spredle is a Python package for predicting RNA splicing from DNA sequences using deep learning. It is designed to first reproduce the SpliceAI model (Jaganathan et al., 2019) in PyTorch, rather than the original TensorFlow (v1) implementation, and to adapt it to a combined classification and regression framework for predicting both splicing site and usage, as demonstrated in Pangolin. The package then explores additional architectures - such as Transformers and Hyena — with the goal of further improving splicing prediction performance.
Ideally, as the project progresses, spredle will be extended to support training on: (1) individual genomes, with splicing outcomes inferred from matched RNA-seq data; and (2) tissue-specific splicing patterns derived from long-read RNA-seq data.
- using conda
git clone git@github.com:HaniceSun/spredle.git
cd spredle
conda env create -f environment.yml
conda activate spredle
# applied to human data
spredle download-training-data
spredle preprocess --nt 5000 --flank 40
spredle torch-dataset --input Homo_sapiens.GRCh38.115_seq_nt5000_flank40.txt
spredle train --config_file config.yaml --model_name SpliceAI-1 --train_file Homo_sapiens.GRCh38.115_seq_nt5000_flank40_train.pt --val_file Homo_sapiens.GRCh38.115_seq_nt5000_flank40_val.pt --lr_lambda "1,1,1,1,1,1,0.5,0.25,0.125,0.0625,0.03125,0.015625"
spredle train --config_file config.yaml --model_name Hyena-1 --train_file Homo_sapiens.GRCh38.115_seq_nt5000_flank40_train.pt --val_file Homo_sapiens.GRCh38.115_seq_nt5000_flank40_val.pt
spredle predict --config_file config.yaml --model_name SpliceAI-1 --epoch 9 --pred_file predict.txt
# extended to mouse data
spredle download-training-data --genome_reference https://ftp.ensembl.org/pub/current/fasta/mus_musculus/dna/Mus_musculus.GRCm39.dna.primary_assembly.fa.gz --gene_annotation https://ftp.ensembl.org/pub/current/gtf/mus_musculus/Mus_musculus.GRCm39.115.gtf.gz
spredle preprocess --nt 5000 --flank 5000 --input Mus_musculus.GRCm39.115_seq.txt
spredle torch-dataset --input Mus_musculus.GRCm39.115_seq_nt5000_flank5000.txt
spredle train --config_file config.yaml --model_name SpliceAI-4 --train_file Mus_musculus.GRCm39.115_seq_nt5000_flank5000_train.pt --val_file Mus_musculus.GRCm39.115_seq_nt5000_flank5000_val.pt --lr_lambda "1,1,1,1,1,1,0.5,0.25,0.125,0.0625,0.03125,0.015625"
spredle predict --config_file config.yaml --model_name SpliceAI-4 --epoch 9 --pred_file predict.txt
Author: Han Sun
Email: hansun@stanford.edu
License: MIT License