Open exchanges output in minimal_media() could be inconsistent across machines

[anmole17]

I was trying to test how many alternative minimal_medium() are possible with open_exchanges as True for each model. #589 is what I was testing for different models.
I first tried my code on windows machine and then ported to my server with podman for faster results but the results were not replicating.
But if I am getting different outputs on different machines (Windows and Podman
container [Ubuntu]) with same code and python packages, the concept of limited solutions for minimal_media() seems wrong. @sunilnagpal :)

I am not sure if this should go as a question or an issue, so have made an issue just in case.

My questions are:

• Why does cobrapy act in this way?
• Is this just the difference in machines?
• If so are there ways to improve this? Can I do anything about it?
• How does the minimal_medium() function works? Why does it limit to just 5 output as in the example case, when we can see there are 5 more solutions also feasible. How and why does it limit the output?
• If the same model shows very different outputs which one is correct? Does it still hold any biological relevance if I know there are possibly many other alternative medias which are possible other than the outputs I observe.

These versions are same for both machines as I thought maybe those could be the reason:
Cobrapy version: 0.29.1
swiglpk: 5.0.12
optlang: 1.8.3

The solution i got on different machines:
Windows:

	0	1	2	3	4
EX_cpd00012_e0	0.216872	0.217619	0.186794	7.023511	5.270229
EX_cpd00028_e0	0.003096	0.003096	0.003096	0.003096	0.003096
EX_cpd00030_e0	0.001548	0.001548	0.001548	0.001548	0.001548
EX_cpd00034_e0	0.001548	0.001548	0.001548	0.001548	0.001548
EX_cpd00048_e0	0.109715	0.109715	0.109715	0.109715	0.109715
EX_cpd00051_e0	0.124897	0.124897	0.124897	0.124897	0.124897
EX_cpd00058_e0	0.001548	0.001548	0.001548	0.001548	0.001548
EX_cpd00063_e0	0.001548	0.001548	0.001548	0.001548	0.001548
EX_cpd00099_e0	0.001548	0.001548	0.001548	0.001548	0.001548
EX_cpd00108_e0	501.252	2.655737	322.5745	2.997655	2.267766
EX_cpd00149_e0	0.001548	0.001548	0.001548	0.001548	0.001548

Podman:

	0	1	2	3	4
EX_cpd00012_e0	22.43154	9.681643	15.375	8.121669	5.49252
EX_cpd00028_e0	0.077686	0.003096	0.003096	0.003096	0.003096
EX_cpd00030_e0	0.038843	0.001548	0.001548	0.001548	0.001548
EX_cpd00034_e0	0.038843	0.001548	0.001548	0.001548	0.001548
EX_cpd00048_e0	2.752599	0.109715	0.109715	0.109715	0.109715
EX_cpd00051_e0	3.133482	0.124897	0.124897	0.124897	0.124897
EX_cpd00058_e0	0.038843	0.001548	0.001548	0.001548	0.001548
EX_cpd00063_e0	0.038843	0.001548	0.001548	0.001548	0.001548
EX_cpd00099_e0	0.038843	0.001548	0.001548	0.001548	0.001548
EX_cpd00108_e0	59.81598	665.8628	4.887436	500.05	2.378912
EX_cpd00149_e0	0.038843	0.001548	0.001548	0.001548	0.001548
EX_cpd00182_e0	119.3432	2.619473	4.954539	2.526269	0
EX_cpd00205_e0	0.038843	0.001548	0.001548	0.001548	0.001548
EX_cpd00220_e0	0.077686	0.003096	0.003096	0.003096	0.003096

To replicate the problem:
I got GSMM models for testing from: https://github.com/danielriosgarza/MAMBO, but this should be applicable for any model(I do not think it is a problem of models)
Code:

# making the media complete for the models setting bounds(-1000,1000) then finding all the possible medias
def min_medias(cmodel):
    growth=0.5
    model=cobra.io.read_sbml_model(folder+cmodel)
    model.id=cmodel.replace('.xml','')
    add_media(model,model.id,complete=True)
    min_comp_open=minimal_medium(model, growth, minimize_components=100, open_exchanges=True)
    return(min_comp_open)

min_medias("Zunongwangia_0.xml")

Podman image replication:

FROM python:3.10

WORKDIR /home

COPY requirements.txt .
RUN pip install --upgrade pip && pip install -r requirements.txt

requirements.txt:

optlang
swiglpk
micom
Cython
biom-format
optlang
modelseedpy
ncbi-genome-download
memote
pyscipopt
carveme

Please inform if anything else is required to address the issue.

[cdiener]

So for minimize component it only minimizes the number of different consumed metabolites, the actual fluxes for the imports are somewhat arbitrary here, so I would expect you get different solutions on different machines or software versions. Also note that in general for constraint optimization like FBA or minimal media, the objective is guaranteed to have a unique solution but the variables/fluxes do not. So there can be many import fluxes that yield the optimum.

And regarding the questions:

Why does cobrapy act in this way?

Because in this case the fluxes can be anything that fits the constraints, so many non-unique solutions are possible.

Is this just the difference in machines?

Yes somewhat. It is the difference in bases the solver estimates to solve the problem. Those are affected by many things including floating point accuracy.

If so are there ways to improve this? Can I do anything about it?

Either don't use minimize_components or follow this up with another "normal" minimal_medium call. Meaning, first run it with minmize_components, then set all other imports to 0 and run minimal_medium again with open_exchanges set to False.

How does the minimal_medium() function works? Why does it limit to just 5 output as in the example case, when we can see there are 5 more solutions also feasible. How and why does it limit the output?

Because it is computationally expensive to generate more MIP solutions. In many cases one only wants one minimal solution.

If the same model shows very different outputs which one is correct? Does it still hold any biological relevance if I know there are possibly many other alternative medias which are possible other than the outputs I observe.

All of them are correct in the sense that they yield a medium that fulfills the growth rate requirement and does contain the minimum number of different metabolites.