Multi-Modal Foundation · Sentinal4D
PhenoSeq
Predicts single-cell scGPT RNA-seq embeddings from multi-channel fluorescence microscopy imaging features.
What PhenoSeq is
PhenoSeq is a conditional Gaussian diffusion model developed by Sentinal4D for image-to-transcriptome translation at single-cell resolution. It predicts single-cell transcriptomic states directly from fluorescence microscopy imaging features, bridging cellular imaging assays with single-cell gene expression representations.
The model accepts visual feature vectors extracted from fluorescence microscopy wells or individual cells and generates 512-dimensional scGPT transcriptomic embeddings. This enables researchers to infer transcriptomic profiles from visual phenotypes without carrying out single-cell RNA sequencing on every sample.
The generated RNA embeddings represent single-cell expression states on the original scGPT embedding scale, allowing their direct application to downstream biological analyses such as cell-type identification, population clustering, and developmental trajectory inference.
How it works
PhenoSeq employs a cross-attention transformer denoiser with approximately 168 million parameters. Its imaging encoder uses a two-layer self-attention transformer to project 5,120-dimensional image vectors—extracted across five fluorescence channels using a ViT-L/14 backbone—from an input shape of (B, N, 5120) to a (B, N, 1024) conditioning representation, typically using 16 imaging cells per well.
The generative diffusion backbone combines noisy 512-dimensional scGPT embeddings with sinusoidal timestep embeddings before passing them through six cross-attention blocks. In each block, RNA queries attend to the imaging context alongside self-attention and adaptive layer normalisation conditioned on the diffusion timestep. The network predicts noise to reverse a 1,000-step cosine schedule, trained with exponential moving average weight averaging and supporting 50-step DDIM inference.
Key facts
- Architecture
- Cross-attention diffusion denoiser
- Parameters
- ~168 million
- Input dimensions
- (B, N, 5120) ViT-L/14 features across 5 fluorescence channels
- Output dimensions
- (B, 512) scGPT RNA-seq embeddings
- Training dataset
- scGeneScope
- Diffusion schedule
- Cosine beta schedule (1,000 steps)
- Sampling steps
- 50 (DDIM default) or 1,000 (DDPM)
What you can use it for
- Microscopy-to-transcriptome translation
- Predicting 512-dimensional scGPT single-cell transcriptomic embeddings from 5-channel fluorescence microscopy images.
- Cell-type classification
- Providing predicted single-cell transcriptomic embeddings for downstream classification pipelines to determine cell identities.
- Single-cell clustering and trajectory inference
- Generating transcriptomic vectors from imaging assays to perform cellular clustering and trajectory modelling.
Run it here
We list this model but cannot run it yet: no library and no architectures — nothing can load it.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Translates multi-channel microscopy image features directly into single-cell transcriptomic embeddings at single-cell resolution.
- Supports fast inference in 50 steps using DDIM sampling alongside standard 1,000-step DDPM sampling.
- Achieved a validation mean squared error loss of 0.1683 at epoch 87 on the scGeneScope dataset.
Watch out for
- Requires pre-extracted 5,120-dimensional ViT-L/14 embeddings from five fluorescence channels rather than raw microscopy image inputs.
- Requires recomputation of feature normalisation statistics when applied to imaging datasets outside scGeneScope.
- Conditioning is designed for well-matched multi-cell image contexts, typically requiring 16 imaging cells per sample.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use PhenoSeq commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
What inputs are required to run PhenoSeq?
The model expects ViT-L/14 image feature embeddings extracted across five fluorescence channels, with 1,024 dimensions per channel yielding 5,120 features per imaging cell, shaped as (B, N, 5120) where N is typically 16 cells per well.
What do the output embeddings represent?
PhenoSeq outputs 512-dimensional vectors un-normalised to the scGPT embedding space, representing single-cell transcriptomic states.
Which sampling schedules does PhenoSeq support?
The model defaults to 50-step DDIM sampling for fast inference, but can also run full 1,000-step DDPM sampling.
Can PhenoSeq be applied to microscopy datasets other than scGeneScope?
Yes, but users must compute and provide per-feature mean and standard deviation statistics for their own dataset, as the default normalisation file is derived from the scGeneScope training split.
How to cite it
Naidoo, R., Hu, J., Tripodi, G., Bakal, C., & Chakraborti, T. (2026). Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion. ICML 2026 Workshop on Multi-modal Foundation Models and Large Language Models for Life Sciences (FM4LS). https://openreview.net/forum?id=ACHMa1e8J1
Written from PhenoSeq's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.