Data and Code for: Salinity converges green-leaf bacteria but diverges them as leaves senesce: a common-garden test of the Scale–Spectrum–Stage framework

This archive contains the data and code supporting a three-layer quantitative synthesis of the 3S framework, accompanying the manuscript "Salinity converges green-leaf bacteria but diverges them as leaves senesce: a common-garden test of the Scale–Spectrum–Stage framework" by Lele Liu, Wenyi Sheng, Yuhui Wang, Huijia Song, Yaolin Guo, and Weihua Guo. It is organized into four top-level components: a data/ directory of unified comparison tables, a scripts/ directory holding the full analysis pipeline, a d_tier/ directory with recomputation inputs for three of five D-tier studies, and documentation of the repro chain, software requirements, licensing, and citation. A path note clarifies that the archive uses its own layout rather than the upstream repository structure, bundling files that upstream lived under config/, integration/, meta_analysis/, public_reanalysis/, and main_study/ into data/ and scripts/, while deliberately excluding rendered figures, per-panel figure tables, and manuscript or Supporting-Information documents because those regenerate from the archived tables. The data/ directory holds the unified comparison tables and related outputs. The central file, unified_comparisons.csv, contains 151 stress-versus-control comparison units spanning three layers (main, meta, pub), with columns identifying study, comparison, row type, kingdom, organ, stage, stress type, 3S cell, effect size, variance, and model inclusion. Alongside it are extraction_data.csv (a 59-row three-tier literature extraction table, including four focal-study rows) and extraction_data_edwards2025.csv (six companion D-tier units for Edwards et al., 2025). The directory also contains three_layer_table.csv with M2 per-cell × per-layer estimates, sensitivity_leave_layer.csv with M5 leave-one-layer-out refits, three_layer_results.txt (the full M1–M6 console report and the only copy of six of the nine rows of SI Table S4), predictions_3S.csv (the pre-registered prediction matrix), sensitivity_independence.txt, and main_study/ with gzip-compressed processed OTU and taxonomy tables plus shared sample and genotype metadata. The scripts/ directory provides the full analysis pipeline. recompute_experimental_layer.R re-derives the experimental layer's 24 units from the bundled main_study/ tables via rarefaction, chloroplast/mitochondria filtering, median pairwise Bray–Curtis computation, and a 9,999-replicate cluster bootstrap; it is an archive port differing from its upstream copy only inside marked PACKAGE PORT blocks. build_unified_table.R converts the three layers into the unified 151-unit table, while three_layer_analysis.R runs the joint M1–M6 metafor models and produces all per-cell estimates. na_lineage_sensitivity.R re-estimates the below-floor NA lineage without the n ≥ 5 guard, and sensitivity_independence.R performs cluster-robust and study-blocked independence checks. The public-reanalysis layer is served by 00_fetch_sra.py, 02_dispersion_lnrr.R (shared PERMDISP and lnRRβ bootstrap functions used by all three layers), and 03_integrate.R. The d_tier/ directory supplies recomputation inputs for three of the five D-tier studies: Ning 2020 (grassland warming, PRJNA626428), Wang 2024 (invasion rhizosphere, PRJNA1152501), and Halo 2026 (cowpea root drought, author-provided ASV table). Each study bundles a treatment coding table plus a VSEARCH script, or the script and its input table where processed data are available; raw FASTQ files are not bundled and must be re-downloaded from the stated BioProjects. The remaining two D-tier inputs—Bintarti 2022 (ShadeLab GitHub processed table, rarefied to 11,138 reads per sample) and Edwards 2025 (Zenodo 10.5281/zenodo.15186007)—are noted but not included. Primary sequencing data for the common-garden experiment are deposited under GSA BioProject PRJCA050192. The repro chain documents five steps. Steps 1–3 run directly from the package as distributed and were verified to produce byte-identical outputs: step 1 reruns the joint models M1–M6 and rewrites three_layer_table.csv, sensitivity_leave_layer.csv, three_layer_results.txt, and the two support figures; step 2 re-derives the experimental layer's 24 units and writes per-kingdom lnRRbeta_v2_2.csv and group_distance_summary.csv matching unified_comparisons.csv; step 3 performs the NA-lineage sensitivity quoted in Methods, self-validating by first reproducing the five above-floor lineages. Steps 4 and 5 are reference implementations requiring external inputs: rebuilding the 151-unit table from layer outputs and running the D-tier VSEARCH and dispersion analyses (with paths resolvable within the package, VS override for vsearch, and --spectrum F for the WangY 2024 entry). Steps 1–3 require R, with steps 1 and 4 also needing metafor (models run under R 4.5.0 with metafor 5.0.1) and steps 2 and 3 requiring vegan (2.6-10 here); step 5 needs vsearch (2.30.5 here), python3, and vegan for the reanalysis layer. The archive is released under CC-BY-4.0, with the full text in LICENSE and a machine-readable citation in CITATION.cff, credited to the six authors and versioned 3.0.0 (2026-09-20). Finally, the three-tier extraction table is designed to be extendable: new studies can be appended using the documented schema and incorporated by rerunning the M1–M6 script.