Muhammad Muneeb · GitHub · Project repository
A reproducible, source-aware workflow for extracting, harmonising, integrating, and evaluating drug–disease relationships from 11 biomedical resources. The project retains original source identifiers, relationship and evidence classes, and supporting-event provenance. It is intended for biomedical data integration, computational research, evidence-aware retrieval, and hypothesis generation—not as a catalogue of uniformly approved or clinically effective treatments.
Manuscript: A source-aware drug–disease relationship dataset integrating public biomedical resources — Muhammad Muneeb and David B. Ascher.
Published dataset (24 October 2026): The validated eleven-source dataset is publicly available as Zenodo v3 (DOI: 10.5281/zenodo.23627171). The archive contains 14 files, including the event, assertion and strict-identity pair datasets, integration and QC manifests, data dictionaries, source/licensing documentation, selected supplementary outputs, and SHA-256 checksums. The concept DOI links all versions. This is the published Zenodo v3 record for the internally validated eleven-source integration; a final tagged code release is not yet asserted. Third-party data redistribution and CC BY 4.0 compatibility remain subject to rights review.
- Validated dataset summary
- Sources and versions
- Data layers and schema
- Installation
- Full processing workflow
- Analysis scripts
- Technical validation
- Licensing and redistribution
- Data availability and release provenance
- Citation and acknowledgements
These statistics are from the locked manuscript-analysis inputs processed and audited on 24 October 2026.
| Measurement | Validated value |
|---|---|
| Included original resources | 11 |
| Harmonised source evidence events | 10,895,395 |
| Distinct source-specific assertions | 4,534,385 |
| Additional supporting events represented by assertion counts | 6,361,010 |
| Strict-identity drug–disease pairs | 4,470,601 |
| Distinct conservatively normalised drug-name keys | 158,364 |
| Distinct conservatively normalised disease/condition-name keys | 72,563 |
| Distinct conservatively normalised drug–disease name pairs | 4,313,189 |
| Normalised-name pairs present in 2+ sources | 83,878 (1.945%) |
| Potentially conflicting strict-identity pairs | 87 |
| Original relationship types / evidence types | 18 / 28 |
| Analytical evidence families | 14 |
Interpretation: Events, assertions, strict-identity pairs, and normalised-name pairs are different counting units. The 6,361,010 additional supporting events were retained in the event ledger and accounted for through assertion support counts; they are not 6.36 million deleted or independently validated biological observations. Conservatively equal drug/disease names do not establish chemical or ontology-level identity.
| Source | Evidence events | Assertions | Additional supporting events |
|---|---|---|---|
| AACT | 688,745 | 485,060 | 203,685 |
| Broad Drug Repurposing Hub | 3,892 | 3,885 | 7 |
| ChEMBL | 60,055 | 59,751 | 304 |
| CTD | 9,735,698 | 3,592,638 | 6,143,060 |
| DrugCentral | 42,515 | 42,479 | 36 |
| Hetionet | 1,145 | 1,145 | 0 |
| MEDI | 222,412 | 222,412 | 0 |
| Open Targets | 53,950 | 53,950 | 0 |
| PrimeKG | 42,631 | 42,631 | 0 |
| repoDB | 13,558 | 12,118 | 1,440 |
| SIDER | 30,794 | 18,316 | 12,478 |
| Total | 10,895,395 | 4,534,385 | 6,361,010 |
CTD's large reduction from evidence events to assertions mainly reflects multiple inferred-gene and publication-associated events underlying the same chemical–disease assertion. AACT similarly retains distinct trial-level contexts. The source-event ledger, rather than the assertion layer, should be used when those contexts are needed.
Pairwise drug–disease name-pair overlap across the eleven included databases. The graphic summarises shared conservatively normalised drug–disease name pairs and pairwise Jaccard similarity. Lexical overlap does not establish ontology-level equivalence, independent evidence, or validated therapeutic efficacy. The PNG figure is stored in the repository root; its generation code is Analysis_09_Cross_source_drug_disease_pair_ovelap_diagram.py. For reproducibility, consult the analysis scripts and deposited dataset version.
