raven-toolbox release notes#

Unreleased#

Breaking changes#

get_model_from_homology with preferred_order treats a template gene that maps to a gene an earlier template already maps as a subunit without an ortholog, and complex_policy decides what happens to it (was: the gene and every part of the rule that needs it were removed). For the rule (a and b) or x with a taken by an earlier template, flag gives (OLD_T2_a and nb) or nx, keep gives nb or nx and drop gives nx. A reaction that has genes, none of which keeps a mapping, is not transferred.
get_model_from_homology with complex_policy="drop" removes each complex that has a subunit without an ortholog, and removes the reaction only when no complex remains (was: the reaction was removed as soon as one subunit had no ortholog). For the rule (A and B) or (C and D) without an ortholog of B, the draft has C and D.
manipulation.contract_model is the new name of remove_duplicate_reactions, after RAVEN's contractModel, and the old name is removed. contract_model takes dist_reverse (default True), the distReverse option of contractModel: with False, reactions with the same metabolites and a different reversibility are duplicates.

Data#

The KEGG assets are the kegg120 release of toolbox-assets (was: kegg118): the reference model and tables, the taxonomy and the profile HMM libraries of 4246 prokaryotic and 2724 eukaryotic KOs. The baked registry and data/manifest.json list them.

3.0.0b2 (2026-11-01)#

Second beta toward 3.0. The main changes are RAVEN-format model export, SuppOrth homology input, and a closer match to RAVEN in homology, gap-filling, curation tables and ftinit.

Breaking changes: fill_gaps_kumar_milp is renamed fill_gaps_milp, omics.hpa moved to init.hpa, connect_blocked_reactions renames eps to threshold, ftinit takes rxn_scores as a keyword, manipulation.remove_genes defaults to blocked_reactions="constrain", manipulation.merge_compartments defaults to drop_single_metabolite_reactions=False, init.ftinit defaults to big_m=1000, and export_for_git writes its .mat file in the RAVEN structure.

Model I/O#

New: io.write_matlab_model and read_matlab_model write and read the RAVEN model structure, field for field as RAVEN's exporter does. read_matlab_model reads a COBRA Toolbox .mat through cobra's reader. export_for_git writes its .mat file with write_matlab_model (was: cobra's COBRA Toolbox layout).
The SBML file of export_for_git writes subsystems as SBML groups and rxnNotes, rxnReferences and rxnConfidenceScores under the labels that RAVEN's importModel parses. It does not write the Python repr of metaData as a note (was: it did); only string-valued notes are written, and the model version and date are their own notes.
read_yaml_model reads the SBO term as a string and whole-number charges as ints, so charges survive an SBML write.
The Excel export no longer lists InChI twice in the MIRIAM column.
New: get_model_metadata and set_model_metadata. export_for_git takes version and date and stamps them on the exported copy.

Homology and curation#

New: reconstruction.homology.hits_from_supporth and run_supporth read SuppOrth consensus orthologs as homology input.
get_model_from_homology now matches RAVEN's getModelFromHomology: it keeps every hit tied for the best bitscore at strictness 3, applies max_evalue exclusively, and filters only_genes_in_models on the template gene. Hit values are parsed with round-trip float precision, so a hit at the cutoff is no longer let through.
get_model_from_homology gained keep_gene_free and preserve_notes (default False) to keep gene-free reactions and the notes of the template. It drops duplicate genes, raises on gene ids that cobra would split, on an emptied search direction and on organism ids that differ only in case, and warns for review_identity on SuppOrth hits. It follows the complex policy flag or drop, the preferred order and the order of the templates. The draft model does not carry the version and date of the template.
curation.batch_curate stores notes, references, confidence score and InChI under the keys the YAML layer uses.

Gap-filling and analysis#

Renamed gapfilling.fill_gaps_kumar_milp to fill_gaps_milp (module gapfilling.kumar_milp to gapfilling.milp, KumarGapFillResult to MILPGapFillResult), matching RAVEN's gapFillMILP. It adds template reactions only; reversed_reactions and weights are removed.
fill_gaps_milp runs with integrality tolerance 1e-9 and returns exit_status="not_restored" when the selection does not restore growth. An infeasible MILP on Gurobi no longer crashes.
gap_report raises a ValueError naming reactions with a forced flux when the model is infeasible (was: a bare cobra Infeasible).
biomass.scale_biomass gained target_total and raises before changing the model when the balancing component cannot absorb the change.

