MOMA predicts most genes to be essential

[wshao1]

Hi, I was curious to try out MOMA as an alternative to FBA for identifying essential genes. I applied cobra.flux_analysis.single_gene_deletion to Human-GEM, and the result is that almost all genes are identified as essential (2807/2897). The same holds true for contextualized human models I have generated. Is this expected behavior for MOMA?

Many of the post-gene KO growth-rates are zero, indicating that MOMA opted for zero growth rate compared to a suboptimal one. I reasoned that this might be because compensatory pathways are much longer and thus require activation of more fluxes, making it more "ideal" from MOMA's standpoint to simply have zero flux through the system. But how can it be that almost all gene KOs lead to zero growth rate?

[cdiener]

What medium/import bounds did you use here? I would imagine that this makes a difference. I just ran Recon3D and on completely open bounds there is only a single essential gene with MOMA, but the growth rates are also way too high. Another issue could be sink or demand reactions.

[wshao1]

I did not alter any bounds for the generic Human-GEM model, all exchange reactions are completely unbounded. For the contextualized models, the exchange reactions are also unbounded. Neither have demand/sink reactions, upon inspecting model.sinks or model.demands.

Can you explain why sink/demand reactions might pose an issue?

[wshao1]

OK it took a long time to run, but CPLEX does give a different result: 80 essential genes for Human-GEM. I was also able to replicate your result of 1 essential gene for Recon3D.

How concerning is the discrepancy between the two solvers?

[cdiener]

Yeah somewhat. MOMA depends a lot on the reference solution. So the first thing I would check is whether the pFBA solution is the same with both solvers (since pFBA is not guaranteed to be unique). I don't have a Gurobi license so I can't really debug it though. If the reference solution is indeed the same, than one of the two solver interfaces is wrong, because the quadratic problem in MOMA has a unique solution. And that would be something we would need to fix for sure.

[wshao1]

You were right, the pFBA solutions are indeed different (about half of the fluxes differ between the two solvers). Thanks for pointing out the sensitivity of MOMA. Interpretation-wise, is this just an inherent limitation of an under-constrained problem? If both pFBA solutions generated by the two solvers are equally valid, and MOMA is working as designed, then I don't have any reason on theoretical grounds to choose one solver over the other correct?

[cdiener]

You were right, the pFBA solutions are indeed different (about half of the fluxes differ between the two solvers). Thanks for pointing out the sensitivity of MOMA.

Phew, that is a bit less concerning 😅

Interpretation-wise, is this just an inherent limitation of an under-constrained problem? If both pFBA solutions generated by the two solvers are equally valid, and MOMA is working as designed, then I don't have any reason on theoretical grounds to choose one solver over the other correct?

Yeah kind of. MOMA just minimizes the distance to a reference solution, but does not help to pick a good reference solution. Instead of choosing a solver, it would be more about choosing a reference flux distribution/solution here. For instance, does one seem more correct than the other? Maybe one produces a metabolite one would expect and the other doesn't. Also, the constraints are probably an issue as well. pFBA might give you a more unique solution when you apply realistic import bounds (like a medium).

[wshao1]

OK that makes sense. Thank you so much for explaining! 😊