MCGA: a multi-strategy conditional gene-based association framework integrating with isoform-level expression profiles reveals new susceptible and druggable candidate genes of schizophrenia

Linkage disequilibrium and disease-associated variants in non-coding regions make it difficult to distinguish truly associated genes from redundantly associated genes for complex diseases. In this study, we proposed a new conditional gene-based framework called MCGA that leveraged an improved effective chi-squared statistic to control the type I error rates and remove the redundant associations. MCGA initially integrated two conventional strategies to map genetic variants to genes, i.e., mapping a variant to its physically nearby gene and mapping a variant to a gene if the variant is a gene-level expression quantitative trait locus (eQTL) of the gene. We further performed a simulation study and demonstrated that the isoform-level eQTL was more powerful than the gene-level eQTL in the association analysis. Then the third strategy, i.e., mapping a variant to a gene if the variant is an isoform-level eQTL of the gene, was also integrated with MCGA. We applied MCGA to predict the potential susceptibility genes of schizophrenia and found that the potential susceptibility genes identified by MCGA were enriched with many neuronal or synaptic signaling-related terms in the Gene Ontology knowledgebase and antipsychotics-gene interaction terms in the drug-gene interaction database (DGIdb). More importantly, nine susceptibility genes were the target genes of multiple approved antipsychotics in DrugBank. Comparing the susceptibility genes identified by the above three strategies implied that strategy based on isoform-level eQTL could be an important supplement for the other two strategies and help predict more candidate susceptibility isoforms and genes for complex diseases in a multi-tissue context.


Introduction 40
Genome-wide association studies (GWASs) have been used to identify novel genotype-phenotype 41 associations for more than a decade, and thousands of single-nucleotide polymorphisms (SNPs) 42 have been revealed for their associations with hundreds if not thousands of complex human 43 diseases 1,2 . Nevertheless, conventional GWAS analyses have limited power to produce a complete 44 set of susceptibility variants of complex diseases 3 . Because most susceptibility SNPs only have 45 small effects on a complex phenotype, conventional SNP-based association tests are generally 46 underpowered to reveal susceptibility variants after multiple testing corrections. Moreover, the 47 susceptibility variants scattering randomly throughout the genome are often in strong linkage 48 explore the susceptibility genes associated with a complex phenotype using GWAS summary 126 statistics and eQTL summary statistics of SNPs. MCGA has two main advantages over DESE. 127 First, MCGA is based on a new effective chi-squared statistic (ECS), with which the type I error 128 could be controlled within a proper level. Second, MCGA can perform conditional gene-based 129 association analysis using different SNPs sets, i.e., physically nearby SNPs, gene-level eQTLs and 130 isoform-level eQTLs. 131

132
To evaluate the performance of MCGA, we performed extensive simulations and a real data 133 application to schizophrenia. Specifically, we organized the present study in four sequential parts 134 that cover the optimizing the exponent of chi-squared statistics to control type I error rates, Our simulation results showed that the type I error rate was controlled within a reasonable level by 153 using the new effective chi-squared statistics (ECS) with the favorable exponent. Another 154 simulation study pointed out that association analysis based on isoform-level eQTLs was more 155 powerful than gene-level eQTLs. As for predicting the potential susceptibility genes of 156 In the above simulation study, we demonstrated that the conditional gene-based analysis based on 244 the improved ECS was more powerful than the likelihood ratio test in each simulation scenario. 245 Here to further evaluate the performance of MCGA in the real-world data, we used a recent 246 human tissues 21 to identify the susceptibility genes of schizophrenia. Here, MCGA_Dist was first 248 used, i.e., the variants were assigned to genes if the variants were in the window of +/-5kb around 249 the gene boundary (see details in Materials and Methods). We found that 221 of 34,159 genes 250 identified by MCGA_Dist had significantly statistical p-values smaller than 2.5E-6 (see details in 251 Table S1). 252

