Research on Molecular Property Prediction Model Base on Random Forest and LSTM Network

In today's scientific research, the application of advanced technologies such as machine learning, deep learning and Bayesian optimization is increasingly widespread. They are integrated into the theoretical laws and experimental evolution in the field of chemistry, which greatly improves the intelligence level of chemical science. This study aims to explore the molecular property prediction model based on random forest and Long Short-Term memory (LSTM) network, in order to improve the prediction accuracy and enhance the generalization ability of the model. Firstly, this study preprocessed the experimental case data of the physicochemical properties of 200 thousand chemical molecules, including missing value checking and outlier handling, and used the Z-score method to effectively identify and remove 1132 outliers. Subsequently, the nonlinear relationship between molecular id and other variables was established by visual analysis and variable relationship analysis. In the feature selection stage, the SelectKBest method is used to select the 10 most relevant features to the target variable, which effectively reduces the complexity of the model and improves the computational efficiency. Next, two models are constructed in this study: random forest model and LSTM model. The random forest model generates decision trees by Bootstrap sampling and CART algorithm, while the LSTM model uses its unique gating mechanism to process time series data and capture long-term dependencies. In addition, this study also adopted a model optimization strategy, combined with the advantages of LSTM and random forest, extracted features through the deep learning model and used the random forest classifier for accurate classification prediction. Finally, the results show that the model combining LSTM and random forest performs well in the classification task, with an accuracy of 0.9986, which is significantly better than the performance of the random forest model alone. The results of this study not only provide a new methodology for chemical molecular property prediction, but also provide a valuable reference for complex data analysis and modeling in other fields.