New functionality, dependencies, and error handling#

1. Gap-filling and context-specific model extraction#

	RAVEN 2	RAVEN 3
`fillGaps` algorithm	one built-in MILP connectivity-maximization strategy	same strategy is still the default; `algorithm` option adds `'fastLP'`/`'swiftLP'` (FASTCORE/SWIFTCORE-style LP relaxation), `'gapfillMILP'` (growth-floor MILP with directionality repair), `'topological'` (BFS producibility pre-screen, no model modification)
`fillGaps` positional arguments	`allowNetProduction, useModelConstraints, supressWarnings, rxnScores`	unchanged; new options are name-value-only
`ftINIT` tie-breaking	depends on solver seed	optional `resolveTies` pins degenerate MILP optima to a deterministic answer
`ftINIT` MILP optimality proof	escalating relative-gap search only	optional `proveAbsGap` proves each stage to a fixed absolute gap in one solve
`ftINIT` positional arguments	preserved	unchanged; `resolveTies`/`proveAbsGap` are name-value-only, default off
`getINITModel`/`runINIT`	supported, no runtime notice	supported, now under `INIT/tINIT/`; one-time "legacy, use `ftINIT`" notice (see §5)
`fitTasks` gap-fill mode	one behavior (equivalent to today's `'merge'`)	`gapFillMode`: `'merge'` (default, same as before) or `'preMerged'` (faster, ftINIT-style)
`fitTasks` outputs	`[outModel, addedRxns]`	`[outModel, addedRxns, failedTasks]`
Task requiring net production of a metabolite with no reactions	boundary constraint silently dropped before gap-filling; task could incorrectly report as passing	constraint preserved; task correctly reports failing if the reference model can't produce it

2. New functionality#

All additive; none of this replaces or changes behavior of existing RAVEN 2 functions, except where noted.

Area	New function(s)	What it adds
Flux sampling (`analysis/`)	`sampleACHR`, `sampleCHRR`, `sampleMaxVolEllipse`, `sampleChebyshevCenter`, `sampleWarmupPoints`	`randomSampling` becomes a dispatcher (`'method'` option) across RAVEN 2's original random-objective method (now `'randomObjective'`) plus two new near-uniform MCMC samplers: ACHR (hit-and-run) and CHRR (hit-and-run with rounding, better mixing on ill-conditioned polytopes)
Flux analysis / QC	`compareFluxes`, `getMinimalMedium`, `traceFluxPath`, `walkFluxes`, `modelSummary`	structured flux-vector diffing; MILP-based minimal-medium search; highest-flux-fraction path tracing between two reactions; interactive flux-neighborhood navigation; model/flux summary printing
Model comparison (`comparison/`)	`diffModels`	structured, ID-keyed diff between two models: stoichiometry, bounds, objective, grRule (as logical DNF equality), EC codes, metabolite properties; distinct from the existing overlap-style `compareMultipleModels`
Manipulation	`findDuplicateRxns`, `findPotentialErrors`	standalone duplicate-stoichiometry query; parse-tree-based non-DNF grRule detection (see §8)
Genome-based curation (`curation/`, new folder)	`downloadGenomeData`, `getGeneData`, `processProteinFastaFile`, `renameModelGenes`, `curateModelFromTables`	pipeline from an NCBI genome accession to a gene ID mapping table to bulk gene renaming in a model; `curateModelFromTables` bulk-curates mets/rxns/genes from `.tsv` files against an existing model
Biomass composition (`biomass/`)	`getBiomassFractions`, `scaleBiomassFraction`, `scaleBiomassPseudoreaction`, `setGAM`	organism-agnostic biomass-composition and growth-associated-maintenance (GAM) calibration; `tutorial/tutorial7.m` demonstrates the workflow
Localization (`localization/`)	`assignCompartments`, `defaultCompartmentMap`, `getUniProtScores`	deterministic MILP compartment assignment (alternative to the heuristic `predictLocalization`); `parseScores` gains `'deeploc'`/`'cello'`/`'mulocdeep'`/`'compartments'`/`'uniprot'` predictors (replacing WoLF PSORT; see §3)
Annotation (`annotation/`)	`assignSBOterms`, `loadDeltaGCSV`/`saveDeltaGCSV`	organism-agnostic SBO term assignment; ΔG thermodynamic data CSV round-trip
Conditions (`conditions/`, new folder)	`applyCondition`	applies a data-driven growth/media condition (a YAML file, not code) to a model in a fixed sequence: exchange bounds, optional cofactor-pseudoreaction edits, optional biomass-stoichiometry deltas, per-reaction bound overrides
Internal infrastructure	`ravenModelFields`	central registry of every 1-D model field (entity type, default value), backing `permuteModel`/`removeReactions` internally

3. Toolbox dependencies and installation#

	RAVEN 2	RAVEN 3
MATLAB version	no pinned/documented minimum (guidance: avoid very new MATLAB syntax, to accommodate older installs)	CI is tested against R2024b
Statistics Toolbox	required (`combnk`)	not required (`nchoosek`, base MATLAB)
Bioinformatics Toolbox	required (`fastaread`/`fastawrite`)	not required (`io/readFasta.m`/`io/writeFasta.m`)
Optimization Toolbox	required for `qMOMA`/`solveQP`	not required (both removed)
Excel I/O	bundled Apache POI (Java)	dependency-free OOXML writer
LP/MILP solver	required (bundled `glpk`, or Gurobi/SCIP/COBRA)	unchanged, still required
External binaries (BLAST+, DIAMOND, HMMER, cd-hit, MAFFT)	committed to the repository, available offline immediately after clone	fetched on demand from the `raven-data` GitHub release on first use (per-platform ZIPs); an offline `-full`/`-binaries` release bundle is available for air-gapped installs
KEGG HMM libraries / reference data	manual download or local build required	fetched on demand from `raven-data`, cached locally after first use
Installation self-check	`installation/checkInstallation.m`; ran the full test suite to populate its pass/fail table	renamed `checkRaven`, moved to the repo root; does fast, targeted checks (a small import/export round-trip, a trivial LP per solver) instead of invoking the full test classes, and proactively offers to download missing binaries

Net effect: a bare RAVEN 3 clone is far smaller and installs with fewer MATLAB toolbox prerequisites. The trade-off: it needs network access the first time KEGG- or homology-based reconstruction functions actually run.

4. Error and warning handling#

RAVEN 2's dispEM always raised errors with an empty identifier, so no RAVEN 2 try/catch could have matched on err.identifier; RAVEN 3's move to native error()/warning() with real (mostly RAVEN:*) identifiers is additive from that angle, not breaking. Message text is preserved for most existing checks, so text-matching try/catch blocks are unaffected either way. See §3 for the details and the new utils/ravenList.m message-formatting helper.

One related, purely additive change: queries/checkModelStruct.m, called with an output argument (issues = checkModelStruct(model)), now returns a structured array of findings (category/target/message) instead of only throwing or printing, useful for programmatic model QC. The no-output-argument calling convention (throws/warns directly) is unchanged.