RECOMB 2027 DMI reproducibility repository

This study examines when weak qualitative directional information can further reduce uncertainty after strong quantitative flux constraints, using residual flux distributions in genome-scale metabolic models.

Release status: frozen-input reproduction is available for Figure 2D–G, Figure 3B–D, Figure 4B–C, and S1–S3. The current source is RECOMB 2027 DMI commit 225279129; included compact inputs have source SHA-256 records and pass the population and endpoint checks. Other figure assemblies and supplement lineages remain explicitly marked in the reproducibility matrix. The migration history and remaining scientific decisions are in the migration report.

Analysis settings

• A1: HEX1 is the single strong anchor.
• A2: HEX1 + LDH_L are the two strong anchors.
• A separate weak directional cue has primary strength λ = 0.25. Frozen manuscript sensitivities use λ = 0, 0.5 and 1.0 where retained.

The production operators use dimensionless rank-based soft weighting. The DMI measurements do not establish absolute flux equalities or physical-unit conversion. Gains are in native GEM/model flux units. An operator-lineage issue remains: A20 ranks glucose uptake max(-EX_glc__D_e, 0) as its candidate Vmax observable while the manuscript/contracts describe the glucose anchor through HEX1. Their intended equivalence requires author verification; the migration preserves the operator. The anchor operator audit records the traced distinction and required author decision; the labels above remain under review. The historical label A2-L denotes A2 in internal filenames and frozen table schemas; those schemas are preserved.

The frozen design uses iMAT, GIMME, CORDA and RIPTiDe; four RNA contexts; two tumor contexts; five DMI subjects per tumor; and a full 5 × 5 mouse-comparison grid. An individual evaluation compares one CT2A mouse with one GL261 mouse: 4 methods × 4 contexts × 5 CT2A mice × 5 GL261 mice = 400 evaluations. Each of the 16 method/RNA contexts contains all 25 mouse pairs exactly once. The two sets of five are biological DMI subjects; there is no independent five-draw factor. These comparisons share animals and vectors and are not independent biological replicates. See the evaluation design audit. There are 32 reconstruction ensembles with 20 stored vectors each, and 4,179 common non-anchor reaction coordinates. Finite candidate/support counts are not biological replication.

The final utility comparison uses 836 matched truth pairs, with 1,440,313 distinct non-tie cases and 660,237 eligible reaction–evaluation pairs per anchor. Cases are averaged within pairs before pairs receive equal weight. Original A1 held-out confirmation is distinct from development; A2 uses the same split as post-freeze robustness and is not independent confirmation.

Repository structure

Location	Contents
scripts/	Original Bridge geometry, directional update, evaluation and evidence modules; paths retained for import/hash compatibility
11_recomb_flux_expert_generation/src/	Original reconstruction and candidate-vector modules; audit/upstream code, not a turnkey production release
analyses/	A1/A2 workflow descriptions and scientific boundaries
data/	Frozen plotting inputs, derived summaries, and data-boundary policy
figures/main/	Selected source plotting cells and standalone main-panel plotters
figures/supplementary/	S1 support-diagnostic code; manuscript discrepancies are recorded
manifests/	File mappings, source identities, checksums, rights decisions and figure registry
docs/	Reproducibility, panel map, provenance, author-review items and migration audit
environment/	Minimal plotting dependencies and separate optional upstream dependencies
tests/	Existing synthetic scientific-core regression checks

The native code hierarchy is intentionally retained instead of reorganizing scientific modules into new subsystem packages. Empty template directories are not created, and code is not duplicated to satisfy a layout.

Reproducing figures

Install the plotting environment, then run from this root:

python reproduce.py status
python reproduce.py checksums
python reproduce.py validate
python reproduce.py figures --only fig2DG,fig3BC,fig3D,fig4B,fig4C,s1,s2,s3
python reproduce.py tables

validate, figures, and tables run from included files only. validate checks frozen panel populations and key endpoints. The plotting command renders Figure 2D–G, Figure 3B–D, Figure 4B–C, and the audited supplementary Figures S1–S3; tables exports Figure 2 correlation statistics and frozen pair-weighted Figure 3 summaries. See tables/manuscript/ for a checked-in machine-readable snapshot and generation notes. Plot jobs record code, input, and environment identities under ignored reproduced/ output directories.

Outputs go under ignored reproduced/ directories. Existing outputs are reused only when input/code/environment identities and output hashes agree; incompatible outputs are not overwritten. Each plotting job records logs and a status manifest. All persistent writes, temporary work and plot caches are directed into this repository by the entry point. The source repository is never needed by it.

The numeric panels above are reproducible; this does not establish that each saved composite is the final manuscript assembly. Current source artwork and editable components are retained under figures/released/ with identity and status recorded in manifests/ARTWORK.json. Corrected Figure 3D and current three-method Figure 4C assets are generated from bundled frozen tables. Earlier case-pooled Figure 3D artwork and four-method Figure 4C composite remain provenance-only. Supplementary figures S1–S3 have validated standalone reproduction paths. S4–S7 have current source artwork retained, but lack a validated release reproduction path or have unresolved manuscript mapping; see the panel matrix.

Upstream analysis and data availability

The reproduction hierarchy is external inputs → production computation → frozen processed tables → figure inputs → figures and manuscript statistics. Frozen figure inputs should make expensive upstream regeneration unnecessary.

The preserved upstream modules require external datasets, the parent iMM1865 model, frozen medium/projection inputs, and historical identity artifacts. Some foundation patch files are missing from the source checkout. Their gates have not been bypassed; a complete upstream rerun is not currently supported or scientifically revalidated. No expensive computation was performed during migration.

data/README.md explains omitted upstream datasets, archive boundaries, compressed-artifact policy and included derived inputs. No full raw datasets, parent model, credentials, or bulk upstream caches are included.

Citation and license

Use CITATION.cff. Final title, DOI, ORCIDs and conference citation remain TODOs; author names are taken from the current manuscript draft.

Source code is licensed under MIT. Cleared original research products, figures and documentation are licensed under CC BY 4.0. Third-party datasets, metabolic models, software, reconstruction methods and externally sourced materials retain their original licenses and are not relicensed by this repository. Ambiguous artifacts are withheld pending author review.