Reproducing the public figures of Sugar-rich foods exacerbate antibiotic-induced microbiome injury

This repository reproduces the publicly-shareable figures — those free of protected health information (PHI) — of the dietary-sugar / microbiome manuscript from a small set of de-identified released tables. Every panel is rebuilt by a clean, numbered script under reproduce/; the scripts read only from released_data/ and write one PDF per panel into results/.

The companion data deposit is on Zenodo: 10.5281/zenodo.14538105 — Supplementary Data for "Sugar-rich foods exacerbate antibiotic-induced microbiome injury" (Dai, Anqi; CC-BY-4.0). Three of the released tables here are the same files deposited there (see Released data).

Most data elements have been released publicly; some are subject to data-sharing agreements
Repository layout
Environment setup
Getting the restricted data (internal users)
How to reproduce the figures
Scripts → figures
Released data tables
Running QIIME 2 through Docker
Regenerating the TaxUMAP embedding
Citation
License
Acknowledgements

Code to generate every panel of the manuscript is made available here, along with the necessary data for most panels. Food-item and meal-level data, microbiome-level sample data are available. Patient-level data (e.g. mortality) are subject to institutional data-sharing policies and are available to internal MSKCC users following the instructions below and to users outside of MSKCC via data sharing agreements.

Need a panel whose data isn't shipped here?

Internal MSKCC users pull the PHI-free restricted tables from the lab drive with DVC — see Getting the restricted data.

Everyone else: the patient-level clinical variables and mortality outcomes underlying those restricted panels are available via a data sharing agreement, per institutional policies. Requests should be directed to Dr. Jonathan Peled (peledj@mskcc.org).

Reproducible by anyone

Clone the repo and run — these need only the de-identified tables shipped in released_data/: Figure 1, Figure 2, Figure 4, Extended Data E1 (c–h), E2 (b–e), E3, E4, E5, E7, E8, E9, and all the mouse experiments.

Reproducible by internal users (restricted data required)

These panels are fully implemented here, but they read de-identified tables that carry no PHI yet are not cleared for public release. Those tables are not shipped: they live in a gitignored restricted_data/ folder that internal users fetch with DVC (see Getting the restricted data). The scripts live under reproduce/restricted/ and skip cleanly — no error — for anyone who has only released_data/.

Figure	Script	Restricted input
Extended Fig. E2a	`reproduce/restricted/60_e2a_abx_heatmap.R`	`R21_meds_updated_all_medication_classified.csv` — each patient's full daily antibiotic time course (the released `Data_S4` only carries the 2-day window prior to each stool sample)
Data S6	`reproduce/restricted/61_dataS6_pt_timecourse.R`	`df_main_clinical_outcome.rds` — supplies the per-patient engraftment day (the green dashed line); the other three inputs are in `released_data/`
Fig. 3 a,b · E6 c,d,j · Supp. Tables 1–6	`reproduce/restricted/63_fig3_e6_clinical.R`	`df_main_clinical_outcome.rds` — the cleaned, merged clinical-outcome table (de-identified survival + sugar-density summary + covariates). It is built upstream in the dev repo and shipped here only in cleaned form; 63 draws all the panels and tables from it
Extended Fig. E6 e,f,g	`reproduce/restricted/64_e6efg_cluster_intake.R`	`df_main_clinical_outcome.rds` — supplies the diet-pattern cluster (`modal_diet`); daily calorie/macronutrient intake over HCT day by cluster (the diet table itself is released)
Extended Fig. E6h	`reproduce/restricted/66_e6h_alpha_trajectory.R`	`df_main_clinical_outcome.rds` — supplies the diet-pattern cluster (`modal_diet`); fecal microbiota alpha-diversity trajectory over HCT day by cluster (the diversity table `153_combined_META.csv` is released)
Extended Fig. E6i	`reproduce/restricted/65_e6i_discharge.R`	`df_main_clinical_outcome.rds` — supplies the diet-pattern cluster and the discharge/engraftment landmark; cumulative incidence of hospital discharge after engraftment by cluster, with the adjusted-Cox HR (1.54, p=0.023)
Extended Fig. E1b	`reproduce/restricted/67_e1b_covariates_contribution.R`	`df_main_clinical_outcome.rds` — supplies the clinical covariates (source/intensity/age/sex/disease); per-covariate microbiome variance explained (`vegan::envfit` r²) bar chart (the metadata `153_combined_META.csv` and amplicon sequence variant (ASV) counts `63_asv_count_relab_res.csv` are released)