The included source processors write harmonised CSV files for exactly eleven resources. Source download/release labels are not automatically equivalent to the original acquisition timestamp.
| Resource | Inspected input / version | Processor | Primary source |
|---|---|---|---|
| AACT / ClinicalTrials.gov | Local archive labelled aact_20260114 |
aact.py |
AACT |
| Broad Drug Repurposing Hub | Annotation file labelled 20200324 |
Broad.py |
Repurposing Hub |
| ChEMBL | ChEMBL 37 | chembl.py |
ChEMBL downloads |
| Comparative Toxicogenomics Database (CTD) | Chemical–disease report generated 2026-04-30 | ctd.py |
CTD |
| DrugCentral | Database version 56, frozen extraction 2026-10-24 | drugcentral.py |
DrugCentral |
| Hetionet | v1.0 | hetionet.py |
Hetionet |
| MEDI | MEDI-2 (ICD and UMLS inputs) | MEDI.py |
MEDI |
| Open Targets Platform | 26.03 | opentargets.py |
Release directory |
| PrimeKG | Harvard Dataverse kg.csv snapshot |
PrimeKG.py |
PrimeKG dataset |
| repoDB | Local full export; exact upstream revision unconfirmed | repoDB.py |
repoDB |
| SIDER | SIDER 4.1 indication files | sider.py |
SIDER |
Not included: TTD.py and openFDA.py are exploratory processors and do not contribute to the locked eleven-source integration. Do not add their outputs without a documented and revalidated change to the source manifest.
The source-specific files share ten leading columns:
| Column | Meaning |
|---|---|
drug_name |
Source-derived drug, substance, chemical, or intervention label |
drug_identifier |
Reported drug identifier, if supplied |
drug_identifier_type |
Drug identifier namespace |
disease_or_condition_name |
Reported disease, condition, phenotype, or indication label |
disease_or_condition_identifier |
Reported condition identifier, if supplied |
disease_or_condition_identifier_type |
Condition identifier namespace |
relationship_type |
Original/harmonised relationship classification |
evidence_type |
Source-defined evidence classification |
source |
Contributing database |
internal_source |
Internal source or evidence subset |
Additional source-specific columns retain available trial IDs, PubMed IDs, inferred genes, original identifiers, clinical stages, source rows, and other provenance. Not every source has the same optional fields.
File under data/processed/ |
Unit of analysis | Meaning |
|---|---|---|
ALL_INCLUDED_SOURCES_drug_disease_merged.csv |
Evidence events | Pre-deduplication harmonised source event ledger; this is not an untouched raw-source dump |
ALL_INCLUDED_SOURCES_drug_disease_deduplicated.csv |
Source-specific assertions | Unique 9-field source-semantic states, with supporting-event counts and preserved evidence links |
ALL_INCLUDED_SOURCES_drug_disease_source_collapsed.csv |
Strict-identity pairs | Exact agreement across the six represented drug/disease name, identifier and namespace fields; evidence and source support remain available |
The assertion identity uses the six drug/disease identity fields plus relationship_type, evidence_type, and source. Strict-identity pairs use the six identity fields alone. A separate descriptive lexical analysis lowercases, trims, and collapses whitespace in names; it does not resolve synonyms, stereoisomers, salts, drug classes, brand names, disease ontologies, or chemical equivalence.
flowchart TD
A[11 frozen biomedical inputs] --> B[11 source-specific Python processors]
B --> C[Harmonised events with original evidence metadata]
C --> D[Event ledger: 10,895,395]
D --> E[Source-specific assertions: 4,534,385]
E --> F[Strict-identity pairs: 4,470,601]
D --> G[Integrity and join checks]
E --> H[Evidence semantics and lexical overlap]
F --> I[Retrieval and validation analyses]
# Clone the project
git clone git@github.com:MuhammadMuneeb007/drug-disease-mapping.git
cd drug-disease-mapping
# Create an environment
python3 -m venv .venv
source .venv/bin/activate
# Dependencies used across the processors, merger, and analyses
python -m pip install 'polars==1.44.1' tqdm pandas pyarrow duckdb requests psycopg2-binary rich numpy matplotlibPython must have access to the original compatible source files, enough disk space for the event ledger and work partitions, and sufficient RAM. Several source processors assume project-relative raw files or an HPC-specific path. Check each processor's --help and existing source inventory before running. On shared clusters, schedule large acquisition, merging, and QC work on a compute node rather than a login node.
