SpliceAI: A deep learning-based tool to identify splice variants

This package annotates genetic variants with their predicted effect on splicing, as described in Jaganathan et al, Cell 2019 in press. The annotations for all possible substitutions, 1 base insertions, and 1-4 base deletions within genes are available here for download. These annotations are free for academic and not-for-profit use; other use requires a commercial license from Illumina, Inc.

License

SpliceAI source code is provided under the GPLv3 license. SpliceAI includes several third party packages provided under other open source licenses, please see NOTICE for additional details. The trained models used by SpliceAI (located in this package at spliceai/models) are provided under the CC BY NC 4.0 license for academic and non-commercial use; other use requires a commercial license from Illumina, Inc.

Installation

The simplest way to install SpliceAI is through pip or conda: ```sh pip install spliceai

or

conda install -c bioconda spliceai ```

Alternately, SpliceAI can be installed from the github repository:

sh
git clone https://github.com/Illumina/SpliceAI.git
cd SpliceAI
python setup.py install

SpliceAI requires

tensorflow>=1.2.0

or

conda install tensorflow ```

Usage

SpliceAI can be run from the command line: ```sh spliceai -I input.vcf -O output.vcf -R genome.fa -A grch37

or you can pipe the input and output VCFs

cat input.vcf | spliceai -R genome.fa -A grch37 > output.vcf ```

Required parameters: -

-I

-O

ALLELE|SYMBOL|DS_AG|DS_AL|DS_DG|DS_DL|DP_AG|DP_AL|DP_DG|DP_DL

-R

-A

grch37

grch38

spliceai/annotations/grch37.txt

Optional parameters: -

-D

-M

Details of SpliceAI INFO field:

| ID | Description | | -------- | ----------- | | ALLELE | Alternate allele | | SYMBOL | Gene symbol | | DSAG | Delta score (acceptor gain) | | DSAL | Delta score (acceptor loss) | | DSDG | Delta score (donor gain) | | DSDL | Delta score (donor loss) | | DPAG | Delta position (acceptor gain) | | DPAL | Delta position (acceptor loss) | | DPDG | Delta position (donor gain) | | DPDL | Delta position (donor loss) |

Delta score of a variant, defined as the maximum of (DSAG, DSAL, DSDG, DSDL), ranges from 0 to 1 and can be interpreted as the probability of the variant being splice-altering. In the paper, a detailed characterization is provided for 0.2 (high recall), 0.5 (recommended), and 0.8 (high precision) cutoffs. Delta position conveys information about the location where splicing changes relative to the variant position (positive values are downstream of the variant, negative values are upstream).

Examples

A sample input file and the corresponding output file can be found at

examples/input.vcf

examples/output.vcf

T|RYR1|0.00|0.00|0.91|0.08|-28|-46|-2|-31

19:38958362 C>T

Similarly, the output

CA|TTN|0.07|1.00|0.00|0.00|-7|-1|35|-29

2:179415988 C>CA

Frequently asked questions

1. Why are some variants not scored by SpliceAI?

SpliceAI only annotates variants within genes defined by the gene annotation file. Additionally, SpliceAI does not annotate variants if they are close to chromosome ends (5kb on either side), deletions of length greater than twice the input parameter

-D

2. What are the differences between raw (

-M 0

-M 1

The raw files also include splicing changes corresponding to strengthening annotated splice sites and weakening unannotated splice sites, which are typically much less pathogenic than weakening annotated splice sites and strengthening unannotated splice sites. The delta scores of such splicing changes are set to 0 in the masked files. We recommend using raw files for alternative splicing analysis and masked files for variant interpretation.

3. Can SpliceAI be used to score custom sequences?

Yes, install SpliceAI and use the following script:

from keras.models import load_model
from pkg_resources import resource_filename
from spliceai.utils import one_hot_encode
import numpy as np

input_sequence = 'CGATCTGACGTGGGTGTCATCGCATTATCGATATTGCAT'
Replace this with your custom sequence
context = 10000
paths = ('models/spliceai{}.h5'.format(x) for x in range(1, 6))
models = [load_model(resource_filename('spliceai', x)) for x in paths]
x = one_hot_encode('N'(context//2) + input_sequence + 'N'(context//2))[None, :]
y = np.mean([models[m].predict(x) for m in range(5)], axis=0)
acceptor_prob = y[0, :, 1]
donor_prob = y[0, :, 2]

Contact

Kishore Jaganathan: [email protected]