Data efficient molecular image representation learning using foundation models

Area of Science:

• Chemistry
• Artificial Intelligence
• Materials Science

Background:

• Deep learning (DL) in chemistry is advancing, but faces challenges with limited labeled data and feature extraction.
• Molecular representation learning (MRL) addresses these by separating feature extraction and property prediction.
• Current MRL models are typically trained from scratch, limiting their efficiency.

Purpose of the Study:

• To investigate the utility of foundation models as a starting point for MRL.
• To develop a novel MRL framework, MoleCLIP, leveraging a vision foundation model.
• To evaluate MoleCLIP's performance against state-of-the-art models and its robustness to distribution shifts.

Main Methods:

• Utilized OpenAI's CLIP, a vision foundation model, as the backbone for MoleCLIP.
• Trained MoleCLIP for molecular image representation learning.
• Benchmarked MoleCLIP on standard datasets and homogeneous catalysis data.

Main Results:

• MoleCLIP achieved performance comparable to state-of-the-art models with significantly less pretraining data.
• MoleCLIP demonstrated superior performance on homogeneous catalysis datasets.
• The framework showed robustness to distribution shifts, enabling effective adaptation to diverse tasks.

Conclusions:

• Foundation models offer an advantageous approach for developing efficient MRL models.
• MoleCLIP represents a significant advancement in molecular representation learning, requiring less data and showing improved performance.
• This work highlights the potential of general foundation models to drive innovation in synthetic chemistry and molecular property prediction.