Repository layout

released_data/      de-identified input tables (the only data that ships)
restricted_data/    PHI-free but non-public inputs, internal only (gitignored, not shipped)
reproduce/          numbered scripts, one analysis "family" per number block
reproduce/restricted/  scripts that need restricted_data/ (skip cleanly when it is absent)
intermediate_data/  cached model fits, QIIME ordinations, etc. (created on first run)
results/            output PDFs, one per panel (created on first run)

Scripts are organised by analysis family, not by printed figure: an expensive model is fit and cached once, then consumed by the panels that need it.

10–19 human Figure 2 diversity family (+ procrustes, sensitivity, diet Faith PD)
20–27 Figure 1 and the Extended-Data E1/E2/E4 panels
30 Extended Fig. E3 (compositional PCoA)
40–44 taxon-abundance family (F4, E7)
50–54 mouse experiments (CFU, 16S, RNA-seq), driven by 00_run_mouse.R
60+ (under reproduce/restricted/) panels requiring restricted, non-public inputs

Environment setup

1. R and packages (`renv`)

The project pins its packages with renv. Tested on R 4.5.3.

# from the project root, in R:
renv::restore()      # restores renv itself + any pinned packages

# analysis packages (install on first use, then snapshot):
renv::install(c("tidyverse","readxl","here","ggpubr","scales","ggrepel",
                "vegan","ashr","msigdbr","seriation"))
renv::install(c("bioc::Maaslin2","bioc::DESeq2","bioc::fgsea","bioc::edgeR",
                "bioc::org.Mm.eg.db","bioc::AnnotationDbi"))     # mouse seq panels
# human Bayesian + plotting family:
renv::install(c("brms","tidybayes","broom.mixed","posterior","ggtext","wesanderson",
                "cowplot","ggsci","furrr","dendextend","ggdendro","ggh4x","ggforce",
                "ggbeeswarm","lmerTest","ggpmisc"))
renv::snapshot()

2. Stan backend (for the brms models)

The Bayesian diversity / CLR models (10, 16, 40, 27, 44, …) use brms with the cmdstanr backend (rstan 2.32 does not compile against the current Apple clang). Install CmdStan once:

renv::install("stan-dev/cmdstanr")
cmdstanr::install_cmdstan()     # builds CmdStan (the helper expects ~/.cmdstan)

The fitting scripts cache every fit to intermediate_data/*.rds (file =), so they compile/sample once and are instant on rerun.

3. Docker (for the QIIME 2 steps)

A few scripts call QIIME 2 inside a Docker container (UniFrac / Bray-Curtis / Faith PD on the food tree). Install Docker Desktop and pull the image used by the pipeline — see Running QIIME 2 through Docker.

4. Python / TaxUMAP (optional, Figure 1 e–h only)

The TaxUMAP embedding for F1 e–h ships precomputed (released_data/taxumap_embedding.csv), so you do not need Python to draw the figure. To regenerate it, see Regenerating the TaxUMAP embedding.

All public scripts resolve their input from released_data/ (override with the NUTRITION_DATA env var). Scripts under reproduce/restricted/ additionally read restricted_data/ (override with RESTRICTED_DATA) and skip cleanly when it is absent.

Getting the restricted data (internal users)

The public figures need only released_data/, which ships with the repo — you can skip this section entirely unless you are reproducing one of the restricted panels (E1b, E2a, E6 c–j, Fig 3 a/b, Data S6, Supp. Tables 1–6). Those need the PHI-free-but-non-public tables in restricted_data/, which is not in git.

restricted_data/ is version-controlled with DVC and stored on the Peled-lab drive (content-addressed, not a browsable copy). The repo tracks only the small restricted_data.dvc manifest; the actual files are pulled from the drive.