253
To further study the functional annotations of the 221 potential susceptibility genes, we performed 254 Gene Ontology (GO) enrichment analysis. Interestingly, we found that most GO:BP and GO:CC 255 enrichment terms were neuronal-, dendrite-or synaptic signaling-related terms. Besides, the  Table S2). Systematic text-mining method was used to search the PubMed 258 database to find papers that had reported the potential susceptibility genes of schizophrenia. The 259 results showed that 87 of the 221 (~ 39.4%) potential susceptibility genes had at least one search 260 hit (see details in Table S3). The GO enrichment analysis and the text-mining results both implied 261 the utility of MCGA_Dist. A common assumption is that genes close to significant variants are more likely to be the 273 susceptibility genes, but the reality is that some potentially associated genes are not closest to the 274 significant variants 16 . Molecular Quantitative Trait Loci (molQTL) is a genetic variant associated 275 with a molecular trait, such as a gene-level eQTL and isoform-level eQTL, and can associate a 276 +/-5kb window around the gene boundary. Next, we assigned a variant to a gene (or isoform) if 278 the variant is a gene-level or isoform-level eQTL to broaden the application of MCGA. Since the 279 isoform-level and corresponding gene-level expression profiles were quantified based on the same 280 RNA-sequencing data, we wanted to test whether the power of association analysis based on the 281 gene-level eQTLs was higher than that based on isoform-level eQTLs or not. We first performed a 282 simulation study to evaluate the power of association analysis based on gene-level eQTLs and 283 isoform-level eQTLs, respectively. 284

285
We considered the simplest case for simplicity, i.e., variants affected phenotype only through 286 regulating the gene expression. We simulated genotype data, isoform-level expression profiles and 287 corresponding phenotype data (see details in Materials and Methods). Specifically, we simulated 288 four scenarios, i.e., association analysis using all variants (phenotype-associated isoform-level 289 eQTLs and the other isoform-level eQTLs, denoted as Allvar in Table 1), association analysis 290 only using phenotype-associated isoform-level eQTLs (denoted as isoform eQTL in Table 1), 291 association analysis using gene-level eQTLs which were computed by the gene expression profiles 292 derived by the average value of multiple isoforms belonging to the gene. As for scenarios of genes 293 with multiple isoforms, we specifically simulated two new scenarios (denoted as eQTL_3isoform 294 and eQTL_6isoform in Table 1), i.e., a gene with three (eQTL_3isoform) and six different 295 isoforms (eQTL_6isoform). The expression value of the gene with three isoforms was averaged by 296 the following three isoforms, i.e., one isoform associated with phenotype and the other two 297 random isoforms simulated by the standard normal distribution N(0,1). The expression value of 298 the gene with six isoforms was averaged by the following six isoforms, i.e., one isoform 299 associated with phenotype and the other five random isoforms simulated by the standard normal 300 distribution N(0,1) (see details in Materials and Methods). Based on the four scenarios 301 mentioned above, we used six different parameter combinations to simulate six different cases, 302 and each parameter combination was simulated 100 times to evaluate the statistical power. As 303 shown in Table 1, the power of the association test based on phenotype-associated isoform-level 304 eQTLs was the highest of all cases. The simulation results implied that isoform-level eQTLs were 305 more powerful than gene-level eQTLs in association analysis. 306 307 Broaden the application of MCGA by using gene-level eQTLs and isoform-level eQTLs 312

(MCGA_eQTL and MCGA_isoQTL) 313
In the previous simulation study, we demonstrated that association analysis based on isoform-level 314 eQTLs was more powerful than gene-level eQTLs in each simulation scenario. To further test this 315 conclusion in real data and identify more potential susceptibility genes for schizophrenia, we first 316 gene-level eQTLs of the top-five tissues based on gene-level expression profiles and isoform-level 321 eQTLs of the top-five tissues based on transcript-level expression profiles, respectively. 322 Hereinafter, the gene whose expression is associated with at least one SNP was denoted as eGene, 323 and the gene with an isoform whose expression is associated with at least one SNP was denoted as 324 sGene. Then we performed the improved conditional gene-based association analysis based on 325 gene-level eQTLs and isoform-level eQTLs resulted from the corresponding tissues. In each of the 326 top-five tissues, we found the number of potential susceptibility sGenes identified by 327 MCGA_isoQTL was larger than that of potential susceptibility eGenes identified by 328 MCGA_eQTL under the same filter cutoff 2.5E-6 (Figure 7b, see details in Table S4 and S5). 329 Besides, we found a considerable number of common genes between the estimated eGenes set and 330 sGenes set in each of the top-five tissues (Figure 7b). 331

332
We also performed the GO enrichment analysis to further investigate the functional annotations of 333 these potential susceptibility eGenes and sGenes. For the eGene set in each of the top-five 334 associated tissues, we found only the eGenes identified based on the gene-level eQTLs of 335 Brain-FrontalCortex (BA9) had GO enrichment terms (Figure 7c). For the sGene set in each of 336 the top-five associated tissues, we found the potential susceptibility sGenes in 337 Brain-FrontalCortex(BA9), Brain-Anteriorcingulatecortex (BA24) and Brain-Hippocampus all 338 biologically sensible GO enrichment results and the PubMed search results both implied that the 347 potential susceptibility sGenes and eGenes might have strong associations with schizophrenia. Brain-FrontalCortex(BA9), Brain-Anteriorcingulatecortex (BA24) and Brain-Hippocampus. BF: 360 Brain-FrontalCortex(BA9).