Context-specific models#

omics.hpa moved to init.hpa; the omics package is removed.
init.ftinit / run_ftinit gained feas_tol, opt_tol, int_feas_tol and presolve (defaults unchanged).
init.ftinit skips the tie-break when the primary solve stops at the time limit. init.run_ftinit, and so ftinit, warns when it solves with GLPK. init.merge_linear and manipulation.merge_compartments keep the solver of the source model.
With resolve_ties, the tie-break phases of init.ftinit and init.fill_tasks use 0.5 of time_limit. init.fill_tasks gained mutate_in_place.

Alignment with MATLAB RAVEN#

read_yaml_model turns a usage_prot_* reaction forward when its prot_pool coefficient is positive and prot_pool_exchange when it is negative, as GECKO's upgradeEcModel does, and changes the sign of the objective coefficient of a reaction it turns (was: a negative lower bound decided, and the coefficient stayed, which reversed the objective of the model).
write_yaml_model writes an ec reaction without enzymes with no enzymes entry (was: enzymes: !!omap []). read_yaml_model reads a missing enzymes entry, a content-less enzymes: !!omap and enzymes: !!omap [] as an empty enzyme map, so files written by earlier versions of GECKO, RAVEN and raven-toolbox still load.
init.fill_tasks with resolve_ties=True on Gurobi restores MIPGap and MIPGapAbs of reference_model after each gap-fill (was: 0.0 and 0.4 stayed set).
tasks.find_task_essential_reactions writes its cache_path as JSON (was: pickle) and ignores, with a warning, an existing file that is not a JSON cache of the current version, such as a pickle. checkpoint_every below 1 raises ValueError (was: a write after every task).
init.hpa_gene_scores uses the default level scores also for an empty level_scores mapping, and init.rna_gene_scores compares tissue case-insensitively and takes the highest TPM over tissue names that differ in case, as RAVEN's scoreModel does.
localization.predict_localization with apply=True returns a copy of the model in LocalizationResult.model also when no gene of the reactions to place has a score (was: the input model itself).
localization.assign_compartments takes the default growth floor (min_growth=None) as 10 % of the maximum flux of biomass_reaction (was: of the optimum of the model's current objective).
reconstruction.kegg.organisms_in_domain raises ValueError for an empty domain (was: it matched every organism).
conditions.load_condition returns an empty dict for a YAML file without content (was: None; other empty values such as [] are returned as they are), which apply_condition applies as a condition without steps, as RAVEN's parseYAML and applyCondition do.
analysis.reporter_metabolites leaves a scored gene that has no entry in gene_fold_changes out of both the "up" and "down" subsets (was: it counted as fold change 0 and joined "up"), like a gene with a NaN fold change.
manipulation.constrain_reversible_reactions skips the Gurobi tolerance setting only when the solver has no Params attribute (was: any exception was ignored).
gapfilling.connect_blocked_reactions raises a RuntimeError that names the non-optimal MILP (the message said that the blocked reactions cannot carry flux with all templates). comparison.compare_models annotates tasks as also accepting os.PathLike.
manipulation.set_exchange_bounds is annotated as returning list[str | Metabolite]: the entries of mets that are not exchanged, as supplied (ids, names or Metabolite objects).
scripts/analyze_init_params.py runs the mip_gap, force_on and prep_scale sweeps with the ftinit default big_m (1000, was: 100) and takes --big-m to set it. The big_m sweep starts at 1000, which is its reference row. The cached sweep results are keyed on big_m, so a resumed pickle from a run with another big_m is not reused.
Models read with io.read_yaml_model and io.read_matlab_model have model.tolerance set to 1e-9, the feasibility and optimality tolerance RAVEN's solveLP gives the solver (cobrapy's default is 1e-6 for Gurobi and 1e-7 for GLPK). utils.solver.set_raven_tolerance applies it to any other model.
manipulation.remove_genes defaults to blocked_reactions="constrain" (was: "remove"), as RAVEN's removeGenes does: a reaction whose rule loses its last enzyme stays, with the rule dropped and bounds (0, 0). Pass blocked_reactions="remove" to delete such reactions.
manipulation.remove_genes also removes every gene that no remaining reaction named in its rule before the removal, also when not requested (was: only the requested genes), as RAVEN's removeGenes does, when at least one of the requested genes is in the model; when none is, the model is not changed. A gene kept only by a constrained reaction stays.