One-time: install DVC.

pip install dvc            # or: brew install dvc / conda install -c conda-forge dvc

Each time you need the restricted data:

Mount the lab drive, then point DVC at it. The remote lives on the Peled-lab share; its path is deliberately kept out of the repository, so each user configures it once locally. Ask the lab for the mounted share path (<lab-drive> below).

On macOS, mount the share via Finder → Go → Connect to Server (⌘K), enter the lab SMB URL and authenticate; it then appears under /Volumes/. Confirm the remote exists, then register it:
```
ls <lab-drive>/Projects/Reproduce_nutrition/restricted_dvc

# writes to .dvc/config.local, which is gitignored
dvc remote add --local -d lab_restricted <lab-drive>/Projects/Reproduce_nutrition/restricted_dvc
```
Pull the data from the repo root — DVC reads restricted_data.dvc and materializes the files into restricted_data/:
```
dvc pull
```
You should now have restricted_data/ populated (4 files, ~5.8 MB, including df_main_clinical_outcome.rds).
Run the restricted script(s) — they now find their inputs:
```
Rscript reproduce/restricted/63_fig3_e6_clinical.R
```

Notes:

If the drive is not mounted (or you lack access), dvc pull fails and restricted_data/ stays empty; the reproduce/restricted/* scripts then skip cleanly (they guard on the folder's presence and quit 0). Public reproduction is unaffected.
restricted_data/ is gitignored, so pulled files never get committed. To point at the data without DVC, set RESTRICTED_DATA=/path/to/folder.
Contributors who change the restricted tables re-run dvc add restricted_data && dvc push, then commit the updated restricted_data.dvc. Most users only ever dvc pull.

How to reproduce the figures

All commands run from the project root.

Mouse figures (F4c, E8, E9)

Rscript reproduce/00_run_mouse.R                 # all of 50–54
RUN_SEQ=false Rscript reproduce/00_run_mouse.R   # CFU panels only, skip the slow seq fits

Human figures

Run the fitting scripts first (they cache the models), then the plotting scripts that consume them. Independent families can be run in any order.

# Figure 2 diversity family
Rscript reproduce/10_fit_diversity_models.R      # caches the F2 brms fits
Rscript reproduce/11_fig2_diversity.R            # F2 c–h, E5i
Rscript reproduce/12_diagnostics_diversity.R     # E1g, E1h
Rscript reproduce/13_fig2_procrustes.R           # F2a   (QIIME — see Docker section)

# Figure 1
Rscript reproduce/19_diet_faith_pd.R             # rebuilds diet-alpha-diversity.tsv (QIIME)
Rscript reproduce/20_fig1_taxumap.R              # F1 e–h
Rscript reproduce/21_fig1_diet_timecourse.R      # F1 b,c,i–m
Rscript -e 'rmarkdown::render("reproduce/22_fig1_food_tree.Rmd")'   # F1d (GraPhlAn/Docker)
Rscript reproduce/23_fig1_beta_diversity.R       # F1 n,o (QIIME)

# Extended Data
Rscript reproduce/24_e1cde_random_intercepts.R   # E1 c,d,e   (needs the F2d fit from 10)
Rscript reproduce/25_e1f_alpha_breakdown.R       # E1f
Rscript reproduce/26_e2bc_abx_exposure.R         # E2 b,c
Rscript reproduce/26b_e2de_fluoroquinolone.R     # E2 d,e   (fits the simplified fluoroquinolone model)
Rscript reproduce/restricted/60_e2a_abx_heatmap.R  # E2a (needs restricted_data/; skips cleanly if absent)
Rscript reproduce/restricted/61_dataS6_pt_timecourse.R  # Data S6 (needs restricted_data/; skips cleanly if absent)
Rscript reproduce/restricted/63_fig3_e6_clinical.R      # Fig 3 a,b + E6 c,d,j + Supp. Tables 1-6 (reads the cleaned df_main)
Rscript reproduce/restricted/64_e6efg_cluster_intake.R  # E6 e,f,g (daily intake by diet-pattern cluster)
Rscript reproduce/restricted/65_e6i_discharge.R         # E6i (discharge cumulative incidence by cluster, adjusted HR)
Rscript reproduce/restricted/66_e6h_alpha_trajectory.R  # E6h (fecal alpha-diversity trajectory by cluster)
Rscript reproduce/restricted/67_e1b_covariates_contribution.R  # E1b (per-covariate microbiome variance explained)
Rscript reproduce/16_fit_e4_models.R             # caches E4 fits
Rscript reproduce/17_fig_e4.R                    # E4 a–e,i,j
Rscript reproduce/27_e4h_fndds_zscored.R         # E4h
Rscript reproduce/30_extdata_pcoa.R              # E3 a,b,c

