A Psychiatric Drug Prescription Framework to Predict Drug-Drug Interaction Risks with Artificial Intelligence

The drastic escalation in the mental health crisis affects 53 million Americans, resulting in a tremendous demand for psychiatric medications. However, only 37% of treatments are successful. A leading cause of clinical failure is the complexity of multiple drug usage, i.e., polypharmacy, which exacerbates drug-drug interactions (DDIs) and adverse drug reactions (ADRs), ranging from mild symptoms to life-threatening consequences. Thus, it is essential to predict the side effects of concomitant drug use before their clinical complications. To streamline drug prescriptions, we developed a machine-learning pipeline to calibrate the severity of DDIs tailored to psychiatric patients. We leveraged 581 unique drug targets and ADRs from DrugBank, SIDER, and MedDRA to construct a data set capturing mechanisms of action. Jaccard index measured drug similarity, and two classification machine-learning models were trained to predict clinically severe outcomes. We identified 95,088 DDIs, 14,502 drug-target interactions, and 23,597 ADRs, of which 1,146 involved a severe side effect. Comparisons with clinical psychiatry case reports and cross-validation tests indicate that this pipeline achieved excellent Area Under the ROC Curve performance levels (0.87 and 0.90). Using computed similarity scores, we derived a composite risk matrix indicating DDI seriousness to bring more transparency to safe drug prescriptions. Further independent tests with commonly used psychiatric drugs validated this algorithm's accuracy. Overall, we developed a novel method to quantify DDIs—thus, increasing therapeutic efficacy by preventing severe ADRs. This framework may aid clinicians in making the right decisions while using concomitant drugs.