[edkerk]

geckopy gives the same results as GECKO on the cases the comparison in CompMetEng/raven-gecko-parity#141 turned up, and works on the scipy (HiGHS) solver. Needs CompMetEng/raven-toolbox#185 (fseof(..., equality=True), the pFBA top level and the YAML readers); the minimum raven-toolbox version is raised once a release contains it.

Solver status.

• flexibilize_enz_concs and sensitivity_tuning decide an infeasible start from the solver status, so Gurobi, GLPK and scipy behave alike. flexibilize_enz_concs warns and returns an empty result, as flexibilizeEnzConcs does. sensitivity_tuning still raises ValueError at the start and warns and stops inside the loop.

scipy. optlang's scipy interface keeps a dense matrix that it copies on every edit, renumbers columns wrongly after a variable removal, removes several constraints at once incorrectly, ignores model.tolerance and returns no duals. The new geckopy.solver_support module works around this:

• A sparse problem class replaces the dense one. EcModel.solver repairs a stale index map when it is read. load_conventional_gem and load_ec_model build on GLPK and convert at the end, and get_subset_ec_model and apply_kcat_constraints (for 100 or more ec reactions) edit on GLPK and convert back.
• Shadow prices come from HiGHS row marginals after a primal and dual optimality check. get_enzyme_bottlenecks, Enzyme.shadow_price and relax_proteomics_greedy use them. relax_proteomics_greedy takes one price vector per iteration and breaks ties by lowest ID.
• model.tolerance reaches HiGHS, and a solve that ends with numerical difficulties is repeated at the default tolerance. An infeasible solve returns a solution with its status.
• Light Human-GEM now builds in about 7 minutes on scipy (it took 2 h 17 min) and get_subset_ec_model finishes in about 5 minutes (it took 65 min). ecYeastGEM loads in 25 to 36 s. ec_fva on ecYeastGEM takes 52 to 81 minutes on scipy, against 9 minutes on Gurobi, and its bounds differ from Gurobi's by up to 0.15 percent because HiGHS is less accurate on these LPs.

Same results as GECKO.

• get_subset_ec_model removes genes as removeGenes does, keeps the ec rows of a light ecModel (the 001_ prefix) and drops the metabolites and genes no reaction uses. It raised ValueError: <gene> not found on the light Human-GEM ecModel.
• assign_standard_kcat compares subsystem names without regard to letter case, and a reaction without a subsystem gets the model-wide standard kcat.
• load_uniprot_tsv and get_uniprot_ids_from_table read the cells like textscan with %q: quotes are removed and tabs inside quotes belong to the cell. A quoted EC cell such as "1.1.1.25; 2.5.1.19" kept its quotes and changed the BRENDA matches of 44 reactions in the Yarrowia build.
• load_ec_model reads a file with an empty enzymes: !!omap and with unescaped quotes in a value, as loadEcModel does. This stays in geckopy until the minimum raven-toolbox version contains the same repair.
• pfba_enzymes holds the objective at 0.999999 of its optimum, as getPfbaEnzymes does. Pass fraction_of_optimum=1.0 for the previous behaviour.
• ec_fseof fixes the target flux at each level by equality, as ecFSEOF does. The targets table lists reaction targets before transport targets, each group in score order, and has a new boolean column transport. This changes the row order for callers that take the top N rows.