For the validated integration, the observed Polars version was 1.44.1. Hash-derived IDs can be version-dependent. Use the recorded environment, source checksums, and the final tagged Git commit to reproduce an exact deposited release.
Obtain the correct source versions from the official publishers in accordance with their terms. The expected source locations include:
data/raw/aact_20260114/aact_export.zip
data/raw/broad_repurposing/repurposing_drugs_20200324.txt
data/raw/chembl/chembl_37_sqlite.tar.gz
data/raw/ctd/CTD_chemicals_diseases.tsv.gz
data/raw/hetionet/hetionet-v1.0-nodes.tsv
data/raw/hetionet/hetionet-v1.0-edges.sif.gz
data/raw/medi/MEDI-2.csv
data/raw/medi/MEDI-2_UMLS.csv
data/raw/repodb/repodb_full_download.tsv
data/raw/sider/drug_names.tsv
data/raw/sider/meddra_all_indications.tsv.gz
Other inputs include the Open Targets 26.03 Parquet files, the PrimeKG kg.csv download and the frozen DrugCentral relationship snapshot. These large upstream datasets are not bundled with the GitHub code, and a successful installation does not automatically download or license them.
Run these scripts only after their corresponding source inputs are available. They can overwrite the relevant files under data/processed/. Check each script's source-specific options before rerunning a locked integration.
python aact.py
python Broad.py
python chembl.py
python ctd.py
python drugcentral.py --offline # Reuses the frozen local DrugCentral snapshot
python hetionet.py
python MEDI.py
python opentargets.py --offline
python PrimeKG.py
python repoDB.py
python sider.pyThe eleven files required by the merger must be present at these precise relative paths:
data/processed/aact/aact_drug_disease.csv
data/processed/broad_repurposing/broad_repurposing_drug_disease.csv
data/processed/chembl_drug_disease.csv
data/processed/ctd_aggregate_chemical_disease.csv
data/processed/drugcentral_drug_disease.csv
data/processed/hetionet/hetionet_drug_disease.csv
data/processed/medi/medi_drug_disease.csv
data/processed/opentargets_drug_disease.csv
data/processed/primekg/primekg_drug_disease.csv
data/processed/repodb/repodb_drug_disease.csv
data/processed/sider/sider_drug_disease.csv
These are the input filenames hardcoded in Analysis_00_MergeDatabases.py; do not rename them or substitute raw archives. Some scripts use pinned input files by default and will fail rather than silently download missing data. drugcentral.py needs an existing snapshot for --offline; otherwise consult its PostgreSQL options.
Warning: this step regenerates/replaces final integration outputs. Skip it when merely inspecting the validated frozen files.
python Analysis_00_MergeDatabases.py --processed-dir data/processedThe merger validates eleven source inputs, writes the complete harmonised event ledger, constructs source-specific assertion and exact-identity pair views, and generates the manifest with counts and SHA-256 checksums. By default it expects the locked source counts; --allow-count-drift exists for deliberate exploratory runs and must not be used to claim reproduction of the published dataset.
python Analysis_00C_MergeQuality.py \
--processed-dir data/processed \
--verify-hashes --verify-source-hashesThe auditor checks the available merged layers, source and output checksums, event support, pair identities, and potential semantic conflicts. Use the same Polars version and bucket count as the merger.