init.merge_linear numbers the merged groups 1, 2, ... in the order of the first reaction of each group in the original model, as RAVEN's mergeLinear does, so group_ids no longer depends on the merge order.
manipulation.merge_compartments gives a merged metabolite the id of the copy already in the merged compartment, if any, otherwise the id that sorts first (ACCOA_c for ACCOA_c and ACCOA_m), as RAVEN's mergeCompartments does. It rewrote the first member's compartment suffix to the merged compartment (ACCOA_s). The merged metabolite also keeps the annotation of that copy (was: dropped), and drop_single_metabolite_reactions defaults to False, as RAVEN's deleteRxnsWithOneMet (was: True).
manipulation.merge_compartments stores the base id of each merged metabolite in notes["merged_base_id"] (ACCOA for the kept ACCOA_c; the own id when two merged metabolites would get the same base id). The default metabolite key of localization.assign_compartments, apply_assignment, relocate_reactions and curation_priority is that stored base id, otherwise the id without the suffix of the metabolite's own compartment, so the merged ACCOA_c in s and a universal ACCOA_m share the key ACCOA. The four functions raise ValueError when two metabolites in one compartment have the same key (was: one of them stood for both).
manipulation.merge_compartments documents where notes["merged_base_id"] survives: YAML written and read by raven-toolbox, copy(), cobra JSON and SBML. RAVEN's readYAMLmodel shows it in metNotes (merged_base_id: ACCOA note: free text note) and no RAVEN function uses it. A round trip through RAVEN YAML or a .mat file loses it, and the localization functions then key a merged ACCOA_c in s as ACCOA_c, without an error.
io.read_yaml_model puts the notes that write_yaml_model writes as a notes mapping (notes without a RAVEN field) back into notes (was: nested under notes["note"]).
reconstruction.homology.run_blast, run_diamond and run_supporth, and reconstruction.kegg.run_hmmsearch, get_kegg_model_from_sequences, build_ko_hmm and build_hmm_library, use every CPU the process may run on when threads is not given, as RAVEN's getBlast, getDiamond, runSuppOrth and getKEGGModelForOrganism do (was: the machine's CPU count minus one).
gapfilling.analyse_topology takes RAVEN's gapFillTopological defaults. Seeds: the metabolites of every reaction with only substrates and lower_bound < 0 or only products and upper_bound > 0, whatever its number of metabolites (was: one-metabolite reactions with lower_bound < 0); a UserWarning when there is none (was: a message only when verbose). Targets: the metabolites of the first objective reaction, substrates and products (was: every metabolite of the model); ValueError when the model has no objective.
New: utils.parse_gr_rule reads a grRule in the grammar of RAVEN's parseGrRule (and/or in any case, &, &&, |, ||, operators as whole words). manipulation.gpr_to_dnf now also takes a rule string and reads it with this parser, as RAVEN's grRuleToDNF does; utils.is_dnf reads a string with it too, and both raise ValueError with the message of RAVEN's RAVEN:badGrRule for a malformed rule. G1 && G2, G1 || G2 and G1 Or G2 are read as rules, G1&G2 is one gene id. Whitespace is what MATLAB's isspace counts, so a no-break space (U+00A0), U+0085, U+2007 and U+202F are part of a word.
curation.batch_curate (and batch_curate_from_tsv) stores a gene rule as parse_gr_rule reads it, so the reaction gets the genes that were checked: G1&G2 or G3 has the genes G1&G2 and G3, as in RAVEN's curateModelFromTables (was: G1, G2 and G3).
io.read_yaml_model reads gene rules with parse_gr_rule, so G1&G2 written by write_yaml_model is read back as one gene (was: G1 and G2, cobra's reading). A rule that parse_gr_rule refuses is read by cobra as before.
Functions that copy or build a gene rule keep RAVEN's reading of it, so G1&G2 stays one gene (was: the rule was copied as a string and cobra split G1&G2 into G1 and G2): manipulation.merge_compartments, copy_to_compartment, convert_to_irreversible, merge_models, add_reactions_from_model, expand_model, change_gene_reaction_rules, add_reactions_from_equations, localization.apply_assignment, init.fill_tasks, init.remove_low_score_genes, reconstruction.homology.get_model_from_homology, reconstruction.kegg.assemble_model_from_ko_genes and io.read_matlab_model. A rule string is read with parse_gr_rule; one that parse_gr_rule refuses is read by cobra. cobra's JSON reader (cobra.io.load_json_model) still splits G1&G2; cobra's SBML reader keeps it.
