LaMGen: LLM-based 3D molecular generation for multi-target drug design

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence

Background:

• Current multi-target drug design methods lack biological context and generalizability.
• Ligand-based approaches are limited to specific target pairs, hindering complex disease treatment.

Purpose of the Study:

• Introduce LaMGen, a general-purpose framework for multi-target drug design.
• Leverage large language models (LLMs) for accurate and efficient drug candidate generation.

Main Methods:

• Utilized MTD2025 dataset with quantum-accurate molecular conformations and multi-target associations.
• Integrated ESM-C protein embeddings, rotation-aware ligand tokens, and TriCoupleAttention module.
• Developed a large language model (LLM)-powered framework for direct generation of energy-favorable conformations.

Main Results:

• LaMGen demonstrated superior performance over diffusion-based models in independent benchmarks.
• Generated molecules with high conformational plausibility in an average of 0.44 seconds.
• Successfully reproduced known active molecules and generated novel candidates with enhanced binding affinities.

Conclusions:

• LaMGen offers a powerful and generalizable approach for multi-target drug design.
• The framework achieves quantum-level accuracy in predicting molecular conformations.
• LaMGen accelerates drug discovery by efficiently generating novel, high-affinity drug candidates.