Rag2Mol: structure-based drug design based on retrieval augmented generation.

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Artificial Intelligence

Background:

• Drug discovery faces challenges in identifying compounds with optimal properties.
• Structure-based drug design (SBDD) is promising but limited by data biases and lack of synthetic accessibility.
• AI offers advancements but current SBDD models often disconnect from practical drug discovery.

Purpose of the Study:

• To develop novel AI-driven methodologies for designing small molecules tailored to specific 3D protein pockets.
• To address limitations of existing SBDD models, including data biases and synthetic accessibility.
• To identify potent drug candidates for challenging therapeutic targets.

Main Methods:

• Exploration of two retrieval-augmented generation methodologies: Rag2Mol-G and Rag2Mol-R.
• Rag2Mol-G: Generates molecules and searches a database for purchasable, similar compounds.
• Rag2Mol-R: Generates molecules by creating new ones from a database of existing molecules that fit a 3D pocket.

Main Results:

• Rag2Mol methods consistently produced drug candidates with superior binding affinities and drug-likeness.
• Rag2Mol-R demonstrated broader chemical landscape coverage and more precise targeting than advanced virtual screening models.
• Both workflows successfully identified promising inhibitors for protein tyrosine phosphatases PTPN2, a previously undruggable target.

Conclusions:

• The developed Rag2Mol framework represents a significant advancement in AI-driven SBDD.
• The methodologies provide a highly extensible platform for integrating diverse SBDD approaches.
• This work advances the design of small molecules with improved drug-like properties and therapeutic potential.