The only phytochemical dataset combining USDA botanical records, PubMed citation counts, ClinicalTrials.gov study counts, ChEMBL bioactivity scores, and USPTO patent density — in production-ready JSON + Parquet.

Free 400-Row Sample · Full Dataset (€699) · Quickstart Notebook

Records	Compounds	Species	Enrichment Layers
104,388	24,771	2,315	4

Schema (v2.0)

Column	Type	Nulls	Description
`chemical`	`string`	0%	Standardised compound name (USDA Duke's nomenclature)
`plant_species`	`string`	0%	Binomial Latin species name
`application`	`string`	~40%	Traditional medicinal application (e.g. "Antiinflammatory")
`dosage`	`string`	~55%	Reported dosage, concentration, or IC50 value
`pubmed_mentions_2026`	`int32`	0%	Total PubMed publications mentioning this compound (March 2026 snapshot)
`clinical_trials_count_2026`	`int32`	0%	ClinicalTrials.gov study count per compound (March 2026)
`chembl_bioactivity_count`	`int32`	0%	ChEMBL documented bioactivity measurement count
`patent_count_since_2020`	`int32`	0%	US patents since 2020-01-01 mentioning compound (USPTO PatentsView)

Why Not Build This Yourself?

Normalising and cross-referencing 24,771 phytochemicals against four authoritative databases is not a weekend project:

Task	Hours	Cost @ $85/hr
USDA data cleaning + deduplication	12h	$1,020
ClinicalTrials.gov async enricher	8h	$680
ChEMBL REST + PubChem fallback pipeline	10h	$850
PatentsView API integration	8h	$680
Parquet export + SHA-256 manifest	4h	$340
QA, assertions, null-count validation	6h	$510
Total	48–60h	~$4,080–$5,100

This dataset: €699 one-time. No subscription. No API calls. Instant download.

Why This Dataset Exists

Large language models hallucinate botanical taxonomy. A biotech team's RAG pipeline confidently outputting "Quercetin found in 450 species at 2.3 mg/g" sounds plausible — but the real number of species in our data is 215, and dosage varies by three orders of magnitude depending on the plant part.

The raw USDA Dr. Duke's database is spread across 16 relational tables. Joining them correctly requires understanding non-obvious foreign keys, handling >40% null values in application fields, and normalising species names against accepted binomial nomenclature. Most teams give up after a week.

Quickstart

Python — Load 400-row sample

import pandas as pd

url = "https://raw.githubusercontent.com/wirthal1990-tech/USDA-Phytochemical-Database-JSON/main/ethno_sample_400.json"
df = pd.read_json(url)
print(f"{df.shape[0]} records, {df['chemical'].nunique()} unique compounds")
df.head()

PyArrow — Parquet (full dataset, after purchase)

import pyarrow.parquet as pq

table = pq.read_table("ethno_dataset_2026_v2.parquet")
print(f"Schema: {table.schema}")
print(f"Rows: {table.num_rows}  Memory: {table.nbytes / 1e6:.1f} MB")

DuckDB (analytical queries, no install required)

import duckdb

result = duckdb.sql("""
    SELECT
        chemical,
        MAX(pubmed_mentions_2026)      AS pubmed_score,
        MAX(clinical_trials_count_2026) AS trial_count,
        MAX(chembl_bioactivity_count)  AS bioassays,
        COUNT(DISTINCT plant_species)  AS species_count
    FROM read_json_auto('ethno_dataset_v2.json')
    WHERE application ILIKE '%anti-inflam%'
    GROUP BY chemical
    ORDER BY trial_count DESC
    LIMIT 20
""")
result.show()

HuggingFace Datasets

from datasets import load_dataset

# Load the free 400-row sample directly from HuggingFace Hub
ds = load_dataset(
    "wirthal1990-tech/USDA-Phytochemical-Database-JSON",
    split="sample",
    trust_remote_code=False
)
df = ds.to_pandas()
print(f"Records: {len(df)} | Columns: {list(df.columns)}")
df.head()

Note: The split="sample" loads ethno_sample_400.json (400 rows, 8 columns). The full 104,388-row dataset is available at ethno-api.com.

