Reviewers 1.1 / 1.4: Euler for Fig 1B, a legend for Fig 2A, a pathway panel, and the growth basis

[freiburgermsu]

Update: the whole paper is now on the live-only population

Per the corresponding author's decision on the growth-basis question below, this PR no longer just fixes the 55% figure — it converts the entire manuscript to the live-only (non-obsolete) basis.

The abstract headline changes: ~56,000 reactions becomes ~48,000. That is the number people quote, so it is called out here rather than buried. Growth becomes 34% in compounds and 34% in reactions, measured identically. Every count in M09, M11, M12, M13 and Supplementary Table S1 moves with it, and so do all three figures — none of which filtered obsolete records before, so they were silently on the other population.

analysis/population_basis_table.py prints every quantity on both populations. Its all-records column reproduces the submitted draft exactly (84.3% / 42.9% / 41.1%, 30,157 and 25,855, 95.0%, 4,248 and 13,289, the full grade table), which is what makes the live column trustworthy.

Two things the conversion exposed, both now disclosed in the text:

• The anchor set is duplicated. 806 records carry a stereo-exact measurement, but 441 are obsolete records that each link to a live replacement — duplicates of the same chemistry. There are 365 distinct anchored reactions. The calibration was fitted over all 806 and has not been refit (that changes published grades and needs its own PR; grade_thermo_sources.py also currently dies under the pinned pandas). S03 and M12 now state that the effective sample size is smaller than the count implies.
• M13 atom-mapping numbers were stale regardless of population: the draft said 32,877 / 25,058 / 7,819 against a shipped 32,378 / 25,058 / 7,320 — the 2026-09-14 chain-salvage withdrawal. Live: 26,256 / 19,810 / 6,446.

---

Answers reviewer 1's comments 1 (where is new information still arriving) and 4 (the figures), and fixes three things neither reviewer raised but both would have hit.

Reviewer 1, comment 4 — the figures

"The figures could use a little improvement. There is a bit of barplot abuse (smiley face). For instance, Fig 1B would be more informative as a Venn diagram. Fig 2A needs a color legend (and maybe a different color scheme)."

Fig 1B → Euler diagram

The reviewer is right, and for the reason implied. The stacked bar reported that 67% of MetaCyc's reactions are unique — but never said who the other 33% are shared with, which is exactly the question integrating Rhea raises. The regions are well separated:

region	reactions		region	reactions
MetaCyc only	21,258		MetaCyc ∩ Rhea	2,943
Rhea only	8,339		MetaCyc ∩ KEGG	2,115
KEGG only	5,756		KEGG ∩ Rhea	706
all three	5,489		in none of the three	9,406

Circle areas are not proportional — three-set areas cannot be solved exactly — so the counts carry the quantity and the circles carry only membership. The caption says this rather than leaving a reader to assume proportionality. The 9,406 reactions in none of the three are labelled rather than silently dropped.

matplotlib-venn is not in the environment, so the panel is drawn directly with matplotlib patches — no new dependency for one panel.

Fig 2A → a legend, and a colour that means something

The panel had four segments and identified two; the caption was doing the legend's job in prose. There is now a four-entry key.

On the colour scheme, the reviewer's parenthetical was well aimed. Both non-Marvin segments were grey, which visually merged two different things:

• no ionizable site — a chemistry result: Marvin ran and reported no dissociable proton between pH −2 and 16, for 1,242 compounds
• no structure — a curation gap: there is nothing to run on

These are not two shades of one thing. The result now takes its own categorical slot; only the genuine absence stays grey.

On "barplot abuse"

Fair. Figure 1B stops being a bar, and the new 1D earns its bars by carrying a comparison rather than three numbers.

Figure 2B is the remaining weak panel — it carries three numbers (39%, 53%, 53%) that the text already states, and folding it into 2C would be the right call. It is deliberately not done here: removing it renumbers 2C→2B, which edits the same line of M05 that PR #297 rewrites, and a guaranteed conflict is not worth a cosmetic gain. Flagged as a follow-up once #297 lands.

