atanna / hmm

Variable Length Hidden Markov Models (VLHMM)

The repository concludes two modules vlhmm, chipseq, which can solve next problems:

• vlhmm:
  • building, training, plotting context transition tree
  • building, training VLHMM
• chipseq:
  • applying VLHMM to ChIp-seq data
  • conversion for genome browser fromat http://genome.ucsc.edu/

Language: Python 3.x

Dependencies:

• Cython
• NumPy, SciPy
• Pylab
• datrie
• PyGraphviz

Instalation:

git clone https://github.com/atanna/hmm.git

Cython-files compilation:

python setup.py build_ext --inplace

---

Examples

The directory vlhmm_/examples/ concludes test examples with training VLHMM and context trees on simulated data.

chipseq/real_test.py shows example VLHMM on ChIP-seq data

---

Description

Context transition tree

• Context transition tree defines a stochastic process.
• Context state -- any preffix from previous states (the process moves from right to left, i.e. states go in descending order of time).
• Vertex ~ context.
• Edge ~ state.
• Outdegree of internal vertex -- number of states.
• The leaf defines the distribution of the current state.

• Examples

  1. "Unfair Coin"
     

  2. Markov chain
     

  3. Second-order Markov chain
     

  4. Variable length Markov chain
     

*Context transition tree defines a variable leghth Markov stochastic process.

VLHMM

• The same as HMM.
• Hidden layer is defined by variable leghth Markov stochastic process (which we can define by context transition tree).