HARVEST

HARVEST (High-throughput Automated extRaction of bioactivity from patentS using agenTic LLMs) is a multi-agent LLM system for automated extraction of protein-ligand bioactivity data from USPTO patents. The pipeline processes patent documents through specialized agents for assay identification, bioactivity extraction, compound name resolution, chemical structure resolution, and protein mapping to UniProt identifiers.

From 43,187 USPTO patents, HARVEST extracted 3.15 million activity records, identifying 326,342 novel molecular scaffolds and 967 protein targets absent from existing databases like BindingDB.

Links

• Paper: biorxiv:10.64898/2026.03.15.711910
• Dataset: h-bench

Dataset Structure

The released H-bench dataset (data/h_bench/) contains one CSV per protein target (48 targets), each a curated subset of HARVEST results filtered for valid activity values and compounds absent from BindingDB.

Column	Description
Sequence	Full amino acid sequence of the protein target, multiple entries separated by ;
Ki (nM)	Inhibition constant in nanomolar (if reported)
IC50 (nM)	Half-maximal inhibitory concentration in nanomolar (if reported)
Kd (nM)	Dissociation constant in nanomolar (if reported)
EC50 (nM)	Half-maximal effective concentration in nanomolar (if reported)
relation	Activity value relation operator (=, <, >, ~)
original_range	Original range string if the patent reported a range (e.g. "10-20 nM")
patent_number	USPTO patent identifier (e.g. US20200299264A1)
chemical_id	Internal chemical identifier from the patent XML
compound_IUPAC_name	IUPAC name of the compound (when available in the patent text)
original_alias	Compound alias as used in the patent (e.g. "Example 1", "Compound 160a")
protein_target_name	Name of the protein target as stated in the patent
gene	Gene symbol (e.g. ESR1, JAK1)
organism_scientific	Scientific name of the source organism (e.g. Homo sapiens)
uniprot_acc	UniProt accession of the protein target, multiple entries separated by ;
UniProt ID	UniProt entry name (e.g. ESR1_HUMAN), multiple entries separated by ;
mutations	Protein mutations noted in the assay (if any)
protein_modification	Protein modification type (e.g. Degradation, Inhibition)
assay	Assay type classification (e.g. Functional, Binding)
assay_description	Full assay description extracted from the patent
year	Publication year of the patent
SMILES	Canonical SMILES representation of the compound
cluster_id	Tanimoto-based cluster identifier from scaffold clustering
final_label	Split label: A = novel (harvest-unique cluster), C = buffer region between HARVEST and BindingDB
nearest_BDB_smiles	SMILES of the most similar compound in BindingDB
tanimoto_sim	Tanimoto similarity to the nearest BindingDB compound
InChIKey	InChIKey identifier for the compound

Limitations

• English USPTO patents only: The current release covers only US patents in English. EPO, WIPO, and non-English patents are not included.
• h_bench is a raw subset, not a ready-to-use benchmark: h_bench contains molecules from HARVEST that are absent from BindingDB, but it is not a fully curated validation set. To create a high-quality benchmark for a specific ML task (e.g. virtual screening, binding affinity prediction, activity cliff detection), additional processing is required — such as filtering by activity type, selecting active/inactive thresholds, deduplicating by scaffold, and ensuring temporal or structural separation from training data.

Training Data Allocation

When using H-bench as a test set, your training data must be checked for leakage. The allocate_training.py script compares your training dataset against H-bench and adds a final_label column indicating whether each molecule is safe to use:

• B — safe for training, no leakage risk
• A — exact duplicate of an H-bench test compound, must be excluded
• C — too structurally similar to test compounds (within the Tanimoto buffer zone), must be excluded

Only keep rows with final_label == "B" in your training set.

Protein matching uses UniRef90 clusters by default: if your training data contains a protein that shares a UniRef90 cluster (>=90% sequence identity) with an H-bench protein, the script will check for compound-level leakage across all cluster members.

Usage

# Basic usage (UniRef90 protein grouping, recommended)
python allocate_training.py \
    --new-data your_training_data.parquet \
    --split-dir data/h_bench/ \
    -o output/ \
    --uniref-cache uniref_cache.json

# With custom column names
python allocate_training.py \
    --new-data your_training_data.csv.gz \
    --split-dir data/h_bench/ \
    -o output/ \
    --smiles-col-new SMILES \
    --uniprot-col-new uniprot_id

# Without UniRef grouping (exact UniProt match only)
python allocate_training.py \
    --new-data your_training_data.parquet \
    --split-dir data/h_bench/ \
    -o output/ \
    --uniref none

The script outputs:

• Labeled copy of your data with the final_label column (same format as input: parquet or csv.gz)
• allocation_report.csv — per-protein breakdown of removed molecules