---

Reviewer 1, comment 1 — where new information is still arriving

"Assuming that central pathways are already well characterized for many years, it would be curious to see where are we still gaining new information."

The reviewer's hypothesis is right and now quantified. Only 41 of the 12,261 reactions added since 2020 fall in a central-carbon or energy-metabolism class. New panel 1D shows the eight MetaCyc classes that gained most, against what was already held:

class	new	already held	enrichment
antibiotics	265	740	2.0×
O-antigen	218	29	41×
polyketides	169	214	4.3×
toxins	142	192	4.1×
lipids	140	281	2.7×
branched fatty acids	98	8	67×
fatty acids	77	358	1.2×
alkaloids	53	196	1.5×

Named pathways make it concrete: mycolate biosynthesis (57), arachidonate metabolites (37), colibactin (34), apratoxin A, platensimycin, acylsucrose, H. pylori O-antigen. Mycobacterial cell wall, a bacterial genotoxin, eicosanoid signalling, two antibiotics.

A data gap this exposed

Building the panel required a join the released files cannot currently do. Unique_ModelSEED_Reaction_Pathways.txt stops at rxn48568 and Unique_ModelSEED_Reaction_ECs.txt at rxn48573, both just below the 2020 boundary (rxn48575) — so not one of the 12,261 post-2020 reactions carries a pathway or EC annotation, in the alias files or in the per-reaction records. The annotation step was never re-run over the MetaCyc and Rhea intake.

figure_common rebuilds the mapping from the shipped MetaCyc source table instead, putting both eras through the same one-level parent join so the two bars are comparable — reading the pre-2020 side out of the alias file would compare a full ancestor closure against a direct-parent lookup, which is meaningless. The underlying annotation gap is a data defect and gets its own PR; this panel does not depend on it.

Two traps worth recording for whoever does that PR: MetaCyc_pathways.tsv mixes individual pathways and ontology classes in one id column (Degradation, Biosynthesis and Energy-Metabolism are all ids), so treating every id as a pathway finds zero classes; and Super-Pathways is organisational rather than biological and would otherwise top the chart.

---

Three things neither reviewer raised

1. M12 cites a Figure 3 panel that does not exist

\ref{fig:direction}D — Figure 3 has A, B and C. Checking the whole manuscript turned up something worse: panels 3B and 3C were never cited by letter anywhere, so the only references into Figure 3 were A and a panel that isn't there. The grading sentence now points at B and C, which are the two panels that actually show what it describes.

2. "55% growth in reactions" is not reproducible on any single population

basis	compounds	reactions
all records — the paper's basis elsewhere	33,992 → 45,708 = +34.5%	43,774 → 56,012 = +28.0%
live only (obsolete dropped both sides)	+34.5%	+33.7%
2026 all records vs 2020 live	—	+54.7%

The 55% is the third row: 2026's totals against a 2020 baseline with its 7,577 obsolete rows removed. The compound figure is the first row and is correct.

This PR quotes 28%, keeping the whole paper on one population — the abstract's "~56,000 reactions", Figure 2A's 56,002, M12's 25,855 and Supplementary Table S1 are all all-records figures, and Figure 1's source counts (which apply no obsolete filter — verified in _reaction_sources()) become consistent rather than wrong.

This is a corresponding-author call, flagged rather than settled. Live-vs-live gives a tidier 34%/34%, but adopting it means converting every other number in the paper, including the abstract's headline. Say the word and I will do that instead.

(A correction on my own part: I first reported that the tree no longer reproduced the manuscript's grade counts — 27,355 graded against 33,099, 365 anchors against 806 — and attributed it to the September rebuilds. That was my own obsolete filter, not a regression. On the manuscript's basis the tree reproduces Supplementary Table S1 exactly, 806 anchors included, matching the stereo_exact rows in opentecr_comparison.csv. Nothing needs re-measuring.)

3. No figure in this paper could be regenerated

figure_common.py raised FileNotFoundError on import in a clean checkout. _silver_sigma_band() read two intermediates under results/thermo_grades/ that are .gitignore'd (Biochemistry/Thermodynamics/SourceGrading/.gitignore line 12) and absent, and grade_thermo_sources.py cannot rebuild them — it dies on ValueError: Encountered all NA values under the pinned pandas.

