Complete data record for the manuscript "An enzyme-resolved dataset linking drug interaction mechanisms to clinical severity grades" (submitted to Scientific Data). Everything a reader, reviewer or reuser needs is here: all 17 data tables, a machine-readable data dictionary, generated figures, worked examples, and a self-contained interactive browser.
Data availability. All files are in this public repository. No registration, login, licence request, fee or embargo applies — anyone can download everything anonymously and immediately, either from this page or with
git clone https://github.com/adeebnoor/ddi-enzyme-severity-resource.git. Released under CC BY 4.0.
Cite this dataset
Noor, A. An enzyme-resolved dataset linking drug interaction mechanisms to clinical severity grades. Submitted to Scientific Data (2026). Dataset: https://github.com/adeebnoor/ddi-enzyme-severity-resource
BibTeX, APA and other formats are generated automatically from
CITATION.cff— use the "Cite this repository" button in the sidebar. ORCID 0000-0002-8251-1853.
- What this resource is
- What it shows
- The interactive browser
- Repository layout
- File inventory
- How to use it
- Verifying the record
- Scope, caveats and known limitations
- Provenance of source data
- Licence and citation
Mechanistic drug–drug interaction (DDI) catalogues tell you how two drugs interact. Clinical catalogues tell you how badly. The two vocabularies have historically not been joined, so a pharmacologist asking "which enzyme is carrying the clinically severe interactions?" has had no dataset to ask it of.
This resource harmonises a mechanistic DDI corpus (GoldD3R, D3-derived, UMLS CUI-indexed) with a severity-graded clinical corpus (DDInter 2.0), and — the step that makes it useful — resolves each surviving interaction to the specific cytochrome P450 enzyme or membrane transporter responsible for it.
The result is a pair-level table where mechanism, responsible enzyme and clinical severity grade sit on the same row, plus per-enzyme association statistics and four independent external validations: the DrugBank hard-proven set, the manually curated KEGG DDI set, clinical-trial adverse-event reports, and the trueDDI gold set.
In one line: 1,900 enzyme-annotated drug pairs, 240 drugs, 35 enzymes and transporters, four external validations, one interactive browser, everything open.
Every figure below is generated directly from the CSV files in data/ by
scripts/make_figures.py — nothing is hand-drawn, so
the figures cannot drift from the data. Full commentary on each is in
figures/FIGURES.md; PNG and SVG versions of all of them
are in figures/.
CYP3A4 handles more interactions than any other enzyme in the resource (234 pairs) and is usually treated as the main driver of clinical DDI risk. It is in fact depleted of Major-severity interactions (OR 0.49, q = 0.017). The enzymes that are enriched are SLCO1B1 (OATP1B1) (OR 5.24, q = 0.022) and CYP2C9 (OR 2.86, q = 0.001) — a transporter and a comparatively low-traffic CYP.
Six of the nine enzymes do not survive FDR correction. They are shown in grey
and should be read as null results, not as trends. The SLCO1B1 estimate rests
on 12 pairs and its confidence interval runs from 1.65 to 16.64 — and it is
sensitive to a labelling artefact: the same transporter also appears under a
bare SLCO1B1 label, and merging the two gives OR 3.92 (1.33–11.57),
p = 0.018. The direction survives, the magnitude does not. See
limitations.
Severity grades inherited through harmonisation are only worth anything if they predict something external. They do, in two gold standards that were not used to build the resource. The DrugBank gradient is monotone across all four levels (trend p = 1.8 × 10⁻¹⁸). In KEGG, Major and Moderate are effectively indistinguishable (35.1% vs 35.4%) — an honest limitation — but graded pairs still confirm far more often than ungraded ones (OR 4.82).
70% of pairs are Unknown: DDInter records no clinical grade for them. This is
a statement about coverage, not about safety, and it is the single most
important thing to understand before reusing the data. Treating Unknown as a
negative class will bias any model trained on it.
 |
 |
| Inhibition dominates induction; CYP-mediated dominates transporter-mediated. A pair may carry several classes. | The strongest of 434 FDR-significant enzyme × adverse-event signatures. Drug-class-mediated and descriptive — not causal. |
docs/index.html is a single self-contained HTML file:
all data and code are embedded, there are no external scripts, stylesheets, CDN
calls or network requests of any kind. It works offline, from a USB stick, or
behind an air-gapped firewall.
