Target-aware molecule SMILES generation using a large language model with retrieval-augmented generation, multi-turn...

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• High-throughput screening is costly and time-consuming for drug discovery.
• Efficient computational methods are crucial for prioritizing drug candidates.
• Large language models (LLMs) show potential for molecular generation.

Purpose of the Study:

• To evaluate an open-weight LLM with retrieval-augmented generation (RAG) for target-aware molecule generation.
• To assess if RAG, multi-turn memory, and a drug-target interaction predictor can generate molecules without retraining.
• To explore resource-efficient computational strategies for drug discovery.

Main Methods:

• Utilized protein-ligand-pKi data from BindingDB, Davis, and KIBA for contextual guidance.
• Employed DeepPurpose for optimization signal during iterative SMILES refinement.
• Integrated retrieval-augmented generation (RAG) with multi-turn memory and a pretrained drug-target interaction predictor.

Main Results:

• Achieved a statistically significant increase in predicted pKi across multi-turn memory iterations.
• Demonstrated high novelty (100%), diversity (up to 0.882), and uniqueness (up to 1.0) in generated molecules.
• Observed trade-offs in molecular validity and drug-likeness with increased predicted affinity.

Conclusions:

• The RAG-augmented LLM framework offers a flexible and resource-efficient strategy for target-aware de novo molecular design.
• Retrieval and memory-guided refinement show promise for improving target-conditioned molecular generation.
• Further multi-objective optimization and independent validation are necessary to refine the approach.