DrugBankDDI: open-ended multi-document questions about drug–drug interactions

DrugBankDDI is a small evaluation set of open-ended questions whose answer requires combining two PubMed abstracts: one establishing how drug X relates to a protein M (an enzyme, transporter or target) and one establishing how drug Y relates to the same protein. The interaction between X and Y follows from the two facts. Every drug pair comes from a MedHop (QAngaroo) training question and every interaction is confirmed by a DrugBank interaction sentence; the shared protein and the drugs' roles come from DrugBank's protein annotations, and the two evidence abstracts are DrugBank's own citations, kept only when the abstract text names the drug, the protein and the role.

Pairs	109
Questions	327 (3 tiers per pair)
Distinct drugs	97
Bridge proteins	25
Bridge types	enzyme 74, transporter 18, target 17
DrugBank sentence classes	pk_transport_or_other 48, metabolism 44, adverse_additive 12, activity_additive 3, efficacy 2
Most common bridges	ABCB1 (15), CYP3A4 (13), CYP2D6 (12), CYP2C9 (10), CYP1A2 (10), CYP2C8 (8), CYP2C19 (7), CYP3A5 (6)
Documents per question	2 gold + up to 8 distractors
Evidence	every gold abstract LLM-verified to state the claimed role (see Evidence review)
Negative pairs	109 (config negatives) × 3 tiers = 327 questions, no known interaction
Unverified extra	49 pairs / 147 questions with a weak side, in the unverified_* configs

Question tiers

Each pair yields three questions sharing the same documents and gold:

• bridge – "Do X and Y have any protein target, enzyme or transporter in common? If so, which?"
• mechanism – "A clinician plans to prescribe X together with Y. Is there a documented basis for expecting an interaction between them, and if so what is the mechanism?"
• consequence – "A patient stabilized on X is started on Y. Would you expect a change in X's exposure or effect, or an added risk? If so, what change, and through what mechanism?"

No question refers to the documents supplied with it, so the same items can be used closed-book, with retrieval, or with the bundled documents as context. All three are phrased so that "no" is a legitimate answer, and the negatives config uses the same three templates verbatim, so question wording never reveals the label. "Documented" in the mechanism template is deliberate: the negative label means no interaction is recorded in DrugBank, not that none can exist. Positive answers open with "Yes.", negative answers with "No.". Every tier can therefore be scored as a balanced yes/no decision plus a justification.

Example (mechanism, enzyme bridge):

Q: A clinician plans to prescribe Tramadol together with Fluphenazine. Do these documents provide a basis for expecting an interaction between them, and if so what is the mechanism? A: Yes. Fluphenazine is an inhibitor of Cytochrome P450 2D6 (CYP2D6), which metabolizes Tramadol; co-administration therefore makes Tramadol exposure rise.

Schema

questions (default config), one row per question:

column	type	meaning
id	str	<pair_id>:<tier>
tier	str	bridge, mechanism or consequence
question, answer	str	natural-language question and templated reference answer
gold	str (JSON)	structured answer: interaction (true here, false in negatives), bridge_gene, bridge_name, bridge_uniprot, substrate_drug, modulator_drug, modulator_role, affected_drug, direction, effect
documents	list of struct	pmid, title, journal, year, text, role (gold:X-M, gold:Y-M, distractor:drug_other_protein, distractor:protein_other_drug, distractor:unrelated); shuffled
pair_id, medhop_id	str	pair key and the MedHop training id the pair came from
drug_X_id, drug_X_name, drug_Y_id, drug_Y_name	str	DrugBank accession and generic name
drugbank_context	str	DrugBank's interaction sentence
description_class	str	template class of that sentence
bridge_type, bridge_gene, bridge_name, bridge_uniprot	str	the shared protein. DrugBank protein-group entries (e.g. "NMDA receptor") have no gene symbol or UniProt id, so those fields are empty and answers name the protein without a parenthetical
X_M_actions, Y_M_actions	str	DrugBank action terms of each drug on the protein
X_M_pmid, Y_M_pmid	str	the gold evidence papers
mechanism_consistent	bool	roles agree with the sentence's direction (always true)
X-M_strong_evidence, Y-M_strong_evidence	bool	LLM review verdict: the gold abstract explicitly states the claimed role of that drug at the bridge protein; always true in the default configs, per side in unverified_*
n_alternative_bridges	int	other qualifying shared proteins for the pair
split	str	drug-level split: train / dev / test / mixed (see below)

pairs: one row per pair with the same metadata and gold, without questions or documents.

Evidence review

Every gold abstract was reviewed by an LLM (Claude) against a fixed rubric asking whether the abstract explicitly states the role the answer asserts for that drug at that protein. Where the first-choice paper was judged weak, every other paper on the drug's DrugBank protein card that mentions both the drug and the protein was reviewed the same way, and the first paper judged strong replaced it as the gold document (39 of 316 sides). A supporting quote and a one-line reason were recorded for every side during review.

