidence tiers, quantitative Tier 1 records, secondary mechanistic records, pharmacokinetic genes, and target-engagement audit are provided in Supplementary Table S21.

The primary direct-overlap analysis matched HGNC symbols between the strict BD mapping set and quantitative Tier 1 targets. A direct overlap was defined as an identical approved symbol in both sets. Overlap with the functionally expanded and broad-locus BD sets and with Tier 2, Tier 3, and PGx layers was evaluated secondarily.

Predefined bridge relations were analyzed separately from exact overlap. GSK3B met the quantitative Tier 1 rule through direct lithium inhibition, but it was not present in any BD mapping set and had been selected a priori because of the AKAP11-GSK3B-lithium hypothesis. It was therefore retained as a predefined bridge rather than counted as an independently discovered overlap.

An illustrative one-sided hypergeometric calculation contextualized exact overlap using the protein-coding subsets of the strict BD and Tier 1 lists and an HGNC protein-coding universe of 19,296 approved genes downloaded on 22 July 2026. This broad background was not matched for druggability or annotation density and was not treated as a definitive convergence test. Complete overlap calculations based on the method-specific BD sets and quantitative Tier 1 targets are provided in Supplementary Table S22.

The initially submitted workflow used the source-level BD classification described in Section 2.2. Exact overlap was evaluated between the 95-gene combination of 39 core and 56 high-confidence functional genes and 21 unique direct pharmacodynamic targets. Enrichr

[31][32]

A subsequent g:Profiler sensitivity stage tested direct pharmacodynamic targets (n = 21), PK/PGx genes (n = 6), mechanistic nodes (n = 9), drug-associated genes after excluding antidepressants (n = 29), the same set after additionally excluding CYP genes (n = 27), and two combined BD-drug sets (n = 115 and n = 121), using the annotated-domain background and g:SCS correction within each query. The methodological review identified two principal limitations: the 95-gene BD set did not record a uniform mapping rule for every gene, and the 35-gene drug set combined direct pharmacodynamic targets, mechanistic nodes, pharmacokinetic genes, and association markers. These features could affect gene attribution and make pathway signals dependent on heterogeneous annotation classes. Therefore, exact-overlap and enrichment analyses were fully recomputed using the strict, functionally expanded, and broad-locus BD sets and the quantitative drug-evidence tiers. Supplementary Table S12 and Tables S16–S19 preserve the initial sensitivity and enrichment outputs as an analytical record; they were not used for primary inference. Tables S20-S24 contain the method-specific mapping, evidence tiers, recomputed primary results, candidate assessment, and figure source data.

The revised BD queries were the strict mapping set (80 input symbols; 68 recognized), functionally expanded set (307; 249 recognized), and broad-locus set (331; 273 recognized). The primary drug query was Tier 1 (22 input and recognized genes); secondary queries were Tier 2 (n = 2), Tier 3 PK (n = 5), the separate PGx marker layer (n = 1), and Tier 1 plus Tier 2 (n = 24). Primary pathway enrichment was recomputed on 22 July 2026 using Gene Ontology, KEGG, and Reactome

[33][34][35][36]

Representative heatmap terms were selected to show the strongest nonredundant layer-specific signals and the calcium, monoaminergic, and xenobiotic themes defined for this analysis. Grey cells denote terms that did not meet g:SCS-adjusted p < 0.05, and white cells denote terms not returned for the corresponding query. No formal semantic pruning of redundant GO terms was performed. Enrichr adjusted p-values from the initial analysis and g:Profiler g:SCS-adjusted p-values from the revised analysis were not compared directly because the services use different correction procedures and gene-domain definitions.

Candidates were ranked with a transparent rule-based 0-11 score comprising BD genetic evidence (0-3), drug evidence (0-3), cross-layer relation (0-4), and identified human CNS target engagement (0-1). Bands denoted exact overlap (A), a predefined bridge (B), or thematic/one-sided evidence (C). The score is not a probability, clinical score, or treatment-response biomarker. The Venn diagram used raw symbol counts and schematic circles; the heatmap displays layer-specific g:SCS-adjusted p-values. The scoring audit is in Supplementary Table S23, and figure source data are in S24

3. Main results

The genomic literature reviewed for this study shows a clear expansion from early individual loci to large-scale multi-ancestry maps of BD risk architecture. Early GWAS implicated ANK3 and CACNA1C, followed by larger studies identifying additional loci such as ODZ4/TENM4, TRANK1, and NCAN. Stahl et al. (2019) identified 30 genome-wide significant loci, while Mullins et al. (2021) expanded this to 64 loci and reported enrichment in synaptic signaling pathways and drug target gene sets [5], [6]. O'Connell et al. (2025) further expanded BD genomic discovery to 298 genome-wide significant loci and prioritized credible genes across ancestries [7]. Palmer et al. (2022) identified AKAP11 as a rare-variant risk gene with relevance to GSK3B biology [9]. Table 1 summarizes the principal genomic studies and their roles in gene prioritization.

