Poisson-based HMM Integer Classifier
PHIC is designed to solve a simple recurring problem in bioinformatics. Sequencing read depth is often used to infer something from copy-number variants to R-loops, to ChIP-seq peaks. PHIC reads the sequence of read depths and classifies these into any number of numeric classes. For example, most of the genome may be 0-1x coverage, but some places are 4-6x, and others are 50 or more. PHIC can classify coverage into 3 classes of 1, 5, and 50.
PHIC assumes input data are integer values within a limited range
(0-255). Read depth files are sometimes normalized into floating point
values, and some values may be out of range. For this reason, there is a
python wrapper phic.py that performs a variety of pre- and
post-processing activities on the raw data. It is generally better to
use phic.py than phic itself.
In order to compile phic you must first build the genomikon.a library.
git clone https://github.com/KorfLab/genomikon.git
cd ik
make
Place the genomikon directory at the same level as the phic directory. That
is, there should be a parent directory containing both the genomikon and
phic directories. This is necessary because the phic Makefile
references the parent directory to get to genomikon.
After you've build genomikon.a you can go to the phic directory and
build the phic executable.
git clone https://github.com/iankorf/phic.git
cd phic
make
There is a sample file of integers called values.txt. Try running the
programs with that.
python3 phic.py --file values.txt --x 5 --classes 1 5 20
phic values.txt 5 1 5 20
There is also test-phic.py that creates a random file of 1 million
integers and decodes with phic.py.
The PHIC HMM is fully connected, meaning that you can get from any state to any other state. The transition probabilities are the same for each state. The chance to leave a state is called X. The chance to stay in a state should therefore be 1.0 - kX, where k is the number of states. But since X is generally small, the self-returning transition probability is near 1.0. So, for simplicity, and because I don't give a phuck, the self-returning probability is 1.0 and the transition probabilities don't quite sum to 1.0.
Unlike most HMMs which have tables of discrete emission probabilities, the emissions in PHIC are drawn from a Poisson distribution. The emissions are the number of expected counts. All of the calculations are done in log space, so long sequences of numbers should be fine.