capice 4.0.0rc3

Consequence Agnostic Pathogenicity Interpretation of Clinical Exoma variations. State of the art machine learning to predict SNVs and InDels pathogenicity.

CAPICE : a computational method for Consequence-Agnostic Pathogenicity Interpretation of Clinical Exome variations

CAPICE is a computational method for predicting the pathogenicity of SNVs and InDels. It is a gradient boosting tree model trained using a variety of genomic annotations used by CADD score and trained on the clinical significance. CAPICE performs consistently across diverse independent synthetic, and real clinical data sets. It ourperforms the current best method in pathogenicity estimation for variants of different molecular consequences and allele frequency.

The software can be used as web service, as pre-computed scores or by installing the software locally, all described below.

Use online web service

CAPICE can be used as online service at http://molgenis.org/capice

Requirements

The list below is a complete list. Depending on whether GRCh37 and/or GRCh38 is used and whether all mentioned features are used, some items in the list below can be skipped.

• VEP v107
  • Including VEP cache (which needs to be unarchived!):
    • homo_sapiens_refseq_vep_107_GRCh37
    • homo_sapiens_refseq_vep_107_GRCh38
  • Including plugin(s):
    • Grantham
    • SpliceAI
  • Including additional data (GRCh37) available here:
    • gnomad.total.r2.1.1.sites.stripped.vcf.gz
    • gnomad.total.r2.1.1.sites.stripped.vcf.gz.csi
    • hg19.100way.phyloP100way.bw
    • spliceai_scores.masked.indel.hg19.vcf.gz
    • spliceai_scores.masked.indel.hg19.vcf.gz.tbi
    • spliceai_scores.masked.snv.hg19.vcf.gz
    • spliceai_scores.masked.snv.hg19.vcf.gz.tbi
  • Including additional data (GRCh38) available here:
    • gnomad.genomes.v3.1.2.sites.stripped.vcf.gz
    • gnomad.genomes.v3.1.2.sites.stripped.vcf.gz.csi
    • hg38.phyloP100way.bw
    • spliceai_scores.masked.indel.hg38.vcf.gz
    • spliceai_scores.masked.indel.hg38.vcf.gz.tbi
    • spliceai_scores.masked.snv.hg38.vcf.gz
    • spliceai_scores.masked.snv.hg38.vcf.gz.tbi
• BCF tools v1.14-1
• Python >=3.8

Install

The CAPICE software is also provided in this repository for running CAPICE in your own environment. The following sections will guide you through the steps needed for the variant annotation and the execution of making predictions using the CAPICE model.

UNIX like systems

Note: performance of CAPICE has been tested on Python 3.8, 3.9 and 3.10. Performance on other Python versions is not guaranteed.

1. Download and installation

Preferred

pip install capice

Optional

git clone https://github.com/molgenis/capice.git
cd capice
pip install .

Developers

git clone https://github.com/molgenis/capice.git
cd capice
pip install --editable '.[test]'

Windows

Installation on Windows systems is as of current not possible. Please refer to UNIX like systems (macOS or Linux) or use the Windows subsystem for Linux. You may also use the Singularity image of CAPICE found here.

SpliceAI

CAPICE requires additional VEP plugin SpliceAI. Files for the SpliceAI VEP plugin can be found here after creating an account (for free). In order to obtain the SNV and Indel files you must apply for the Predicting splicing from primary sequence project (should be free). The link to apply can be found within the VEP SpliceAI plugin description. The files can then be found within the Predicting splicing from primary sequence project -> ANALYSES -> genome_scores_vX -> FILES -> genome_scores_vX (where X is the latest version).

