Unveiling the Immunomodulatory Potential of Medicinal Plants: A Machine Learning Framework Coupled with Experimental...

Medicinal plants are known for their beneficial health effects through their ability to modulate the immune system, offering promising therapeutic options for various diseases. However, traditional methods for identifying these properties can be time-consuming, labor-intensive, and expensive. This study aims to (i) use machine learning (ML) to characterize the immunomodulatory effects of medicinal plants and (ii) validate the ML model via in vitro assays to assess the predicted immunomodulatory effects. To achieve this goal, data on medicinal plants were collected from the literature, structured and encoded using pre-trained language models to generate vector representations capturing semantic relationships among botanical, chemical, and microbiological entities. Several classifiers were benchmarked through a rigorous cross-validation strategy with optimal threshold tuning, and the best-performing model (logistic regression on BiomedBERT embeddings) reached a mean cross-validated F1-score of 0.787 ± 0.007 and, on the held-out test set (n = 37), an accuracy of 0.757, an F1-score of 0.743 and an AUC of 0.835; three of the four external validation plants were correctly classified. The results were validated experimentally using selected medicinal plants by assessing the modulation of inducible nitric oxide synthase (iNOS) and key inflammatory cytokine genes, including IL-1β, IL-6, and TNF-α. C. spinosa extract up-regulated IL-1β and TNF-α and down-regulated IL-6 transcripts in peritoneal macrophages, while E. fragilis extract induced a significant down-regulation of these cytokine genes and of the expression of iNOS. These preliminary results suggest that pre-trained biomedical embeddings provide a useful prioritization tool for investigating the immunomodulatory properties of medicinal plants.