# Taxon family (F4, E7)
Rscript reproduce/45_quality_genus_relab.R       # rebuilds 45_quality_asv_relab_pident97_genus.csv from the 63 tables
Rscript reproduce/40_fit_taxon_models.R          # caches the CLR fits (derives genus CLR inline)
Rscript reproduce/41_fig4_human.R                # F4a, F4b
Rscript reproduce/42_extdata_taxon.R             # E7 a,b,e
Rscript reproduce/43_extdata_enterococcus_asv.R  # E7 c,d
Rscript reproduce/44_e7f_abxclass.R              # E7f

Common environment toggles:

RUN_QIIME=false — reuse the cached QIIME ordinations instead of re-running Docker.
QIIME2_IMAGE=... — override the QIIME 2 image tag.
RUN_SEQ=false — (mouse) skip the slow MaAsLin2 / DESeq2 / fgsea fits.

Scripts → figures

Script	Panel(s)
`21_fig1_diet_timecourse.R`	F1 b, c, i, j, k, l, m
`22_fig1_food_tree.Rmd`	F1 d
`20_fig1_taxumap.R`	F1 e, f, g, h
`23_fig1_beta_diversity.R`	F1 n, o
`13_fig2_procrustes.R`	F2 a
`11_fig2_diversity.R`	F2 c, d, e, f, g, h (and E5i)
`41_fig4_human.R`	F4 a, b
`50_figure_4c.R`	F4 c (mouse)
`24_e1cde_random_intercepts.R`	E1 c, d, e
`25_e1f_alpha_breakdown.R`	E1 f
`12_diagnostics_diversity.R`	E1 g, h
`26_e2bc_abx_exposure.R`	E2 b, c
`26b_e2de_fluoroquinolone.R`	E2 d, e (simplified diversity model with an explicit prophylactic-fluoroquinolone term)
`30_extdata_pcoa.R`	E3 a, b, c
`17_fig_e4.R`	E4 a, b, c, d, e, i, j
`17b_e4fg_wweia.R`	E4 f, g (What We Eat in America, WWEIA, nomenclature)
`27_e4h_fndds_zscored.R`	E4 h
`28_e5abcd_added_sugars.R`	E5 a, b, c, d (added vs other sugars)
`29_e6ab_sweet_grains.R`	E6 a, b (Sweet vs Other Grains; E6c–j are restricted)
`14_sensitivity_diversity.R`, `15_robustness_subsampling.R`, `18_fig_e5jk_ons.R`	E5 e, f, g, h, j, k, l, …
`42_extdata_taxon.R`	E7 a, b, e
`43_extdata_enterococcus_asv.R`	E7 c, d
`44_e7f_abxclass.R`	E7 f
`53_extdata_8_16s.R`	E8 a–f (mouse 16S)
`51_extdata_8.R`	E8 g, h, j, k, l, m (mouse)
`52_extdata_9.R`	E9 a, b, c, d, h (mouse)
`54_extdata_9_rnaseq.R`	E9 e, f, g (mouse RNA-seq)

Fitting/data scripts (no panel of their own): 10, 16, 40 (cache the brms fits); 19 (rebuilds diet-alpha-diversity.tsv); 13d/13e (procrustes sensitivity); 00_run_mouse.R (mouse driver).

Released data tables

Everything in released_data/ is de-identified and shareable. Zenodo column links each table to the companion deposit where it is also archived; tables marked derived are de-identified products built in this project's upstream pipeline and are not separately on Zenodo.

