Exploring a Small Molecule Property Prediction Model with Optimal Comprehensive Performance through Multi-Objective Optimization Algorithms

The evolution of artificial intelligence has given rise to numerous machine learning (ML) models for predicting the structural properties of materials, expediting the process of new material development. However, many of these models have parameters set based on empirical knowledge, lacking validation for optimality or multi-objective fitness. In this study, we employ a fusion modeling approach that integrates physical mechanisms with ML to construct a neural network model for predicting multiple properties of small molecules. We utilize the non-dominated sorting differential evolution (NSDE) algorithm to explore the dimensions of the network input data and the structure of the network. The goal is to identify feasible solutions that simultaneously optimize the predictive performance of the model and the training time cost. Experimental results show that the trained model determined by multi-objective optimization (MOP) not only exhibits outstanding predictive capabilities for multiple properties of small molecules but also achieves a balance between predictive performance and training time cost, showcasing excellent comprehensive performance.