It now derives the same band from files that ship: best_grade from the committed reaction_grades.tsv, and each source's σ from the released Biochemistry/reaction_*.json, which is where the published uncertainty lives anyway. It reproduces the published band exactly — eQuilibrator 0.27–1.81, group contribution 1.62–25.12, dGPredictor 2.09–38.84 — once the sentinels the Results section already documents are excluded. Leaving them in put eQuilibrator's p95 at 23,901 kcal mol⁻¹, which is how the bug announced itself.

---

Also included

• analysis/pathway_distribution_of_growth.py and analysis/review_transport_and_llm.py — the two scripts behind every number in this PR series, registered in analysis/README.md. Both read only Biochemistry/ and this repository's git history, so unlike their siblings they need no EQUILIBRATOR_DIR.
• reviewer_response_plan.md — every reviewer comment mapped to its origin in the manuscript, the measurement, and the change.

Verification

• latexmk -pdf clean for main.tex and supplementary.tex; zero undefined references or citations.
• Both figures regenerated from the scripts in this PR and inspected at render size for label collisions.
• Body text ends on page 6; the reference list spills to page 7, as it does in Reviewers 1.2 / 1.3: transport scoring limits, and what template scope actually means #297 and Reviewer 2.2: give the LLM ensemble a measured role, and disclose its forward bias #298. \lastpage{6} and the combined count need one pass after all four merge.

🤖 Generated with Claude Code

---

Merge note (added after the live-basis conversion)

This branch and #297 both touch M13_results_mapping.tex, which has a single-line body, so git cannot auto-merge them:

• Reviewers 1.2 / 1.3: transport scoring limits, and what template scope actually means #297 appends a \paragraph{Reconstruction scope.} after the atom-mapping sentence
• Reviewers 1.1 / 1.4: Euler for Fig 1B, a legend for Fig 2A, a pathway panel, and the growth basis #299 rewrites the atom-mapping sentence itself (stale numbers + live basis)

The resolution is to keep both, in this order:

Atom mappings are published for 26,256 reactions, 54\% of the database: 19,810 clean, and 6,446 that required salvage (which are tagged). As the mappings can be used to span a metabolic reconstruction, reactions needing one repair are likely to carry others.

\paragraph{Reconstruction scope.} A template carries far fewer reactions than the database holds ...

No other pair of branches in this series conflicts — verified with git merge-tree across all four.

---

Review pass (2026-09-23)

Six defects found and fixed, two substantive:

1. Panel 1D's "already held" bars were still all-records. _pathway_classes iterates the alias file, which has no obsolete flag, and was the one loader LIVE_ONLY didn't reach — antibiotics read 824 where the live count is 740. The table above is corrected; "added" bars are unchanged (every addition is live).
2. Figure 2C's histograms came from figure_data_sigma.json, a static all-records cache with no generator. Its drawn medians (0.63 / 10.41 / 17.01) contradicted M11's live text (0.62 / 10.83 / 17.56). Now derived from the records; cache deleted; medians match the text exactly.
3. Caption and M09 did not say panel D can only place MetaCyc-sourced additions — the 8,409 Rhea reactions have no pathway ontology and are absent from it. A reviewer would have asked. Both now say so.
4. S03's "unpaired" sentence had lost its denominator (now 370 of 4,690).
5. cover_letter.md still said 56,000.
6. The _growth docstring called the live-basis 2020 counts "wrong" (they differ from all-records by exactly the 34 obsolete compounds).

Known and deliberately not fixed here, for the authors: S02's "~34,400 of ~56,000 reactions (61%) contain a magnesium-binding participant" is all-records and needs the eQuilibrator cache's 510-compound list to recompute. And M12's "eQuilibrator resolves 83.3% of the reactions it covers" uses a denominator that includes its refusal sentinels (20,890) while the "~17,900" two sentences later excludes them — a pre-existing inconsistency in the draft; with sentinels excluded eQuilibrator resolves 97.4%, which changes the narrative, so it is flagged rather than changed.