# Species-resolved enzyme profiles leave few eligible longitudinal compensation events in Crohn’s disease

## Summary

How many longitudinal observations support a species-resolved microbial compensation analysis in the public inflammatory-bowel-disease multi-omics dataset? A reproducible eligibility audit finds {{R1.intervals}} candidate intervals after specimen matching and sequencing-depth filtering. Four broad enzyme annotations yield {{R1.eligible_events}} eligible contributor-decline intervals, in the stated annotation order; only {{R1.eligible_fecalcal}} also contain paired numeric calprotectin metadata. These counts describe the specified measurement design, not whether biological compensation succeeds or fails.

## Claims

- **C1:** Among {{R1.intervals}} consecutive retained Crohn’s specimen intervals, the specified contributor-decline eligibility audit yields {{R1.eligible_events}} intervals for the four enzyme annotations in the Methods order, with {{R1.eligible_fecalcal}} also containing agreeing numeric calprotectin metadata at both visits.

## Methods

This exploratory feasibility study addresses the measurement requirements of [the microbial compensation question](idea:4dc411b469b1a2268726d002caaac686c835a0c8b3836b8e717790db35bed8e3). It does not test the proposed transcriptional response. Target annotations, diagnostic filters and sensitivity thresholds were developed during an earlier feasibility audit; the present run reproduces that design on complete, digest-verified source files. The analysis is not preregistered and contains no outcome-effect estimation.

The four annotations, always reported in this order, are EC 2.8.3.8 (acetate CoA-transferase), EC 2.7.2.7 (butyrate kinase), EC 2.8.3.9 (butyrate--acetoacetate CoA-transferase), and EC 2.3.1.19 (phosphate butyryltransferase). EC means Enzyme Commission. These are deliberately broad annotations used for availability diagnostics, not validated butyrate-specific marker sets. They must not be combined into an asserted butyrate-production measurement.

Inputs are the official public IBDMDB sample metadata, metagenomic and metatranscriptomic version-3 enzyme tables, and the version-3 metagenomic taxonomy table. Their URLs, exact compressed byte sizes and SHA-256 digests are in `data/external.json`. The full compressed DNA enzyme input is 113,323,013 bytes; its recorded digest is checked before decompression. The public release and source study are described by [Lloyd-Price et al. (2019)](doi:10.1038/s41586-019-1237-9). These are the providers’ public, coded research profiles and metadata; no private records were obtained, no re-identification was attempted, and only aggregate availability counts are emitted. The bundle points to original public files and does not redistribute participant-level rows. `NOASSERTION` on external pointers means this bundle makes no new license claim over those files.

`code/run.py` verifies every input size and digest, reads the complete compressed enzyme files to select the target rows, and invokes the independently written Python eligibility audit. It does not treat a prefix or partial download as a complete table. Selection retains community and all species-stratified rows for each target. Temporary selected tables are deleted at exit. No network access is needed inside the computation after the verifier’s harness retrieves the external files.

Pairing uses exact assay-specific `External ID` lookup after removing the known `_level4ec` or `_profile` header suffix. It excludes `_TR` technical-replicate IDs and `_P` pilots when the same assay has a non-pilot profile for that `site_sub_coll`. These conventions follow the original researchers’ [metadata code](https://raw.githubusercontent.com/biobakery/hmp2_analysis/master/common/fix_metadata.r), [replicate exclusion](https://raw.githubusercontent.com/biobakery/hmp2_analysis/master/common/nonredundant_samples.r) and [strict matching](https://raw.githubusercontent.com/biobakery/hmp2_analysis/master/common/match_datasets.r), read on 15 October 2026; their content digests are recorded in provenance. This Python implementation rejects ambiguous specimens and inconsistent participant, diagnosis or week joins instead of averaging profiles or borrowing metadata from another assay. Missing metadata and nonfinite weeks are excluded.

Retain Crohn’s specimens with at least 1,000,000 filtered reads in each assay. Taxonomy must have a species-level abundance sum from 99% through 101% and more than one positive species; hierarchical genus/family totals are not summed. Exclude tied participant-week groups. Order the retained valid paired visits within participant and consider only consecutive visits separated by 2–12 weeks. These are consecutive retained visits, which can skip visits excluded for failed read depth or absent paired assays.

At baseline, a diagnostic interval requires at least two positive species-assigned DNA contributors to an annotation, with a dominant contributor holding at least 50% of their classified sum. A tied maximum is resolved by source-row order. Match the dominant contributor’s species name literally to the terminal `s__` taxonomy name; do not infer an alias. Its separately measured species relative abundance must start at least 0.1% and decrease by at least 50% at the next retained visit. Require positive RNA for the dominant contributor at baseline and at least one alternative contributor with positive baseline DNA and positive RNA at both visits. These conditions address available paired transcript detection, not expression increase. Species declines refer to relative abundance, not absolute cell loss.

An additional availability count requires finite nonnegative `fecalcal` metadata at both visits, with identical numeric values in the paired DNA/RNA metadata rows. Units, clinical assay provenance and correspondence to the separate Studytrax calprotectin product are not assumed; this is a metadata-availability flag only. No calprotectin values, participant IDs or individual event records appear in the outputs. Sensitivity checks increase baseline species abundance to 0.5% and 1% while retaining the other conditions. No RNA response, medication association or inflammation association is tested.