analysis.fseof (and so geckopy.ec_fseof) solves each level with pFBA that holds the objective at 0.999999 of its optimum at that level, as RAVEN's solveLP does, where it held the full value. On the scipy solver (HiGHS) the optimum is only accurate to the solver's feasibility tolerance, so the second step was infeasible at the top level of ecYeastGEM and the scan returned 9 of 10 levels; it now returns all 10, as on Gurobi and GLPK. A level at which the biomass maximisation is infeasible still ends the scan.
analysis.fseof classifies a reaction as a target when its absolute flux exceeds 1e-10 at some level and its standard deviation is above 0, and takes slope and correlation on the absolute flux (was: the signed flux). A knock-out target has a last absolute flux of at most 1e-8. flux_eps gates the check against the theoretical maximum.
analysis.fseof takes the top of the scan as max_fraction of the target flux in the pFBA solution that maximises the target, as RAVEN's FSEOF and GECKO's ecFSEOF do, where it took the plain maximum (about 1e-6 relative difference).
analysis.fseof lists targets of equal score in model order (a stable sort), as MATLAB's sort does, where their order was arbitrary.
analysis.fseof gained equality: True fixes the target flux at each level as GECKO's ecFSEOF does, False (the default) keeps the lower bound of RAVEN's FSEOF.
io.read_yaml_model reads an ec reaction written as - enzymes: !!omap with nothing under the tag, the form GECKO and RAVEN write for a reaction without enzymes, as a reaction with no enzymes, as readYAMLmodel does. It raised AttributeError: 'ScalarNode' object has no attribute 'flow_style'.
io.read_yaml_model reads a scalar written as ""text"" (invalid YAML, written by older RAVEN and GECKO for a note that carried its own quotes) as "text", as readYAMLmodel does. Files that parse as YAML are read unchanged.
io.read_yaml_model reads a double-quoted scalar with an unescaped quote inside, such as - notes: "a "b" c", as a "b" c: it drops the outer pair of quotes and unescapes \\ and \", as readYAMLmodel does. This repair runs only when the strict parse fails.
io.read_yaml_model reads a single-quoted scalar with an apostrophe inside, such as - name: 'it's', as it's: it drops the outer pair of quotes and reads each '' as ', as readYAMLmodel does. This repair runs only when the strict parse fails; a comment after the closing quote is not handled.
init.ftinit and init.run_ftinit give each scored reaction its own big-M, its flux capacity 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; its default is 1000 (was 100).
tasks.find_task_essential_reactions (and so init.prep_init_model) re-solves a knock-out from a cold start when the warm-started solve does not report a feasible status, and counts the reaction as essential only if that solve is not feasible either, as RAVEN's solveLP does. Gurobi could report a feasible knock-out as infeasible when started from the previous knock-out's basis, which added reactions to the essential set.
init.ftinit and init.run_ftinit with resolve_ties take each step's fluxes from the least-total-flux solution for the tie-broken on set, so the direction and magnitude at which the next step forces a reaction no longer depend on the solver's vertex. That phase stops at a relative or absolute gap of 1e-6 and has the full time_limit, as in RAVEN.
init.ftinit and init.run_ftinit with resolve_ties: among the least-total-flux solutions, each step takes its fluxes from the one with the least id-rank-weighted backward flux, so the direction of a flux that carries the same total either way no longer depends on the solver. The id-rank phases rank a merged reaction by the smallest original reaction id in its group (new run_ftinit argument rank_keys), so they no longer depend on which member names it.
init.ftinit and init.run_ftinit with resolve_ties raise a UserWarning naming the phase (phase2c-minflux, phase2d-direction) when the minimal-flux or the direction phase finds no solution, as RAVEN does, and keep the fluxes of the phase before it. A phase that stops at the time limit is adopted with a UserWarning that its fluxes are an unproven incumbent.
init.ftinit with resolve_ties: the last step skips the minimal-flux and direction phases, as no step reads its fluxes, and reports the fluxes of the id-rank phase. The reaction set does not change.
Docstrings are rewritten in impersonal, present-tense wording with British spelling; the code is unchanged. Statements that differed from the code are corrected, for example find_good_reactions, sum_biomass, diff_models, compare_models, set_gene_rule, constrain_reversible_reactions, ftinit (essential forcing, mip_gap, determinism figures), fast_lp metabolite matching and batch_curate index matching.

