FAME: Fragment-based Conditional Molecular Generation for Phenotypic Drug Discovery.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Genomics and bioinformatics

Background:

• De novo molecular design is crucial but challenging in drug discovery due to complex chemical spaces.
• Existing deep generative models often focus on molecular distribution learning or target-based design, limiting real-world applications.
• Phenotypic molecular design, especially gene expression-based, offers advantages for discovering first-in-class drugs.

Purpose of the Study:

• To propose the first deep graph generative model (FAME) for phenotypic molecular design, specifically using gene expression profiles.
• To address the challenges of learning the molecular generation distribution from noisy gene expression data.
• To generate novel molecules with desired biological activity through a gene expression-driven approach.

Main Methods:

• Developed a gene expression denoising (GED) model using a contrastive objective function to reduce noise in gene expression data.
• Designed FAME, a conditional variational autoencoder, to generate molecules from denoised gene expression profiles.
• Employed a fragment-based autoregressive generation strategy within FAME to enhance molecular generation.

Main Results:

• FAME successfully generates novel molecules with high validity and desired biological activity from gene expression data.
• The proposed GED model effectively reduces noise in gene expression data, improving downstream molecular generation.
• FAME outperforms existing SMILES-based and graph-based deep generative models in phenotypic molecular design tasks.

Conclusions:

• FAME represents a significant advancement in deep graph generative models for phenotypic molecular design.
• The fragment-based generation strategy and denoising approach are key to FAME's success.
• The developed noise reduction technique for omics data has broader implications for phenotypic drug discovery.