Multi-level feature Genomic Variants Estimator (ML-GVP)

ML-GVP is an XGBoost-based model for predicting pathogenicity of variants in exons and introns. It was originally designed for the Sherloc challenge in the 6th Critical Assessment of Genome Interpretation. The model is one of the two best-performing models in the Sherloc challenge, ranking the first for overall and indel prediction. This repository extended the original implementation (chenkenbio/CAGI6-Sherloc-model) and provided scripts for additional analysis and experiments.

System Requirements and Installation

ML-GVP was trained on Ubuntu. The key requirements are Python (>3.8) and XGBoost (xgboost-gpu>1.6.2). The databases used for collecting features were list in ./config.sh.

• ANNOVAR
• blast: NCBI BLAST
• MMSplice
• bedtools
• pyBigWig
• XGBoost
• CrossMap
• skopt(required only during training process)
• tabix (required to builed the .tbi file for ensemble predictors)

Configure config.sh

Set the paths in ./config.sh:

• blastp: path of the tool blast: NCBI BLAST
• hg19/hg38: GRCh37(hg19)/GRCh38(hg38) reference genome, can be visited from here
• hg19_gtf: hg19 gene annotation in GTF format
• PSSM_SPOT_PATH: path to PSSM_SPOT_DISORDER folder (decompress PSSM-HHM-SPOT_DISORDER.tgx in [data/PSSM-HHM-SPOT_DISORDER]
• ROADMAP_DB: path to RoadMap ChIP-seq/DNase-seq processed data(download using the following commands in data/roadmap)

for eid in `seq 63 113`; do
    eid=`python -c "print('E%03d' % $eid)"`
    for assay in DNase H3K27ac H3K27me3 H3K36me3 H3K4me1 H3K4me3 H3K9me3; do
        wget -c https://egg2.wustl.edu/roadmap/data/byFileType/signal/consolidated/macs2signal/pval/${eid}-${assay}.pval.signal.bigwig
    done
done

• ENCODE_CHIPSeq_DB: path to data/encode_chipseq folder (unzip data/encode_chipseq/encode_chipseq.tgz)
• ENCODE_eCLIP_DB: path to RNA binding protein (RBP) binding sites data, tar -xf ENCODE_RBP.tar in data/eCLIP
• gtex_eQTL_loci: path to data/GTEx.eQTL.v8.hg38to19.merged.bed.gz
• gnomAD_AF_DB: path to gnomAD.v3 folder (tar -xf gnomad.v3.AF.vcf.tar in data/gnomAD.v3)
• uniprot_fasta: uniprot protein fasta file (unzip data/uniprot_human.fa.gz)
• hg19_phastCons_100way: download
• hg19_phyloP_100way: download
• hg19_phastCons_46way_primates: download
• hg19_phyloP_46way_primates: download
• hg19_phastCons_46way_vertebrate: download
• hg19_phyloP_46way_vertebrate: download
  Or download the above files in bigWig format in data/ucsc/hg19 using

wget -c http://hgdownload.cse.ucsc.edu/goldenpath/hg19/phyloP46way/vertebrate.phyloP46way.bw
wget -c http://hgdownload.cse.ucsc.edu/goldenpath/hg19/phyloP46way/primates.phyloP46way.bw
wget -c http://hgdownload.cse.ucsc.edu/goldenpath/hg19/phyloP100way/hg19.100way.phyloP100way.bw
wget -c http://hgdownload.cse.ucsc.edu/goldenpath/hg19/phastCons46way/primates.phastCons46way.bw
wget -c http://hgdownload.cse.ucsc.edu/goldenpath/hg19/phastCons46way/vertebrate.phastCons46way.bw
wget -c http://hgdownload.cse.ucsc.edu/goldenpath/hg19/phastCons100way/hg19.100way.phastCons.bw  

• MMSplice: install mmsplice and set mmsplice_python to the python interpreter with MMSplice installed
• ensemble_dir: the directory including CADD, MVP, M-CAP, PrimateAI, AlphaMissense score (VCF file and .tbi file)

Demo

1. Prepare features:

./prepare_features.sh input.vcf hg19 output

After executing the command, the features file curated for variants can be found at output/input.std.avinput.features-ALL.tsv

2. Make prediction

./run_prediction.sh output/input.stc.avinput.features-ALL.tsv predict_dir

The prediction scores can be found at input.std.avinput.features-ALL_rfecv.output.txt. The scores of ML-GVP+ can be found at input.std.avinput.features-ALL_rfecv.meta.output.txt

Repository Structure and Usage

./train included all codes used for training an ML-GVE model.
./src included all other scripts except those for model training.
./cagi6-sherloc inlcuded all training and test data which were provided by CAGI6 competition.