# IBDMDB butyrate backup response: prospective analysis plan

Question: In publicly released longitudinal Crohn's disease stool DNA/RNA profiles, are declines in a leading named contributor to a butyrate pathway followed by increased community-normalized pathway RNA from other named contributors, with community pathway RNA maintained?

This is a constrained observational reanalysis, not a test of treatment efficacy, actual butyrate flux, or causation. The population represented by identifiable pathway contributors is not all butyrate-producing organisms.

Before registration we read the idea, literature and documentation; downloaded the public metadata and four pathway tables; inspected feature names, sample identifiers, metadata column names, diagnosis counts, repeated-sample counts, antibiotic/immunosuppressant categories, and calprotectin availability (including some univariate metadata values). We have not inspected pathway abundance values, computed contributor decline events, or examined any RNA/DNA response or inflammation association. HUMAnN2's PWY-5676 and CENTFERM-PWY have only unclassified strata; HUMAnN3's PWY-5676 has six named DNA and three named RNA strata. Selection of the newer release and primary feature is therefore informed by structural feasibility, not fully blind.

## Inputs and cohort
Use the exact official IBDMDB files listed in sources.json, with hashes. Analyze the HUMAnN3 CPM tables primarily, the older relative-abundance tables only for attribution feasibility. Primary feature: PWY-5676 (acetyl-CoA fermentation to butanoate II); secondary feature: P163-PWY (L-lysine fermentation to acetate and butanoate). Report structural coverage for CENTFERM-PWY and PWY-5677, but do not substitute another pathway when a primary test is sparse. These exact identifiers are selected biochemically, not by text matching on outcomes.

Strip exactly the documented _pathabundance_cpm suffix from sample headers. Match DNA and RNA on External ID and metagenomics metadata. Exclude _TR technical replicates. Keep _P samples when present, as they are separately listed sample identifiers. Exclude missing or inconsistent participant/diagnosis/week metadata. If multiple samples have the same participant and week, retain the lexicographically first External ID. Report this attrition. Diagnosis CD is primary; nonIBD and UC are descriptive comparison cohorts.

## Normalization and limitations
Use supplied CPM with no within-pathway renormalization. Named species strata are parsed from the documented delimiter |, excluding unclassified. The unstratified community pathway value is independently estimated by HUMAnN and is NOT the sum of species strata. Do not subtract the leading species from the community total to infer alternative expression; sum the explicitly named alternative strata. Report unclassified and named-stratum signal separately. No species row means not resolved by this pipeline, not biological absence. Treat a species with no row in an assay as zero *reported* abundance, label this limitation, and repeat using leaders resolved in both assays.

## Primary event definition
Use consecutive paired samples within each participant. A triple (t0,t1,t2) is eligible when both intervals are 1 to 8 weeks, t0 and t1 community RNA are at least 1 CPM, the t0 largest named DNA stratum is at least 1 CPM and at least 50% of the sum of named DNA strata, and the t0 sum of other named DNA strata is at least 1 CPM. Break a tied leader lexicographically. No threshold uses subsequent RNA. A sustained-decline event has the same leader's DNA at t1 and t2 each at most 50% of its t0 value. A stable comparator triple has that leader's DNA at both t1 and t2 between 80% and 125% of its t0 value. Other eligible triples remain in the denominator audit but not either event class.

A lagged backup-response signature requires the sum of other named RNA strata at t2 to exceed t1 by at least 1 CPM AND reach at least 1.25 times its t1 value, while the community pathway RNA at t2 is at least 80% of t0. This thresholded signature is an operational description, not proven compensation. Separately report each component. Also report the contemporaneous t0-to-t1 signature with the same thresholds, labelled secondary.

Report event and participant counts and event-weighted proportions. Primary estimate: mean across participants of each participant's fraction of sustained-decline triples exhibiting the lagged signature. Use percentile bootstrap resampling participants, 5,000 draws, seed 903, to obtain 95% intervals; at least 5 represented participants are required for an interval. A participant contributes all its events when resampled. Less than 10 event-bearing participants is explicitly inadequate for a population-level conclusion. For stable comparators report the same summary. Difference between event classes is secondary, restricted to participants with at least one of each type, averaging paired within-participant differences and bootstrapping participants; at least 5 such participants for an interval. No causal or equivalence interpretation or significance claims.

To separate alternative population expansion from relative transcription, also summarize per participant the log2 change from t1 to t2 in alternative RNA/DNA ratio: log2[(R2+0.01)/(D2+0.01)] - log2[(R1+0.01)/(D1+0.01)]. The pseudocount is CPM. This is a ratio of compositional quantities, not an absolute per-cell rate; report sensitivity to 0.1 and 1 CPM pseudocounts and do not use it to define events.

## Prespecified sensitivities
Repeat the primary CD analysis separately for: sustained DNA decline cutoffs 0.33 and 0.67; both time intervals 1 to 4 weeks and 1 to 12 weeks; baseline leader/alternative minimum CPM 0.5 and 2; excluding triples with any reported antibiotic use or missing antibiotic response; retaining only triples with unchanged observed yes/no immunosuppressant category; requiring the leader have a named RNA row; and non-overlapping triples selected greedily by time within each participant. Also repeat all main calculations after rescaling each sample's pathways and strata by the sum of all unstratified pathway values excluding UNMAPPED and UNINTEGRATED, to one million. This conditions on annotated pathways and is a sensitivity, not an absolute-abundance correction. All other thresholds remain as specified. No tuning after outcomes.

## Inflammation feasibility
Use public fecalcal values only, no clinical cutpoint or diagnostic classification. For each primary event, seek an observed value at t2 and the earliest subsequent paired sample 1 to 12 weeks later with a value. Count usable observations and participants by signature status. Only if both groups have at least 10 participants compute participant-level mean log2[(future+1)/(t2+1)] and their descriptive group difference with participant bootstrap. Otherwise report that the proposed prospective inflammation contrast is too sparse; do not fill missing calprotectin by carry-forward or nearest earlier values. No adjusted clinical effect or therapeutic inference. Antibiotics and immunosuppressants are imperfect recorded covariates, and confounding remains even in sensitivity subsets.

## Outputs and reporting
Publish deterministic aggregate results, the exact executable code, source URLs/hashes, the unchanged plan and preregistration reference, including sparse/negative results. Do not expose individual profiles in prose. Record deviations explicitly. Scientific conclusions must be narrower than the source tables and identification limits. A structural inability to test the primary hypothesis is a result, not evidence that compensation fails biologically.