Usage

VEP

In order to score your variants through CAPICE, you have to annotate your variants using VEP by using the following command:

vep --input_file <path to your input file> --format vcf --output_file <path to your output file> \
--vcf --compress_output gzip --sift s --polyphen s --numbers --symbol \
--shift_3prime 1 --allele_number --refseq --total_length --no_stats --offline --cache \
--dir_cache </path/to/cache/107> --species "homo_sapiens" --assembly <GRCh37 or GRCh38> \
--fork <n_threads> --dont_skip --allow_non_variant --use_given_ref --exclude_predicted \
--flag_pick_allele --plugin Grantham \
--plugin SpliceAI,snv=<path/to/spliceai_scores.masked.snv.vcf.gz>,indel=</path/to/spliceai_scores.masked.indel.vcf.gz> \
--custom "<path/to/gnomad.total.sites.stripped.vcf.gz>,gnomAD,vcf,exact,0,AF,HN" \
--custom "<path/to/phyloP100way.bw>,phyloP,bigwig,exact,0" \
--dir_plugins <path to your VEP plugin directory>

IMPORTANT: Ensure the right files are used based on GRCH37 or GRCH38!!!

Note: Certain arguments might not be needed if training/predicting without using all possible features offered by CAPICE.

Then you have to convert the VEP output to TSV using our own BCFTools script: ./scripts/convert_vep_vcf_to_tsv_capice.sh -i </path/to/vep_output.vcf.gz> -o </path/to/capice_input.tsv.gz>

CAPICE

CAPICE can be run by using the following command:

capice [-h] [-v] [--version] {module} arguments

• -h: Print help and exit.
• -v: Verbose flag. Add multiple v to increase verbosity (more than 2 v does not further increase verbosity).
• --version: Print current CAPICE version and exit.

CAPICE currently has the following available modules:

• predict
• train
• explain

For all modules predict, train and explain, the following arguments are available:

• -i / --input (required): The path to the input VEP annotated dataset using the tab separator (can be both gzipped or not). Example input data can be found in the resources directory (based on genome build 37 with VEP105). The non-raw input files can be used directly with CAPICE. VEP outputs can be converted using convert_vep_to_tsv_capice.sh in the scripts directory (requires BCFTools).
• -o / --output (optional): The path to the directory, output filename or output directory and filename where the output is placed (will be made if it does not exists). If only a filename is supplied, or no output is supplied, the file will be placed within the directory of which CAPICE was called from. The file will always be gzipped with a .gz extension!

For instance:

-i input.tsv becomes input_capice.tsv.gz

-i input.tsv -o output.txt becomes output.txt.gz

-i input.tsv -o path/to/output.tsv becomes path/to/output.tsv.gz

-i input.tsv -o path/to/output becomes path/to/output/input_capice.tsv.gz

• -f / --force: Overwrite an output file if already present.

The following argument is specific to predict:

• -m / --model (required): The path to a custom pickled CAPICE model that includes attributes CAPICE_version (str) and model_features (list). Models can be found as attachments on the GitHub releases page.

The following arguments are specific to train:

• -m / --impute (required): The path to a JSON containing the features desired for training. Each key is a training feature, each value is ignored and can be left NULL. Please note that CAPICE is value type specific!
• -s / --split (optional): Percentage of input data that should be used to measure performance during training. Argument should be given in float from 0.1 (10%) to 0.9 (90%), default = 0.2.
• -t / --threads (optional): The amount of processing cores the training protocol can use. Default = 1.

You can also use capice {module} --help to show help on the command line.

Output of CAPICE prediction files

A file will be put out containing the following columns:

• chr: column containing the chromosome of a variant
• pos: the position of the variant
• ref: the reference of the variant
• alt: the alternative of the variant
• gene_name: The gene name of the variant as supplied.
• gene_id: The id of the gene name.
• id_source: The source of the gene id.
• feature: The feature of the variant as supplied.
• feature_type: The type of the feature of the variant as supplied.
• score: The predicted CAPICE score for the variant. The higher the score, the more likely that the variant is pathogenic.
• suggested_class: Suggested output class of the variant keeping in mind the score and gene. Currently VUS only. Work in progress.

Usage for making new CAPICE like models

Outside of Predictions, this repository also provides users the availability to create new CAPICE like models according to their specific use case. Since the input file features are not validated apart from 6 features (%CHROM, %POS , %REF, %ALT, %sample_weight, %binarized_label (case sensitive, % or # optional)), user can provide their own features. Please note that performance is validated on natively supported features. Performance is not guaranteed for custom features. Sample weight can be 1 for all samples if no sample weight should be applied. Binarized label should be either 0 or 1, depending on your labels of classification. Train is optimized for a 2 class problem. Performance is not guaranteed for more than 2 classes.

