Inferring heterozygous SNP positions from tumor samples without a matched normal using Hidden Markov Models
Authors: Melody Choi, Metin Balaban, Ben Raphael
HetDetect allows for the inference of hetSNPs. It takes as input
- any VCF file (bcftools, cellSNPlite),
- a specified output path,
- any additional arguments,
uses cyvcf2 to parse the input VCF files, and finally outputs a re-genotyped VCF with the inferred hetSNPs.
It uses a Hidden Markov Model consisting of:
- a user defined number of hidden states,
- a fixed, low transition probability matrix (tau = 3*10^-4),
- a 1D Gaussian emission probability matrix, and
- LAF as observed states.
It also produces a colored BAF scatterplot, as such:
where blue points show the false positive het SNPs that are filtered from the VCF. These plots can be found in the specific outdirectory provided as input to run_hetdetect.py.
- Clone the repository and change directory
- Run
pip3 install -e .
- GPU usage. For accelerated performance using pomegranate, PyTorch, and CUDA GPU, users can specify if they would like to run the model with tensors on GPU with the option
--g. - User defined number of hidden states. Users can customize the HMM to define any number of hidden states corresponding to the number of mean/covariance pairs that the model will infer.
- Binomial test. To filter out SNPs that may have been falsely labeled as heterozygous by the HMM, a binomial statistical test is performed to re-label false het-SNPs as homozygous (either 0/0 for homozygous REF or 1/1 for homozygous ALT).
To run HetDetect, call python run_hetdetect.py -i [input file path] -o [output file path] along with any other arguments as necessary.
Run python run_detect.py -h to see all the options.