The advantages of MCGA_isoQTL versus MCGA_Dist and MCGA_eQTL 364
The connectivity score of a gene in the weighted co-expression network might imply its real 365 association with other genes, and highly connected genes are often defined as hub genes. These 366 hub genes are located in or near the center of corresponding co-expression modules and might 367 (d) play important roles in trait development 23 . We built an unsigned weighted co-expression network 368 for each top-five tissue and investigated the normalized intra-module connectivity of potential 369 susceptibility genes in each co-expression module. We found that the normalized intra-module 370 connectivity scores of potential susceptibility genes in Brain-FrontalCortex_BA9 and 371 Brain-Nucleusaccumbens(basal ganglia) identified by MCGA_eQTL were significantly larger 372 than that of non-susceptibility genes. Interestingly, the normalized intra-module connectivity 373 scores of potential susceptibility genes identified by MCGA_isoQTL were significantly larger 374 than that of non-susceptibility genes in all the top-five schizophrenia-associated tissues (Wilcoxon 375 rank-sum test p<0.05) ( Table 2). 376 377

381
We next compared the potential susceptibility genes predicted by MCGA_Dist, MCGA_eQTL and 382 MCGA_isoQTL. As shown in Figure 8, twenty-three genes were collectively predicted to be 383 susceptible to schizophrenia by the three models of MCGA. As MCGA_isoQTL could output 384 susceptibility isoforms of twenty-three genes in corresponding tissues (see details in Table S9). 386 Interestingly, we found that susceptibility isoforms for a gene varied greatly in different tissues. 387 For example, ABCC8 and LINC01415 both had one susceptibility isoform in the top-five tissues. 388 ENST00000529967 of ABCC8 was significantly associated with schizophrenia only in 389 Brain-Hippocampus, while ENST00000587320 of LINC01415 was significantly associated with 390 schizophrenia in Brain-Cortex, Brain-FrontalCortex(BA9) and Brain-Nucleusaccumbens(basal 391 ganglia). We also found that different isoforms of the same gene were predicted to be significantly 392 associated with schizophrenia in different tissues, such as ENST00000377600 of BTN2A1 393 significantly associated with schizophrenia in Brain-Cortex and ENST00000312541 of BTN2A1 394 significantly associated with schizophrenia in Brain-FrontalCortex(BA9). MCGA_isoQTL can 395 help predict potential susceptibility genes and isoforms of corresponding phenotype-associated 396 tissues at a more precise level. Except for the advantage of identifying the potential susceptibility isoforms, we found that some 403 of the susceptibility genes exclusively predicted by MCGA_isoQTL were also biologically 404 sensible. We found 478 susceptibility genes exclusively predicted by MCGA_isoQTL (Figure 8).

418
Taken together, from the perspective of WGCNA, the statistical significance of the comparison 419 between potential susceptibility genes and non-susceptibility genes implied that these 420 susceptibility genes identified based on isoform-level eQTLs might play more important roles in 421 the weighted gene co-expression network of corresponding tissues. Our results also suggested that 422 incorporating with isoform-level eQTLs can help predict more potential susceptibility genes than 423 gene-level eQTLs in each potentially phenotype-associated tissue. Our results pointed that 424 MCGA_isoQTL could help find some novel and important susceptibility genes which cannot be 425 found by MCGA_Dist and MCGA_eQTL. Moreover, based on the isoform-level eQTLs of each 426 phenotype-associated tissue, the MCGA_isoQTL strategy can also predict the potential 427 susceptibility isoforms in the corresponding tissues. 428

