A fingerprints based molecular property prediction method using the BERT model.

Area of Science:

• Computational chemistry
• cheminformatics
• drug discovery

Background:

• Molecular property prediction (MPP) is crucial for drug discovery and repositioning.
• Deep learning models analyze molecular features from various representations for MPP.
• Existing methods require effective feature extraction from molecular data.

Purpose of the Study:

• To propose a novel molecule sequence embedding and prediction model for MPP tasks.
• To introduce Fingerprints-BERT (FP-BERT) for semantic representation of compound fingerprints.
• To enhance the accuracy of molecular property prediction through a hybrid deep learning approach.

Main Methods:

• Pre-trained a bi-directional encoder representations from Transformers (BERT) encoder for compound fingerprints, creating Fingerprints-BERT (FP-BERT) via self-supervised learning.
• Inputted FP-BERT encoded molecular representations into a convolutional neural network (CNN) for higher-level feature extraction.
• Utilized a fully connected layer for final classification or regression in distinct MPP tasks.

Main Results:

• The proposed FP-BERT model demonstrated high prediction performance across various molecular property prediction tasks.
• Achieved superior results in both classification and regression tasks compared to baseline models.
• Validated the effectiveness of the combined BERT and CNN architecture for molecular property prediction.

Conclusions:

• The FP-BERT model offers a powerful and effective approach for molecular property prediction.
• This method significantly advances capabilities in drug discovery and repositioning through improved predictive accuracy.
• The hybrid deep learning strategy provides a robust framework for complex cheminformatics challenges.