Integrating machine learning and SHAP analysis to advance the rational design of benzothiadiazole derivatives with tailored photophysical properties

Abstract

The design of organic fluorophores with tailored optical properties is critical for advancing technologies in optoelectronics and bioimaging. Among them, 2,1,3-benzothiadiazole (BTD) derivatives stand out due to their strong π-conjugation, high fluorescence efficiency, and ability to form ordered structures. However, rationally predicting their photophysical properties remains challenging due to the complex interplay between substituent effects, solvent polarity, and electronic transitions. Traditional methods relying on quantum chemistry are accurate but computationally expensive, limiting their scalability for high-throughput compound screening. To address this gap, this work proposes a data-driven strategy combining machine learning (ML) and explainable artificial intelligence to accelerate the prediction and interpretation of photophysical properties in BTD derivatives. The primary objective was to develop robust ML models for predicting maximum absorption and emission wavelengths and to derive mechanistic insights via SHapley Additive exPlanations (SHAP). The methodology involved curating a dataset of BTD derivatives, generating molecular descriptors through Morgan fingerprints, and training models using Random Forest, LightGBM, and XGBoost algorithms. Model validation employed 10-fold crossvalidation and y-randomization, with performance assessed via R2, Q2, MAE, and RMSE metrics. The results demonstrated that ML models achieved high predictive accuracy, with the Random Forest model outperforming others for both photophysical endpoints. SHAP analysis identified electron-donating groups, particularly tertiary amines, and solvent polarity as key determinants of optical behavior. The deployed web tool enables real-time property prediction and substructure attribution, thus bridging computational chemistry and practical compound design. These findings underscore the potential of interpretable ML to guide the rational design of organic materials, reducing reliance on expensive calculations and fostering accessibility for chemists and material scientists.