Another implementation of SpliceAI. This one is designed to better utilise GPU-parallelism than the original implementation. In our benchmarks we get a 3-3.5 times performance improvement over the original implementation using an Nvidia RTX 2080. Trained models were taken directly from the original implementation, so you are going to get the same predictions. It only makes sense to use this implementation if you are planning to use a GPU to predict a lot of variants. If you are going to use this implementation in your research, don't forget to reference the original paper Jaganathan et al, Cell 2019 in press.
Although we've added an automatic pip installer that grabs all the required dependencies, the best way to install spliceai-reforged is to create a new conda environment and grab dependencies from conda repositories rather than pip's repository. The following instructions assume you are planning to use a GPU. If that's not the case, replace tensorflow-gpu with tensorflow (though in that case you might as well use the original implementation).
$ conda create -y -n spliceai python=3.7 numpy pandas tensorflow-gpu=2.2 click
$ conda activate spliceai
$ conda install -y -c bioconda pysam pyfaidx ncls
$ pip install --no-cache-dir git+https://github.com/skoblov-lab/spliceai-reforged.git
You can run use spliceai.py
Example run with bundled data.
spliceai.py -r grch37.fna -a grch37 -i testdata/input.vcf -o testdata/output.vcf
Usage: spliceai.py [OPTIONS]
Annotate a VCF file with SpliceAI predictions
Options:
-i, --input FILE Path to the input VCF file [required]
-o, --output PATH Path to the output VCF file [required]
-r, --ref_assembly FILE Path to a reference assembly. For best
performance the assembly should be stored on
a fast-access drive (an SSD or a RAM-disk).
Chromosome identifiers in the assembly must
follow the same notation format as
chromosome values in the annotation file.
Built-in annotations use short-format
chromosome notation (that is there are no
"chr" prefixes in chromosome identifiers).
[required]
-a, --annotations TEXT Reference genome annotation. You can specify
a path to an annotation file or specify one
of built-in annotations: "grch37" or
"grch38". An annotation file must contain
the following columns: #NAME (gene name),
CHROM (chromosome), STRAND, TX_START
(transcript start), TX_END (transcript end)
EXON_START (exon start positions), EXON_END
(exon end positions). The values in CHROM
must follow the same chromosome naming
format as sequence identifiers in the
reference assembly. Built-in annotations use
short-format chromosome notation (that is
there are no "chr" prefixes in chromosome
identifiers). TX_START and TX_END use
1-based indexing, end positions are
inclusive. EXON_START and EXON_END are
comma-separated lists of corresponding exon
start and end locations. We use GENCODE
annotations, wherein EXON_START values point
to the last intron position right next to an
exon. Your custom annotation files must
follow the same conventions. Annotation
files must be tab-separated. [required]
-d, --distance INTEGER RANGE Maximum distance between the variant and
gained/lost splice site. An integer in the
range [0, 5000]. Defaults to 50.
--mask Activate the masking of scores representing
annotated acceptor/donor gain and
unannotated acceptor/donor loss
--preprocessing_threads INTEGER RANGE
The number of preprocessing threads to use.
If your reference assembly is located on a
fast-access drive (an SSD or a RAM-disk),
using up to 4 preprocessing threads
significantly cuts down preprocessing time.
Defaults to 1.
--preprocessing_batch INTEGER RANGE
The number of variant records in a VCF file
to process at a time. Using larger batches
tends to cut down preprocessing overhead. It
is safe to keep the default value. Defaults
to 10000
--prediction_batch INTEGER RANGE
The batch size to use during inference. This
option is useful if you are using a GPU and
want to exploit its full potential. It is
best to use powers of 2. If you are getting
out of memory errors, you should reduce the
batch size. Defaults to 64
--help Show this message and exit.