PhenoSeq

A Cell Painting plate photographs millions of cells for what it costs to sequence a few thousand, but mechanism-of-action signal lives in the transcriptome, not the image. Both assays are destructive, so the two readouts rarely come from the same cell — the pairing exists only at the level of a treated well. PhenoSeq takes that mismatch as its premise: if a treated population's morphological fingerprint carries recoverable structure in transcriptomic space, every imaging experiment already run becomes a latent source of molecular readout.

The mechanism is a conditional Gaussian diffusion model that denoises in a representation space rather than in pixels or gene counts. Its target is the 512-dimensional cell embedding produced by scGPT; its conditioning signal is a set of ViT-L/14 features extracted from five-channel fluorescence microscopy. A cross-attention denoiser lets the noisy RNA embedding query the imaging context at every step, so morphology steers the reverse diffusion trajectory toward the transcriptomic state consistent with the observed phenotype. Two frozen backbones therefore define the model's interface contract — ViT-L/14 features in, scGPT-space embeddings out — and that constraint is as much a part of the design as the denoiser itself.

PhenoSeq was built by Reed Naidoo and colleagues in the Bakal lab at The Institute of Cancer Research, London, working with the lab's spinout Sentinal4D and with Tapabrata Chakraborti's group at The Alan Turing Institute, with support from the Turing–Roche strategic partnership. It was presented at the ICML 2026 workshop on Multi-modal Foundation Models and Large Language Models for Life Sciences, with weights published under the Sentinal4D account.

#Key Features

• Diffusion in embedding space: The generative target is a 512-dimensional scGPT vector, so outputs feed directly into cell-type classification, clustering, and trajectory inference without a decoding step.
• Cross-attention conditioning: Six transformer blocks let the noisy RNA embedding attend to a morphological context of 16 imaging cells per well, with adaptive layer normalization conditioned on the diffusion timestep.
• Population-level supervision: Training matches imaging and sequencing at the well level rather than the cell level, which is what makes the task learnable from assays that cannot be run on the same cell.
• Orchestration over frozen backbones: Neither the vision transformer nor the transcriptomic language model is fine-tuned; PhenoSeq learns only the conditional map between their spaces.
• Fixed checkpoint inference: A single EMA checkpoint runs end to end from a published pipeline; adapting to a new imaging dataset requires recomputing input normalisation statistics, not retraining weights.

#Technical Details

The denoiser is a ~168M-parameter cross-attention transformer. A two-layer self-attention imaging encoder projects the (B, N, 5120) conditioning tensor — five fluorescence channels × 1,024 ViT-L/14 dimensions — down to a 1,024-dimensional context, which six cross-attention blocks with eight heads attend to while denoising a 512-dimensional RNA embedding carrying a sinusoidal timestep embedding. The objective is ε-prediction over a cosine beta schedule with T = 1,000 steps and EMA weight averaging at decay 0.9999; sampling defaults to 50 DDIM steps, with full DDPM available. Training used the scGeneScope dataset from Altos Labs, which pairs Cell Painting imaging features with scGPT embeddings of single-cell RNA-seq matched by sample identifier, and reached a best validation MSE of 0.1683 at epoch 87.

Evaluation used a 28-compound treatment-identification benchmark. In the single-profile setting, generated embeddings reached 0.293 weighted balanced accuracy against 0.270 for raw imaging, and fusing the two reached 0.315 — recovering roughly 29% of the gap to the 0.425 achieved with real transcriptomics. In the multi-profile setting, fusing synthetic RNA with imaging more than doubles imaging-only balanced accuracy, and embedding-space analysis confirms correct treatment localisation for most conditions.

#Applications

The intended use is phenotypic drug discovery at screening scale, where morphological profiling is already routine and transcriptomic profiling is not. Generating a transcriptomic representation for each imaged well gives a second, complementary profile for mechanism-of-action clustering and compound triage, and the results indicate it is most useful fused with imaging rather than as a replacement for it. Because the model consumes ViT-L/14 features rather than raw images, integrating it into an existing pipeline means matching that feature extractor and recomputing the per-feature normalisation statistics for the new screen.

#Impact

PhenoSeq is a demonstration that cross-modal generation from morphology into transcriptomic representation space is architecturally tractable and downstream-useful, and the released artifacts make it reproducible: checkpoint and inference pipeline under Apache 2.0, training code under MIT. The honest limits are equally clear. The fused result remains well short of real transcriptomics, supervision is population-level rather than per-cell, evaluation covers a single 28-compound benchmark on one dataset, the underlying scGeneScope corpus carries a non-commercial licence, and the work is a workshop paper rather than a peer-reviewed article. What is most likely to generalise is the pattern it prototypes: treating a cheap assay as the conditioning signal for a generative model of an expensive one.