Machine Learning QSPR and Materials Informatics

J-OCTA’s MI-Suite features a QSPR (Quantitative Structure–Property Relationship) function for predicting the relationship between molecular structure and physical properties using machine learning. Descriptors are extracted from molecular structures described in SMILES notation, and methods such as graph convolutional networks (GCNs) are used to predict properties including density, glass transition temperature, and dielectric constant. All operations can be executed in Python scripts, from modeling and data generation to training and prediction. In addition to pre-trained models, users can build their own QSPR models using their own property data, enabling flexible property prediction tailored to specific objectives and contributing to efficient materials development.

Use Cases Highlights

Structure–property correlation using deep learning
Construction of predictive models using molecular structure information (SMILES) as input
Improvement of generalization performance through machine learning

Structure–property correlation analysis using deep learning

An automated workflow from molecular modeling to property prediction using J-OCTA’s modeling API is shown. Python scripts allow efficient generation of large datasets.

Automation of molecular modeling and property prediction using Modeling API

Construction of predictive models using molecular structure information (SMILES) as input

The concept of machine learning-based QSPR and a structure–property correlation learning model are shown. Molecular descriptors are used to predict property values, and training is performed using neural networks.

Construction of structure–property correlation model using machine learning QSPR

Improvement of generalization performance by machine learning

The density prediction results of amorphous polymers are shown. The left shows machine learning QSPR predictions, and the right shows predictions by the conventional Bicerano method, indicating that machine learning correlates more strongly with experimental values.