This is GREMLIN v2.01 - a version that learns a Markov Random Field from an input alignment with parameters optimized for accuracy and speed of contact prediction.
Once you've installed gremlin (see INSTALL)
1) Within matlab
gremlin(<input_msa_file>, <output_file>, 'option_name', 'option_value',...)
where supported options are:
lambda_l2 (float): specifies the strength of the regularization parameter for the edge coupling (def:1)
lambda_bias(float): strength of regularization parameter for node weights (def:0.01)
lambda_l2_scale_by_length(boolean): to scale lambda_l2 by length of protein (def:1)
apc (boolean): Do APC correction at the end(def:1)
MaxIter(int): Number of iterations to run optimization (def:40).
verbose(boolean): To give detailed output during minimization (def:0)lambda_matf (string): file containing symmetric matrix of lambda values (from priors).
lambda_mat_multiplier(float): weight to give lambda_matf. If lambda_matf is log(prob), then this value should be a negative number (def: 1)
--using priors-- We are working on making distributable versions of the code we used to generate the priors. Until then, please use our webserver: http://gremlin.bakerlab.org if you'd like to use the ss/svm prior information.
--other options (not used in recommended settings) -- method:'cg' or 'LBFGS' have been tested (def:'cg'). CORR (int): Number of iterations to use to compute approximate Hessian in LBFGS(def:40, if method is 'LBFGS'). multiplicative_lam(boolean): to add or to multiply lambda_l2 with lambda_matf (def:0) if multiplicative_lam==0: lambda(i,j) = [lambda_l2+lambda_mat(i,j)] else: lambda(i,j) = [lambda_l2*lambda_mat(i,j)] Note:when lambda_l2_scale_by_length is turned on, it is applied at the end of this to all lambda(i,j).
2) From command line using compiled executableFor this to work, you'll need the pre-compiled mex executables compiled for your architecture. We have include mex executables for an x86_64 architecture.
Same as above except
a) First argument to run_gremlin.sh should be the path to the MCR
b) Replace commas with whitespace.
Example:
$run_gremlin.sh <MCR_PATH> <input_msa> <output_file> <option_name> <option_value>Notes: contact prediction accuracy does not appear to be sensitive to most of these options except lambda_l2(increasing it can worsen accuracy) and apc (the code assumes apc will be done at the end and therefore does not correct for entropy in anyway; therefore turning it off will almost consistently make things worse). If you find that changing the options consistently improves behavior on your inputs, we'd love to know more about your inputs!
1) Input format:
Input msa file is plain text and can be
(a) a list of aligned sequences, one in each line. This is the same format as PSICOV.
or,
(b) a list of
2) Output format: Outputs a symmetric matrix as a text file. mat[i][j] is strength of interaction between col i and col j of the msa. Diagonal is set to 0.
This is the protocol we've been using to generate alignments
1) run hhblits -n 4 -maxfilt 100000 -neffmax 20 -cpu 20 -nodiff
2) remove rows that have more than 25% gaps.
3) remove colums that have more than 25% gaps.
4) generate a non-redundant set of sequences (using hhfilter -id 90) and convert format
4a) (optional) generate priors
5) run gremlin
1) GREMLIN v2 and the older v1 (serial version) use C implementations of lbfgs minimizer and pseudo-likelihood computation and associated matlab wrappers from Mark Schmidt's software page. See http://www.di.ens.fr/~mschmidt/Software/index.html 2) If you use this code in your research, please cite the following papers: Hetunandan Kamisetty, Sergey Ovchinnikov, and David Baker. "Assessing the utility of co-evolution based residue-residue contact predictions in a sequence and structure-rich era." PNAS (2013) Sivaraman Balakrishnan, Hetunandan Kamisetty, Jaime G. Carbonell, Su-In Lee, and Christopher James Langmead. "Learning generative models for protein fold families." Proteins: Structure, Function, and Bioinformatics 79, no. 4 (2011): 1061-1078.