Genetic Data Can Lead to Medical Discrimination: Cautionary tale of Opioid Use Disorder

Using genetics to predict the likelihood of future psychiatric disorders such as Opioid Use Disorder (OUD) poses scientific and ethical challenges. This report illustrates flaws in current machine learning (ML) approaches to such predictions using, as an example, a proposed genetic test for OUD derived from 16 candidate gene variants. In an independent sample of OUD cases and controls of European and African descent, results from five ML algorithms trained with purported "reward-system" candidate variants demonstrate that ML methods predict genomic ancestry rather than OUD. Further, sets of variants matched to the candidate SNPs by allele frequency produced similarly flawed predictions, questioning the plausibility of the selected candidate variants. We conclude that the genetic prediction of OUD (and by extension other highly polygenic psychiatric diseases) by ML has high potential to increase the likelihood of medical discrimination against population subgroups, with no benefit of accurate prediction for early intervention.