Skip to content

This package simplifies evolutionary algorithms for applications in systems biology. It can be used to generate oscillating reaction networks, networks that match input timeseries data, or fit parameters (rateconstants).

License

Notifications You must be signed in to change notification settings

ltatka/ReactionNetworkEvolution.jl

Repository files navigation

ReactionNetworkEvolution.jl

Purpose

This module allows the user to easily run an evolutionary algorithm to create mass-action chemical reaction networks with specific behaviors or to match time series data. It is intended as a research tool for computational systems biology.

There are three primary use cases for this packge:

  1. Evolving reaction networks with specific behaviors. The default is oscillation, but other behaviors may be added in the future. Additionally, users can input time series that reflect those desired behaviors.
  2. Evolving reaction networks (or ensemble models) to fit time series data. Often systems biologist will have timeseries data from experiments and will want to search for a reaction network (or group of them) that can recapitulate the time series data. This is useful for building models, informing future experiments or looking at output models for common reaction themes which may indicate the presence of those reactions in the ground truth.
  3. Parameter (rate constants) fitting. In the case where model topology is known (for example, the biologist has already built the model from first principles) and wants to generate parameter values for those reactions that will recapitulate experimental time series data.

How to use

Installing Julia

This module uses the julia programming language. Installation instructions can be found here.

Quick Start

  1. Download the package
    From a julia console:

    using Pkg
    Pkg.add("ReactionNetworkEvolution")
    
  2. Run evolution with default settings Remaining in the julia console:

    import ReactionNetworkEvolution as rne
    rne.run_evolution()
    

    NOTE: If an output directory is not supplied, then one will be created in the current working directory. Make sure that the working directory allows new directories to be made.

  3. Customize Evolution (optional) The run_evolution() function takes several optional keyword arguments:

  • ntrials: the number of trials to run (100 by default)
  • ngenerations: the number of generations per batch (800 by default)
  • population_size: the number of networks in a population
  • pathtosettings: path to a json file storing additional custom settings
  • outputpath: path to a directory where evolution output will be written
  • seed: specify a seed for the random number generator

For example, to change the population_size to 200 via the command line:
rne.run_evolution(population_size=200)

Additional Custom Settings

Specify Additional Settings in a JSON File

Additional settings can be specified in a JSON file and supplied to the run_evolution() function as a keyword argument.
run_evolution(pathtosettings="/home/name/path/to/your/settings.json")

An example of a JSON file specifying custom settings:

{"chemical_species_names": ["A", "B", "C"],
"initial_concentrations": [5.0, 3.0, 6.0],
"reaction_probabilities": [0.2, 0.3, 0.3, 0.2],
"ngenerations": 500,
"population_size": 200,
}

Any settings that are not specified in the JSON file will be set to the default value.

Description of Settings and Default Values

Setting Name Default Value Description
ngenerations 800 The number of generations
population_size 100 The number of reaction networks in the population
nreactions 5 The number of reactions in each network at the beginning of evolution (this number may be different in final networks)
chemical_species_names ["S0", "S1", "S2"] List of the chemical species names
initial_concentrations [1.0, 5.0, 9.0] The initial concentrations for the chemical species in chemical_species_names
seed Random seed
portion_elite 0.1 Portion of best networks to copy to the next generation
portion_delete 0.1 Portion of worst networks to remove
tournament_select false If true, use tournament selection to choose non-elite networks
reaction_probabilities [0.1, 0.4, 0.4, 0.1] Probability of adding uni-uni, uni-bi, bi-uni, and bi-bi reaction types
rateconstant_range [0.1, 50] Min and max values for rate constants in new reactions or when choosing a completely new rate constant (values can be mutated to outside this range for existing reactions)
p_rateconstant_mutation 0.6 Probability of changing a rate constant (as opposed to adding/deleting a reaction)
p_new_rateconstant 0.15 If mutating a rate constant, probability that a completely new rate constant will be selected from the rate constant range
p_crossover 0 Probability of crossover
same_fitness_crossover false If true, networks with similar fitness will be considered equally fit and crossed over more leniently
same_fitness_percent_range 0.05 If using same fitness crossover, networks with fitnesses within 1-x and 1+x will be considered the same fitness
p_mutation 1 Probability of mutating a network by modifying rate constants or adding/deleting reactions
exclusive_crossover_mutation false If true, networks will EITHER be crossed over OR mutated but never both
enable_speciation true If true, networks will be divided into species and compete amongst themselves
starting_delta 0.65 The speciation threshold at the beginning of evolution
delta_step 0.1 The amount by which delta is adjusted to achieve the target number of species
target_num_species 10 The threshold to split reaction networks into species will be adjusted up and down to attempt to maintain this number of species.
use_seed_network false If true, the first generation will be populated with copies of a seed network loaded from the specified path
seed_network_path If using a seed network, the path to the network
randomize_seed_network_rates true If true and using a seed network, the reaction rate constants will be selected randomly from the rate constant range
average_fitness false If true, each species fitness will be the average of its networks - otherwise species fitness is the top network’s fitness
rateconstant_distance_weight 0.0 Weight given to different paramters values when calculating distance between two networks - if 0, parameters will not be considered when calculating distance
objective_data_path If generating networks to fit data, path to .csv file with data to fit - if no path is given, ReactionNetworkEvolution.jl will attempt to generate oscillators
track_fitness true Track information about evolution such as top fitness, number of species, etc for each generation
writeout_threshold 0.05 If the best network reaches this fitness or higher, evolution will be stopped
process_output_oscillators true If true, automatically test if evolved networks are oscillators or not

Use Cases

1. Evolving reaction networks with specific behaviors

This package was primarily designed to generate mass-action chemical reaction networks with oscillatory behavior. With the defualt settings, it will generate oscillators with 3 chemical species.

  1. Download the package
    From a julia console:
    using Pkg
    Pkg.add("ReactionNetworkEvolution")
    
  2. Run evolution with default settings Remaining in the julia console:
    import ReactionNetworkEvolution as rne
    rne.run_evolution()
    

2. Evolving reaction networks to fit time series data

  1. Supply the time series data as a CSV file.
  2. Create a JSON file with at least the following setting to direct the package to the time series CSV and disable sorting oscillators
{"objective_data_path": "/path/to/timeseries_data.csv",
"process_output_oscillators": false}
  1. Run evolution supplying the JSON file:
  import ReactionNetworkEvolution as rne
  rne.run_evolution(pathtosettings="/path/to/settingsfile.json")

3. Evolve paramters with fixed topology network

  1. Supply the time series data as a CSV file
  2. Supply the network topology as an antimony file.
  3. Create a JSON file with at least the following settings:
{"objective_data_path": "/path/to/timeseries_data.csv",
"process_output_oscillators": false,
"seed_network_path": "/path/to/antimonyfile.txt",
"rateconstant_distance_weight": 1.0,
"p_rateconstant_mutation" 1.0}

The rateconstant_distance_weight setting enables the package to consider differences in reaction weight constants when comparing twp networks. This is necessary if all networks have the same topology, although some adjustments to the value may be necessary.

Setting p_rateconstant_mutation to 1.0 prevents the algorithm from adding or deleting reactions during evolution.

  1. Run evolution supplying the JSON file:
  import ReactionNetworkEvolution as rne
  rne.run_evolution(pathtosettings="/path/to/settingsfile.json")

About

This package simplifies evolutionary algorithms for applications in systems biology. It can be used to generate oscillating reaction networks, networks that match input timeseries data, or fit parameters (rateconstants).

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published