A novel method to disentangle tightly linked risk and resilience genes in psychiatric disorders

Genetic risk factors for neuropsychiatric disorders are well documented yet some individuals with high genetic risk remain unaffected. These resilient individuals may carry protective genes that mitigate risk. However, resilience genes are difficult to detect because protective variants may be masked by risk variants. We developed an adversarial learning (AL) approach that explicitly separates resilience from risk by training a model to learn patterns that discriminate high-risk controls from risk-matched cases while actively “unlearning” patterns present in low-risk groups. We aimed to evaluate whether AL can generate genetic resilience scores and identify resilience-associated variants across real and simulated data. We tested this approach in simulated data, Alzheimer’s disease, and ADHD. In simulated data, in which we generated limited resilience single nucleotide polymorphisms (SNPs) among many risk SNPs, our AL approach had significantly higher mean sensitivity (0.91 vs 0.81) and specificity (0.95 vs 0.29) across a range of trials. Applied to real data, the method identified 5 resilience-associated SNPs in Alzheimer’s disease and 45 resilience-associated SNPs in ADHD. In both datasets, we observed significant negative interactions between polygenic risk and polygenic resilience scores, indicating that higher resilience scores attenuated the effect of genetic risk. We plan to expand this approach to larger genomic data sets and other disorders, as well as adapt the approach for use in transcriptomic and neuroimaging data. The markers identified in this work and in ongoing interaction tests could inform efforts into new diagnostic tools and therapies.