Core inputs

File	Used for	Zenodo
`152_combined_DTB.csv`	the diet tracker — every food item each patient ate; the source of all diet exposures, the food tree, and diet diversity	same file on Zenodo
`153_combined_META.csv`	the per-stool-sample analysis table (1009 samples / 158 patients): diversity outcome, antibiotic/TPN/EN exposure, prior-2-day food-group & macronutrient intake	same file on Zenodo
`Data_S4_Medication_Exposures…csv`	medications in the 2 days before each stool sample, by drug class — antibiotic-exposure panels (E2b/c), the fluoroquinolone diversity model (E2d/e) and the antibiotic-class CLR model (E7f)	Zenodo "Data S4"
`R59_meta_expanded.csv`	`153` plus the E. faecium CLR outcome and extra covariates — used by the taxon CLR models (F4b, E7e, E7f)	derived
`FPED_1516.xls`, `FPED_1720.xls`	USDA Food Patterns Equivalents Database (2015-16 and 2017-20) — the added-sugars content (teaspoon equivalents per 100 g) per food code, used to split total sugars into added vs other for E5a–d (`28`); 1516 is preferred, the two salad-dressing codes that exist only in 2017-20 are filled from 1720; related to Zenodo "Data S5", the nutrient values from the USDA Food and Nutrient Database for Dietary Studies (FNDDS)	reference (USDA)
`2015-2016 FNDDS At A Glance…xlsx`, `2019-2020 FNDDS At A Glance…xlsx`	USDA FNDDS — each food code's WWEIA food category (and per-100 g nutrients), used to re-derive the diversity model under WWEIA nomenclature for E4f–g (`17b`); 2015-16 preferred, 2019-20 fills the rest	Zenodo "Data S5" (FNDDS nutrient values)

152_combined_DTB.csv — one row per food item consumed:

column	meaning
`pid`	de-identified patient ID
`fdrt`	food day relative to transplant (day 0 = transplant)
`Meal`, `Food_NSC`, `Unit`, `Por_eaten`	meal, food name, unit, portions eaten
`Food_code`, `description`	USDA FNDDS food code and description
`Calories_kcal`, `Protein_g`, `Fat_g`, `Carbohydrates_g`, `Fibers_g`, `Sugars_g`	nutrient content of the portion
`total_weight`, `dehydrated_weight`	wet and dehydrated weight (g); `dehydrated_weight` is the intake measure used throughout

153_combined_META.csv / R59_meta_expanded.csv — one row per stool sample:

column	meaning
`pid`, `sampleid`	patient and stool-sample ID
`sdrt`	stool day relative to transplant
`simpson_reciprocal`	inverse Simpson microbiome alpha-diversity (outcome)
`empirical`	antibiotic-exposed in the prior 2 days (TRUE/FALSE)
`intensity`	conditioning intensity (nonablative / reduced / ablative)
`EN`, `TPN`	enteral / total parenteral nutrition exposure
`fg_fruit … fg_legume`	prior-2-day average intake of the 9 FNDDS food groups (g/day)
`ave_Calories_kcal … ave_Sugars_g`	prior-2-day average macronutrient intake
`timebin`	7-day transplant-time bin (random-effect grouping)
(R59 only) `asv_1_clr`, `ci_cleaned_numeric`, `disease_lineage`, `PCA`, `exposure_type`	E. faecium CLR outcome + extra covariates

Data_S4_Medication_Exposures…csv — one row per medication exposure: sampleid, pid, sdrt, class, drug_name_clean, route_clean, drug_category_for_this_study (broad_spectrum / fluoroquinolones / other_antibacterials / not_antibacterial).

