Artificial Intelligence-driven Biomarker Discovery for Precision Psychiatry

Poor phenotype quality and small cohorts remain major constraints in psychiatric biomarker research, and recent methodological advances increasingly address them. Language models turn clinical text into analysable variables at a scale manual chart review cannot reach. Self-supervised models learn from the unlabelled signals psychiatry already collects. Variant-effect prediction makes molecular interpretation computable, and research agents have generated and experimentally confirmed candidates outside psychiatry. These capabilities act early, on what is measured, what can be learned from it and which candidates are proposed. None of the artificial intelligence capabilities reviewed here yet reaches the stages at which a candidate becomes a clinical test. Producing candidates is becoming cheap while qualifying them is not. This review is organised around that asymmetry rather than by disorder or marker catalogue. It separates the measured characteristic, the prediction model and the intended clinical use, and treats method family as distinct from evidential stage. Across the selected evidence, discrimination is supported more strongly in development data than in independent evaluation, and prospective evidence that model-guided use improves care is scarce. Many reported metrics also address a simpler task than the intended decision: diagnosis instead of differential diagnosis, general prognosis instead of differential treatment benefit. The review proposes a six-stage framework linking intended use, measurement validity, model evaluation, comparative utility and monitored deployment, and places each capability on it, clarifying where methodological progress can accelerate discovery and where conventional clinical validation remains indispensable.