Search by drug name; filter by enzyme/transporter, severity grade and trial-observation status; sort any column; export the current filtered view to CSV.
| How to open | What to do |
|---|---|
| Locally — always works | Download the repository and double-click docs/index.html. Any modern browser. |
| Live on the web | https://adeebnoor.github.io/ddi-enzyme-severity-resource/ — enable once under Settings → Pages → Source: main / /docs. |
| Without Pages | https://htmlpreview.github.io/?https://github.com/adeebnoor/ddi-enzyme-severity-resource/blob/main/docs/index.html |
.
├── README.md this file — data descriptor and file inventory
├── README.ar.md ملخص بالعربية
├── LICENSE CC BY 4.0
├── CITATION.cff machine-readable citation metadata
├── CHECKSUMS.sha256 SHA-256 for every released file
├── validate.py re-derives all 219 documented figures from the data
├── data/ 17 CSV tables + DATA_DICTIONARY.md
├── docs/
│ └── index.html self-contained interactive browser
├── figures/ generated PNG + SVG, and FIGURES.md
├── tables/ Tables 1-8 as CSV + a formatted workbook
├── scripts/
│ └── make_figures.py regenerates every figure from data/
└── examples/
├── quickstart.py six worked recipes
└── README.md what each recipe demonstrates
All files are UTF-8, comma-separated, with a single header row. Row counts
exclude the header and were computed from the released files.
Column-by-column definitions are in
data/DATA_DICTIONARY.md.
| File | Rows | Description |
|---|---|---|
data/ddi_enzyme_database.csv |
1,900 | The primary resource. One row per drug pair: responsible enzyme(s)/transporter(s), mechanism class(es), DDInter clinical severity grade, external-validation flags. |
Columns: drug_A, drug_B (cleaned display names) · enzymes
(comma-separated gene symbols) · mechanism (one or more of CYP inhibition,
CYP induction, transporter inhibition, transporter induction) ·
severity (Major / Moderate / Minor / Unknown) · in_trial
(Yes/No) · trial_AEs (count of trial adverse-event reports) ·
in_trueDDI (Yes/No).
At a glance: 240 distinct drugs, 35 distinct enzymes and transporters,
562 pairs carrying a graded severity (93 Major, 419 Moderate, 50 Minor) and
1,338 Unknown. Mechanism-class occurrences: CYP inhibition 1,547, transporter
inhibition 667, CYP induction 630, transporter induction 228. 79 pairs appear
in the clinical-trial DDI set, 229 in the trueDDI gold set.
| File | Rows | Description |
|---|---|---|
data/enzyme_pair_severity.csv |
3,072 | Long format: one row per pair × enzyme × direction. Columns CID_A, drug_A, CID_B, drug_B, enzyme_gene, uniprot, direction (inhibition / induction for CYP-mediated, transporter_inhibition / transporter_induction for transporter-mediated), severity. |
data/enzyme_pair_drugbank_flagged.csv |
1,799 | The same pair table with DrugBank accessions (DB_A, DB_B) and a per-row drugbank_hard_proven flag. Covers 1,172 unique pairs. |
| File | Rows | Description |
|---|---|---|
data/enzyme_severity_stats.csv |
9 | Per-enzyme association with Major severity for the nine enzymes with ≥10 non-Unknown, inhibition-direction pairs: n_pairs, n_major, pct_major, odds ratio with 95% CI, raw p, FDR q, E-value, direction of effect. |
data/enzyme_phenotype_enrichment.csv |
434 | FDR-significant enzyme × adverse-event (HPO term) associations across 7 enzymes and 338 distinct HPO terms. |
| Enzyme / transporter | Pairs | % Major | OR (95% CI) | FDR q | E-value |
|---|---|---|---|---|---|
| SLCO1B1 (OATP1B1) | 12 | 50.0 | 5.24 (1.65–16.64) | 0.022 | 9.96 |
| CYP2C9 | 94 | 30.9 | 2.86 (1.71–4.76) | 0.001 | 5.16 |
| ABCC2 (MRP2) | 18 | 22.2 | 1.44 (0.46–4.48) | 0.669 | — |
| ABCB1 (P-gp) | 149 | 16.8 | 1.08 (0.66–1.76) | 0.900 | — |
| ABCC1 | 13 | 15.4 | 0.90 (0.20–4.13) | 1.000 | — |
| CYP1A2 | 39 | 10.3 | 0.55 (0.19–1.59) | 0.559 | — |
| CYP3A4 | 234 | 10.7 | 0.49 (0.30–0.80) | 0.017 | 3.48 |
| CYP2C19 | 37 | 8.1 | 0.42 (0.13–1.40) | 0.316 | — |
| CYP2D6 | 35 | 5.7 | 0.29 (0.07–1.22) | 0.220 | — |
Bold = survives FDR correction. The other six rows are null results.
