Source code for Cross-modal Graph Contrastive Learning with Cellular Images.

MIGA requires anaconda with python 3.7 or later, cudatoolkit=11.1 and below packages

torch                     1.7.1+cu110
torch-cluster             1.5.9
torch-geometric           1.6.3
torch-scatter             2.0.7
torch-sparse              0.6.10
torch-spline-conv         1.2.1
torchvision               0.8.2+cu110

MIGA has been tested on Ubuntu 18.04, with eight GPUs (Nvidia RTX4090). Installation should take no longer than 20 minutes on a modern server.

Please refer to environment.yml

CIL dataset originally consists of 919,265 cellular images collected from 30,616 molecular intervention.

The current CIL data in data includes 50 molecules and 1270 corresponding images, we would release the full version after being accepted.

Check the following scripts: GIN:

    python submit.py --config config/miga/miga_gin.yaml

Graph Transformer:

    python submit.py --config config/miga/miga_graphTrans.yaml

Check the following scripts:

    finetune_classification.sh
    finetune_regression.sh
    finetune_clinical.sh

import torch
from core.network import MIGA
from dataset import process_data

model = MIGA('graph_transformer', is_eval=True)
model.eval()
checkpoint = torch.load('models/model.pth', map_location='cpu')
model.load_state_dict(checkpoint, strict=False)

smiles = 'c1ccc(cc1)C2=NCC(=O)Nc3c2cc(cc3)[N+](=O)[O]'
data = process_data(smiles, 'graph_transformer')
molecule_embeddings = model.get_graph_embedding(data)

Please cite the following paper if you use this code in your work.

S. Zheng, J. Rao, J. Zhang, L. Zhou, J. Xie, E. Cohen, W. Lu, C. Li, Y. Yang, Cross-modal Graph Contrastive Learning with Cellular Images. Adv. Sci. 2024, 2404845. https://doi.org/10.1002/advs.202404845