Prediction of Molecular Critical Properties Based on Boruta Feature Selection: Balancing Accuracy, Applicability and Explainability
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The critical point as the endpoint of vapor-liquid equilibrium for a pure substance and play a pivotal role in the calculation of thermophysical properties. For Quantitative Structure-Property Relationship modeling, we used the Boruta algorithm to extract the optimal feature subset from 2705 initial features and applied regression algorithms including Gaussian Processes Regression, Random Forest and Support Vector Regression. Comparisons with GNN and other models demonstrate the good applicability of the current model. All models were tested on experimental data of 148 new molecules. The AARD for the predictions of Tc and Pc were below 2.5% and 5.7%, respectively, surpassing traditional group contribution methods and recent machine learning models. Furthermore, SHAP was employed to conduct an interpretability analysis of feature importance, and the top k features were selected for modeling. The results indicate that the features screened by Boruta yield superior modeling performance. This validates the compatibility of the Boruta feature selection algorithm with SHAP-derived feature importance. Finally, the UMAP-based visualization analysis indicates that the selected feature set can effectively distinguish high and low Critical Points, validating the effectiveness of the Boruta algorithm in feature selection.
Fields of Study & Subjects
Machine Learning in Materials Science
Phase Equilibria and Thermodynamics
Computational Drug Discovery Methods
Publication details
- Journal / source
- ChemRxiv
- Publication date
- 1 April 2026
- Language
- EN
- Document type
- Preprint