UVAE: Integration of Unpaired and Heterogeneous Clinical Flow Cytometry Data

UVAE: Integration of Unpaired and Heterogeneous Clinical Flow Cytometry Data This repository contains training data, metadata, and results for the above publication. 'data.zip' contains the subsampled flow cytometry data and clinical metadata for the lineage and chemokine panels. 'toy-data.zip' contains the synthetic data (in pickle and csv formats) used for benchmarks and the associated source flow file. 'data-cycombine-comp.zip' contains the processed DFCI and van Gassen datasets used for model comparison, the scripts used to obtain the .RDS files from raw flow data, the derived imputation datasets (created by splitting the data and hiding 5 markers from each panel), and the results of training imputation UVAE and cyCombine models. 'models-integration.zip' contains the trained UVAE models for lineage and chemokine panels, as well as imputed marker data. 'models-severity.zip' contains the trained COVID severity regression models, as well as their scores and gradient attribution plots. Note that 'models-integration' and 'models-severity' are trained on the entire CIRCO data (the severity models regression was performed on 4 held-out folds, but the initial integration and cluster definition were performed on the entire dataset). 'crossvalidation.zip' contains the models used to obtain test scores, which were trained by fully holding out the test data for each fold. Inside, there are 4 separate UVAE models for lineage (in 'lineage-split') and chemokine (in 'chemokine-split'). Each contains a separate embedding file containing both training and test data used to train the severity models. The severity models are in 'severity-reg-held' folder, with full results in 'scores/combined.csv'.