Predicting Antiviral Inhibitory Activity of Dihydrophenanthrene Derivatives Using Image-Derived 3D Discrete Tchebichef Moments: A Machine Learning-Based QSAR Approach

Making advancements in Quantitative Structure-Activity Relationship (QSAR) modeling is crucial for predicting biological activities in new compounds. Traditional 2D-QSAR and 3D-QSAR methods often face challenges in terms of computational efficiency and predictive accuracy. This study introduces a machine learning approach using 3D Discrete Tchebichef Moments (3D-DTM) to address these issues. The 3D-DTM method offers efficient computation, robust descriptor generation, and improved interpretability, making it a promising alternative to conventional QSAR techniques. By capturing global 3D shape information, this method provides better representation of molecular interactions essential for biological activities. We applied the 3D-DTM model to a dataset of 46 molecules derived from the Dihydrophenanthrene scaffold, screened against the enzymatic activity of 3-chymotrypsin-like protease, a key antiviral target. Principal Component Analysis and k-means clustering refined descriptors, followed by stepwise Multiple Linear Regression (step-MLR), Partial Least Squares Regression (PLS-R), and Feed-Forward Neural Network (FFNN) techniques for 3DTMs-QSAR model development. The results showed high correlation and predictive accuracy, with significant validation from internal and external tests. The step-MLR model emerged as the optimal method due to its balance of predictive power and simplicity. Validation through y-Randomization and applicability domain analysis confirmed the model’s robustness. Virtual screening of 100 novel compounds identified 32 with improved pIC50 values. This study highlights the potential of 3D-DTMs in QSAR modeling, providing a scalable and reliable tool for computational chemistry and drug discovery. A user-friendly software tool was also developed to facilitate 3D-DTM extraction from input 3D molecular images.