Molecular Contrastive Pretraining with Collaborative Featurizations.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Molecular pretraining learns representations from unlabeled data for chemistry tasks.
• Current methods use various featurizations like SMILES, graphs, or 3D structures.
• The impact of featurization choice on pretraining is understudied.

Purpose of the Study:

• To investigate how different molecular featurizations impact pretraining.
• To develop a pretraining framework that leverages complementary featurizations.
• To improve performance on molecular property prediction tasks.

Main Methods:

• Analyzed information conveyed by different featurization techniques through case studies (chirality, aromatic rings).
• Proposed the MOlecular pretraining with COllaborative featurizations (MOCO) framework.
• Evaluated MOCO on diverse molecular property prediction tasks.

Main Results:

• Different featurizations capture distinct chemical information.
• The MOCO framework integrates multiple complementary featurizations.
• MOCO surpasses existing models relying on single or dual featurizations.

Conclusions:

• Collaborative featurizations offer a more comprehensive approach to molecular pretraining.
• The MOCO framework represents a significant advancement in computational chemistry and drug discovery.
• Leveraging diverse molecular representations is key to enhancing predictive accuracy.