COATI: Multimodal Contrastive Pretraining for Representing and Traversing Chemical Space

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Drug discovery is a complex optimization problem within a vast chemical space.
• Generative models offer a promising approach to enhance molecular design and discovery.
• Existing methods often lack efficiency and comprehensive optimization capabilities.

Purpose of the Study:

• To develop advanced generative optimization methods for small molecule drug design.
• To integrate key features for effective multiparameter molecular optimization.
• To present a novel model and algorithm for accelerated therapeutic inference.

Main Methods:

• Collected a large dataset of 2 billion quantitative binding measurements.
• Developed COATI, a pretrained, multimodal encoder-decoder model using contrastive learning.
• Implemented a novel metadynamics algorithm for generative optimization.

Main Results:

• COATI demonstrates universal molecular embedding properties: fixed-dimension, invertibility, autoencoding, accurate regression, and low computational cost.
• The generative optimization successfully designed molecules with desired potency, solubility, and drug-likeness for carbonic anhydrase.
• The approach enables simultaneous optimization of multiple molecular properties.

Conclusions:

• Integrated generative models and large quantitative datasets can significantly improve small molecule drug design.
• COATI provides a powerful, versatile tool for molecular representation and downstream tasks.
• This work paves the way for fully integrated generative molecular design and optimization pipelines.