Perturb-OT

Cross-modality matching and prediction of perturbation response with labeled Gromov-Wasserstein Optimal Transport.

Repository layout

├── .gitignore
├── README.md
├── perturbot
│   ├── cv      # Cross-validation experiments
│   └── perturbot
│       ├── match
│       │    ├── ot_labels.py   # Label-constrained Entropic Optimal Transport
│       │    ├── ott_egwl.py    # Label-constrained Entropic Gromov-Wasserstein
│       │    ├── cot_labels.py  # Label-constrained Co-OT (Redko et al., 2020)
│       │    └── cot_feature.py # Feature-feature OT based on Co-OT concept
│       ├── predict
│       └── eval
├── scvi-tools              # forked `scvi-tools` with label adaptation
└── ott                     # forked `ott` with label adaptation

Installation

perturbot/ uses the modified scvi-tools and ott submodules which can be installed with pip install.

cd scvi-tools/
pip install .
cd ../ott/
pip install .
cd ../perturbot
pip install .

Usage

import numpy as np
from sklearn.decomposition import PCA
from perturbot.match import (
    get_coupling_cotl, 
    get_coupling_cotl_sinkhorn, 
    get_coupling_egw_labels_ott,
    get_coupling_egw_all_ott,
    get_coupling_eot_ott,
    get_coupling_leot_ott,
    get_coupling_egw_ott,
    get_coupling_cot, 
    get_coupling_cot_sinkhorn, 
    get_coupling_gw_labels,
    get_coupling_fot,
)
from perturbot.predict import train_mlp

# Generate data
n_samples = 300
labels = [0,1,2,3]
X_dict = {k: np.random.rand(n_samples,1000) for k in labels}
Y_dict = {k: np.random.rand(n_samples,2000) for k in labels}
pca = PCA(n_components=50)
X_reduced = {k: pca.fit_transform(X_dict[k]) for k in labels}
Y_reduced = {k: pca.fit_transform(Y_dict[k]) for k in labels}


# Learn matching in the latent space
T_dict, log = get_coupling_egw_labels_ott((X_reduced, Y_reduced)) # Other get_coupling_X methods be used

# Train MLP based on matching
model, pred_log = train_mlp((X_dict, Y_dict), T_dict)

# Learn feature-feature matching
T_feature, fm_log = get_coupling_fot((X_dict, Y_dict), T_dict)

See documentation and manuscript for more details.

Support

Please submit issues or reach out to jayoung_ryu@g.harvard.edu.

Authors and acknowledgment

Jayoung Ryu, Romain Lopez, & Charlotte Bunne

Citation

If you have used Perturb-OT for your work, please consider citing:

@misc{ryu2024crossmodality,
      title={Cross-modality Matching and Prediction of Perturbation Responses with Labeled Gromov-Wasserstein Optimal Transport}, 
      author={Jayoung Ryu and Romain Lopez and Charlotte Bunne and Aviv Regev},
      year={2024},
      eprint={2405.00838},
      archivePrefix={arXiv},
      primaryClass={q-bio.GN}
}