CROSS TRIAL IDENTIFICATION OF A PREDICTIVE ENRICHMENT SUBGROUP FOR ESMETHADONE (REL 1017) IN MAJOR DEPRESSIVE DISORDER USING EXPLAINABLE MACHINE LEARNING

Background

High patient heterogeneity and placebo response contribute to frequent late-stage clinical trial failure in major depressive disorder (MDD). Regulatory guidance on enrichment strategies and model-informed drug development (MIDD) emphasizes prospectively identifying patients most likely to benefit from therapy. We evaluated an explainable machine learning–based framework to derive and validate a predictive enrichment strategy for esmethadone (REL-1017) across multiple randomized trials.

Methods

Patient-level data from three independent randomized, double-blind, placebocontrolled trials of esmethadone were analyzed: Phase 2a (analysis set, N=61), Phase 3 RELIANCE I (analysis set, N=205), and Phase 3 RELIANCE III (analysis set, N=208). The program included both adjunctive and monotherapy study designs. Baseline variables included routinely collected demographic, clinical, and symptom-level measures. Subgroup discovery was conducted using an explainable machine learning platform designed for small clinical trial datasets and aligned with MIDD principles. The approach employs unsupervised feature embedding and constrained evolutionary search to identify patient subgroups associated with differential treatment response, while limiting overfitting and preserving interpretability. Subgroup definitions were identified in the Phase 2a trial and locked to assess replicability in Phase 3 trials. Treatment-by-subgroup interactions were assessed using change from baseline in Montgomery–Åsberg Depression Rating Scale (MADRS) total score at study endpoint.

Results

A single Model-Derived Subgroup (MDS) met all replication criteria across trials. The MDS was defined by marked depressive weight loss, a score of 2 (range 0-2) on the Hamilton Depression Rating Scale weight loss item indicating definite weight loss by selfreport and baseline body weight that is ≥76.3 kg. Within this subgroup, esmethadone demonstrated consistent and clinically large antidepressant effects across all 3 studies, with drug–placebo MADRS differences of 10.4–12.5 points and effect sizes of Cohen’s d = 0.90– 1.16 (all nominal p < 0.05). This treatment separation exceeded that observed in the full Phase 2a population by 4.7 MADRS points and was 8.2–9.5 points > in the full Phase 3 populations. The prevalence of the MDS differed substantially across trials (38% in Phase 2a versus 12–14% in Phase 3), accounting for attenuation of the treatment signal in unenriched confirmatory studies.

Conclusions

This analysis demonstrates a regulatorily aligned, model-informed enrichment methodology that emphasizes interpretability, cross-trial replication, and prospective feasibility. By grounding subgroup definitions in routinely collected clinical features and validating them across independent trials, this framework addresses key regulatory concerns surrounding post hoc subgroup analyses. Prospective application of this predictive enrichment strategy may improve trial efficiency and therapeutic signal detection for esmethadone and other novel antidepressants, consistent with FDA guidance on enrichment strategies and MIDD.