GEM-GPT Enables Personalized Cell Type-Resolved Therapeutic Design for Systems Pharmacology

Area of Science:

• Computational biology
• Artificial intelligence in medicine
• Pharmacology

Background:

• Current generative AI for drug discovery often uses simplified models, struggling with complex, heterogeneous diseases.
• A gap exists in AI tools tailored for systems pharmacology and personalized drug design.
• Omics data offers a path to understanding disease complexity, but integrating it with AI for drug design is challenging.

Purpose of the Study:

• Introduce GEM-GPT, a novel transcriptomics-based molecule generation framework.
• Enable personalized therapeutic compound design for cell type-specific disease states.
• Address limitations in current AI-driven drug discovery for complex diseases.

Main Methods:

• Developed GEM-GPT, a framework integrating a single-cell RNA sequencing (scRNA-seq) foundation model with a molecular GPT model.
• Employed a biology-inspired deep fusion architecture to model cell type-specific gene-chemical interactions.
• Validated GEM-GPT's performance against state-of-the-art baselines in molecule generation.

Main Results:

• GEM-GPT successfully generated distinct molecules tailored to different cell types.
• The framework demonstrated robust generalization to novel cell types.
• Achieved superior performance compared to existing state-of-the-art methods.
• Identified potential therapeutic compounds and FDA-approved drugs for opioid use disorder (OUD) in a case study.

Conclusions:

• GEM-GPT represents a significant advance in AI-driven systems pharmacology.
• The framework effectively bridges single-cell omics data with molecular generation for personalized drug design.
• GEM-GPT supports systems-aware therapeutic strategies for complex and chronic diseases.