429
The druggability of the potential susceptibility genes identified by MCGA 430 Since drug target genes with genetic support are twice or as likely to be approved than target genes 431 with no known genetic associations 24,25 , we searched the DrugBank 5.0 database 26 and found that 432 nine potential susceptibility genes identified by MCGA were the target genes of multiple 433 FDA-approved antipsychotics ( Table 4). Several most popular target genes of approved 434 antipsychotics, i.e., DRD2, DRD1 and ADRA1A, were identified by different MCGA models and 435 the results suggested that the three models could complement each other to identify more potential 436 target genes. 437

441
To further investigate the druggability of the potential susceptibility genes, we searched the Drug 442 Gene Interaction database (DGIdb v4.2.0) 27 and filtered the drug-gene interaction terms with at 443 least one supported PubMed paper. After the filtration, we kept 30,072 unique drug-gene 444 interaction terms and found 679 unique drug-gene interaction terms for 34 FDA-approved 445 antipsychotics (see details in Table S10). Then we put the full list of potential susceptibility genes 446 (by MCGA_Dist, MCGA_eQTL and MCGA_isoQTL, respectively) into DGIdb to investigate if 447 the "antipsychotic"-"susceptibility gene" interactions were enriched in DGIdb. As shown in Table  448 5, we found that "antipsychotic" -"potential susceptibility genes" identified by the three models 449 of MCGA were all significantly enriched in DGIdb. Moreover, as shown in Figure 8, 372 out of 450 578 and 478 out of 696 potential susceptibility genes were exclusively identified by 451 MCGA_eQTL and MCGA_isoQTL, respectively. We found 253 unique drug-gene interaction 452 terms for susceptibility genes exclusively predicted MCGA_eQTL (see details in Table S11), and 453 17 of 253 interaction terms were antipsychotics-gene interactions (hypergeometric distribution test 454 p-value= 7.05E-5). We also found 291 unique drug-gene interaction terms for susceptibility genes 455 and the application of eQTLs (especially the isoform-level eQTLs) could aid MCGA to identify 465 more potentially druggable genes. 466 467

472
In this study, we proposed a multi-strategy conditional gene-based association framework, MCGA, 473 based on a new correlation matrix of chi-squared statistics to identify the potential susceptibility 474 genes and isoforms for complex phenotypes. Comparing with the unconditional association test 475 and likelihood ratio test, MCGA showed a lower type I error rate and higher statistical power. 476 Since MCGA is a gene-based method, in this study, we adopted three strategies to map a variant to 477 a gene, i.e., mapping based on physical position, gene-level eQTLs and isoform-level eQTLs. We 478 implemented these three mapping strategies in corresponding three conditional gene-based 479 association models, i.e., MCGA_Dist, MCGA_eQTL and MCGA_isoQTL, to predict the potential 480 susceptibility genes for schizophrenia. MCGA_Dist with MCGA_eQTL and MCGA_isoQTL. We performed a simulation study and 493 demonstrated that isoform-level eQTLs were more powerful than gene-level eQTLs in association 494 analysis. Moreover, we found in real data that the size of the susceptibility gene set for 495 schizophrenia predicted by MCGA_isoQTL was larger than MCGA_eQTL in each 496 advantages over MCGA_eQTL and MCGA_Dist. First, several important potential susceptibility 498 genes were exclusively predicted by MCGA_isoQTL. For example, fifteen potential susceptibility 499 genes exclusively predicted by MCGA_isoQTL each had at least ten search hits in PubMed, which 500 implied these genes were popular in schizophrenia studies. Second, to our best knowledge, 501 MCGA_isoQTL was the first conditional gene-based association approach to produce a list of 502 phenotype-associated isoforms (or transcripts). 503

504
In addition, we investigated the druggability of the susceptibility genes for schizophrenia 505 identified by MCGA. Several susceptibility genes identified by MCGA were also the popular 506 target genes of multiple FDA-approved antipsychotics. Besides, the "susceptibility gene"-507 "antipsychotics" interactions were enriched in DGIdb. The druggablilty of the important 508 susceptibility genes, especially the sGenes identified based on isoform-level eQTLs, provided 509 more credible supports for the utility of MCGA. 510 511 Our framework might have three potential applications. First, MCGA_Dist can be used to predict 512 potential susceptibility genes and isoforms for other complex phenotypes. Second, based on the 513 assumption that the distribution of expression profiles of true susceptibility genes might change 514 before and after therapeutic drug treatment, MCGA_Dist can be used to perform drug 515 repositioning analysis based on the drug perturbed expression profile. Third, since MCGA_eQTL 516 and MCGA_isoQTL can help predict potential susceptibility genes in each potential 517 phenotype-associated tissue, our framework can help perform synergistic drug combination 518 that five brain regions were involved in the present study, and each brain region might have very 527 different dysfunctional genes associated with schizophrenia. We also used MAGMA to identify 528 the susceptibility genes of schizophrenia with the same GWAS summary statistics and found that 529 MAGMA also identified ~ 600 potential susceptibility genes with the basic parameter setup (see 530 details in Table S14). Susceptibility genes identified by MCGA_eQTL and MCGA_isoQTL had 531 many biologically meaningful annotations (such as neuronal-or synaptic signaling-related terms) 532 in the GO databases, and some susceptibility genes were the target genes of multiple 533 antipsychotics, and more than 20% of the susceptibility genes had been previously reported by 534 other schizophrenia research in the PubMed database. Though these potential susceptibility genes 535 were lack of systematically experimental validation, we shared the potential susceptibility genes in 536 Table S1, S4 and S5 and encouraged follow-up studies to evaluate the function and roles of these 537 susceptibility genes in the development of schizophrenia. 538

539
In conclusion, in this study, we proposed a new statistical framework to predict potential 540 gene/isoform-level eQTLs in a multi-tissue context. The application of our framework to 542 schizophrenia revealed many novel susceptible and druggable genes. Besides, the usage of 543 isoform-level eQTLs can be an important supplement for the conventional gene-based approach. 544 The framework was packaged and implemented in our integrative platform KGGSEE 545 (http://pmglab.top/kggsee/#/). We hope our framework can facilitate researchers to gain more 546 insights into the phenotype-associated genes and isoforms of complex phenotypes. 547 548

The new effective chi-squared statistics (ECS) for conditional gene-based association 550 analysis 551
We improved our previously proposed effective chi-squared test 10 for a more efficient conditional 552 gene-based association analysis based on a new correlation matrix of chi-squared statistics. The 553 improved effective chi-squared statistics had two methodological advances to address the potential 554 inflation issue, i.e., a type I error-controlled correlation matrix of the observed chi-squared 555 statistics and a non-negative least square solution for the independent chi-squared statistics. The 556 reasoning process was as follows. Suppose there were n loci in a set of genes. One wanted to 557 calculate the association p-value of another physically nearby gene (containing m loci) 558 conditioning on the set of genes (n loci). The first step of the conditional analysis was to produce 559 effective chi-squared statistics for the set of genes (n loci) and all the genes (n+m loci in total). 560 Each locus had a p-value for phenotype association in the GWAS. The p-values were converted to 561 corresponding chi-squared statistics with the degree of freedom 1. According to Li et al. 10 , each 562 locus could be assumed to have a virtually independent chi-squared statistic. An observed 563 marginal chi-squared statistic of a locus was equal to the summation of its virtually independent 564 chi-squared statistic and the weighted virtually independent chi-squared statistic of nearby loci. 565 The weight was related to the chi-squared statistics correlation, which was a key parameter of the 566 analysis. The correlation of chi-squared statistics between two loci was approximated by the 567 absolute value of genotypic correlation to the power of c, i.e., |r| c . Here, we derived that the key 568 parameter, i.e., exponent c, ranged from 1 to 2, corresponding to different non-centrality 569 parameters of a non-central chi-squared distribution (See the derivation in the next section). 570 According to Li et al. 10 , the n virtually independent chi-squared statistics of the gene set could be 571 approximated by a linear transformation of the n observed chi-squared statistics (Formula (1) (1) . 582

583
The effective chi-squared statistics (́+ ) and degree of freedom (́+ ) of the n+m loci could be 584 calculated in the same way. 585 The effective chi-squared statistics of the m loci conditioning on the n loci was then approximated 586 Because the virtually independent chi-squared statistics and degrees of freedom were expected to 597 be larger than 0, we adopted a sequential coordinate-wise algorithm to approximate them 29 . This 598 algorithm avoided unstable solutions in the above linear Formula (1) due to stochastic errors in 599 the correlation matrix and observed chi-squared statistics. 600

