The current implementation of GapFind and GapFill is in GAMS, which charges a significant license fee from each user, even though many solvers are free for academic purposes. To promote the usage of this computational tool, we developed this open-source Julia package, GapFindFill.jl, to enable researchers to use GapFind and GapFill for free by harnessing the power of academic free solvers provided by Gurobi and IBM.
GapFindFill.jl is implemented in Julia and makes use of the high-level interface JuMP.jl.
JuMP is a domain-specific modeling language for mathematical optimization embedded in Julia. With JuMP, it is easier for users to specify and call different optimizers to solve optimization problems in GapFindFill.jl than using interfaces provided by solvers directly. Built upon the generic high-level programming language Julia, users can embed GapFindFill.jl in their complex workflows to simplify task processing. GAMS, as a specific optimization tool, does not provide support for processing other tasks or integrate with other programming languages.
To use this package, the user needs to install Julia first. This version is built on Julia v1.1. Gurobi.jl is used in GapFindFill.jl as the default solver for both gap finding and filling. Gurobi provides a free academic license for non-commercial use. Users can also choose to use GLPK.jl or CPLEX.jl for gap finding. GLPK is free for all users, while CPLEX provides a free academic license. But note that compared to Gurobi and CPLEX, it takes GLPK much more time and memory to solve the same problem. For gap filling, Gurobi is set as the only solver since it worked way better than the other two in our tests.
For Julia v1.0 users, any package compatibility issues while testing GapFindFill.jl can be resolved by pinning CPLEX.jl and Gurobi.jl to a specific version by running the following commands in pkg> mode:
pin Gurobi@0.6.0
pin CPLEX@0.5.0For Julia v1.2 users, GapFindFill.jl will be compatible with Julia v1.2 once CPLEX.jl is updated for Julia v1.2.
Installation.
Within Julia, press ] to enter pkg> mode.
To install GapFindFill.jl, issue the command:
add GapFindFillor issue
add https://github.com/varnerlab/GapFindFill.jl.gitTo use GapFindFill in your project simply issue the command:
using GapFindFillTest.
To test GapFindFill installation, use the following command in pkg> mode:
test GapFindFillwhich runs two examples under test directory. The expected outcomes from this test set are illustrated in Example section.
Uninstallation.
To delete GapFindFill package use the following command in pkg> mode:
rm GapFindFillTwo examples are provided under test showing how to use GapFindFill.jl to solve two problems in Manaras paper. testGapFind.jl and testGapFill.jl demonstrate how to set up GapFind and GapFill models, respectively. We reported our experimental results here for users' reference. All experiments were run on an Intel Core i7-6700 CPU with Ubuntu 18.04.3 LTS.
For find_gaps, the testing example contains 1668 compounds and 2383 reactions, which is of the size of real problems.
The expected outcome is to find 115 blocked metabolites in the network, namely, 1553 non-blocked metabolites. The following table shows a run-time comparison between GAMS and GapFindFill.jl on gap finding.
| Software | Solver | Running time (s) |
|---|---|---|
| GAMS | CPLEX | 3.3 |
| GAMS | Gurobi | 0.6 |
| GapFindFill.jl | CPLEX | 68.4 |
| GapFindFill.jl | Gurobi | 64.5 |
GAMS is fast in solving large-size problems like the testing example, but GapFindFill.jl finished the job in less than 2 minutes, which is also acceptable for real problems as large as this example. Using either Gurobi or CPLEX in GapFindFill.jl did not make much difference in running time, while GLPK was unable to solve the testing example within 1 hour. Thus, although find_gaps allows users to specify GLPK as the solver, it is recommended only for small-scale problems.
For fill_gaps_min, the testing example contains 1822 compounds and 2888 reactions. The outcomes are summarized in the following table:
| No-production-metabolite | Solution 1 | Solution 2 |
|---|---|---|
| 2doxg6p[c] | EX_2DOXG6P(e) & 2DOXG6P_t | |
| 2dglc[c] | EX_2DGLC(e) & 2DGLC_t | EX_2DOXG6P(e) & 2DOXG6P_t |
| alatrna[c] | TRNAALA_t & EX_TRNAALA(e) | ALATRNA_t & EX_ALATRNA(e) |
| 2dr5p[c] | DRIB_t & EX_DRIB(e) | 2DR5P_t & EX_2DR5P(e) |
| 4gudbutn[c] | 4GUDBD_t & EX_4GUDBD(e) | 4GUDBUTN_t & EX_4GUDBUTN(e) |
The following table shows a run-time comparison between GAMS and GapFindFill.jl on gap filling:
| Software | Solver | Running time (s) |
|---|---|---|
| GAMS | Gurobi | 6.2 |
| GapFindFill.jl | Gurobi | 28.7 |
The difference between GAMS and GapFindFill.jl is smaller than solving gap-finding tasks. This is largely because GapFindFill.jl benefited from Julia's just-in-time feature, as similar problems were solved repeatedly.