Outputs for training a new model:

A file will be put out containing the following element:

• xgb_classifier: Custom Pickled instance of a XGBClassifier instance that has successfully trained on the input data, containing additional attributes CAPICE_version and model_features.

Note: To load in a pickled instance of a model, use the following commands:

import pickle
with open(path/to/model_file.pickle.dat, 'rb') as model_file:
    model = pickle.load(model_file)

Usage for the explain module

Is only supported for CAPICE models that have been created using CAPICE v3.0.0 or greater!

The explain module takes a model and exports the feature importances of said model. Each of the importance type is described here.

Do note that the output of the explain module will always a gzipped TSV, just like the output of predict.

FAQ:

Question:
Will CAPICE support CADD 1.6 and Genome Build 38?

Answer:
No (CADD 1.6) and yes (Genome Build 38). CADD has moved on to Snakemake and is quite slow. It also limits us on updating VEP for improved and bugfixes on features. However, CAPICE will support genome build 38.

---

Question:
These scores are nice and all, but what do they really mean for this particular variant?

Answer:
CAPICE bases it's scoring on the training that it was provided with. A score is assigned based on features the model learned to recognize during training. There are plans to make a "Capice Explain Tool" which will tell how a score came to be.

---

Question:
Training a new model failed with an error in joblib.externals.loky.process_executor.RemoteTraceback with ValueError: unknown format is not supported. Why?

Answer:
This could possibly originate from a low sample size during the cross validation in RandomSearchCV. Please contact us for further help.

---

Question:
I'm on Windows and installing XGBoost fails with the PIP error “No files/directories in C:\path\to\xgboost\pip-egg-info (from PKG-INFO)”. Am I doing something wrong?

Answer:
Unfortunately, XGBoost does not cooperate well with Windows. You might want to try to install Setuptools before you attempt to install the dependencies. If that does not work either, we suggest you use either a Unix style virtual machine or, if you are using Windows 10, the Windows Subsystem for Linux is also available in the Windows Store for free, which is guaranteed to work.

You could also use CAPICE through the Singularity Container.

---

Question:
I'm trying to run CAPICE, but I'm getting the following error: xgboost.core.XGBoostError: XGBoost Library (libxgboost.dylib) could not be loaded.

Answer:
This error is caused on (likely) macOS when the package "OpenMP" is not installed. Please install libomp to get XGBoost running.

---

Question:
I'm trying to run the CAPICE explain module, but I'm getting the following error: _pickle.UnpicklingError: invalid load key, '\x0a'.

Answer:
This error is caused because the model you try to explain is created using a (much) older XGBoost version that is currently supported. You will have to manually extract the model features from that model using the correct XGBoost version. For the original publication model XGBoost version 0.90 should work.

---

Question:
I'm getting the following error: ModuleNotFoundError: No module named 'sklearn'. What is going wrong?

Answer:
sklearn is a module that should be installed when scikit-learn is installed. Either install sklearn manually though pip install sklearn or try to re-install scikit-learn.

---

Question:
I'm getting the warning /usr/local/lib/python3.8/dist-packages/joblib/_multiprocessing_helpers.py:45: UserWarning: [Errno 2] No such file or directory. joblib will operate in serial mode warnings.warn('%s. joblib will operate in serial mode' % (e,)) when using the CAPICE Singularity image. What's going wrong?

Answer:
This is likely due to the fact that the Singularity image searches for shared memory, which is different for Windows style operating systems. This means that any and all multiprocessing parts of CAPICE will perform in single threaded mode. Other than that, CAPICE should work just fine.

---

Question:
I want to use the standalone SpliceAI instead of the VEP plugin. Is this possible?

Answer:
We are investigating options to include the standalone SpliceAI since this requires a lot less resources for the precomputed scores that the VEP plugin uses. You could try to use it, but proceed at your own risk.

---

Question:
CAPICE gives an error that the model version does not match with the CAPICE version.

Answer:
There are certain restrictions regarding what model versions can be used with CAPICE. For regular releases, the major version must be identical. For pre-release versions (with rc<number> in the version), the entire version number (major, minor, patch & pre-release) must be identical.

Overview of code

If you're lost in the code, a map can be found here.