Multi-domain model predicts future substance use disorder before heavy use takes hold
A new prognostic model built from longitudinal cohort data suggests that substance use disorder risk can be estimated before full-blown illness emerges, using a blend of genetic, behavioral, and environmental information. The analysis pulled together four large studies from the United States, the United Kingdom, and Finland, then asked a practical clinical question: can a model identify who is more likely to develop a moderate-to-severe DSM-5 SUD later in life?
The answer, in this preprint, is yes with meaningful but imperfect accuracy. The integrated model reached an area under the receiver operating characteristic curve of 0.78 for any SUD, with sensitivity of 69% and positive predictive value of 45%. Those are not trivial numbers for a prevention-focused tool, especially in a field where validated prognostic instruments remain scarce. They also make clear that prediction is possible without pretending to be deterministic.
Four cohorts, one harmonized framework
Investigators assembled data from the National Longitudinal Study of Adolescent to Adult Health, the Avon Longitudinal Study of Parents and Children, the Collaborative Study on the Genetics of Alcoholism, and FinnTwin12. Together, these cohorts supplied a broad range of risk: from population-based samples to family-based cohorts enriched for substance-related vulnerability. Participants had genomic and behavioral data available, and follow-up stretched from early adolescence into young adulthood.
The final analytic sample included 16,362 participants, 54.7% of them female, with a mean age of 26.73 years. SUD prevalence in the dataset was 22.0%. Genetic similarity group composition was reported as 70.4% European, 16.8% African, 9.8% admixed American, 2.2% East Asian, and 0.7% South Asian. Those proportions matter because risk algorithms often degrade when they leave the ancestry groups on which they were trained. Here, performance remained reasonably stable across groups, although it did vary.
What went into the model
The prediction framework combined polygenic scores for substance use-related phenotypes with 10 early-life behavioral and environmental risk factors. The study did not rely on genetics alone, and that design choice is central to the result. Many people at elevated genetic liability never develop a disorder, while many others with modest inherited risk do so because of family context, conduct problems, peer exposure, early use, or other developmental pressures. A risk model that ignores those dimensions will miss a large part of the clinical picture.
By integrating multiple domains, the model attempted to capture both inherited susceptibility and the circumstances that shape whether that susceptibility becomes clinically visible. In practical terms, this is the difference between a purely statistical score and a risk estimate that reflects adolescent behavior, social environment, and biological loading at the same time.
Performance stayed fairly strong across substances and groups
The integrated model performed with AUC values ranging from 0.74 to 0.82 for specific substances, including alcohol, cannabis, and tobacco. Across genetic similarity groups, AUC ranged from 0.61 to 0.82. The lower end of that range signals a familiar problem in predictive medicine: accuracy is not uniform across all populations, and the model clearly performed better in some groups than others. Even so, the investigators demonstrated that a harmonized approach could retain discriminative ability across diverse cohorts rather than collapsing outside a single training sample.
Sex-stratified performance also differed. The model showed higher AUC in females than in males, suggesting that the same feature set may not encode risk equally well across sex-based subgroups. The paper does not solve that problem, but it identifies it cleanly. In risk prediction, subgroup performance often determines whether a model can ever move from publication into real-world use.
One result stands out: the model classified future risk before disorder onset, not after treatment failure or during an established illness course.
Why this kind of prediction matters in addiction medicine
SUD prevention has long suffered from a timing problem. Clinicians usually see the consequences after escalation has already occurred: missed school, repeated intoxication, withdrawal, injuries, legal trouble, or psychiatric comorbidity. A model that flags risk earlier could shift intervention upstream, when counseling, family-based prevention, school supports, monitoring, and targeted behavioral strategies may still change trajectory.
The biological rationale is straightforward. Addiction risk does not arise from a single pathway. Genetic liability can shape reward sensitivity, impulsivity, stress reactivity, and exposure to co-occurring psychiatric traits, while early-life environment influences access, normalization, coping, and exposure patterns. When those factors accumulate, the probability of progression rises. A model that captures several of those layers can therefore outperform a score built from one domain alone.
Limits of the preprint remain important
This work is promising, but the boundaries are obvious. It is a preprint and has not undergone peer review. The article also evaluates predictive discrimination rather than a completed clinical implementation strategy, so the results do not yet tell clinicians how to act on a given score in routine care. The positive predictive value of 45% means a substantial share of high-risk classifications would still be false positives, which is a major issue if the tool is ever used to guide counseling, surveillance, or preventive services.
The analysis used existing cohort data, with varying degrees of access restrictions across studies. It also drew from participants who had available genomic and behavioral information, which can introduce selection effects. Some ancestry groups were represented by relatively small proportions, and the reported AUC spread across genetic similarity groups suggests further calibration work will be necessary before the model could be considered broadly deployable.
Even with those caveats, the study provides a concrete framework for combining genetic and developmental information in addiction risk prediction. The next step is not simply higher accuracy. It is proving whether a tool like this can be calibrated, interpreted, and used in a way that changes prevention decisions without widening disparities or overcalling risk in patients who would never progress to disorder.