Two interfaces are provided, find_gaps() and fill_gaps_min(), for gap finding and filling, respectively.
The find_gaps() interface:
function find_gaps(isRev::Array{Bool}, isCyt::Array{Bool}, isExt::Array{Bool},
stoiMatrix::Array{Float64}, fluxLB::Array{Float64}, fluxUB::Array{Float64};
epsilon::Float64 =0.001, bigM::Float64 =1000.0, nonZero::Float64 =1e-8,
solver::Module=Gurobi)Input arguments:
| Argument | Required | Description |
|---|---|---|
| isRev | yes | true if the corresponding reaction is reversible |
| isCyt | yes | true if the corresponding metabolite is in the cytosol |
| isExt | yes | true if the corresponding metabolite is in the extracellular compartment; |
| stoiMatrix | yes | stoichiometric matrix, |compounds| * |reactions| |
| fluxLB | yes | flux lower bound; |
| fluxUB | yes | flux upper bound; |
| epsilon | optional | minimum amount to be considered active; |
| bigM | optional | constant used in MILP model; |
| nonZero | optional | minimum stoichiometric coefficient to be considered non-zero; |
| solver | optional | CPLEX, Gurobi, or GLPK. |
Outputs Description:
| Argument | Description |
|---|---|
| m | the JuMP model; |
| objVal | objective value, i.e., number of non-blocked compounds; |
| status | termination status; |
| binX | 1 if the corresponding compound is non-blocked; |
The fill_gaps_min() interface:
function fill_gaps_min(isMd::Array{Bool}, isDb::Array{Bool}, isRev::Array{Bool},
isCyt::Array{Bool}, isExt::Array{Bool}, noProdID::Array{Integer},
stoiMatrix::Array{Float64}, fluxLB::Array{Float64}, fluxUB::Array{Float64};
epsilon::Float64 =0.001, bigM::Float64 =1000.0, nonZero::Float64 =1e-9)Inputs description:
| Argument | Required | Description |
|---|---|---|
| isMd | yes | true if the corresponding reaction is in the model; |
| isDb | yes | true if the corresponding reaction is in the database, i.e., not in the model; |
| isRev | yes | true if the corresponding reaction is reversible |
| isCyt | yes | true if the corresponding metabolite is in the cytosol |
| isExt | yes | true if the corresponding metabolite is in the extracellular compartment; |
| noProdID | yes | indices of no-production-metabolites; |
| stoiMatrix | yes | stoichiometric matrix, |compounds| * |reactions| |
| fluxLB | yes | flux lower bound; |
| fluxUB | yes | flux upper bound; |
| epsilon | optional | minimum amount to be considered active; |
| bigM | optional | constant used in MILP model; |
| nonZero | optional | minimum stoichiometric coefficient to be considered non-zero. |
Outputs Description:
| Argument | Description |
|---|---|
| results | a dictionary, each key is an index of no-production-metabolites, and each value is a Set containing all possible gap-filling ways with a minimum number of added reactions. |
- Maranas, Costas D., and Ali R. Zomorrodi. Optimization methods in metabolic networks. John Wiley & Sons, 2016.
- Satish Kumar V, Dasika MS, Maranas CD. Optimization based automated curation of metabolic reconstructions. BMC Bioinformatics. 2007 Jun 20;8:212. doi: 10.1186/1471-2105-8-212. PMID: 17584497; PMCID: PMC1933441.
- GapFind/GapFill in GAMS
Need help with installation or function? Feel free to contact VarnerLab or authors.
The work described was supported by the Center on the Physics of Cancer Metabolism at Cornell University through Award Number 1U54CA210184-01 from the National Cancer Institute. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute or the National Institutes of Health.