Transcriptome-conditioned molecule generation via gene interaction-aware fragment modeling with a GPT-based...

Area of Science:

• Computational chemistry
• Genomics
• Drug discovery

Background:

• Phenotype-driven drug discovery uses cellular responses to guide therapeutic molecule design.
• Transcriptomics provides data on gene expression changes due to chemical stimuli, enabling links between molecular generation and cellular phenotypes.
• Challenges exist in connecting transcriptomic data to chemical structure generation due to complex gene interactions and chemical feasibility.

Purpose of the Study:

• To develop a novel generative model, GGIFragGPT, that integrates transcriptomic data with gene-gene interaction embeddings for fragment-based molecular generation.
• To guide the molecular generation process using biologically relevant genes highlighted by cross-attention mechanisms.
• To generate chemically feasible, novel, and diverse molecules aligned with biological context.

Main Methods:

• Developed GGIFragGPT, a generative model using an autoregressive transformer architecture.
• Integrated transcriptomic perturbation profiles with biologically informed gene-gene interaction embeddings.
• Employed cross-attention mechanisms to link gene relevance to molecular generation.

Main Results:

• GGIFragGPT generated chemically feasible, novel, and diverse molecules.
• Generated compounds were aligned with the biological context indicated by transcriptomic data.
• Gene-level interpretability analysis identified key target genes, validating the model's biological relevance.
• Case studies demonstrated the generation of plausible inhibitors, such as for CDK7.

Conclusions:

• The integration of biological insights, particularly transcriptomic data, into chemical generation processes is promising for phenotype-driven drug discovery.
• GGIFragGPT offers a novel approach to link gene expression profiles to the design of targeted therapeutic molecules.
• This work highlights the potential of AI in accelerating the discovery of novel drug candidates based on cellular phenotypes.