| File | Rows | Description |
|---|---|---|
data/drugbank_validation.csv |
73 | Pairs independently confirmed in the DrugBank hard-proven set. |
data/drugbank_validation_summary.csv |
8 | 1,172 pairs DrugBank-mapped, 73 hard-proven confirmed (11 Major, 37 Moderate, 2 Minor, 23 Unknown); 72 Micromedex pairs mapped, 35 in the intersection. Confirmation by grade: Major 25.0%, Moderate 15.0%, Minor 7.7%, Unknown 2.7% (trend p = 1.8 × 10⁻¹⁸). |
data/kegg_ddi_validation.csv |
1,243 | Testable pairs against the curated KEGG DDI set. kegg_flag is documented as CI = contraindication / P = precaution, but all 204 flagged rows are P — no contraindication-level pair survives the mapping, so the KEGG check rests entirely on precaution-level pairs. |
data/kegg_enzyme_concordance.csv |
86 | Per-drug comparison of our enzyme attribution against KEGG METABOLISM annotation. 86.0% agreement. |
data/kegg_validation_summary.csv |
13 | 228 drugs mapped, 3,034 KEGG DDI pairs, 1,243 testable. Major 35.1%, Moderate 35.4%, Minor 14.3%, Unknown 9.6% (z = 10.49, p = 9.6 × 10⁻²⁶); graded-vs-Unknown OR 4.82. |
data/trial_validation.csv |
79 | Enzyme-annotated pairs observed in clinical-trial adverse-event reports. 41 of the 79 carry a graded severity (7 Major, 29 Moderate, 5 Minor); the remaining 38 are Unknown. |
data/trial_validation_summary.csv |
8 | 20,618 enzyme pairs total, 545 (2.6%) in the trial DDI set, 4,233 (20.5%) in trueDDI. Caveat: the metric labelled graded_and_in_trials = 79 counts all trial-observed pairs, not only graded ones — 41 are graded. The 20,618-pair superset is not part of this deposit. |
data/liddi_coverage_comparison.csv |
7 | Overlap with the independent LIDDI corpus: 4,070 LIDDI pairs vs 12,493 GoldD3R pairs, 169 shared drugs, 0 shared interaction pairs — LIDDI is entirely complementary. |
tables/ holds Tables 1–8 of the Data Descriptor as UTF-8 CSV, plus
SupplementaryTables.xlsx with one sheet per table. Tables 4, 6, 7 and 8 are
recomputable from data/ and are asserted by validate.py; Tables 1, 2, 3 and 5
derive from intermediate tables upstream of this deposit and are included so the
published values can be read and cited without transcription. See
tables/README.md.
| File | Rows | Description |
|---|---|---|
data/cid_severity_bridge.csv |
448 | PubChem CID ↔ DDInter ID bridge with the authoritative cleaned display_name. |
data/cid_drugbank_crosswalk.csv |
255 | PubChem CID → DrugBank accession mapping used for the DrugBank validation. |
data/ddinter2_mechanism_annotations.csv |
11,298 | DDInter 2.0 mechanism layer: identifiers, severity, seven binary PK/PD mechanism flags, free-text mechanism_text. Every one of the 11,298 rows is Major — this is a Major-only slice, not a general DDInter export. Documentation and linkage only. |
data/mimic_coprescription_examples.csv |
13 | Illustrative real-world co-prescriptions from the MIMIC-IV demo cohort (100 patients). |
git clone https://github.com/adeebnoor/ddi-enzyme-severity-resource.git
cd ddi-enzyme-severity-resource
python3 examples/quickstart.py # six worked recipes, needs pandasimport pandas as pd
ddi = pd.read_csv("data/ddi_enzyme_database.csv")
# All Major-severity pairs mediated by CYP2C9
ddi[(ddi.severity == "Major") & ddi.enzymes.str.contains("CYP2C9")]
# Long format, one row per pair x enzyme x direction — best for joins
pairs = pd.read_csv("data/enzyme_pair_severity.csv")
pairs.groupby(["enzyme_gene", "severity"]).size().unstack(fill_value=0)
# Only the rows suitable for supervised learning
graded = ddi[ddi.severity != "Unknown"] # 562 pairsJoin keys. Use CID_pubchem or DrugBank accessions for machine linkage.
