Machine learning augmented genome-wide meta-analysis of prescription opioid use in 860,000 individuals

Opioid analgesics are widely prescribed for pain, yet individuals vary markedly in their patterns of medical opioid use, influencing the risk of prolonged exposure and misuse. To investigate the genetic basis of prescription-derived opioid intake, we analyzed 859,675 individuals of European ancestry from three population biobanks. Prescription records were harmonized to total cumulative oral morphine equivalents (OME), enabling the construction of three complementary outcomes: any opioid prescription, total cumulative dose among users, and population-level cumulative dose including non-users. Genome-wide meta-analyses identified 78, 20, and 135 loci for these traits, respectively, comprising a total of 234 independent signals at 145 distinct genomic regions. All phenotypes were highly correlated and shared strong genetic overlap with pain-related traits, although cumulative dose among users captured a more distinct component of dose intensity. To identify deviations from expected medical use, we trained gradient-boosted machine learning models to derive two refined phenotypes reflecting early onset of prescribing and dose levels exceeding those predicted by clinical history. The early-onset phenotype showed no genome-wide significant associations and mirrored pain-driven genetic architecture. In contrast, the excess-dose phenotype yielded a genome-wide significant signal at rs58099562 in the CYP2D6/CYP2D7 region, in high linkage disequilibrium with the CYP2D6*4 loss-of-function allele, and exhibited stronger genetic correlations with psychiatric and substance-use traits than with pain. These findings indicate that standard prescription-derived measures of opioid exposure predominantly reflect pain biology, whereas disproportionately high dosing captures distinct genetic liability linked to neuropsychiatric and pharmacokinetic pathways. Machine-learning–based phenotyping can therefore reveal dosage-specific genetic influences not detectable through conventional prescription traits. (a) Three opioid prescription phenotypes were defined: binary prescription status (RxExpPop), cumulative dose among users (RxDoseUser), and combined dosage plus binary prescription (RxDosePop). (b) Genome-wide association studies were performed across multiple biobanks and meta-analyzed using METAL. (c) Machine learning was applied to refine overuse phenotypes, resulting in early onset (RxOverUse_Onset) and high dose (RxOverUse_Amount) subtypes. (d) Downstream analyses included genetic correlation, gene annotation, biological interpretation, and cross-study comparisons.