Biophysical Potts-Model inference on the GPU using OpenCL.
This program solves the "Inverse Ising problem" using a parallelized Monte-Carlo sequence generation algorithm. That is, it solves for the real-valued coupling parameters J^{ij}_{\alpha \beta} of an infinite-range q-state Potts model with Hamiltonian H(s) = \sum_{i < j}^L J^{ij}_{s_i s_j}, given bivariate marginals which might be obtained from a multiple sequence alignment containing sequences s with length L and q residue types.
For more information see the supplementary text in the publication:
Haldane, Allan, William F. Flynn, Peng He, R. S. K. Vijayan, and Ronald M. Levy (2016). Structural propensities of kinase family proteins from a Potts model of residue co-variation. Protein Science, 25, 1378-1384. DOI: 10.1002/pro.2954.
Licensed under GPLv3, see source for contact info.
Required packages: Python2, scipy/numpy, pyopencl, mwc64x.
The helper module "seqtools" must be compiled with make seqtools (requires gcc).
The mwc64x package must be obtained (available at http://cas.ee.ic.ac.uk/people/dt10/research/rngs-gpu-mwc64x.html) and placed in a directory named mwc64x next to the IvoGPU.py script.
The IvoGPU.py script can then be run in a variety of modes, incuding inverse ising inference, Monte-Carlo sequence generation, and Potts energy evaluation. Run IvoGPU.py -h to see a list of modes (see "positional arguments"). To get more info on each mode, use the -h option:
./IvoGPU.py inverseIsing -h
To check that the script is correctly detecting the system's GPUs, run:
./IvoGPU.py --clinfo
An example PBS script showing a typical set of arguments for inverse ising inference is in the file example_pbs.sh, which will fit the bivariate marginals from the file example_bimarg_pc.npy, computed from an HIV dataset for sequences of length length 93 with 4 residue types. The log file will contain many details about how the program is running. The script attempts to deduce some arguments from other supplied arguments (eg the sequence length L can be deduced from any supplied sequence file).
The seqmodel argument deserves more detail: If set to the string 'logscore' it will initialize the coupling values accorging to the uncorrelated (logscore) model and generate corresponding initial sequences. It may also be set to a directory name corresponding to a directory containing the output of a previous run from which it will load the couplings and sequences.
Helper scripts are also included: changeGauge.py transforms the Potts parameters between different gauges, and pseudocount.py adds different forms of pseudocount to the bivariate marginals.