Further changes#

curation.batch_curate and batch_curate_from_tsv follow curateModelFromTables: reaction tables are grouped by rxnIdx and name, the header spellings of RAVEN are accepted, a new reaction without bounds gets lower bound -inf and the stored defaultUB of the model (+inf when none is stored), new ids use the width of the largest existing id, an empty MIRIAM cell removes the namespace, and an empty charge or score is 0 for a new entity. The new argument overwrite_with_empty (default True) lets an empty cell of an existing entity clear the value. All tables are validated before the model changes: cells are trimmed as MATLAB strtrim, numbers follow the plain decimal rule of RAVEN, a charge must be a finite whole number, rev must be 0, 1, true or false, rxnIdx must be a whole number of at least 1, a malformed gene rule, a repeated rxnIdx and a gene listed twice raise, and repeated coefficients are summed.
manipulation.merge_models keeps all metabolites, genes and compartments of every model, concatenates the objectives, keeps the compartment name of the first model and appends new items in order (was: only items used by reactions, the objective dropped, the last name won). add_transport_reactions sets the subsystem to "Transport between and " for a reversible reaction and "Transport from to " otherwise (was: empty).
manipulation.simplify_model and remove_duplicate_reactions follow simplifyModel and contractModel: the order of the modes, the lowest-index survivor, reserved reactions and orphan metabolites; duplicates are grouped by stoichiometry and reversibility, with bounds and objective merged, and remove_duplicate_reactions has no reserved option. remove_zero_interval_reactions and remove_no_flux_reactions no longer remove genes. expand_model keeps the first isozyme copy at the original position.
New: utils.sort_model (counterpart of sortModel) and utils.ids.next_ids. utils.generate_new_ids takes the numeric maximum, as generateNewIds does.
biomass.set_gam stores the exact GAM value, checks the NGAM reaction before it edits the biomass reaction, and refuses an NGAM value without an NGAM reaction. The rebalancing of the charge refuses unset charges.
init.ftinit takes tissue, celltype and hpa by position, and transcr and rxn_scores as keywords, and has the new arguments remove_genes, use_scores_for_tasks, no_gene_score, isozyme_scoring, complex_scoring, multiple_cell_scoring, hpa_level_scores and data_precedence. Scoring follows scoreModel (level None scores -8, tissue and cell type names are compared case-insensitively, an unknown level raises). A call that passes rxn_scores by position needs the keyword. The output is built from the reference model with the bounds and directions of the template and has no gene or metabolite without a reaction.
init.remove_low_score_genes takes the new argument tied_reactions, a list that receives the ids of the reactions cut to one of tied isozymes.
init.prep_init_model prunes unused genes as prepINITModel does, keeps the metabolites that tasks name, accepts repeated and empty task ids and removes a task that cannot be performed by its position. rescale_for_init leaves bounds alone, and prep_init_model resets them to ±1000. get_init_steps with the full series ignores no metabolites and carries the single MILP run of RAVEN. comparison.compare_models fills the task matrix by position.
tasks.parse_task_list leaves a bound unchanged when its CHANGED LB or UB cell is empty (was: -1000 and 1000), raises when the ID column is first, as RAVEN does, and refuses a binary .xls file with an error that names .xlsx. check_tasks warns when a name[comp] matches several metabolites, and duplicate or conflicting task inputs and outputs raise ValueError (was: outputs overwrote input bounds).
gapfilling.connect_blocked_reactions renames eps to threshold (default 1.0), has the new argument min_flux=0.1, prices renamed template reactions by their template id, and restores the steady-state bounds after the net-production check. connect_blocked_reactions and fill_gaps_fast_lp rename a template reaction whose id is already a draft id to <id>_<template id>. FVA and blocked-reaction searches in gap-filling run in the calling process (was: a worker process, which hung on Python 3.14).
analysis.reporter_metabolites raises ValueError for p-values outside [0, 1] (was: they were dropped), centres the z-scores on the first one before taking moments, and returns only all with std_z 0 for a single gene or zero fold changes.
io.write_yaml_model quotes the YAML 1.1 booleans (y, n, yes, no, on, off in any case), writes a single ec code or SMILES annotation as a list, omits empty reaction names, and warns about metabolites with an empty or duplicate name. It writes defaultLB and defaultUB each when the stored metadata has it.
io.export_for_git writes the formats yml, xml, mat, xlsx and txt by default, the txt table headed Rxn id without padding, the gene name as fbc:label in SBML, and a dependencies.txt with the versions of python, cobra, libSBML, ruamel.yaml, scipy, openpyxl, raven_toolbox, and the model version and date. main_branch_flag raises when a raven_toolbox checkout is on a branch other than main or master, and the flags must be bool.
write_fasta writes only the header line for an empty sequence. manipulation.add_reactions_from_equations adds new genes in sorted order. The essential set of tasks.find_task_essential_reactions follows the reaction order of the model and does not depend on the hash seed.
localization parses UniProt locations by whole words, with isoform blocks, all columns and duplicate genes. init.parse_hpa_rna reads the Sample and Value columns, raises when the tissue or value column is missing, and keeps the last of repeated rows.
The package supports Python 3.14 and 3.15, and the CI tests both.