The default questions / pairs configs contain only pairs whose two gold abstracts were both judged strong (109 pairs, 327 questions), so X-M_strong_evidence and Y-M_strong_evidence are always true there. The 49 pairs with at least one weak side (147 questions) are kept in unverified_questions / unverified_pairs with the flags set per side. For those, no paper cited on the DrugBank card states the claimed role in its abstract: the drug appears only in a screening or ligand list, the protein is discussed only at family level, or the abstract supports a different role. DrugBank still asserts the interaction, so the claims may be correct, but the documents do not carry the reasoning chain and the questions should not be used for evaluation as they stand.

Negatives (negatives config)

109 drug pairs drawn from 102 drugs for which DrugBank records no interaction in either direction and which share no DrugBank protein at all — no target, enzyme, carrier or transporter in common, in any role. The stricter condition (rather than merely "no interacting role combination") is what makes the bridge tier's negative answer true as well as the interaction answer. Pairs that MedHop itself lists as interacting are excluded even where current DrugBank has no row for them, so the label does not contradict either source.

Columns: id, tier (bridge / mechanism / consequence), pair_id, question, answer, gold, drug_X_id, drug_X_name, drug_Y_id, drug_Y_name, X_pmids, Y_pmids, documents, label (always no_known_interaction). question uses the same template as the positive tier of the same name and gold has the same keys with interaction: false and every mechanism slot null, so the two configs concatenate on their shared columns into a balanced 654-item task (327 yes / 327 no), or one balanced pair of tiers at a time. documents (6–8 per pair, repeated across the pair's three tiers) are abstracts from the same PubMed pool as the positive set: up to four that mention X and four that mention Y, preferring abstracts that also name one of the bridge proteins used in the positive release, so a model cannot separate positives from negatives by noticing that no protein is discussed. No document mentions both drugs, and every document mentions the drug whose side it is on.

Intended use: concatenate with the mechanism tier of the positive set to test whether a model asserts a mechanism only when the documents support one. The correct response is that the documents give no basis for an interaction between X and Y. Note the label means no known interaction in DrugBank, not proof of safety; absence of evidence in one database is not evidence of absence.

Grading

answer is a templated sentence; grade against gold slot-wise (bridge protein, roles, affected drug, direction, effect) with a rubric or LLM judge rather than by string match. Each tier shares its template with the matching tier of negatives, so scoring gold.interaction over the union of the two configs measures whether a model asserts a mechanism only when the documents support one; score the mechanism slots on the positive half. Negative answers are identical apart from drug names, so string-similarity metrics will reward parroting — judge the justification separately. The answers describe the direction of an effect, not its magnitude; some bridges are minor pathways for the drug in question. Report the three tiers separately and, ideally, common versus rare bridges separately.

Splits

split assigns 30 % of drugs to test, 10 % to dev and the rest to train; a pair whose two drugs fall in different groups is mixed. The drug graph is dense, so most pairs are mixed. The set is intended for evaluation; use the split column only if a leak-free train/test subset is needed.

Construction

1. 1,620 MedHop training pairs (338 drugs) → DrugBank interaction sentence for each pair (DrugBank, retrieved 2026-09-22).
2. DrugBank target / enzyme / carrier / transporter annotations for all 338 drugs, with DrugBank's article references.
3. Shared proteins per pair, filtered by a role rule (substrate on one side, inhibitor/inducer on the other for enzymes and transporters; an action on both sides for targets). Carrier (albumin) bridges and pairs without a DrugBank sentence were dropped.
4. DrugBank article ids resolved to PMIDs; abstracts from NCBI E-utilities.
5. Citation check: for each side, the cited abstract must name the drug, the protein (with synonyms such as P-glycoprotein for ABCB1) and the specific role the answer asserts for that drug (substrate words on the substrate side; inhibitor or inducer words on the modulator side; agonist- or antagonist-type words for targets). Both sides must pass with two different papers; pairs without such evidence are dropped.
6. Bridge/sentence compatibility (enzyme and transporter bridges only for pharmacokinetic sentences, target bridges only for pharmacodynamic ones), mechanism consistency, one bridge per pair, a cap of 20 pairs per bridge protein, then templated questions, distractors and the drug split.
7. LLM evidence review (below); pairs with a weak side moved to the unverified_* configs.

Limitations

• Evidence is DrugBank's curated citation confirmed by string matching, not by expert review; some bridges are minor pathways for the drug in question.
• Answers are templated and uniform in style.
• The bridge tier is answerable from prior knowledge for common enzymes such as CYP3A4; the cap limits but does not remove this.
• 109 pairs is small; the set is meant for evaluation, not training.

Licensing and sources

Released under CC BY-NC 4.0. Interaction sentences and protein annotations are derived from DrugBank (Wishart et al.), whose terms allow academic, non-commercial use; abstracts are from PubMed/MEDLINE and remain under their publishers' copyright. Drug pairs come from MedHop (Welbl, Stenetorp & Riedel, 2018, TACL). If you use this dataset, please cite those sources as well.