PhenoSeq

Generating single-cell transcriptomic embeddings from Cell Painting morphology using conditional diffusion

---

Overview

PhenoSeq is a conditional diffusion model that generates biologically structured scRNA-seq embeddings directly from Cell Painting microscopy images.

Rather than treating morphology and transcriptomics as independent measurements, PhenoSeq learns a shared latent relationship between the two modalities. Given only imaging-derived features, the model synthesises transcriptomic representations in the pretrained scGPT embedding space, enabling downstream molecular analysis without sequencing.

Key Idea

Cell morphology is a physical manifestation of underlying molecular state.

PhenoSeq learns to translate:

Cell Painting Morphology
            ↓
     Conditional Diffusion
            ↓
   Synthetic scGPT Embeddings
            ↓
Transcriptomic Representation

---

Highlights

• First diffusion model for generating transcriptomic embeddings from Cell Painting morphology
• Cross-attention transformer architecture
• Population-level supervision (no cell-paired data required)
• Generates biologically structured scGPT embeddings
• Improves treatment classification over imaging alone
• Recovers ~29% of the gap to the real-transcriptomics multimodal ceiling

---

Results

Treatment Classification (Single Profile)

Modality	WE Balanced Accuracy
Imaging	0.270
Synthetic RNA	0.293
Imaging + Synthetic RNA	0.315
Imaging + Real RNA	0.425

Synthetic transcriptomic embeddings consistently outperform morphology alone and provide complementary biological signal when fused with imaging.

---

Method

PhenoSeq consists of three components:

1. Imaging Encoder

Cell Painting images are encoded using a frozen ViT-L backbone:

5 channels × 1024 features
          ↓
      5120-d vector

2. Conditional Diffusion

A transformer-based denoiser predicts noise in scGPT latent space while attending to imaging-derived context.

Noisy RNA Embedding
          ↓
 Cross-Attention Transformer
          ↓
 Predicted Noise

3. Transcriptomic Generation

Iterative denoising produces a synthetic transcriptomic embedding:

xT ~ N(0,I)
      ↓
 DDIM Sampling
      ↓
Synthetic scGPT Embedding

---

Architecture

Cell Painting Features (5120)
            │
            ▼
     Self-Attention Encoder
            │
            ▼
      Cross-Attention
            │
            ▼
   Diffusion Transformer
            │
            ▼
 Synthetic scGPT (512)

Key parameters:

Parameter	Value
Diffusion Steps	1000
Transformer Layers	6
Heads	8
Model Dimension	1024
RNA Dimension	512
Imaging Dimension	5120

---

Dataset

We train and evaluate on the scGeneScope dataset.

Modalities

Cell Painting

• 5 fluorescence channels
• ViT-L ImageNet embeddings
• 5120-dimensional representation

Single-cell RNA-seq

• scGPT embeddings
• 512-dimensional representation

Dataset:

(https://huggingface.co/datasets/altoslabs/scGeneScope)

scGeneScope
├── Imaging
│   └── ViT-L embeddings
└── RNA-seq
    └── scGPT embeddings

---

Installation

git clone https://github.com/<username>/PhenoSeq.git
cd PhenoSeq

pip install -r requirements.txt

---

Training

python train.py --config config.yaml

Example:

python train.py \
    --config config.yaml \
    --batch_size 256 \
    --lr 1e-4 \
    --diffusion_steps 1000 \
    --model_dim 1024

---

Project Structure

img2rna/
├── train.py
├── config.yaml
├── requirements.txt
│
├── models/
│   ├── diffusion.py
│   ├── denoiser.py
│   └── model_utils.py
│
├── data/
│   └── dataset.py
│
├── utils/
│   └── train_utils.py
│
└── assets/
    ├── phenoseq_hero.png
    └── architecture.png