Improving VAE based molecular representations for compound property prediction.

Area of Science:

• Chemoinformatics
• Machine Learning
• Computational Chemistry

Background:

• Collecting labeled data for chemoinformatics tasks is costly and time-consuming.
• Machine learning models can learn molecular representations from large unlabeled datasets for transfer learning.
• Variational autoencoders (VAEs) are used for chemical property prediction and molecular generation.

Purpose of the Study:

• To improve chemical property prediction performance of machine learning models.
• To enhance representations learned by VAEs by incorporating correlated molecular descriptors.

Main Methods:

• Proposed a method to incorporate additional information on correlated molecular descriptors into VAE representations.
• Verified the method on three chemical property prediction tasks.
• Explored the impact of descriptor number, descriptor-property correlation, and dataset size.

Main Results:

• The proposed method successfully improved chemical property prediction performance.
• Performance gains were related to the number and correlation of incorporated descriptors.
• A relationship was found between model performance and the distance between datasets in the representation space.

Conclusions:

• Incorporating correlated molecular descriptors into VAEs is an effective strategy to boost chemical property prediction.
• The findings provide insights into optimizing transfer learning in chemoinformatics.
• Understanding dataset relationships in learned representation space is crucial for model performance.