Abstract
<jats:p> Structural determinants of bias at the residue level have yet to be fully identified in the µ-opioid receptor. This study applies machine learning (ML) methods to docking derived features using 15 training ligands across 4 chemotypes. This included fentanyl analogues and emerging nitazene-class compounds on 3 different receptor states, where Ile324 <jats:sup>7.39</jats:sup> and its state-dependent contacts were identified as primary features — inclusive of a Mann-Whitney U=0.0, p = 0.0003 on the active/inactive ratio (r = -0.887, p <jats:sub>adj</jats:sub> =0.0022). Quantum mechanical validation with IGMH was then performed (ORCA/Multiwfn) on 10 ligands which revealed a decision boundary consistent with experimental bias direction (Mann-Whitney U=0.0, p=0.0167), with a strong graded correlation preserved among pharmacologically committed ligands (Pearson r=−0.829, p=0.006, n=9). Computational perturbation of Ile324 through Ala substitution reduced these relationships, consistent with the basis of the ML finding, and suggests Ile324’s conformational selectivity as a structural correlate of MOR signalling bias. A chemotype-bias correlation within the dataset represents a limitation that cannot be fully resolved without more bias data for nitazene G-protein biased or fentanyl β-arrestin biased ligands, though partial evidence against pure scaffold identification is provided by out-of-sample predictions and ensemble-level contact analysis. </jats:p>