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

High-content imaging (HCI) provides a rich snapshot of compound-induced phenotypic outcomes that augment our understanding of how compounds affect cellular systems. Generative imaging models for HCI provide a route towards anticipating the phenotypic outcomes of chemical perturbations in silico at unprecedented scale and speed. Here, we developed Profile-Diffusion (pDIFF), a generative method leveraging a profile-to-image latent diffusion model conditioned on in silico bioactivity profiles to generate high-content images displaying the cellular outcomes induced by compound treatment. We trained and evaluated a pDIFF model using high-content images from a Cell Painting assay profiling 3750 molecules (3375 training compounds and 375 held-out compounds) with corresponding in silico bioactivity profiles. Using the held-out set we demonstrate that pDIFF provides improved visual depictions of phenotypic responses of compounds that are structurally dissimilar to training compounds, compared to a baseline profile-to-image latent diffusion model trained on substructural molecular descriptors only. In a virtual hit expansion scenario, pDIFF yielded statistically significant improvement in expansion outcomes as measured by nearest-neighbor retrieval accuracy, compared to expansions based on compound structural representations, bioactivity profiles, and generative imaging models based only on substructural molecular descriptors, thus showcasing the potential of the methodology to speed up and improve the search for novel phenotypically active molecules.