mohitraiyani27/Human_Gut_Microbiome_Data

Human Gut Microbiome Dataset (40-class)

Pre-processed, train/val/test-split human gut microbiome dataset used to train and evaluate MicrobiomeFM, a Transformer-based foundation model for multi-class disease classification from shotgun metagenomic data.

This release contains the 40-class filtered version of the dataset that the published results were obtained on (test accuracy ≈ 96.05%, weighted F1 ≈ 0.950, macro F1 ≈ 0.691). It is derived from the curatedMetagenomicData (cMD) Bioconductor resource by removing disease classes with fewer than 20 samples and re-stratifying into 70/15/15 train/val/test splits.

---

Dataset Summary

Item	Value
Total samples	14,488
Number of features	2,361 (1,260 categorical + 1,101 numerical)
Number of disease classes	40
Target column	disease (integer label, 0–39)
Train / Val / Test	10,136 / 2,176 / 2,176
Split strategy	Stratified 70 / 15 / 15 (random seed 42)
Source	curatedMetagenomicData (Bioconductor), shotgun stool metagenomics

---

Files

File	Description
train.csv	Training split — 10,136 rows, 2,362 columns (2,361 features + disease).
val.csv	Validation split — 2,176 rows.
test.csv	Held-out test split — 2,176 rows.
metadata.json	Feature schema: lists of categorical/numerical feature names, vocab_sizes for every categorical column, and counts (n_features, n_train, n_val, n_test, n_classes).
disease_mapping.json	name_to_id and id_to_name lookups for the 40 disease labels.

All preprocessing (categorical integer encoding, numerical standardization to zero mean and unit variance, removal of identifier and >95%-missing columns) was performed on the training set only to avoid leakage. Numerical missing values were filled with zero; missing categorical values were treated as a separate category.

---

Disease Classes

40 classes covering metabolic, cardiovascular, gastrointestinal, neurological, autoimmune and infectious conditions, plus the healthy control class. The full list (40 entries) is in disease_mapping.json. Some examples:

0: ACVD          5: CRC           10: IBD          15: T1D          20: cirrhosis
1: BD            6: CRC;T2D       11: MS           16: T2D          ...
2: CAD           7: CRC;hyperten. 12: PD           17: T2D;adenoma  39: respiratoryinf

---

Loading

With pandas

import pandas as pd, json

train = pd.read_csv("train.csv")
val   = pd.read_csv("val.csv")
test  = pd.read_csv("test.csv")

with open("metadata.json")        as f: meta    = json.load(f)
with open("disease_mapping.json") as f: mapping = json.load(f)

X_cols = meta["categorical_features"] + meta["numerical_features"]
y_col  = meta["target_column"]   # 'disease'

With datasets

from datasets import load_dataset
ds = load_dataset("mohitraiyani27/Human_Gut_Microbiome_Data")

---

How the splits were produced

1. Start from the cMD-aggregated dataset (14,855 samples, 131 disease classes, 2,361 features after preprocessing).
2. Drop any disease class with fewer than 20 samples (eliminates rare and single-sample classes that cannot be learned).
3. Remap the remaining 40 class labels to a clean 0–39 range.
4. Stratified 70 / 15 / 15 split with random_state=42.

---

Intended Use

• Training and benchmarking transformer / foundation-model architectures on heterogeneous tabular metagenomic data.
• Multi-class disease classification from gut microbiome profiles.

Not intended for direct clinical decision-making. Class labels are research-grade study annotations from the underlying cMD resource.

---

Reference Code

Training and evaluation code (Chunked Feature Embedding + Masked Chunk Autoencoding pre-training + Attention-Pooling fine-tuning) is published at:

• GitHub: https://github.com/MohitRaiyani-27/Microbiome_FM

Run order:

pip install -r requirements.txt
python Complete_Architecture/pretrain.py
python Complete_Architecture/finetune.py
python Complete_Architecture/evaluate.py

---

Citation

If you use this dataset, please cite the underlying curatedMetagenomicData resource and the MicrobiomeFM paper:

@article{microbiomefm2026,
  title  = {MicrobiomeFM: A Transformer Foundation Model for Microbiome Disease Classification},
  author = {Raiyani, Mohit and Sayed, Khaled},
  year   = {2026},
  institution = {University of New Haven}
}

@article{pasolli2017curatedmetagenomicdata,
  title   = {Accessible, curated metagenomic data through ExperimentHub},
  author  = {Pasolli, Edoardo and others},
  journal = {Nature Methods},
  year    = {2017}
}

---

License

Released under CC-BY-4.0. The underlying metagenomic data is sourced from the publicly available curatedMetagenomicData resource; please also respect the original study-level licenses and citation requirements documented there.