Sample Record

Below is a real record from the dataset — QUERCETIN, one of the most-studied plant compounds:

{
  "chemical": "QUERCETIN",
  "plant_species": "Abelmoschus esculentus",
  "application": "5-Lipoxygenase-Inhibitor",
  "dosage": "IC50 (uM)=4",
  "pubmed_mentions_2026": 31310,
  "clinical_trials_count_2026": 847,
  "chembl_bioactivity_count": 4231,
  "patent_count_since_2020": 312
}

All 8 fields are populated for all 104,388 records in the full dataset. The free 400-row sample contains real values for pubmed_mentions_2026; the three enrichment fields (clinical_trials_count_2026, chembl_bioactivity_count, patent_count_since_2020) contain representative placeholder values pending completion of the full enrichment run.

File Manifest

File	Size	Format	Access
`ethno_sample_400.json`	67 KB	JSON	Free (this repo)
`ethno_sample_400.parquet`	15 KB	Parquet	Free (this repo)
`ethno_dataset_2026_v2.json`	~18 MB	JSON	Commercial (€699)
`ethno_dataset_2026_v2.parquet`	~900 KB	Parquet	Commercial (€699)
`MANIFEST_v2.json`	~1 KB	JSON	Included with purchase
`quickstart.ipynb`	6 KB	Notebook	Free (this repo)

Data Sources & Methodology

Source	Access	Date	Method
USDA Dr. Duke's Phytochemical and Ethnobotanical Databases	Public domain	2026	Full 16-table PostgreSQL import, normalized
NCBI PubMed	E-utilities API	March 2026	`esearch` per compound, total publication count
ClinicalTrials.gov	v2 API	March 2026	Study count per compound name
ChEMBL	REST API (v34)	March 2026	Bioactivity measurement count via molecule search
USPTO PatentsView	REST API v1 (`search.patentsview.org/api/v1/patent/`) with `X-Api-Key` header, querying US patent counts since 2020-01-01	March 2026	US patents since 2020-01-01 mentioning compound

All enrichment scripts are deterministic, checkpoint-resumable, and respect API rate limits. Source code available upon request for enterprise customers.

Use Cases

RAG Pipelines — Ground LLM responses with verified phytochemical data. Each record has a PubMed evidence score — use it to weight retrieval results and filter hallucinations.
Drug Discovery — Prioritise natural product leads by combining PubMed citations, clinical trial presence, ChEMBL bioactivity depth, and patent landscape. One query replaces weeks of manual lit review.
Market Intelligence — Patent density score reveals which compounds are attracting commercial investment. Cross-reference with clinical trials to identify underexplored compounds with IP whitespace.
Academic Research — Pre-computed evidence scores save months of PubMed searching. The BibTeX citation block below makes this dataset citable in peer-reviewed publications.

Dataset Versions

Version	Records	Schema	Status
v1.0	104,388	5 columns (USDA baseline)	Deprecated
v2.0	104,388	8 columns (+ PubMed, ClinicalTrials, ChEMBL, Patents)	Current

The free sample (ethno_sample_400.json) uses the v2.0 schema. Enrichment fields contain representative values pending completion of the full enrichment pipeline.

License & Commercial Access

Free 400-row sample: CC BY-NC 4.0 — use for evaluation, academic research, and prototyping.
Full 104,388-row dataset: Single-entity commercial license, €699 one-time purchase at ethno-api.com. Redistribution, resale, and derivative dataset publication are prohibited.

Citation

@misc{ethno_api_v2_2026,
  title     = {USDA Phytochemical \& Ethnobotanical Database --- Enriched v2.0},
  author    = {Wirth, Alexander},
  year      = {2026},
  publisher = {Ethno-API},
  url       = {https://ethno-api.com},
  note      = {104,388 records, 24,771 unique chemicals, 2,315 plant species, 8-column schema with PubMed, ClinicalTrials, ChEMBL, and PatentsView enrichment}
}

Contact

Website: ethno-api.com
Email: founder@ethno-api.com
GitHub: @wirthal1990-tech

_{Built by Alexander Wirth · PostgreSQL 15 · Python 3.12 · Hetzner CCX33}

Downloads last month: 9