Packaging and docs#

Assets are downloaded from CompMetEng/toolbox-assets; publish_to_raven_data.py is renamed publish_to_toolbox_assets.py.
The package metadata and the README name the GPL-3.0 licence of the LICENSE file (was: MIT).
The docs/ tree and its Read the Docs build are removed; the documentation is on raven-docs. The project URLs point to CompMetEng, and data/deeploc is not part of the source distribution.
Docstrings are corrected for reporter_metabolites, random_sampling, remove_no_flux_reactions, check_tasks, connect_blocked_reactions and find_leak_metabolite.

3.0.0b1 (2026-09-07)#

First beta toward 3.0, matching the upcoming MATLAB RAVEN 3.0 beta.

keep ftinit() as the only extraction pipeline.
analysis.compare_fluxes replaces follow_changed: a compareFluxes port, following RAVEN dropping followChanged/followFluxes/mapCompartments.
comparison.compare_models gained EC-code / metabolite-name / equation overlap matrices, folded from compareRxnsGenesMetsComps.
New: curation.rename_model_genes, ported from renameModelGenes.
New: curation.get_gene_data, download_genome_data, process_protein_fasta_file: the genome-data cluster.
New: manipulation.replace_metabolite, ported from replaceMets.
New: gapfilling.find_leak_metabolite, ported from findLeakMetabolite.
New: manipulation.remove_bad_reactions, ported from removeBadRxns.
New: gapfilling.gap_report, ported from gapReport.
New: analysis.trace_flux_path, ported from traceFluxPath.
New: analysis.get_flux_z / analyze_sampling, ported from getFluxZ/analyzeSampling.
New: analysis.fit_parameters, ported from fitParameters.
New: utils.guess_composition, ported from guessComposition.
New: analysis.walk_fluxes / FluxWalker, ported from walkFluxes.
New: analysis.get_min_nr_fluxes, ported from getMinNrFluxes.
New: utils.generate_new_ids, ported from generateNewIds.
init.ftinit implements the metabolomics production-bonus (metabolomics=, prod_weight=).
init.ftinit reproducibility parameters renamed: strict_gap/canonical are now prove_abs_gap/resolve_ties, with corrected determinism numbers.
ftINIT task gap-filling now edits the reference model in place instead of copying it, and is ~23% faster on a genome-scale extraction.
Fixed three ftINIT/task-essential bugs: a task-essential boundary default, a checkpoint file-handle leak, and a non-finite tie-break objective.
manipulation.close_model boundary-reaction detection is now structural, matching RAVEN.
manipulation.convert_to_irreversible now splits exchange reactions too, carries a negative objective coefficient onto the reverse reaction, and gained an optional rxns restriction.
conditions.apply_condition raises on unknown charge during charge balancing instead of silently miscalculating.
annotation.load_delta_g_csv now always records CSV values literally, matching deltaGCSV.
manipulation.remove_dead_end_reactions dead-end detection is now bounds-blind, matching RAVEN.
analysis.get_min_nr_fluxes no longer crashes with GurobiError reading fluxes from an infeasible solve.
analysis.fseof gained an explicit min_target scan-floor override.
Fixed six redundant-calculation hot spots: merge_linear, batched add_reactions callers, analyse_topology, trace_flux_path, remove_dead_end_reactions, find_task_essential_reactions.
random_sampling and fill_gaps_fast_lp parallelized with ProcessPool.
io.write_yaml_model fixed: an infinite/NaN bound wrote as a bare word instead of the YAML token; sort_ids=True dropped the !!omap tag on compartments; unknown metaData fields are now preserved on write; empty gene_reaction_rule is now dropped. read_yaml_model treats a null genes/reactions/metabolites key as empty.
Parameter defaults measured and unified: BLAST/DIAMOND/HMMER threading, homology min_align_len/evalue, get_init_model.allow_excretion.
Removed io.export_model_to_sif: a mislabeled port of a MATLAB function that never existed.