Run from the repository root after the merged files and manifest exist:
# S1–S7: source composition, relationships, overlap and name/identifier QC
python Analysis_01_Source_level_dataset_composition.py --save-tables
python Analysis_02_Evidence_type_distribution.py --save-tables
python Analysis_03_Relationship_type_distribution.py --save-tables
python Analysis_04_Unique_drug_coverage.py --save-tables
python Analysis_08_Cross_source_drug_disease_pair_overlap.py --save-tables
python Analysis_09_Cross_source_drug_disease_pair_ovelap_diagram.py
python Analysis_11_Drug_disease_identifier_completeness.py --save-tables
python Analysis_13_Excluded_source_analysis.py --save-tables
# S8–S10: worked example, comparator analysis, and selected indications
python Analysis_18_Worked_example_type2_diabetes_indication_use_case.py --save-tables
python Analysis_19_Benchmark_against_PrimeKG_Hetionet.py --save-tables
python Analysis_20_Positive_control_approved_indication_sanity_check.py --save-tables
# S11: build the FDA Orange Book ingredient reference, then evaluate coverage
python Analysis_23_WHO_EML_or_Approved_Drug_Coverage_make_reference.py \
--products EOBZIP_2026_04.zip
python Analysis_23_WHO_EML_or_Approved_Drug_Coverage.py --save-tablesUse the exact filenames as committed. Some module docstrings refer to older filenames; the commands above use the current repository filenames. Analysis_22_Ontology_normalised_cross_source_overlap.py is an empty local placeholder in the inspected working directory, not a completed analytical step.
The committed heatmap PNG is available at the repository root and is displayed above. The figure-generation script can also write plots under figures/ and supporting tables under tables/ when run locally; these generated directories are not necessarily committed to GitHub. The heatmap expresses shared conservatively normalised drug–disease name-pair counts and pairwise Jaccard similarity, not conditional row-overlap percentages. Reproduction should use the matching validated snapshot and documented scripts.
| Script | Purpose |
|---|---|
Analysis_0.1_InspectFiles.py |
Inspect source and processed files |
Analysis_00_MergeDatabases.py |
Eleven-source integration, event ledger, assertion and pair generation |
Analysis_00C_MergeQuality.py |
Independent merger validation and hash integrity |
Analysis_01_Source_level_dataset_composition.py |
Source composition, counts, names, evidence support |
Analysis_02_Evidence_type_distribution.py |
Original evidence labels and analytical evidence families |
Analysis_03_Relationship_type_distribution.py |
Original relationship types and related summaries |
Analysis_04_Unique_drug_coverage.py |
Name-level drug, disease, and pair coverage |
Analysis_08_Cross_source_drug_disease_pair_overlap.py |
Source support and pairwise lexical overlap |
Analysis_09_Cross_source_drug_disease_pair_ovelap_diagram.py |
Heatmap and pairwise Jaccard visualisation |
Analysis_11_Drug_disease_identifier_completeness.py |
Identifier availability, namespace and mapping ambiguity |
Analysis_13_Excluded_source_analysis.py |
Source manifest, schema and name-quality checks |
Analysis_18_Worked_example_type2_diabetes_indication_use_case.py |
Type 2 diabetes retrieval example |
Analysis_19_Benchmark_against_PrimeKG_Hetionet.py |
Comparisons with PrimeKG and Hetionet |
Analysis_20_Positive_control_approved_indication_sanity_check.py |
Selected indication recovery controls |
Analysis_23_WHO_EML_or_Approved_Drug_Coverage_make_reference.py |
FDA Orange Book single-ingredient RX/OTC reference construction |
Analysis_23_WHO_EML_or_Approved_Drug_Coverage.py |
External reference drug-name coverage analysis |
The analysiscode/ directory also holds source-oriented exploratory analyses. Not every script is required to generate the locked eleven-source dataset.
The accompanying manuscript and supplementary analyses describe the following completed analyses on the validated local snapshot:
- Event conservation: 10,895,395 source events reconcile to 4,534,385 source-specific assertions with preserved supporting-event counts.
- Integrity: all 11 processed merger inputs and 23 identified raw/snapshot files passed comparison against recorded SHA-256 hashes.
- Source relationships: 18 original relationship types, 28 original evidence types, 14 broader analytical evidence families; labels are not interchangeable.
- Identifier and name auditing: source identifier namespaces, name multiplicity, identifier-shaped placeholders, and potential mapping conflicts were assessed without asserting global entity resolution.