The method-specific locus-to-gene mapping produced 80 strict symbols, 307 functionally expanded symbols, and 331 broad-locus symbols. The sets were nested by construction. The strict set was restricted to exome evidence, formal colocalization, or multi-evidence fine-mapped variant-to-gene links; the functional and broad sets added progressively less restrictive mapping classes. Of the strict symbols, 63 were approved protein-coding genes for the illustrative overlap calculation and 68 were recognized by g:Profiler. GSK3B was not added to any BD mapping set. Gene-level evidence assignments and membership of all three mapping sets are reported in Supplementary Table S20.

The curated drug-associated dataset contained 76 rows across 14 medications. After evidence adjudication, Tier 1 comprised 45 quantitative direct-interaction rows and 22 unique genes. Tier 2 contained two genes (CACNA1E and CHRNA4), Tier 3 contained 19 PK rows and five unique genes (ABCB1, CYP2B6, CYP2C19, CYP2D6, and CYP3A4), and three PGx rows mapped to one unique marker gene (HTR2A). Four contested mechanistic rows, two insufficiently gene-specific rows, and one PK regulatory node outside the Tier 3 core were not included in the primary target analysis. The main drug classes are summarized in Table 2; the 76-row curated matrix and source-validation records are provided in Supplementary Tables S2 and S3. Complete evidence-tier assignments and quantitative pharmacology details are provided in Supplementary Table S21.

Only one exact overlap was observed between the strict BD set and the 22 quantitative Tier 1 targets: HTR6, supported on the BD side by functional mapping evidence and on the drug side by an olanzapine interaction with a representative Ki of 7.943 nM. GSK3B was a quantitative Tier 1 lithium interaction (Ki = 2 mM) but was not a BD-mapped gene and remained an a priori bridge. Table 3 ranks these and other candidates using the rule-based evidence rubric. The complete ranked candidate table and scoring rubric are provided in Supplementary Table S23.

The Venn analysis used 80 curated strict BD symbols and 22 Tier 1 targets and identified HTR6 as the only shared symbol (Figure 1). For the illustrative protein-coding calculation, 63 strict BD genes and all 22 Tier 1 genes were compared against 19,296 HGNC protein-coding genes; the probability of at least one overlap was p = 0.0695. The result is therefore not statistically compelling and should be interpreted as candidate prioritization rather than target-level convergence. Complete overlap calculations are reported in Supplementary Table S22, and the Venn source counts are reported in Supplementary Table S24.

Primary g:Profiler enrichment compared the separated BD mapping and drug-evidence layers (Figure 2 and Table 4). The strict BD query had no g:SCS-significant term across GO:BP, GO:MF, KEGG, or Reactome. The functionally expanded BD set produced two significant GO:MF terms and two Reactome terms; the calcium-related result emphasized in the primary analysis was high voltage-gated calcium channel activity (GO:0008331; p = 0.0486), driven by CACNA1B, CACNA1C, and CACNB2. Full corrected enrichment outputs are provided in Supplementary Table S22.

The strict-set null result and the sensitivity-only calcium term show that disease-side enrichment depended on locus-to-gene mapping breadth. CACNA1B was present in the strict set, whereas CACNA1C and CACNB2 entered through the functionally expanded mapping. The calcium-channel finding is therefore biologically relevant but not robust to the strictest mapping definition. In the broad-locus set, the high voltage-gated calcium channel activity term showed the same direction but did not remain significant after correction (GO:0008331; g:SCS-adjusted p = 0.0650).

Tier 1 showed strong expected enrichment for monoaminergic receptors, transporters, GPCR signaling, and synaptic processes. Representative terms included Neuroactive ligand signaling (KEGG:04082; p = 3.65 × 10⁻²¹), Neuroactive ligand-receptor interaction (KEGG:04080; p = 1.06 × 10⁻¹¹), and Calcium signaling pathway (KEGG:04020; p = 4.15 × 10⁻⁶). The calcium term was contributed by ADRA1A, ADRA1B, CHRM1, HRH1, HTR2A, HTR2C, HTR6, and HTR7, all receptor genes rather than voltage-gated calcium-channel targets.

