PhenoProfiler : Advancing Morphology Representations for Image-based Drug Discovery

PhenoProfiler is an advanced tool for phenotypic profiling of cell morphology, to efficiently extract phenotypic effects of perturbations from high-throughput imaging.

PhenoProfiler operates as an end-to-end image encoder, converting multi-channel images directly into low-dimensional quantitative features, thus eliminating the need for extensive preprocessing in non-end-to-end pipeline. For more details, please refer to our paper.

---

To be updated:

• Code for all Figures in our paper.
• The PhenoProfiler package, which can be installed via pip install PhenoProfiler.

---

Requirements

git clone https://github.com/QSong-github/PhenoProfiler.git
cd PhenoProfiler
conda env create -f environment.yaml
conda activate PhenoProfiler

Usage

See PhenoProfiler_inference.ipynb.

The trained models of PhenoProfiler can be downloaded at synapse or google drive.

import numpy as np
import torch
from skimage.transform import resize
from PIL import Image
from models import PhenoProfiler

# List of image paths
img_paths = [
    './Sample_imgs/IXMtest_A01_s2_w1AA6B1894-F561-42EE-9D1D-E21E5C741B75.png',
    './Sample_imgs/IXMtest_A01_s2_w3A597237B-C3D7-43AE-8399-83E76DA1532D.png',
    './Sample_imgs/IXMtest_A01_s2_w50F1562CD-EBCF-408E-9D8D-F6F0FDD746C8.png',
    './Sample_imgs/IXMtest_A01_s2_w246FFAEE1-BEB6-4C81-913B-B979EC0C4BC3.png',
    './Sample_imgs/IXMtest_A01_s2_w46657239A-5AFE-4C29-BB9B-49978EFE4791.png',
]

# Load and preprocess images
images = np.stack([resize(np.array(Image.open(path)), (448, 448), anti_aliasing=True) for path in img_paths])
images_tensor = torch.tensor(images).float().cuda()

# Load model
model = PhenoProfiler().cuda()
model.load_state_dict(torch.load('./PhenoProfiler.pt', weights_only=True))

# Generate embeddings
image_features = model.image_encoder(images_tensor.unsqueeze(0))
image_embeddings = model.image_projection(image_features)

# Print the shape of the embeddings
print(image_embeddings.shape)

Training Model

Data Preparation

BBBC022: aws s3 cp s3://cytodata/datasets/Bioactives-BBBC022-Gustafsdottir/ ./ --recursive --no-sign-request
CDRP-BIO-BBBC036: aws s3 cp s3://cytodata/datasets/CDRPBIO-BBBC036-Bray/ ./ --recursive --no-sign-request
TAORF-BBBC037: aws s3 cp s3://cytodata/datasets/TA-ORF-BBBC037-Rohban/ ./ --recursive --no-sign-request
cpg0019: aws s3 cp s3://cellpainting-gallery/cpg0019-moshkov-deepprofiler/ ./ --recursive --no-sign-request

After download, organise the dataset as follows:

dataset/
    bbbc022/
        images/
            20585/
                IXMtest_B15_s8_w49DE56250-C587-48D8-8840-5A322A3F0177.png
                IXMtest_F20_s4_w4ABAD0B1C-D157-496F-96B9-5909F00DEFF7.png
                ...
        embedding/ # from cpg0019
            20585/
                A01/
                    1/
                        embedding.npz
        profiling.csv
    bbbc036/
        images/ ...
        embedding/ ...
        profiling.csv
    bbbc037/
        images/ ...
        embedding/ ...
        profiling.csv

Training from scratch

Step1: training the first model by regression loss:

python train.py --model MSE

Step2: training the final model by all loss:

python train.py --exp_name result/PhenoProfiler --pretrained_model result/PhenoProfiler_MSE

Testing

The phenotype correction strategy-processed features are available by executing the corresponding ipynb file Phenotype correction strategy.ipynb.

python test_22.py
python test_36.py
python test_37.py

Reference

If you find this project is useful for your research, please cite:

@article{PhenoProfiler,

}

Our code is based on the CLIP and BLEEP. Special thanks to the authors and contributors for their invaluable work.