INTERPRETABLE PREDICTION OF MOLECULAR BIOACTIVITY FROM CHEMICAL STRUCTURE PATTERNS

Abstract

Molecular bioactivity prediction supports compound prioritization by linking measurable chemical characteristics with expected biological responses, yet predictive accuracy alone provides limited value when the structural basis of a model decision remains unclear. This study developed an interpretable machine-learning framework for predicting continuous molecular bioactivity from physicochemical descriptors and explicit chemical structure patterns. A structured set of 1,200 molecules was analyzed using molecular weight, lipophilicity, topological polar surface area, hydrogen-bonding properties, molecular flexibility, ring characteristics, atom composition, and nine functional or substructural indicators. Elastic Net, support vector regression, random forest, and gradient boosting models were optimized using five-fold cross-validation and evaluated on a predefined independent test set. Predictive performance was quantified using coefficient of determination, root mean squared error, and mean absolute error, while Shapley Additive Explanations were applied to the strongest tree-based model to identify influential molecular determinants. The analysis demonstrated measurable variation in bioactivity across physicochemical and structural features and showed that nonlinear ensemble learning captured structure-bioactivity relationships more effectively than simpler alternatives. Global explanations further identified a concentrated set of molecular characteristics driving predictions and clarified their directional contributions. The framework therefore combines prediction with chemically interpretable evidence, supporting transparent computational screening while simulated bioactivity values require validation against experimentally measured endpoints before biological generalization. Transparency remained central.