Abstract
<jats:p>Microglia, the resident immune cells of the central nervous system, share myeloid lineage and core innate-immune signaling machinery — including Toll-like receptor 4 (TLR4) and NF-κB pathway activation — with peripheral macrophages such as the RAW 264.7 cell line commonly used in anti-inflammatory drug screening. This overlap raises a practical question for early-stage drug discovery: which compounds showing anti-inflammatory activity in macrophage assays are likely to retain that activity in microglia, and are therefore worth prioritizing for neuroinflammation-focused follow-up? We present a machine learning classifier that addresses this question directly. Bioactivity data were assembled from ChEMBL and PubChem BioAssay, filtered for inflammation-relevant readouts (TNF-α, IL-6, nitric oxide/iNOS, NF-κB, cytokines), and used to label 493 compounds as active in macrophages, microglia, both ("translatable"), or neither. Compounds were represented as Morgan/ECFP4 molecular fingerprints (2048 bits, radius 2) and used to train a Gradient Boosting classifier with class weighting to address the substantial imbalance of the translatable class (64 of 493 compounds). The model achieved an AUC-ROC of 0.811 ± 0.030 in 5-fold cross-validation and 0.803 on a held-out test set. Feature importance analysis identified fingerprint bits corresponding to chemical substructures with documented NF-κB-modulating activity — including oleanolic-acid-type pentacyclic triterpenoids, polyphenolic lignans and chalcones, and protoberberine-type alkaloids — providing mechanistic plausibility for the model's predictions without this knowledge being explicitly encoded during training. We report these results together with an explicit account of the dataset expansion process, the comparison of four class-imbalance strategies, and the limitations of the current approach, as a methodological proof of concept for translatability prediction between peripheral and central myeloid inflammatory contexts.</jats:p>