Molecular Descriptors Property Prediction Using Transformer-Based Approach.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• cheminformatics

Background:

• Predicting molecular properties is crucial for drug discovery.
• Existing methods often require large labeled datasets.
• Machine learning offers potential for efficient property prediction.

Purpose of the Study:

• To develop and evaluate semi-supervised machine learning models for predicting molecular properties.
• To leverage both labeled and unlabeled data for improved model training.
• To explore efficient attention mechanisms for drug candidate prediction.

Main Methods:

• A two-stage approach: pre-training using Masked Language Model on SMILES strings and fine-tuning on downstream tasks.
• Utilizing large datasets of labeled and unlabeled SMILES strings.
• Developing a novel 3D structure-based attention score for end-to-end transformer models.

Main Results:

• The proposed semi-supervised models achieve performance comparable to state-of-the-art methods on MoleculeNet tasks.
• The novel 3D structure-based attention approach enables comparable performance to pre-trained models with reduced computational cost.
• The model successfully predicts anti-malaria drug candidates.

Conclusions:

• Semi-supervised learning effectively predicts molecular properties and aids drug discovery.
• Integrating 3D structural information offers computational efficiency without sacrificing performance.
• The developed models represent a significant advancement in machine learning for cheminformatics.