Gene-level designs contain gene KOs with no reaction-level effect

[VasilisFlouris]

Dear StrainDesign team,

When using SD with gene-level interventions (with sd.ANY), the returned gene_sd solutions sometimes contain gene knockouts that have no metabolic effect. Specifically, a gene KO is included in the gene_sd even though the corresponding reaction_sd contains no reaction KO for any of the associated reactions. The reaction level designs seem to meet the requirements imposed by the suppress and protect modules so the solutions don't appear to be problematic. In case this is relevant, during running straindesign the following message was printed regarding the trimming of some variables "One of the generated reaction names is beyond or close to the limit of 255 characters permitted by GLPK and Gurobi.".

A brief investigation with AI tools points between a mismatch in how the interventions are treated in the optimization vs the reporting space.

Environment used:
SD: v1.18
Solver: gurobi
Model: iML1515 (with modifications that shouldn't affect the gpr rules)
gko_cost, ko_cost and ki_cost included in the problem formulation with an associated cost=1 in every case.
the solutions objects were accessed through sols.gene_sd and sols.reaction_sd with sols being the sd.compute_strain_designs object.

Let me know if any further information is required on my end,

Best,

Vasilis

[VasilisFlouris]

NOTE: The issue was raised based on simulations using gene "names" in the gko_cost dictionary. In subsequent simulations using gene "ids", the problem seems to disappear, no neutral gkos are found. It is unclear to me if this is indeed a solution but the second set of interventions was an order of magnitude larger so it is unlikely that these were not found by chance.

[VonAlphaBisZulu]

Revisited on current main with the setup you provided for #38 and #44 With gene-level interventions (gko_cost by names, Gurobi). Computation with 15 seeds generated 3,000+ gene-KO occurrences and found zero neutral gene KOs (every gene KO maps to a corresponding reaction-level KO), and also no truncated/unresolved gene identifiers. This issue couldn't be reproduced with the current main branch. If you're still seeing it, could you share a minimal failing setup (model + sd_setup + one offending gene_sd) on current main?

In particular, the most likely culprit if it recurs is the reporting/translation of KIs and not-introduced (value-0) entries vs. true gene KOs. A concrete failing case would enable us to tell the difference. Not closing this for now.