README.md

metadata

title: PhenoSeq
emoji: 🧫
colorFrom: green
colorTo: pink
sdk: gradio
sdk_version: 6.26.0
app_file: app.py
short_description: Cell Painting images to single-cell transcriptomes
python_version: '3.12'
startup_duration_timeout: 30m

PhenoSeq — morphology → transcriptome

Interactive demo of Sentinal4D/PhenoSeq, a conditional Gaussian diffusion model that generates single-cell RNA-seq embeddings (scGPT, 512-d) from Cell Painting morphology features (ViT-L/14 features of the five fluorescence channels — DNA, RNA, AGP, Mito, ER — 5 × 1024 = 5120-d per cell).

Given the imaging features of 16 cells from a well, the model samples a population of synthetic transcriptomes for that well via 50-step DDIM.

What the demo shows

• the actual Cell Painting cells the model is conditioned on (4×4 false-colour montage),
• a PCA scatter of the generated transcriptomes against the real, held-out scGPT profiles measured in the same well,
• a per-dimension embedding profile (generated vs. real),
• quantitative agreement — mean-centred cosine to the correct well, with the cosine to the other held-out wells as a specificity baseline — plus a 29-way perturbation retrieval check against treatment centroids computed from the training wells only,
• the raw generated (n_cells, 512) matrix as a downloadable .npy.

You can also upload your own (N, 5120) ViT-L Cell Painting feature array (.npy / .npz) to condition the model on data of your own.

All 15 wells offered in the dropdown are held-out validation wells — they were reconstructed with the authors' own split (RandomState(42), 85/15 over the 94 paired sample IDs) and the eval-mode cell selection (np.linspace(0, n-1, 16)), so nothing shown here was seen during training.

What to expect

scGPT embeddings share a large common mean, so raw cosine between any two of them is ≈0.99 and tells you nothing. Everything here is therefore mean-centred.

Sampling all 15 held-out wells (256 cells, 50 DDIM steps, seed 0):

median centred cosine to the correct well	0.66
median centred cosine to the other 14 wells	0.40
correct well ranked #1 / top-3 (of 15)	5/15 · 9/15 (chance 1/15 · 3/15)
29-way treatment retrieval, median rank	9 / 29 (chance 15), 3/15 at rank 1

Clearly above chance, and far from solved: BAY 11-7082 and (R)-roscovitine are recovered exactly, while TPA in well 202502PA_12 is missed outright.

Data attribution

The bundled example assets (cell crops, imaging features, real RNA-seq embeddings, treatment centroids) are small excerpts derived from altoslabs/scGeneScope, released by Altos Labs under CC-BY-NC-4.0. They are redistributed here for non-commercial research demonstration, with attribution, as the licence requires.

Model weights: Sentinal4D/PhenoSeq, Apache-2.0.