Tier 2 contained only CACNA1E and CHRNA4 and produced no corrected significant term. Tier 3 yielded expected xenobiotic and cytochrome P450 enrichment, including xenobiotic metabolic process (GO:0006805; p = 6.77 × 10⁻⁹). The separate PGx layer contained only HTR2A and produced one marginal GO:MF annotation (p = 0.04999); this one-gene result was not interpreted as pathway convergence.

When Tier 1 and Tier 2 were combined, calcium signaling remained significant (KEGG:04020; p = 4.52 × 10⁻⁷) and added CACNA1E to the eight Tier 1 receptor genes. However, the significant disease-side term was a GO molecular-function annotation for voltage-gated calcium-channel activity, whereas the drug-side term was a broad KEGG signaling pathway. No identical significant calcium annotation was shared across the BD and drug layers (Table 5).

The layer-separated results distinguish three findings: no corrected pathway signal in the strict BD set; a calcium-channel term that appears only after functional expansion of BD mapping; and strong Tier 1 receptor/synaptic enrichment reflecting the pharmacology of the selected drugs. Calcium is the clearest cross-layer theme, but the disease and drug layers do not contribute to the same significant annotation.

Adding Tier 2 to Tier 1 retained the broad KEGG calcium pathway and introduced CACNA1E, whereas Tier 2 alone was not significant. Separating Tier 3 removed CYP/xenobiotic terms from the pharmacodynamic interpretation. These checks support the receptor and metabolism assignments within their own evidence layers, but they do not demonstrate formal disease-drug convergence.

4. Discussion

This integrative bioinformatics analysis examined whether evidence-restricted BD locus-to-gene mappings show biologically interpretable relatedness to quantitatively supported targets of medications used in BD. The primary comparison between the strict BD set and Tier 1 identified only HTR6, and the illustrative overlap test was not significant. Separating mapping methods and drug-evidence tiers also showed that pathway conclusions depend strongly on which layer is analyzed: the strict BD set had no corrected enrichment, the functionally expanded set yielded a calcium-channel term, Tier 1 was dominated by expected receptor and synaptic pharmacology, and Tier 3 isolated the CYP/xenobiotic component.

HTR6 was the only exact strict BD-Tier 1 overlap. BD evidence included formal eQTL colocalization and supporting functional mapping in Mullins et al. (2021); the representative olanzapine interaction had a Ki of 7.943 nM

[6][37]

GSK3B met the quantitative Tier 1 definition through direct lithium inhibition, with a representative Ki of 2 mM

[38][9][10][11][39][40]

The Tier 1 enrichment profile was dominated by monoaminergic receptors, transporters, GPCR signaling, and synaptic terms. This is expected from the pharmacological composition of antipsychotic and antidepressant targets and is not independent evidence of BD pathogenesis. The ranked candidate table consequently gives the highest relational priority to the exact overlap HTR6 and the predefined GSK3B bridge, while receptor-only or disease-only calcium contributors remain in thematic band C. The ranking makes the evidence structure explicit but does not convert annotation enrichment into a clinical probability.

Calcium-related biology remained the most informative cross-layer theme, but the layer-separated analysis narrows its interpretation. The strict BD set did not show corrected calcium enrichment; high voltage-gated calcium channel activity reached significance only in the functionally expanded set and was driven by CACNA1B, CACNA1C, and CACNB2. In contrast, the Tier 1 KEGG calcium pathway was driven by adrenergic, muscarinic, histamine, and serotonin receptor genes rather than voltage-gated calcium-channel targets. Adding Tier 2 introduced CACNA1E and retained the broad KEGG signal, but Tier 2 alone was not enriched. Thus, the disease and drug layers implicate different components of calcium-dependent neurobiology and do not share an identical significant calcium annotation.

Tier 3 PK enrichment highlights a clinically relevant but conceptually separate layer. ABCB1, CYP2B6, CYP2C19, CYP2D6, and CYP3A4 produced expected xenobiotic and cytochrome P450 terms. These findings concern exposure, interactions, tolerability, and pharmacogenomic variability rather than BD genetic architecture. Their separation from Tier 1 prevents metabolism annotations from being presented as evidence of therapeutic target convergence.

The evidence ranking is intended to guide experimental prioritization. HTR6 could be tested through genotype- or expression-stratified pharmacology and target-engagement studies, whereas the AKAP11-GSK3B-lithium axis requires exposure-aware functional validation. Calcium candidates require experiments that distinguish voltage-gated channel mechanisms from receptor-mediated intracellular calcium signaling. None of the present scores or enrichment terms predicts individual treatment response.