0.4.0 (2026-08-28)#

A raven-gecko-parity cross-validation harness closed six MATLAB-parity divergences and brought the YAML writer to byte-parity; ftINIT gained an opt-in deterministic-extraction mode; compartment placement and gap-filling in assign_compartments were fixed for determinism and reliability.

convert_to_irreversible now carries a reaction's annotations, subsystem and notes onto its _REV half, matching convertToIrrev.
ftinit gained opt-in deterministic extraction (strict_gap, canonical) for reproducible MILP tie-breaking, default off.
assign_compartments's gap-fill now uses a flux-based fill (pFBA) instead of cobra's unreliable indicator-MILP gapfill.
assign_compartments reaction placement is now deterministic and score-aligned instead of a solver tie-break (yeast agreement 52.8% → 72.5%).
diff_models compares GPRs as logic (DNF-normalised), not text; operand order no longer registers as a difference.
load_delta_g_csv now stamps every matched value literally by default, matching deltaGCSV.
confidence.annotate_confidence runs every applicable scorer in one call; curation_priority can now drop curator-settled placements from the review queue.
New thiele_palsson_score derives the Thiele & Palsson 2010 confidence score from the gene_association facet.
merge_compartments/copy_to_compartment are now public API; merge_compartments no longer drops single-metabolite reactions or the model objective.
Six MATLAB-parity divergences closed via the new raven-gecko-parity harness: check_tasks boundary semantics, apply_condition exchange-reset direction, remove_duplicate_reactions GPR union, add_reactions_from_equations metabolite naming, export_to_excel default-bound cells.
write_yaml_model brought to byte-parity with writeYAMLmodel.m: numeric coercion, metaData field order, MIRIAM collapsing, compartment defaulting, quoting, indentation. read_yaml_model no longer crashes on a model with no genes.
BLAST/homology defaults matched to RAVEN and measured: run_blast.evalue → 1e-4, min_align_len → 100; new review_identity surfaces near-miss homology candidates.
New set_exchange_bounds, RAVEN's setExchangeBounds port.
New cross-language parity testing harness (tests/parity/): exact, set-level, and statistical comparison tiers against real MATLAB RAVEN output.
Fixed a downloaded binary bundle leaving its non-primary executables non-executable.

0.3.0 (2026-07-16)#

Compartment localisation and per-reaction confidence tracking, new gap-filling and flux-sampling algorithms, ftINIT brought to parity with RAVEN, and KEGG artefact hosting moved to the dedicated raven-data repository.

