Skip to content

Latest commit

 

History

History
84 lines (63 loc) · 3.2 KB

README.md

File metadata and controls

84 lines (63 loc) · 3.2 KB
Raw

spliceai-pytorch

model

Implementation of SpliceAI, Illumina's deep neural network to predict variant effects on splicing, in PyTorch. You can find the Illumina's official implementation here.

Installation

pip install spliceai-pytorch

Usage

import torch
from spliceai_pytorch import SpliceAI

model_80nt = SpliceAI.from_preconfigured('80nt')
# model_400nt = SpliceAI.from_preconfigured('400nt')
# model_2k = SpliceAI.from_preconfigured('2k')
# model_10k = SpliceAI.from_preconfigured('10k')

# ... training ...

x = torch.randn([1, 4, 80 + 5000])  # Predicts Donor/Acceptor probs only for core 5000nt region.

probs = model_80nt(x)  # (1, 5000, 3)

Generating train/test sets

First, download 'SpliceAI train code' directory from here and unzip it to spliceai_train_code directory. Also, download human reference genome (version hg19) to spliceai_train_code/reference directory.

Then, run the following command to generate train/test sets after moving into spliceai_train_code/Canonical.

# Before running `grab_sequence.sh`,
# make sure that the variable CL_max is configured properly in `constants.py` (80, 400, 2000 or 10000)
chmod 755 grab_sequence.sh
./grab_sequence.sh

# Requires Python 2.7, with numpy, h5py, scikit-learn installed
python create_datafile.py train all  # ~4 miniutes, creates datafile_train_all.h5 (27G)
python create_datafile.py test 0     # ~1 minute, creates datafile_test_0.h5 (2.4G)

python create_dataset.py train all   # ~11 minutes, creates dataset_train_all.h5 (5.4G)
python create_dataset.py test 0      # ~1 minute, creates dataset_test_0.h5 (0.5G)

Training

$ python -m spliceai_pytorch.train --model 80nt \  # 80nt, 400nt, 2k, 10k
  --train-h5 spliceai_train_code/Canonical/dataset_train_all.h5 \
  --test-h5 spliceai_train_code/Canonical/dataset_test_0.h5 \
  --use-wandb  # Optional, for logging.

Reproduction status (wip)

Currently on the reproduction of Figure 1E. Results are as below, and you can view model training logs here (W&B).

NOTE: Target results are from ensemble of 5 models, while reproduced results are from a single model.

Model Top-k acc. (target) PR-AUC (target) Top-k acc. (reproduced) PR-AUC (reproduced)
SpliceAI-80nt 0.57 0.60 0.54355 0.56435
SpliceAI-400nt 0.90 0.95 0.87265 0.93160
SpliceAI-2k 0.93 0.97 0.9083 0.9541
SpliceAI-10k 0.95 0.98 0.9286 0.96475

Citation

@article{jaganathan2019predicting,
  title={Predicting splicing from primary sequence with deep learning},
  author={Jaganathan, Kishore and Panagiotopoulou, Sofia Kyriazopoulou and McRae, Jeremy F and Darbandi, Siavash Fazel and Knowles, David and Li, Yang I and Kosmicki, Jack A and Arbelaez, Juan and Cui, Wenwu and Schwartz, Grace B and others},
  journal={Cell},
  volume={176},
  number={3},
  pages={535--548},
  year={2019},
  publisher={Elsevier}
}