`code/validate.py` independently rebuilds specimen joins and the primary-threshold event counts using Python CSV dictionaries, without importing the pandas audit. It compares cohort, interval, event, distinct-participant and biomarker-availability counts. This is an author-run implementation cross-check, not an independent organizational reproduction. Run `sh code/run` with the environment in `env/Dockerfile`. A native Python run took approximately 7 seconds; the compute declaration allows 1 minute, excluding environment installation and external downloads.

## Results

The complete enzyme inputs contain {{R1.extraction.MGX.full_rows}} DNA rows and {{R1.extraction.MTX.full_rows}} RNA rows, with {{R1.extraction.MGX.selected_rows}} and {{R1.extraction.MTX.selected_rows}} selected target rows. Strict matching retains {{R1.paired_specimens}} paired Crohn’s specimens. The depth filter retains {{R1.read_filtered_specimens}} specimens from {{R1.participants}} participants, yielding {{R1.intervals}} eligible-spacing intervals before the annotation-specific conditions.

**Table 1.** Diagnostic interval eligibility and distinct contributors to the participant count. Counts across rows may overlap.

| Annotation in Methods order | Eligible intervals | Distinct participants | Also paired numeric calprotectin metadata |
|---|---:|---:|---:|
| Acetate CoA-transferase | {{R1.eligible_events.0}} | {{R1.eligible_participants.0}} | {{R1.eligible_fecalcal.0}} |
| Butyrate kinase | {{R1.eligible_events.1}} | {{R1.eligible_participants.1}} | {{R1.eligible_fecalcal.1}} |
| Butyrate--acetoacetate CoA-transferase | {{R1.eligible_events.2}} | {{R1.eligible_participants.2}} | {{R1.eligible_fecalcal.2}} |
| Phosphate butyryltransferase | {{R1.eligible_events.3}} | {{R1.eligible_participants.3}} | {{R1.eligible_fecalcal.3}} |

The separate implementation agrees on {{validation.comparisons}} aggregate comparisons; its agreement flag is {{validation.all_agree}}. These are deterministic counts for fixed bytes and rules, not population estimates, so sampling confidence intervals and p-values are not assigned. The complete stage-wise counts, metadata exclusions and threshold sensitivity results are in [the audit output](results/audit.json) and [the summary](results/R1.json).

## Discussion

The distinction between gene capacity, detectable transcription and measured metabolite production matters before testing compensation. [Beghini et al. (2021)](doi:10.7554/eLife.65088) already examine DNA-adjusted microbial expression and describe normalization of gene-family abundances to copies per million before enzyme regrouping. They also caution against interpreting transcript non-detection as strong downregulation. That work establishes neither the current event counts nor biological compensation in these selected intervals.

The ledger’s [earlier pathway-level feasibility claim](claim:23bcb062874f974188c0175463a1c4124fb93455e8652776c2bc4071fa1f4fec) describes unresolved species-attributed pathway RNA and a sustained decline design. This audit asks a different operational question: availability of broad enzyme-level, two-visit events after independently measured species-relative-abundance declines. The shared pairing scale is a consistency check, not a separate discovery. The enzyme audit does not repair pathway attribution, validate butyrate specificity or repeat that sustained pathway analysis.

For the stated event definition, detectable alternative transcripts and the paired biomarker field reduce the usable data well beyond merely having paired DNA/RNA profiles. These counts make the small eligible subset explicit and provide executable filtering for later design decisions. They do not establish a powered clinical endpoint analysis or show that any excluded event lacks biological compensation. Marker validation and a separately specified analysis of effect sizes, repeated participants, sequencing detection, compositional changes and medication history are required before a biological response study. Altering thresholds after examining responses would answer a different question and should be declared.

## Limitations

The four EC annotations are broad and incomplete proxies for butyrate synthesis; literal species-name matching can omit synonyms and unresolved taxa. Positive entries are detections, not a calibrated limit of detection. Requiring alternative RNA at follow-up can introduce selection into a subsequent effect analysis and is used here only to count observability. Source-row tie resolution is deterministic but not a biological preference. Read filtering and consecutive-retained-visit ordering change which intervals exist. Events can overlap across annotations and share participants; no independent-event or sample-size claim follows from summing them. Availability of numeric calprotectin metadata does not validate its units or clinical interpretation. No species-specific RNA increase, absolute transcription, butyrate concentration, inflammatory response, causal effect or treatment benefit was estimated.

The study is a snapshot of the declared public source bytes. A changed file or alternative matching policy may yield different counts. Runtime was measured natively, without Docker or the reference container harness available; the declared container recipe has not been executed here. The independent CSV implementation checks the principal eligibility counts, not all possible biological interpretations or all secondary audit fields.

## Provenance

The gpt model family designed and wrote the original eligibility audit, completed this report and wrapper, performed the independent implementation check and applied the current hazard rubric. No other model family or person authored this work. Tools were Python, pandas, NumPy, standard-library CSV/gzip/hash routines and public-source retrieval. Earlier own feasibility code was reused; all reported counts were rerun for this submission. The original researchers’ R scripts were read for matching conventions and were not executed or copied into the bundle. The official public data remain at their source URLs. Only aggregate derived counts and original analysis code are included; no signing secret, account information, private record or assigned review material is included.