repository for the paper "Removal of Batch Effects using Distribution-Matching Residual Networks" by Uri Shaham, Kelly P. Stanton, Jun Zhao, Huamin Li, Khadir Raddassi, Ruth Montgomery, and Yuval Kluger.

The script Train_MMD_ResNet.py is the main demo script, and can be generally used for calibration experiments. It was used to train all MMD ResNets used for the CyTOF experiments reported in our manuscript. It loads two CyTOF datasets, corresponding to measurements of blood of the same person on the same machine in two different days, and denoises them. The script then trains a MMD-ResNet using one of the datasets as source and the other as target, to remove the batch effects. The script Calibrate_RNA_example.py was used to produce the results in section 4.3.

All the datasets used to produce the results in the manuscript are saved in Data. The labels for the CyTOF datasets (person_Day_) were used only to separate the CD8 population during evaluation. Training of all models was unsupervised. The RNA data set Data2_standardized_37PCs.csv contains the projection of the cleaned and filtered data onto the subspace of the first 37 principal components. To obtain the raw data please contact Jun Zhao at [email protected].

All the models used to produce the results in the manuscript are saved in savedModels.

The following scripts were used to produce more results which are reported in the manuscript: The script evaluate_calibration was used to produce the results of section 4.2.3, as well as some additional results. The MLP nets were trained using the script train_MMD_MLP.py.

The script checkGeneralization.py was used to produce the results of section 4.2.4. The nets N_{d_1},N_{d_2} were trained using the script train_vertical_nets.py.

The script Comparison_linear.py produces the results of section 4.2.5.

All scripts are written in Keras.

The script calibrationData2.R was used to produce the results for section 4.3, the calibration analysis on single cell RNA data, several custom functions loaded from funs.R.

Any questions should be referred to Uri Shaham, [email protected].