Machine learning-augmented network pharmacology prioritizes multi-target anti-inflammatory candidates from the n-hexane fraction of Alstonia boonei

Alstonia boonei De Wild is an ethnomedicinal tree with documented anti-inflammatory activity, yet the molecular drivers of its n-hexane fraction remain incompletely resolved. Here we present a machine learning-augmented network pharmacology workflow that is explicitly hypothesis-generating, prioritizing compound-target hypotheses for experimental testing rather than asserting pharmacological causality. We docked 100 phytochemicals of the fraction against four targets selected from endpoints measured in our prior study: IL-6 (PDB 1ALU), TNF-α (6 ×82), IL-1β (8C3U), and the cyclophilin D gatekeeper of the mitochondrial permeability transition pore, PPIF (8EJX). To guard against false positives, the protocol was benchmarked against native co-crystallized inhibitors as positive controls and non-anti-inflammatory drugs as decoys, separating the classes with target-wise ROC-AUC of 0.77–0.81. Because raw affinity correlated strongly with molecular size (Pearson r = -0.61, p < 1e-10), we normalized by ligand efficiency and replaced the former arbitrary cutoff with decoy-calibrated thresholds. A random forest predicted multi-target engagement with cross-validated ROC-AUC of 0.85, retained signal when size descriptors were removed (0.81) far above a label-permuted null (0.51, p < 0.017), and regressed affinity at sub-kilocalorie error (R-squared 0.58, RMSE 0.97 kcal/mol). Cluster structure exceeded a feature-permuted null (silhouette 0.353, p = 0.010; bootstrap ARI 0.83). After efficiency normalization and prioritization, the genus-characteristic indole alkaloids, including cinegalline, macralstonine, akuammicine, mitraphylline, and vincamine, emerged as the most credible multi-target candidates rather than the bulkier terpenoids dominating raw scores. These results provide a reproducible set of testable hypotheses consistent with our prior bioactivity data.