Microbiome tables (derived, de-identified)

File	Columns	Used for
`63_asv_count_relab_res.csv`	`asv_key, sampleid, count, count_relative`	per-ASV 16S counts and whole-community relative abundance (one row per observed, non-zero count); the ASV-count source for E1b, and joined to the annotation for E7c and the quality-genus / genus-CLR tables
`63_asv_blast_annotation.csv`	`asv_key` + taxonomy + BLAST hit (see below)	per-ASV taxonomic lineage and the BLAST hit that assigned it; one row per ASV
`45_quality_asv_relab_pident97_genus.csv`	`asv_key, sampleid, count_relative, genus`	per-ASV 16S relative abundance with a quality genus (kept only where BLAST `pident > 97`, else NA); rebuilds genus relab for F1n/o, F4a, E7b, the E3 PCoA and the E7a genus models
`mgx_enterococcus_species_relab.csv`	`sample, species, relab`	metagenomic Enterococcus species abundance for E7d

63_asv_blast_annotation.csv columns. Each ASV sequence is the query, aligned against a 16S reference database with BLAST; the row records the assigning hit. Columns marked BLAST are standard BLAST tabular (outfmt 6) fields; the rest are derived by the annotation pipeline.

Column	Meaning
`asv_key`	ASV identifier (e.g. `asv_1`)
`kingdom … species`	assigned taxonomic lineage (`kingdom, phylum, class, order, family, genus, species`); `order` was the raw export's `ordr`
`query_length`	length (bp) of the ASV query sequence (BLAST `qlen`)
`align_length`	length of the query–subject alignment, gaps included (BLAST `length`)
`pident`	percent of identical bases over the alignment (BLAST `pident`)
`nident`	number of identical bases in the alignment (BLAST `nident`)
`score`	raw alignment score (BLAST `score`)
`length_ratio`	`align_length / query_length`: fraction of the query covered by the alignment (~1 = full-length match) (derived)

Diet / food-tree tables

File	Columns / contents	Used for
`072_total_patients_zero_eating_days_pid.csv`	`fdrt, diet_data_status, pid`	documented zero-intake days (distinguishes "ate nothing" from "no record")
`diet-alpha-diversity.tsv`	`(patient-day id), faith_pd`	diet Faith phylogenetic diversity per patient-day (regenerated by `19`)
`taxumap_embedding.csv`	`index_column, taxumap1, taxumap2`	precomputed TaxUMAP 2-D diet embedding for F1 e–h
`output_food_tree_datatree.newick`	rooted food taxonomy tree, Food_code leaves	UniFrac on diet (F2a, F1n, diet Faith PD)
`output_food_tree_datatree_name.newick`	same tree, description leaves	F1d food-tree rendering
`final_table_for_writing_out_to_newick.csv`	`FoodID, description, L1…L7, newickstring…, taxonomy`	the 7-level food taxonomy behind the tree (and the TaxUMAP taxonomy)
`NodeLabelsMCT.txt`	`Level.code, Main.food.description` (the label), plus `n_digits, source, usda_fndds_description, used_in_F1d_tree` (provenance)	food-tree node labelling. Levels 1–3 are official USDA FNDDS subgroups; levels 4–6 are mostly the finer split from the Knights lab food-tree code, which USDA does not publish. See the provenance section of `22_fig1_food_tree.Rmd` for the citations
`food_group_color_key_final.csv`	`fgrp1, fdesc, fg1_name, color, shortname`	food-group → colour / short-name key for the forests and F1d
`annotation.base.txt`	GraPhlAn sector header	F1d circular-tree annotation

Mouse data

File	Contents	Used for
`Dai_mouse_figure_raw_data.xlsx`	Nature source-data workbook, one sheet per mouse panel (CFU, 16S relab + metadata, RNA-seq counts)	all of F4c / E8 / E9 (scripts `50–54`)

Running QIIME 2 through Docker

Scripts 13_fig2_procrustes.R, 19_diet_faith_pd.R, and 23_fig1_beta_diversity.R compute phylogenetic diversity (UniFrac / Faith PD) on the food tree using QIIME 2. They call QIIME inside Docker so you don't need a local QIIME install.

Install Docker Desktop and start it.

Pull the image (one-time, ~10 GB):

docker pull quay.io/qiime2/qiime2:2026.4

Run the script normally — it builds the feature tables in R, then runs one container that does biom convert → import → diversity → export for the whole job (one cold start, not dozens):
```
Rscript reproduce/13_fig2_procrustes.R
RUN_QIIME=false Rscript reproduce/13_fig2_procrustes.R   # reuse cached ordinations
```
The container is invoked as docker run --rm --platform linux/amd64 -v <repo>:/data -w /data <image> …, mounting the repo at /data.