export_for_git can pin the MATLAB .mat variable name via varname.
ftinit/prep_init_model now match RAVEN's model preparation, solver parameters, and gap schedule; new manipulation.simplify_model. Human-GEM prep model sizes now match RAVEN's.
New confidence facets equation and gene_association (mass/charge balance, GPR presence) plus facet_summary; validated against yeast-GEM's own curated confidence levels.
Fixed get_elemental_balance crashing on an unparseable polymer formula (now unknown).
Fixed score_localization_confidence vetoing reactions it simply couldn't measure.
Fixed gap-fill materialising on the wrong compartment metabolite in assign_compartments.
New localization.transport_evidence: per-gene transporter evidence lowers a metabolite's transport cost in the assignment MILPs.
assign_compartments/apply_assignment/AssignmentProposal consolidated into raven-toolbox from the standalone assignCompartments repo.
Benchmarked the compartment-assignment MILP against CarveFungi's own carve-MILP and against RAVEN's predictLocalization, more accurate and deterministic on both.
New triage_localization: flags genes/reactions whose localisation call is shakiest, each with a plain-English reason.
Retuned DEEPLOC_COMPARTMENT_TRUST and the min_confidence/membrane_threshold gates from real yeast validation data.
Cross-species DeepLoc benchmarks (yeast, Arabidopsis, Chlamydomonas, Human-GEM): DeepLoc 2.1 generalises across kingdoms.
load_deeploc gained normalise=False for raw calibrated probabilities (accuracy-neutral for assignment).
New combine_scores (weighted-sum consensus across localisation predictors) and min_confidence=/membrane_split= loader gates.
New prepare_deeploc_input, fetch_protein_sequences, write_fasta: writes a DeepLoc-ready protein FASTA for a model's genes.
New localisation loaders: load_mulocdeep, load_compartments, load_uniprot, fetch_uniprot_localization. Removed load_wolfpsort.
Flux sampling: CHRR and ACHR unified under random_sampling(method=...); ACHR is now the default. Breaking: pass method="random_objective" for the previous default.
New gap-filling algorithms: fill_gaps_fast_lp (+ SWIFTCORE variant), fill_gaps_kumar_milp, analyse_topology.
KEGG artefacts moved from raven-toolbox releases to the dedicated raven-data repository.
Refreshed kegg118 artefact set; fixed 404ing release-asset URLs.
Binary provisioning split into runtime/build sets with an explicit raven-toolbox-binaries fetch CLI; native-Windows HMMER support; RAVEN_PYTHON_AUTOFETCH=0 toggle.
KEGG artefact build is now resumable (skips completed stages) and reports progress via tqdm.
export_to_excel writes populated model.ec data (ENZYMES/ENZRXNS sheets).
CI now runs on macOS and Windows in addition to Linux.
Docstring rendering fixes for set_gam and add_sbo_terms (no API change).

0.2.0 (2026-06-14)#

Project rename plus KEGG-reconstruction and CI improvements.

Project renamed raven-python → raven-toolbox; import package is now raven_toolbox.
get_kegg_model_from_sequences uses hmmsearch instead of hmmscan.
get_kegg_model_for_organism_from_artefacts auto-resolves its taxonomy artefact.
Test data no longer ships real KEGG records; they are generated at runtime instead.
Removed the unimplemented visualization stub and its extra.
CI updated to Node 24 / actions/checkout@v5 / actions/setup-python@v6.

0.1.0 (2026-06-10)#

First release with published, downloadable KEGG artefacts, plus a cobra-aligned hardening pass.

KEGG artefacts published (kegg116): ensure_kegg_data/ensure_kegg_hmm_library fetch version-pinned, SHA256-verified files from the GitHub release; gzip instead of xz so MATLAB and Windows can read them without an external tool.
New reconstruction.kegg.phyl_dist (PhylDist), a port of RAVEN's getPhylDist.
raven_toolbox.__version__ now derives from installed package metadata.
Solver/feasibility failures now raise cobra.exceptions.OptimizationError consistently.
Hardening pass: path-traversal guard on ZIP extraction, random_sampling NaN guard, connect_blocked_reactions non-positive-penalty guard, SHA256 cache re-verification, and fixes for a handful of latent edge-case crashes, each with a regression test.

0.1.0a1 (2026-05-30)#

First alpha release. Covers the functional scope of RAVEN built on cobrapy: de-novo reconstruction (KEGG / homology), context-specific modeling (tINIT / ftINIT), metabolic-task validation, connectivity gap-filling, HPA omics ingestion, sub-cellular localisation, N-model comparison, reporter metabolites, FSEOF, flux sampling, and the RAVEN-style I/O formats (YAML / SIF / Excel). Validated against MATLAB RAVEN on Human-GEM (Jaccard 0.975–0.980).

Licensed under MIT (previously GPL-3.0-or-later).
Sphinx + MyST documentation site.
Not yet implemented: visualization, metabolomics-based (f)tINIT scoring, published binary/KEGG-artefact release bundles.