DAGIP: A bias correction algorithm for cell-free DNA data

Given two groups of matched (preferentially paired) samples sequenced under different protocols, the tool explicitly learns the bias using a neural network. The approach builds on Optimal Transport theory, and exploits sample-to-sample similarities to define how to perform bias correction.

Documentations are available at antoinepassemiers.github.io/DAGIP/

Install the tool

Install dependencies:

pip install -r requirements.txt

To reproduce the results of our manuscript, rpy2 should be installed, as well the following R packages: dplyr, GenomicRanges and dryclean.

Install the package:

python setup.py install --user

Basic usage:

from dagip.core import DomainAdapter

# Let X and Y be two numpy arrays, where the rows are (coverage, methylation, fragmentomic) profiles and columns are features (e.g., DMRs, bins). Y and X have been produced under sequencing protocols 1 and 2, respectively.

# Build a model from the matched groups
model = DomainAdapter()
X_adapted = model.fit_transform(X, Y)

# X_adapted is a numpy array of same dimensions as X

# Save the model
model.save('some/location.pt')

...

# Load the model
model.load('some/location.pt')

# Perform bias correction on new independent samples from sequencing protocol 2
X_new_adapted = model.transform(X_new)

For more advanced usage, please check the documentation.

Reproduce the results from the manuscript

To reproduce the results presented in our paper, first download the datasets from the FigShare repository (DOI: 10.6084/m9.figshare.24459304) and place its content in the data folder. The repository is available at https://figshare.com/s/5f837a93ea2719ffcaf9.

Preprocess the data:

python scripts/preprocess.py
python scripts/50kb-to1mb.py
python scripts/to-numpy.py

Cancer detection

python validate-cnas.py HL
python validate-cnas.py DLBCL
python validate-cnas.py MM
python validate-cnas.py OV
python validate-fragmentomics-multimodal.py

Cross-validation with paired samples