[edkerk]

Part of the behaviour parity review of raven-toolbox against RAVEN. Each change below was found by running the same input through both implementations with the scenarios of CompMetEng/raven-gecko-parity; the matching RAVEN changes are in CompMetEng/RAVEN#13.

• reporter_metabolites: z-scores with rounding-level spread were not treated as no spread, the moments were not taken on centred z-scores, and p-values outside [0, 1] were dropped silently; they now raise ValueError.
• fseof: the track used the signed flux and the knockout threshold differed; it regresses the absolute flux, reports every correlated track with a standard deviation above 0, and labels knockouts at 1e-8, as FSEOF.m does.
• remove_duplicate_reactions: it grouped by stoichiometry only and had a reserved option; it groups by stoichiometry and reversibility, merges bounds and objective, and has no reserved option, as contractModel.
• simplify_model: mode order, survivor choice, reserved reactions and orphan metabolites differed; they follow simplifyModel, and the zero-interval and no-flux modes keep genes.
• prep_init_model: metabolites named by tasks were lost when the first simplification removed their last reaction, bounds were not reset after scaling, and non-text task ids were accepted; task metabolites are kept as in prepINITModel, unused genes are pruned, bounds are reset to ±1000, and non-text ids raise ValueError. Repeated task ids are accepted, failed tasks are dropped by position, and compare_models keeps tasks that share an id.
• ftinit: the output was not built from the reference model with the template's bounds and directions, expression data was taken differently, and metabolites and genes without a reaction stayed in the output; all three follow ftINIT.
• ftinit, run_ftinit: each scored reaction gets its own big-M, max(min(ub, big_m), min(-lb, big_m), force_on) with an infinite bound counted as big_m, as RAVEN's ftINITInternalAlg does; big_m is the cap on that capacity and defaults to 1000 (was a fixed 100). find_task_essential_reactions takes its candidates in the model's reaction order, so the essential set does not depend on the hash seed, and re-solves a knock-out from a cold start before counting it as infeasible, as solveLP does; on Human-GEM the essential set equals that of prepINITModel (204 reactions).
• get_init_steps: the full series had no single-MILP run and passed mets_to_ignore; it carries RAVEN's single run and ignores no metabolites.
• hpa_gene_scores, parse_hpa_rna: level names, tissue and cell type were compared case-sensitively, an unknown level was dropped, and the Sample and Value columns were read differently; they follow scoreModel and parseHPA, and an unknown level raises ValueError.
• batch_curate, batch_curate_from_tsv: the rules of curateModelFromTables are applied: grouping, header names and aliases, plain decimal cells (no comma, hexadecimal or non-ASCII digit), cell trimming as MATLAB strtrim, empty cells overwrite existing annotations unless overwrite_with_empty=False, MIRIAM cells, whole-number charges, rev 0/1/true/false, unique rxnIdx, matching reaction tables, summed coefficient rows, gene rules with the grammar of parseGrRule, and every table is validated before the model is changed. Table files are read with the csv module so that a NUL character stays in its cell.
• export_for_git: it supports all five formats, an exact txt table and fuller dependencies, main_branch_flag checks the raven_toolbox checkout, and its flags must be bool.
• generate_new_ids, next_ids: the numeric maximum is taken and a suffix counts only if it is a finite whole number below 2^52.
• set_gam, scale_biomass: the NGAM reaction is checked before the biomass reaction is changed, an NGAM value without a reaction raises ValueError, and the exact GAM value is stored.
• write_yaml_model: empty reaction names were written, default bounds were written when not stored, and equation coefficients differed from num2str; they follow RAVEN, with a warning for duplicate metabolite names.
• connect_blocked_reactions, fill_gaps_fast_lp, fill_gaps_milp: threshold, forced flux, net production, score handling and renamed template reactions follow fillGaps and gapFillMILP.
• get_model_from_homology, load_uniprot, parse_task_list, write_fasta, expand_model, add_reactions_from_equations, rescale_for_init: small differences found by the scenarios are removed (complex policy and preferred order; whole words, isoform blocks and column order; an ID column in first position raises; an empty sequence gives the header only; the first isozyme copy keeps its position; new genes are added in sorted order; bounds are left alone).
• sort_model: added, the counterpart of sortModel.
• release.yml, parity-nightly.yml: pip is upgraded through python -m pip.

Calls that ran before and now raise ValueError: p-values outside [0, 1]; table cells that are not plain numbers, a fractional or infinite charge, a rev other than 0/1/true/false, a repeated or non-integer rxnIdx, a blank, NaN, zero or infinite coefficient, a malformed gene rule, a repeated gene id, and a new reaction that names an undefined gene; set_gam with a value and no reaction; export_for_git with a non-bool flag. batch_curate now overwrites existing annotations with empty cells by default.

Checks: the suite without tests/parity gives 1685 passed and 5 skipped with Gurobi, and 1676 passed and 9 skipped with GLPK; ruff and mypy are clean. The scenarios of CompMetEng/raven-gecko-parity#140 give 86 MATCH and 2 DIFFER on three seeds, and the two DIFFER are documented in its ledger.

Not checked: CI on GitHub, the tests marked parity on a machine without the full Gurobi licence, other cell lines and seeds of the Human-GEM-scale runs (on Human-GEM with HCT116 the ftinit result still differs from MATLAB ftINIT in the forced direction of reversible essential reactions in step 2), and curation tables from real projects against the stricter reading.