Notes / gotchas:

Apple Silicon (M-series): the QIIME image is x86-only, so it runs under --platform linux/amd64 (Rosetta) emulation — correct but slow. That's inherent to running x86 QIIME on ARM, not the pipeline.
q2-composition / phyloseq: the 2026.4 image's plugin manager aborts trying to import the R package phyloseq (not installed) via q2-composition. The scripts run pip uninstall -y q2-composition at container start (the container is --rm, so it's ephemeral and q2-composition is unused here).
Figure 1d (22_fig1_food_tree.Rmd) uses a different container — GraPhlAn via the shengwei/graphlan image (GraPhlAn runs only under Python 2.7) — for the circular food-taxonomy render. The R steps build the inputs; the notebook's Docker commands draw the panel.

Regenerating the TaxUMAP embedding

Figure 1 e–h is drawn by 20_fig1_taxumap.R from the precomputed released_data/taxumap_embedding.csv, so the figure reproduces without Python. To regenerate the embedding from scratch, the full step-by-step (build the two input tables in R → install TaxUMAP → run it) is in reproduce/taxumap_pipeline_HOWTO.md.

In brief:

# 1. install TaxUMAP (https://github.com/jsevo/taxumap) in its own env
git clone https://github.com/jsevo/taxumap.git && cd taxumap
conda create -n taxumap python=3.9 -y && conda activate taxumap
pip install -e .

# 2. (in R) build food_code_relative.csv + food_taxa.csv from 152_combined_DTB.csv
#    and final_table_for_writing_out_to_newick.csv  — see the HOWTO

# 3. run TaxUMAP with the manuscript's weights
run_taxumap.py \
  -t intermediate_data/food_taxa.csv \
  -m intermediate_data/food_code_relative.csv \
  --agg_levels Kingdom/Phylum/Class/Order \
  -n 173 --weights 0.01/2/10/10 \
  -o intermediate_data

UMAP is stochastic (the embedding can rotate/flip between runs), so for an exact figure match use the shipped taxumap_embedding.csv rather than a fresh embedding.

Citation

Machine-readable metadata lives in CITATION.cff — use GitHub's Cite this repository button.

The accompanying manuscript is under review. Once it is published, please cite the article in preference to this software; this repository will then carry a preferred-citation entry pointing at it.

Two related DOIs, which are not interchangeable:

DOI	What it is
10.5281/zenodo.21290618	Software (all versions) — the concept DOI. It always resolves to the latest release. Cite this one.
10.5281/zenodo.14538105	Data (all versions) — the companion Supplementary Data deposit's concept DOI. Cite this one.

License

Code (everything under reproduce/, and the repository as a whole): MIT.
Data (released_data/): CC-BY-4.0, matching the companion Zenodo deposit.

Third-party reference tables redistributed in released_data/ (the USDA FNDDS "At A Glance" and FPED releases) are US Government works in the public domain.

Acknowledgements

This repository reproduces analyses developed with several colleagues; the clean scripts here are ports of their original work:

William Jogia (Institute for Systems Genetics and Department of Microbiology, NYU Langone Health Grossman School of Medicine) — the WWEIA-nomenclature food-group analysis (E4 f,g), the added-sugars analysis (E5 a–d), and the sweet-grain split (E6 a,b).
Mirae Baichoo (Adult Bone Marrow Transplantation Service, Department of Medicine, Memorial Sloan Kettering Cancer Center) — the per-covariate microbiome variance-explained analysis (E1b).
Teng Fei (Department of Epidemiology and Biostatistics, Memorial Sloan Kettering Cancer Center) — the clinical-outcome and survival analyses (Figure 3, and E6 h,i,j).
Nicholas R. Waters (Adult Bone Marrow Transplantation Service, Department of Medicine, Memorial Sloan Kettering Cancer Center) — the Enterococcus ASV / species analyses establishing that ASV 1 is E. faecium (E7 c,d).

Patient-level clinical variables and mortality outcomes — underlying Figure 3, E1b, E6 c–j and Supplementary Tables 1–6 — are available via data sharing agreement per institutional policies.