Exploring the potential of artificial neural networks in predicting physicochemical characteristics of anti-biofilm compounds from 2D and 3D structural information

The application of machine learning has revolutionized drug discovery process, enabling more accurate prediction of physicochemical properties. In this study, we utilized machine learning models including Artificial Neural Networks (ANN), XGBoost, and AdaBoost to predict properties of selected anti-biofilm drugs. Both 2D and 3D structural analyses were conducted for deeper understanding. By using neighborhood degree sum-based topological indices as an input feature variable, we predicted physicochemical properties and compared them with experimental values. SHAP analysis has identified the most influential indices in the prediction process that have strong correlation with properties and machine learning models with more accurate predictive capabilities. Furthermore, we evaluated our model’s predictions using multiple metrics to ensure robust assessment. Specifically, we employed Mean Squared Error, Root Mean Squared Error and Mean Absolute Error to measure prediction accuracy. Additionally, the R-squared metric was used to gauge the model’s explanatory power, indicating how well our model captures the variance in the target variables. This interdisciplinary approach not only accelerates the screening process, but also increases the accuracy of predictions, thereby facilitating the rapid development of effective anti-biofilm drugs.