Predicting condition-aware drug-induced transcriptional responses via a latent diffusion model

Area of Science:

• Computational biology
• Genomics
• Pharmacology

Background:

• Accurate prediction of drug-induced transcriptional responses is crucial for drug discovery and precision medicine.
• Existing computational models often fail to generalize to new conditions and neglect biological characteristics.
• There is a need for advanced models that can effectively predict condition-aware gene expression changes.

Purpose of the Study:

• To develop and evaluate a novel latent diffusion model for predicting condition-aware drug-induced transcriptional responses.
• To improve the generalization performance of computational models for unseen conditions in drug discovery.
• To enhance the biological relevance and accuracy of predicted gene expression profiles.

Main Methods:

• A latent diffusion model was developed, combining a variational autoencoder (VAE) with a diffusion process.
• The VAE compresses gene expression profiles into a latent space for the diffusion model.
• The model incorporates multiple perturbation conditions (cell line, compound, dose, time) to improve predictions.

Main Results:

• The model achieved state-of-the-art reconstruction performance (Pearson's r=0.870, R2=0.739) on unseen compounds.
• It demonstrated superior preservation of gene-gene correlations and biological relevance, validated by pathway analysis.
• Latent space analysis confirmed the model's ability to capture cell line identity and continuous variations.

Conclusions:

• Latent diffusion models are effective tools for modeling transcriptional responses in drug discovery.
• The proposed model enhances prediction accuracy and generalization for condition-aware gene expression.
• This approach holds significant potential for advancing precision medicine and therapeutic development.