Drug names are cleaned display names and are stable within this resource, but
cid_severity_bridge.csv is the authoritative name source — use it rather than
string-matching across files.
See examples/README.md for what each worked recipe
demonstrates.
Two commands reproduce every claim made in this documentation:
sha256sum -c CHECKSUMS.sha256 # file integrity — 46 files
python3 validate.py # every documented figure, re-derivedvalidate.py uses only the Python standard library. It re-derives all 219
figures quoted in this README and in data/DATA_DICTIONARY.md — row counts,
severity distributions, drug and enzyme counts, every odds ratio and confidence
interval, all validation-summary values — directly from the CSV files. It also
asserts cross-file invariants:
- the pair × enzyme
directioncounts reconcile exactly with the primary table's mechanism-class counts; - E-values are populated for exactly the FDR-significant enzymes and no others;
- the DrugBank confirmation gradient (25.0 / 15.0 / 7.7 / 2.7%) is recomputable from the released pair table, not merely asserted in a summary file;
- the bridge's populated
display_nameset is exactly the primary table's 240-drug set; - no CSV has ragged rows;
- and every data-quality caveat listed below is asserted, so a reader can confirm the limitations are real and still hold (the trial-file severity breakdown, the Major-only DDInter layer, the absent KEGG contraindication flag, the two SLCO1B1 accessions, the 32 raw PubChem titles, the 26 target-only attributions).
It exits non-zero on any mismatch. Figures can be regenerated the same way:
python3 scripts/make_figures.pyStated plainly so that reusers do not over-read the resource.
About the severity grades
Unknownis notno interaction. 1,338 of 1,900 pairs (70%) carryseverity = Unknown, meaning DDInter records no clinical grade — not that the pair is safe. Analyses treatingUnknownas a negative class will be biased.- The KEGG gradient is not monotone. Major (35.1%) and Moderate (35.4%) confirm at indistinguishable rates. The graded-vs-ungraded separation is robust; the Major-vs-Moderate ordering is not supported by KEGG, only by DrugBank.
About the statistics
- Small strata. The SLCO1B1 result rests on 12 pairs (6 Major); its CI runs 1.65–16.64. Six of the nine enzymes do not survive FDR correction and are null results.
- The comparator arm is not fully specified. Recomputing
enzyme_severity_stats.csvfromenzyme_pair_severity.csv(unique non-Unknownpairs, inhibition direction, index enzyme vs all other enzyme-attributed pairs) reproducesn_pairsandn_majorexactly for all nine enzymes, and reproduces the odds ratio, confidence interval and p-value exactly for six of them. Three — CYP3A4, ABCB1 and CYP2C9, the three enzymes with the most multi-enzyme overlap — differ by up to about 10% (CYP3A4 0.443 vs 0.49; ABCB1 1.00 vs 1.08; CYP2C9 2.78 vs 2.86). Direction, significance and every conclusion are unchanged, but how a pair attributed to several enzymes is treated in the comparator arm needs to be stated before the numbers are exactly reproducible. - Enzyme–phenotype fingerprints are descriptive, not causal. The
associations in
enzyme_phenotype_enrichment.csvare mediated by drug class and must not be read as enzyme-attributable adverse-event risk. The released file contains the 434 FDR-significant associations only, not the full underlying observation set.
About identifiers and naming
SLCO1B1appears under two labels with two different UniProt accessions, and this affects the headline result.SLCO1B1 (OATP1B1)(accessionQ4U2R8, 18 rows) and bareSLCO1B1(accessionQ9Y6L6, the reviewed SwissProt entry, 2 rows) denote the same transporter. Only the first was carried intoenzyme_severity_stats.csv. Merging the two labels changes the SLCO1B1 estimate from n = 12, 50.0% Major, OR 5.24 (1.65–16.64), p = 0.0075 to n = 14, 42.9% Major, OR 3.92 (1.33–11.57), p = 0.018. The direction and nominal significance survive; the magnitude does not. Because the two labels carry different accessions, theuniprotcolumn cannot be used to normalise them automatically.- One further label (
SLCO1B3? (OATP)) retains an upstream uncertainty marker. Normalise gene symbols before aggregating. enzyme_pair_severity.csvdoes not carry cleaned display names. 32 of its 240 drug names are raw PubChem systematic titles (Azane;cyclobutane-1,1-dicarboxylic acid;platinumfor carboplatin,(E,Z)-Tamoxifen,Methotrexate, (A+-)-). Join the long-format tables onCID_pubchem, never on names — a name join silently drops 13% of drugs.- A few display names are brand names, not INNs.
