A diffusion model conditioned on compound bioactivity profiles for generating high-content images.

Area of Science:

• Computational Biology
• Drug Discovery
• Artificial Intelligence in Chemistry

Background:

• High-content imaging (HCI) captures compound effects on cells, aiding understanding of cellular systems.
• Generative imaging models can predict chemical perturbation outcomes computationally, offering speed and scale.

Purpose of the Study:

• To develop Profile-Diffusion (pDIFF), a generative method for creating HCI data from in silico bioactivity profiles.
• To evaluate pDIFF's effectiveness in predicting phenotypic outcomes and improving virtual hit expansion.

Main Methods:

• Developed pDIFF, a profile-to-image latent diffusion model conditioned on in silico bioactivity profiles.
• Trained and evaluated pDIFF using Cell Painting HCI data for 3750 molecules with associated bioactivity profiles.
• Compared pDIFF against a baseline model using substructural molecular descriptors.

Main Results:

• pDIFF generated improved visual depictions of phenotypic responses for compounds structurally dissimilar to training data.
• pDIFF demonstrated statistically significant improvements in virtual hit expansion accuracy compared to structure-based and profile-based methods.
• The model outperformed generative imaging models relying solely on substructural descriptors.

Conclusions:

• pDIFF effectively translates in silico bioactivity profiles into realistic high-content images of cellular responses.
• The methodology shows significant potential to accelerate and enhance the search for novel phenotypically active molecules.
• pDIFF offers a powerful tool for in silico prediction and optimization in drug discovery pipelines.