An implementation of AppRoximate Bayesian COmputation by Sequential Monte Carlo for genome scale metabolic models (using COBRApy).

This module undertakes Approximate Bayesian Computation using a Sequential Monte Carlo approach (ABC-SMC) to assess proposed additions and uncertain enzyme/reaction relationships in a COBRA constraint-based metabolic model, and find a combination of additions and removals of reactions and enzyme/reaction relationships to and from an original model that produces a model with maximum parsimony with a particular set of experimental observations.

A specific reaction can be catalysed by one or more (or no) enzymes, which are either known or unknown. The assertion that a particular reaction is catalysed by a particular enzyme, a Reaction/Enzyme Link (or REL), is tested in this algorithm. The algorithm requires as input a full model including all known and proposed RELs, a set of experiments to determine the congruence of model and experiment, and a confidence estimation for the existence of each REL.

The algorithm requires a set of parameters to estimate. In this case the existence of a particular REL is cast as a boolean parameter and prior confidence for the existence of that REL is taken from the confidence estimation provided. The existence of all RELs with a confidence less than 1 are included as parameters in the ABC-SMC and will be either included in or excluded from each proposed model created by the algorithm.

The algorithm will propose a large number of alternative models, calculating for them the congruence with the experimental data provided. Models will be accepted if their distance measure form the experimental data falls below a specified value (epsilon). Once N models have been found that are accepted, epsilon is reduced and new models are proposed by selecting a random accepted model from the previous step and perturbing it. At each step both epsilon and the perturbation strength are reduced until a set of N models is settled upon, having a low epsilon.

The output is therefore a set of N models, each with a high congruence with the observed experimental data. The confidence for the existence of each REL in reality can then be calculated as the proportion of these models containing that REL.

This module does not need to be installed; if the directory containing this file is in the PYTHONPATH, it can be imported using:

import abcsmc

This module has been tested with Python 2.7

Required non-standard Python packages (tested versions specified in brackets):

• COBRApy (0.4.0b4)
• NumPy (1.10.4)
• LibSBML (5.12.1) for SBML import and export
• Pyparsing (2.0.3) for GPR <--> list of enzymes translation

This package also requires the installation of a compatible linear optimisation program. Testing has shown Gurobi to be up to 30 times faster than the free GLPK program, and is available for free with an academic license when used solely for the purposes of research.

Most C{AbcProblem} arguments are optional (or conditionally optional), the following are required:

• An SBML file containing the model including predicted RELs
• An experiment file [1]_ containing results of gene essentiality analyses
• if the 'include_all' variable is set to False (default) then a list of candidate RELs for inclusion/exclusion must be supplied (e.g. 'example/candidate_rels.tsv')

Default values for variables are set for the purposes of testing, rather than full analysis. This example uses non-default values for several optional variables.

Create problem object:

.. code-block:: python

abc_problem = AbcProblem(
    model_file='example/model.xml',
    experiments_file='example/experiments.xml',
    prior_file='example/candidate_rels.tsv',
    particles_per_population=20,
    num_populations_max=6
)

.. [1] The experiment file has a specific format, as set out in the file 'example/experiments.tsv'