Phenotype quality also remains important because BD-MDD misclassification can affect genomic discovery and treatment-response studies

[1]

5. Limitations

Several limitations should be considered. First, the strict, functionally expanded, and broad-locus sets reduce but do not eliminate uncertainty in GWAS locus-to-gene assignment. Fine-mapping, colocalization, TWAS, MAGMA, and proximity methods answer different questions and may prioritize different genes. The absence of corrected enrichment in the strict set and the appearance of the calcium-channel term only after functional expansion demonstrate this mapping sensitivity directly.

Second, quantitative Tier 1 curation improves pharmacological specificity but does not standardize assay species, tissue, construct, endpoint, or experimental conditions. A Ki or IC50 value does not by itself establish CNS penetration, therapeutic free concentration, receptor occupancy, or causal relevance to mood stabilization. Human target-engagement evidence was sparse, and broad or contested protein-group mappings had to be excluded.

Third, exact symbol overlap can underestimate network- or pathway-level relatedness, whereas pathway enrichment can overrepresent highly annotated receptor and CYP families. The layer-separated design reduces the latter problem but does not provide a matched null model for annotation density, druggability, or network degree.

Fourth, g:Profiler used an annotated-domain background and g:SCS correction separately within each query. The analysis did not implement a matched druggable-genome background, permutation testing, degree-matched network proximity, formal pathway-overlap testing, or leave-one-drug-class-out resampling. Enrichr and exclusion-based runs performed before the application of method-specific BD locus-to-gene mapping and quantitative drug-evidence tiering remain secondary because their correction procedures and gene-set definitions differ from the primary g:Profiler workflow. Future work should test alternative mapping thresholds, remove drug classes in turn, and evaluate whether the observed themes exceed matched null expectations.

Fifth, the analysis did not include patient-level treatment response, pharmacokinetic measurements, longitudinal exposure, or ancestry-specific pharmacogenomic stratification. Future studies should integrate these data with transcriptomics, proteomics, fine-mapped regulatory effects, therapeutic exposure, and direct target-engagement measures before evaluating clinical prediction.

Finally, antidepressants were included as a pharmacogenomic and monoamine-transporter block, not as universal BD treatments

[13][18]

Safety and pharmacokinetic genes were excluded from the main overlap analysis because they were not direct therapeutic targets or BD risk genes. Examples include HLA-B*15:02 and HLA-A*31:01 for carbamazepine hypersensitivity

[30][25][29]

6. Data and code availability. Supplementary materials

The supplementary files distinguish foundational curation materials from analyses performed at two stages. Tables S2, S3, S11, and S15 contain, respectively, the curated drug-gene matrix, source-validation records, the PGx and drug-safety gene list, and the DrugBank record audit; these materials support both analytical stages. Table S12 contains the initial local sensitivity analysis. Tables S16-S19 report enrichment analyses performed before the application of method-specific BD locus-to-gene mapping and quantitative drug-evidence tiering; they are provided only to document the analytical history and were not used for primary inference. Tables S20-S24 contain the analyses reported as primary results in this article: S20, evidence-stratified BD locus-to-gene mapping; S21, quantitative and secondary drug-evidence tiers; S22, primary g:Profiler enrichment and overlap results, which replace S16-S19 for primary interpretation; S23, ranked candidate evidence assessment; and S24, source data for Figures 1-2. DrugBank source PDFs are not redistributed because of licensing restrictions. Analysis scripts and machine-readable metadata are available from the corresponding author upon reasonable request.

7. Conclusion

This study integrated evidence-stratified BD locus-to-gene mappings with quantitatively supported and secondary drug-associated gene layers. Exact overlap between the strict BD set and Tier 1 was limited to HTR6 and was not statistically compelling. GSK3B remained a predefined AKAP11-lithium bridge rather than an independently discovered disease-drug overlap.

Tier 1 showed expected monoaminergic, GPCR, and synaptic enrichment, while Tier 3 isolated the xenobiotic-metabolism component. Calcium-related biology was the clearest cross-layer theme, but it arose from voltage-gated calcium-channel activity in the functionally expanded BD set and receptor-mediated KEGG calcium signaling in Tier 1. The layers did not share an identical significant calcium annotation.

The results provide a transparent hypothesis-generating framework, not validated treatment-response biomarkers or formal proof of mechanistic convergence. Future work should combine matched null models, network proximity, exposure and target-engagement data, functional experiments, and patient-level treatment outcomes.

The additional file for this article can be found as follows: