We also applied GPA with multiple annotation datasets to improve its performance. Beside the CNS gene set, we considered eQTL annotation from GTEx [36] and transcription factor binding site (TFBS) annotation by ANNOVAR [37]. Specifically, we downloaded the available eQTL data from the GTEx website (http://www.ncbi.nlm.nih.gov/gtex/GTEX2/gtex.cgi) and took the intersection of these eQTL with the markers of the psychiatric disorders, resulting in 23,505 SNPs annotated as eQTL. To obtain our TFBS annotation, we used the key word “TFBS” to query the ANNOVAR database, resulting in 19,029 SNPs annotated as TFBS (More details can be found at http://www.openbioinformatics.org/annovar/annovar_faq.html#tfbs). We performed joint analysis of BPD and SCZ with these three annotations (CNS, eQTL and TFBS). The estimated parameters in the GPA model and their standard errors are shown in Table 4. Notice that hit the boundary of the parameter space in presence of the eQTL annotation. This made the EM algorithm converge in fewer iterations, resulting in the minor differences between the estimated parameters in Table 4 and those in Table 2. With the local FDR , GPA identified 724 and 977
