Deciphering of Cytotoxicity in Fungal-Derived Cytochalasans Using a Deep Neural Network Framework Supported by Molecular Docking and Molecular Dynamic Simulations

Cytochalasans are a structurally diverse class of fungal-derived natural products with remarkable cytotoxic and anticancer properties. However, the rarity of experimental data and the complexity of their molecular scaffolds hinder systematic pharmacological evaluation. In this study, we introduce a deep learning-based artificial neural network (ANN) framework to predict the cytotoxic activity of cytochalasans using an experimental dataset of 291 compounds. Molecular descriptors, including physicochemical properties (SwissADME, pkCSM) and structural fingerprints (RDKit, Morgan, and pharmacophore-based), were extracted and processed via dimensionality reduction (PCA, retaining 95% variance). A six-layer ANN architecture was trained and evaluated across training (70%), testing (15%), and validation (15%) sets, achieving robust predictive performance (training: accuracy = 0.97, AUC = 0.998; test: accuracy = 0.84, AUC = 0.883; validation: accuracy = 0.90, AUC = 0.893). This ANN model enables classification of compounds into active or inactive categories, further subclassifying actives based on predicted potency (strong, moderate, or weak). The effective virtual screening of 127 previously untested cytochalasan derivatives using the trained ANN illustrates the model’s practical applicability. The compounds prioritized by the ANN during virtual screening. High-confidence predictions were further validated through molecular docking studies, and molecular dynamics (MD) simulations against 15-lipoxygenase-2 (15-LOX-2). Furthermore, three derivatives were prioritized based on strong predicted activity by the deep learning model: cytochalasin Z6 (11), chaetoglobosin W (64), and hydroxy-10-phenyl-[11]-cytochalasa-13, 19-diene-1,21-dione (102). These cytochalasans emerged as the most promising candidates for cancer treatment due to their capacity to inhibit the 15-LOX-2 enzyme, as demonstrated by molecular docking and MD analyses.