This is our Pytorch implementation of RNA Alternative Splicing Prediction with Discrete Compositional Energy Network.

Installation

pip install -r requirements.txt

Data: Context Augmented Psi Dataset (CAPD)

Download the CAPD here, unzip and place the files in dataset/.

Saved weights

Saved weights of all the models in the paper are available in: saved_models.

Training codes

Train spliceai on the splice site classification and psi regression objectives:

sh train_scripts/train_spliceai_clsreg.sh

This training script saves two weights files site_aux_pytorch_model.bin and spliceai_pytorch_model.bin at each checkpoint folder.

Train splicesome net on the psi regression objective with pretrained site_aux_pytorch_model.bin and spliceai_pytorch_model.bin weights from spliceai:

sh train_scripts/train_dcen.sh

Evaluation codes

Evaluate and output pred and label npy files for each patient's gene data, output file is formatted as <tissue_type><patient_id>_chr<chromosome_number>:

python eval_scripts/schedule_eval_dcen.py

The evaluation result is computed from the schedule_eval_* script and is stored in each of these folders only cover one particular tissue_type, patient_id and chromosome_number. Other schedule_eval_* scripts are available for other ablation variants. These scripts run python train_and_infer.py iteratively over all the relevant <tissue_type><patient_id>_chr<chromosome_number>.

python train_and_infer.py key arguments for evaluation:

--spliceosome_model_training_type : determine the ablation variants of the model
--eval_patient_list : patient id(s) to evaluate
--eval_chr_list : chromosome number(s) to evaluate
--eval_output_dir_prefix : output directory of evaluation results, pred and label npy files

To compute evaluation across multiple tissue types, patients or genes use the aggregate_eval* scripts such as the ones below.

Compute evaluation across multiple patient and gene pred and label npy files:

python aggregate_eval_all_models_alltest.py

Compute evalation across all patient and only gene from chromosome numbers excluded from training set:

python aggregate_eval_all_models_exchromosomes.py

Compute evalation across all patient and only gene longer than limit in training set:

python aggregate_eval_all_models_longgenes.py

Prediction codes

Infer and output the splice sites' and transcripts' predictions as a jsonl file, for only C1orf21 TTLL7 genes.

sh predict_dcen.sh

Remove --eval_gene_list arg to infer all genes.

python train_and_infer.py key arguments for prediction:

--valid_data_dir : directory containing samples to infer probabilities
--eval_gene_list : list of genes to infer, spaced apart (e.g. "C1orf21 TTLL7" to infer only these two genes). Default: infer all genes in samples --eval_output_dir_prefix : output directory of prediction results, pred and label npy files
--pred_output_filename : output prediction json file name
--eval_patient_list : patient id(s) to evaluate
--eval_chr_list : chromosome number(s) to evaluate