A hidden Markov model (HMM) to detect archaic segments in modern human genomes. The L-BFGS-B algorithm uses Stephen Becker's implementation in C.
git clone https://github.com/ryhui/hmmarc.git
cd hmmarc
make
(Run ./hmmarc -h for a summary of the options)
./hmmarc -m [model file] vit|post|est|bfgs [data file(s)]
Model file:
Includes parameters in the HMM (see example/model.hmm for an example).
N: number of hidden states;
pi: starting state probabilities;
T: time (in generations) since the admixture, used to calculate transition probabilities;
n: number of observed states;
emit: emission probabilities;
In the decoding modes (vit/post), the model is used for inference; in the training modes (est/bfgs), the model specifies the initial parameter values.
Data file:
Records the observed allele-sharing states (see example/obs.HGDP00780.chr1.0.txt for an example).
The file has three columns:
variant position;
observed state;
genetic distance (cM) from the previous variant.
Mode:
vit: output Viterbi sequence of the second state;
est: output posterior probabilities of the second state at each input variant;
est: Baum-Welch training;
bfgs: numerical optimization using the L-BFGS-B algorithm.
To print the Viterbi sequence of archaic segments:
./hmmarc -m example/model.hmm vit example/obs.HGDP00780.chr1.0.txt
To print the posterior probabilities of the archaic state at each input site:
./hmmarc -m example/model.hmm post example/obs.HGDP00780.chr1.0.txt