MolPACL: Molecular Property Prediction Based on Prompt Augmentation and Contrastive Learning

Area of Science:

• Computational chemistry
• Machine learning
• Artificial intelligence

Background:

• Large language models (LLMs) offer potential for molecular representation learning from text.
• Current contrastive learning methods often use graph or SMILES augmentations that can distort chemical structures.
• There is a need for methods that leverage high-level chemical semantics while maintaining molecular integrity.

Purpose of the Study:

• To develop MolPACL, a prompt-augmentation-based supervised contrastive learning framework.
• To incorporate high-level chemical semantics into molecular representations.
• To preserve molecular identity during the learning process.

Main Methods:

• MolPACL generates semantically consistent prompt views from SMILES strings and physicochemical descriptors.
• It utilizes diverse templates and lightweight lexical perturbations for prompt generation.
• A supervised objective based on Soft Nearest Neighbor loss is employed for training.

Main Results:

• MolPACL achieves strong performance on MoleculeNet benchmarks for classification and regression tasks.
• The framework reduces training costs.
• It requires no additional molecular pretraining and uses a relatively small pretrained LLM.

Conclusions:

• MolPACL effectively learns molecular representations by incorporating high-level chemical semantics.
• The prompt-augmentation approach preserves molecular identity.
• This method offers an efficient and effective alternative for molecular representation learning.