Enhancing Conditional Molecular Generation With Pretrained SMILES Transformer and Contrastive Representation Learning

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Drug discovery and materials science

Background:

• Predicting molecular structures with desired properties and reaction conditions is crucial for drug discovery and materials science.
• Simplified Molecular Input Line Entry System (SMILES) is widely used for molecular generation but its inherent redundancy can confuse models.
• Existing methods struggle with ambiguity in feature mapping from SMILES representations.

Purpose of the Study:

• To develop a method that captures invariant, high-level features from SMILES representations.
• To guide molecular generation models in learning the inherent structure of SMILES.
• To improve the accuracy and validity of generated molecular structures.

Main Methods:

• Proposed a pretrained contrastive learning-based SMILES Transformer Encoder (Contras-STE).
• Employed Gumbel-Softmax sampling for differentiable conversion and computed InfoNCE loss for similarity quantification.
• Tested Contras-STE on MoleculeNet benchmarks and real-world tasks like OSDAs prediction and zeolite property prediction.

Main Results:

• Contras-STE outperformed existing fingerprint-based and RNN-based methods on MoleculeNet benchmarks in most cases.
• The SMILES generation model integrated with Contras-STE significantly improved the validity and novelty rates of generated SMILES.
• Demonstrated effectiveness in predicting molecular properties under specific synthesis conditions and for zeolite production.

Conclusions:

• Contras-STE effectively captures essential features from SMILES, addressing ambiguity issues.
• The proposed method enhances the performance of molecular generation models, leading to more valid and novel structures.
• Contras-STE shows promise for advancing drug discovery and materials science through improved molecular design.