pDIFF

Cell Painting and related high-content imaging assays capture a compound's effect on cells as a rich morphological fingerprint, one of the most informative readouts available for mechanism-of-action work and phenotypic screening. They are also physically bounded: an assay can only be run on molecules that have been synthesized and plated, a vanishing fraction of accessible chemical space. Generative models that render the images a compound would produce offer a way past that bound, but the ones built so far condition on chemical structure — fingerprints or graphs — and extrapolate poorly to chemotypes unlike anything in their training set, which is precisely the regime a screening campaign cares about.

pDIFF (Profile-Diffusion), from Novartis Biomedical Research, changes what the generator is told. Instead of conditioning the denoising network on a molecular structure representation, it conditions on an in silico bioactivity profile: a vector of predicted potencies across thousands of biochemical and cellular assays. Two molecules with unrelated scaffolds that hit the same targets end up with similar conditioning vectors, so the model can anticipate a phenotype for chemical matter unlike anything it has seen. In the original Stable Diffusion the conditioning signal is a text embedding; here the model learns what the authors describe as the language of compound bioactivity.

The work was posted to bioRxiv in October 2024 and revised that December. The catalog's PPDiff is an unrelated protein-binder diffusion model from a different group.

#Key Features

• Bioactivity-profile conditioning: The denoising network is steered by predicted assay activity rather than by structure, which lets predictions hold up for compounds chemically distant from the training set.
• Latent diffusion over microscopy: Images are compressed by a pretrained variational autoencoder, diffusion runs in that latent space, and the decoder reconstitutes a full image, keeping training and sampling tractable at microscopy resolution.
• Stochastic image sets, not single frames: Each compound is sampled from different noise tensors, producing a collection of images that mirrors the well-to-well variability of a real assay.
• Drop-in for existing analysis: Generated images are segmented and featurized by the same pipeline used on real plates, so a chemist can visually review a virtual screen the way they would review docking poses.

#Technical Details

pDIFF was built on the Stable Diffusion architecture and trained on Cell Painting images of 3,750 compounds from the Novartis mechanism-of-action chemogenomic library, assayed in U-2 OS cells in triplicate at 12.5 µM with 24 images per compound; the cytoplasm, mitochondria, and nuclei channels were used. Conditioning profiles came from Profile-QSAR, a massively multitask activity model trained on more than two million compounds across 14,222 internal dose-response assays, whose predicted pAC50 values were reduced to 2,018 assays by a target-focused selection scheme. The baseline was the same diffusion model conditioned on extended-connectivity fingerprints folded to a 2,048-count vector. Training consumed roughly 360 GPU-hours over 30,000 steps on 12 fields of view for each of 3,375 training compounds.

Validation used a cluster-based "realistically novel" split — 3,375 training and 375 held-out compounds, median Tanimoto coefficient 0.11 between the two sets. Across six image- and cell-level features, the Spearman correlation between real and generated compound-aggregated measurements was 0.12 to 0.48 for pDIFF against 0.04 to 0.13 for the fingerprint-conditioned baseline, where splitting the real images in half gives an upper bound of 0.48 to 0.67. In a virtual hit expansion test, retrieving the 50 nearest neighbors of 101 query actives gave a median overlap with ground truth of 14% for random selection, 16% for fingerprint similarity, 16% for the fingerprint-conditioned diffusion baseline, 38% for bioactivity profile similarity, and 50% for pDIFF-generated images.

#Applications

Hit expansion is the clearest use: given a phenotypic screening hit, pDIFF ranks unsynthesized or unassayed compounds by the similarity of their predicted morphology to the hit's, well above structural or profile-only similarity search on the paper's benchmark. The generated images also support mechanism-of-action profiling against annotated reference sets, virtual revisiting of legacy imaging campaigns, and early flagging of toxic phenotypes. The approach is not tied to Profile-QSAR; the same design would accept public bioactivity models or transcriptomic signatures as the conditioning vector.

#Impact

pDIFF demonstrates that the conditioning signal, not the generator, governs whether a phenotype model extrapolates — a point that transfers to any generative assay-prediction model. Its practical reach is limited by what has been released: the public repository is a training and inference framework under an MIT license whose only weights are a small test fixture plus the third-party Stable Diffusion autoencoder and text encoder, so the Cell Painting-trained model described in the paper cannot be downloaded. The training images and the bioactivity model behind the conditioning vectors are both internal, the preprint carries a CC-BY-NC-ND license, and the work has not been peer reviewed.