TAC 2017 ADR Query Skill

Search 200 FDA drug labels annotated with adverse reactions, severity, and MedDRA normalization from the TAC 2017 shared task.

Entity Auto-detection

Input Pattern	Detected As	Match Logic
10019211 (8 digits)	MedDRA ID	exact on meddra_pt_id or meddra_llt_id
ACTEMRA (known drug)	Drug name	exact (case-insensitive) on drug label name
headache (known ADR)	ADR string	exact on ADR reaction string
anything else	Free text	substring on drug names, ADR strings, MedDRA PT/LLT names

API

Function	Input	Returns
search(entity)	single entity string	list[dict] — matching label hit(s)
search_batch(entities)	list of entity strings	dict[str, list[dict]]
summarize(hits, entity)	hit list + query label	compact LLM-readable text
to_json(hits)	hit list	list[dict] (JSON-serializable)
list_drugs()	—	sorted list of all drug names
stats()	—	dataset-level statistics dict

Hit Dict Structure

Each hit returned by search() contains:

Field	Type	Description
drug	str	Drug label name
source_file	str	XML filename
sections	list[str]	Annotated section names (e.g. "adverse reactions")
mention_counts	dict	Count per mention type (AdverseReaction, Severity, …)
num_reactions	int	Total unique reactions in this label
positive_adrs	list[str]	Positive (non-negated, non-hypothetical) ADR strings
reactions	list[dict]	Each with adr, meddra_pt, meddra_pt_id, optional meddra_llt, meddra_llt_id, flag

Usage

See if __name__ == "__main__" block in 37_TAC_2017_ADR.py for runnable examples covering: drug name lookup, ADR string search, MedDRA ID search, batch search, and JSON output.

from importlib.machinery import SourceFileLoader
tac = SourceFileLoader("tac2017", "/path/to/37_TAC_2017_ADR.py").load_module()

# Single drug
hits = tac.search("ACTEMRA")
print(tac.summarize(hits, "ACTEMRA"))

# ADR across all labels
hits = tac.search("headache")
print(tac.summarize(hits, "headache"))

# MedDRA PT ID
hits = tac.search("10019211")

# Batch
results = tac.search_batch(["ENBREL", "nausea", "10002198"])

Data

• Source: TAC 2017 ADR shared task (NLM / FDA)
• Files: gold_xml/ (99 test labels) + train_xml/ (101 training labels), each annotated XML
• Annotations: Mentions (AdverseReaction, Severity, Factor, DrugClass, Negation, Animal), Relations (Negated, Hypothetical, Effect), Reactions (unique ADRs with MedDRA PT/LLT normalization)
• MedDRA version: 18.1
• Path: DATA_DIR variable in 37_TAC_2017_ADR.py
• Reference: https://bionlp.nlm.nih.gov/tac2017adversereactions/