601
After the above multiple approximations, it was still difficult to obtain the analytic solution for the 602 exponent c in Formula (1). We proposed a grid search algorithm to find a favorable value of 603 exponent c to control type I error rates of the effective chi-squared tests. The error rate was 604 examined by divergence from a uniform distribution between an obtained and theoretical top 1% derivation in the Materials and Methods). The c value leading to the minimal MLFC was defined 608 as the favorable c value. We considered in total 84 parameter settings, i.e. a combination of three 609 different sample sizes (10,000, 20,000 and 40,000) and 14 different variant sizes (10, 30, 50, 80, 610 100, 125, 150, 200, 250, 300, 400, 500, 800, and 1000)  Let two normal random variables ~( , ) and ~( , ) have covariance c. Note that 624 a squared normal random variable has non-central chi-square distribution and the squared mean of 625 the former is called noncentrality parameter. The two variables can also be factorized as 626 Overall, the correlation between the two (non-central) chi-square ranges from r 2 to r. 638

The conditional gene-based association analysis for genome-wide association study 639
In a GWAS, all genes were firstly calculated with the p-values of unconditional gene-based 640 association test using the above effective chi-squared statistics. For a given p-value cutoff, the 641 significant genes were extracted and subjected to the conditional gene-based association analysis. 642 When there were multiple significant genes in an LD block, the genes were conditioned one by 643 one in a pre-defined order. In the present conditional analysis, the order of the gene was defined The GTEx project (release v8) created a resource including whole-genome sequence data and 730 RNA sequencing data from ~ 900 deceased adult donors 21 . Four tissues or cell types (i.e., whole 731 blood, cells-Leukemiacellline_CML, pancreas and pituitary) were filtered out and not included in

PubMed search 763
To find supports from published research, we performed a text-mining analysis based on PubMed 764 database on June 3rd, 2021. We searched the PubMed database with the items of 765 MCGA_eQTL/MCGA_isoQTL was that the latter two were based on the gene/isoform-level 829 eQTLs of each tissue, thus can produce the potential susceptibility genes/isoforms in a multi-tissue 830 context. 831 832 Like our previous model DESE, MCGA contained three iterative steps. In the first step, associated 833 genes with smaller p-values of the ECS test were given higher priority to enter the following 834 conditional gene-based association analysis. This step could generate a list of roughly associated 835 genes by removing redundantly associated genes. It should be noted that we dealt with the 836 order of a gene entering the conditional gene-based association analysis was determined by its 838 p-value of the ECS test. For MCGA_isoQTL, assume gene A has m isoforms. Each isoform could 839 get a p-value based on the ECS test, representing the overall statistical significance of all 840 isoform-level eQTLs (simultaneously variants) associated with this isoform. If the isoform with 841 the smallest p-value was isoform a, with its p-value p a, among the m isoforms of gene A, we only 842 kept isoform a of gene A for the following analyses. The adjustment p-value for "gene A : isoform 843 a" pair was adjusted to m* p a to enter the following conditional gene-based association analysis. 844

845
The second step was to compute the selective expression score of genes/isoforms in each tissue by 846 taking all tissues as the background (see details in reference 22 ). The Wilcoxon rank-sum test was 847 then performed by using the selective expression score of the associated gene/isoform set and 848 not-associated gene/isoform set (generated by the first step) in each tissue. 849

850
In the third step, all genes/isoforms, including the not-associated genes/isoforms, were ranked in 851 descending order based on the tissue-selective expression score of each gene/isoform. The 852 tissue-selective expression score of a gene/isoform was computed based on the rank of this 853 gene/isoform-selective expression score and the p-value of the Wilcoxon rank-sum test between 854 the associated gene/isoform set and not-associated gene/isoform set in each tissue. 855 856 In the following iteration, genes/isoforms with higher tissue-selective expression scores (in the 857 third step) were given higher priority to enter the conditional gene-based association analysis (in 858 did not change almost, and then corresponding associated genes/isoforms were deemed to be 860 potentially associated with the phenotype. More details about the iterative procedure can be found     Table S9: The potential susceptibility isoforms of the twenty-three common genes in the top five 953 phenotype-associated tissues of schizophrenia. 954 Table S10: The FDA-approved antipsychotics in DGIdb. 955 Table S11: Drug-gene interaction terms in DGIdb for susceptible genes exclusively predicted 956 MCGA_eQTL. 957 Table S12: Drug-gene interaction terms in DGIdb for susceptible genes exclusively predicted 958

MCGA_isoQTL. 959
Table S13: The number of potentially druggable categories for the susceptibility genes of 960 schizophrenia identified by MCGA. 961