Tengyu Li

Background: Chronic low-grade inflammation underlies numerous chronic diseases, yet the molecular targets of dietary bioactive compounds remain largely unknown. Methods: We developed a deep learning model integrating heterogeneous network propagation, stacked denoising autoencoders, and a CNN classifier to predict inflammation-related targets of food-derived compounds. The model was validated by cross-validation and on eight compounds from distinct structural categories. Avn A, Avn B, and Avn C were then input into the model to predict their anti-inflammatory targets, followed by network analysis, docking, and cellular assays for validation. Results: The model achieved a mean AUC of 0.9587 and an AUPR of 0.9590, with validation on eight structurally diverse compounds confirming its generalizability. For Avn A, Avn B, and Avn C, 118 overlapping anti-inflammatory targets were predicted. Network analysis and KEGG pathway enrichment identified ERK1/2 as a key node within the MAPK and PI3K-AKT signaling pathways, while COX-2 was prioritized for further validation based on its high prediction score. Molecular docking revealed that Avns bound at the cyclooxygenase channel entrance of COX-2 with binding energies of −7.7 to −8.0 kcal/mol, and at the ATP-binding pockets of ERK1/2 with binding energies of −7.0 to −7.7 kcal/mol. In LPS-stimulated macrophages, Avns dose-dependently inhibited COX-2 enzymatic activity and PGE2 production, attenuated ERK1/2 and c-Fos phosphorylation, and suppressed IL-6 and IL-1β expression. Conclusions: Our deep learning model successfully identified COX-2 and ERK1/2 as key mediators of the anti-inflammatory effects of Avns, revealing a dual mechanism involving COX-2 inhibition and ERK/c-Fos pathway blockade. This study provides new evidence for the molecular targets of Avns and demonstrates the utility of deep learning-based approaches for target discovery of dietary bioactive compounds.

Compound–target interaction (CTI) prediction plays a critical role in drug discovery and the functional study of food-derived bioactive compounds. However, traditional experimental methods for CTI identification are limited by high costs, long cycle times, and high false-positive rates, highlighting an urgent need for more efficient approaches. Machine learning (ML) has become a revolutionary tool to address these challenges. In this review, we focus on recent developments in ML-based CTI prediction. We first systematically outline the commonly used public databases and feature extraction methods for both compounds (molecular fingerprints) and proteins (sequence-derived features), followed by elaborating on four types of ML approaches, including classical supervised learning, matrix factorization, graph topology-based inference, and deep neural network frameworks. In particular, this review explores the emerging application of these computational approaches in identifying targets of food-derived bioactive compounds, underscoring its significant potential to advance functional food research. Moreover, we analyze key challenges, such as limited model interpretability, high data dependency, and insufficient multi-source information integration, and put forth future prospects to improve the prediction of food-derived CTIs, thereby facilitating their application in functional food research.