Arthrotec,AnzemetandRespulesreach the primary table as PubChem compound titles.Arthrotecis a combination product (diclofenac + misoprostol), which is a meaningful wrinkle in a pair-level DDI resource — treat those rows with care. - 16 PubChem CIDs map to more than one DDInter ID. The bridge keeps all of them semicolon-separated rather than picking one arbitrarily; downstream joins should decide explicitly how to handle the one-to-many case.
display_nameis populated only for the 240 drugs that reach the primary table. The other 208 bridge rows are candidate compounds retained for provenance and leave it blank — including 73 rows markedjoined_to_ddinter = yes. Filter on a non-emptydisplay_namewhen you want the drugs actually used.
About what the enzyme attribution means
- A small number of pairs are attributed to a shared pharmacodynamic target
rather than to a metabolising enzyme or transporter. 26 of 1,900 pairs
(1.4%) rest solely on such an attribution —
PTGS2(COX-2) for celecoxib + ketorolac,CYP19A1for anastrozole + letrozole (both aromatase inhibitors, so the shared target is not a pharmacokinetic mechanism),HPRT1for azathioprine + 6-mercaptopurine. Two of the 26 are graded Major. - A few "interactions" are prodrug–metabolite pairs. 5-fluorouracil + capecitabine is attributed to CYP2C9 and carries the largest trial adverse-event count in the file (465), but capecitabine is a prodrug of 5-fluorouracil; azathioprine + 6-mercaptopurine is the same situation. These co-occurrences are real in the source data but are not drug–drug interactions in the mechanistic sense, and should be excluded from benchmark sets.
About the supplementary layers
- The DDInter 2.0 mechanism layer is a Major-only slice. All 11,298 rows
carry
severity = Major. It is provided for documentation and identifier linkage and carries no analytical claim; do not compute severity distributions or mechanism frequencies from it. - The 20,618-pair enzyme-attribution superset is not deposited. The rates
quoted in
trial_validation_summary.csv(2.6% trial, 20.5% trueDDI) are computed over that superset. Within this deposit the corresponding rates are 79/1,900 = 4.2% and 229/1,900 = 12.1%; the two are not comparable. - MIMIC-IV co-prescriptions are illustrative. 13 rows from a 100-patient demo cohort; they demonstrate real-world co-occurrence and support no prevalence estimate.
- Micromedex is commercial and is therefore represented only as mapped counts in the validation summary (72 pairs mapped, 35 in the hard-proven intersection), never as redistributed content.
About the resource as a whole
- This is a research resource, not a clinical decision-support tool. It must not be used as a sole basis for prescribing decisions.
| Source | Role | Identifier |
|---|---|---|
| GoldD3R (D3-derived) | Mechanistic DDI corpus, UMLS CUI-indexed | https://doi.org/10.1093/jamia/ocz061 |
| DDInter 2.0 | Clinical severity grades and mechanism layer | doi:10.1093/nar/gkae726 |
| PubChem | Compound identifiers and titles | doi:10.1093/nar/gkac956 |
| UniProt | Enzyme/transporter protein accessions | doi:10.1093/nar/gkac1052 |
| Human Phenotype Ontology | Adverse-event vocabulary | doi:10.1093/nar/gkaa1043 |
| DrugBank 5.0 | Independent validation (hard-proven set) | doi:10.1093/nar/gkx1037 |
| KEGG DRUG | Independent validation (curated DDI set) | Kanehisa et al. |
| MIMIC-IV demo | Illustrative co-prescription | doi:10.13026/dp1f-ex47 |
| LIDDI | Coverage comparison | doi:10.1007/978-3-319-25010-6_18 |
| Micromedex (Merative) | Commercial cross-check — counts only | not redistributed |
Redistribution here is limited to derived, harmonised tables. Where a source licence does not permit redistribution, only aggregate counts are reported.
All data files, documentation, figures and the interactive browser are released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). You may share and adapt the material for any purpose, including commercially, provided you give appropriate credit. Source datasets remain under their own terms — see the provenance table.
Machine-readable citation metadata is in CITATION.cff. Until
the article is published, please cite:
Noor, A. An enzyme-resolved dataset linking drug interaction mechanisms to clinical severity grades. Submitted to Scientific Data (2026). Data: https://github.com/adeebnoor/ddi-enzyme-severity-resource
Adeeb Noor — questions and corrections via the issue tracker.