Background: Genome-wide association studies (GWAS) identify genetic variants predictive of common diseases but this does not directly inform on molecular mechanisms. The recently developed deep learning-based method DeepSEA uses DNA sequences to predict regulatory effects for up to 1000 functional units, namely regulatory elements and chromatin features in specific cell-types from the ENCODE project. Results: We here describe "DeepWAS", a conceptually new GWAS approach that integrates these predictions to identify SNP sets per functional units prior to association analysis based on multiple regression. To test the power of this approach, we use genotype data from a major depressive disorder (MDD) case/control sample (total N=1,537). DeepWAS identified 177 regulatory SNPs moderating 122 functional units. MDD regulatory SNPs were located mostly in promoters, intronic and distal intergenic regions and validated with public data. Blood regulatory SNPs were experimentally annotated with methylation quantitative trait loci (QTLs), expression quantitative trait methylation loci and expression QTLs and replicated in an independent cohort. Joint integrative analysis of regulatory SNPs and the independently identified annotations were connected through transcription factors MEF2A, MEF2C and ATF2, regulating a network of transcripts previously linked to other psychiatric disorders. In the latest GWAS for MDD, the MEF2C gene itself is within the top genome-wide significant locus. Conclusions: DeepWAS is a novel concept with the power to directly identify individual regulatory SNPs from genotypes. In a proof of concept study, MEF2C was identified as a master-regulator in major depression, a finding complementary to recent depression GWAS data, underlining the power of DeepWAS.