Learned Conformational Space and Pharmacophore Into Molecular Foundational Model

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Large-scale chemical pre-trained models advance molecular structure-function understanding.
• Existing models often lack explicit integration of chemical knowledge like conformation and pharmacophore information.
• Dynamic molecular behavior and pharmacophore complementarity are crucial for accurate modeling.

Purpose of the Study:

• To develop a molecular foundational model incorporating conformational and pharmacophore information.
• To regularize the representation space using these chemical features.
• To create a unified framework for molecular representation learning and generation.

Main Methods:

• Developed a molecular foundational model with an Ouroboros-like architecture.
• Utilized a graph neural network for encoding molecular graphs into 1D vectors.
• Employed an autoregressive Transformer module for reconstructing SMILES sequences.
• Integrated conformational-space and pharmacophore-similarity projections during pre-training.

Main Results:

• The model effectively captures complex molecular relationships.
• Demonstrated improved performance in similarity-based virtual screening.
• Showcased capabilities in targeted poly-pharmacology design and chemical property prediction.
• Enabled directed molecular optimization within a unified latent space.

Conclusions:

• The proposed model offers a flexible and extensible framework for molecular foundational modeling.
• Integrating explicit chemical knowledge enhances representation learning and generative capabilities.
• The model shows significant potential for addressing practical challenges in drug discovery and chemical design.