Predict splice donor and acceptor sites based on DNA sequence.
The dataset and dataset processing code from original Illumina SpliceAI was obtained from basespace project 66029966 (note: authorization required). The dataset was processed using modified versions of create_dataset.py and create_datafile.py scripts to return .csv.gz files with sequence instead of .h5 files with 1-hot encoded data.
original data split (train/test = 162706/16505 samples) was not changed.
Since original data does not contain validation split, we use 10% of training set as validation set:
python split_train_valid.py --data_path ./dataset_train_all.csv.gz --valid_ratio 0.1 --seed 42As result, we obtained train/valid/test = 146436/16270/16505.
Set paths to the data, set hyperparameters in example script finetune_spliceai.sh and run training on two GPUs:
CUDA_VISIBLE_DEVICES=0,1 NP=2 ./finetune_spliceai.sh