This project is a collection of files to allow users to reproduce the model development and benchmarking in "Dawnn: single-cell differential abundance with neural networks" (Hall and Castellano, under review). Dawnn is a tool for detecting differential abundance in single-cell RNAseq datasets. It is available as an R package here. Please contact us if you are unable to reproduce any of the analysis in our paper. The files in this collection correspond to the code used to train Dawnn's neural network model (following the model selection and hyperparameter optimisation in 10.5522/04/22634416/nn_model_choice.py) and the resulting trained model.

FILES: Training code

train_nn_regen_seed_123.py Python script to train final selected Dawnn model (Loads Keras and Tensorflow; Loads training data in labels_df_autogen4_rerun.csv; Creates neural network architecture; Trains model using early stopping, learning rate 0.00001, and batch size 32; Saves trained model as final_model_dawnn_rerun.h5). train_final_model_regen_seed_123_job_sub.sh Bash script to submit train_nn_regen_seed_123.py to SGE cluster.

Trained model

final_model_dawnn_rerun.h5 Final trained Dawnn model output by train_nn_regen_seed_123.py.