Abstract Opioid use disorder (OUD) is a significant public health problem with opioid-associated overdoses and deaths reaching epidemic levels in recent years. Methadone and buprenorphine are widely used and generally effective medications for OUD (MOUDs). However, relapse and nonadherence rates remain high and average treatment retention is suboptimal (e.g., <6 months in 30-50% of cases). As risk of overdose is highest following relapse and treatment dropout, improved mechanistic understanding of risk and resilience factors in individuals in early MOUD (i.e., <6 months) is urgently needed. This application uses network-based analysis, with built-in cross- validation, to identify brain networks associated with (i) patterns of illicit opioid use during early MOUD and (ii) medication adherence during early MOUD in a diverse sample of individuals (N=240, 50% female, 50% male, 50% receiving methadone, 50% receiving buprenorphine). This is critical to improve understanding of mechanisms and predictors of MOUD response and is an essential precursor to development of improved, evidence-based interventions grounded in known neurobiology. This work builds on our prior work identifying brain connections prospectively associated with future relapse to illicit opioids during sustained MOUD. The identified ‘opioid abstinence network’ included connections between frontoparietal, salience, sensorimotor and default mode regions and was robust to analyses controlling for relevant clinical variables (e.g., MOUD dose, years of opioid use). In AIM 1, we seek to externally validate and extend this finding via collection of neuroimaging data from a larger, more diverse sample of individuals early in MOUD treatment. In AIM 2, we propose to collect additional, multi-task neuroimaging data to determine the impact of different, task-induced brain states on network identification. Finally, in AIM 3, we will use our recently developed approach, in which variation in model accuracies are assessed as a function of core sources of clinical diveristy (e.g., sex, medication dose, co-occurring disorders), to identify sources of model bias and neurobiological heterogeneity. Assessing sources of model bias and neurobiological heterogeneity embraces the clinical complexity that is inherent to the current opioid epidemic. These clinical sources of variance have typically either been excluded for or ignored (e.g., treated as covariates of no interest) in prior neuroimaging studies of MOUD. All acquired data will be shared via the NIMH Data Archive.