- Duplicate multiplicity and joins: CTD retained inferred-gene source records; an AACT join-cardinality check covered 202,623 trials with 0 mismatches in the evaluated 688,745 intervention–condition combinations.
- Source overlap: 83,878 conservatively normalised name pairs occurred across at least two included databases; this is lexical recurrence, not independent confirmation.
- Positive-control retrieval: 20 of 20 selected known indication examples had at least one strict indication-oriented evidence event. This is a selected sanity check, not a population-level precision, specificity, or approval estimate.
- FDA Orange Book active-ingredient-name coverage: 1,467 / 1,543 (95.075%) single-active-ingredient RX/OTC reference names were strictly matched. An additional 25 formulation-review cases and 51 unmatched names were not counted as covered; this is name coverage, not indication-specific validation.
Further external negative controls, manual semantic-adjudication precision estimates, and regulatory-status verification would provide stronger independent accuracy evidence; they are not claimed as completed by the analyses listed above.
The repository LICENSE applies the MIT licence to project-created source code, not automatically to records obtained or derived from third-party biomedical sources. Public downloads do not by themselves confer permission to republish the integrated database under MIT, CC0, or another unrestricted data licence.
The following table reports the licensing statements and concerns documented in the supplementary source-licensing assessment. This is an attribution/reuse summary, not a legal clearance opinion. Verify the current provider terms and constituent-data rights for the actual release before distribution.
| Data source | Reported terms / licence | Conditions to check before redistributing integrated records |
|---|---|---|
| AACT / ClinicalTrials.gov | Public trial data; no blanket redistribution licence established for complete derivative compilation | Cite ClinicalTrials.gov/AACT; confirm constituent-data reuse rights |
| Broad Drug Repurposing Hub | CC BY 4.0 for published compound annotations; separate portal restrictions may apply | Attribution; commercial-use and redistribution scope |
| ChEMBL | CC BY-SA 3.0 Unported | Attribution and ShareAlike requirements for derived content |
| CTD | Source-specific data-use requirements | Attribution, linking, notification, redistribution authority; commercial terms |
| DrugCentral | CC BY-SA 4.0 reported | Attribution and ShareAlike conditions |
| Hetionet | Original content reported as CC0 | Check upstream-derived data and preserve relevant source citations |
| MEDI | CC BY-NC-SA 3.0 Unported | NonCommercial and ShareAlike obligations; review licence compatibility |
| Open Targets | Platform-level CC0, with third-party rights reservations | Confirm contributing-provider conditions |
| PrimeKG | Harvard Dataverse dataset reported CC0 1.0 | Review underlying graph input/terminology rights and cite dataset |
| repoDB | CC BY 4.0 reported | Attribution; review embedded DrugBank/UMLS and trial-derived rights |
| SIDER | CC BY-SA 4.0 reported for indication data | Attribution, ShareAlike and underlying terminology rights |
Rights review for the published Zenodo v3 dataset: The public Zenodo record displays CC BY 4.0, but the eleven contributing sources have potentially incompatible conditions, including non-commercial and share-alike terms. The release documentation reports these terms; it does not establish blanket redistribution clearance. Preserve source attribution, obtain institutional/source-owner guidance, and if required correct access, licence metadata, or the released content using an authorised approach. The MIT software licence applies to project-created code only.
- Published validated dataset (Zenodo v3, 24 October 2026): DOI 10.5281/zenodo.23627171
- All versions (concept DOI): 10.5281/zenodo.18308459
- Earlier versions: v1 and v2
- Code: GitHub repository
- Frozen code tag / archived code DOI: Not yet established in this README; do not substitute an audit-time checkout for a verified release tag.
The 14 deposited files comprise three compressed CSV datasets (ALL_INCLUDED_SOURCES_drug_disease_merged.csv.gz, ALL_INCLUDED_SOURCES_drug_disease_deduplicated.csv.gz, and ALL_INCLUDED_SOURCES_drug_disease_source_collapsed.csv.gz) plus ALL_INCLUDED_SOURCES_merge_summary.json, Analysis_00C_merge_quality_report.json, data_dictionary.csv, data_dictionary.md, data_dictionary_summary.json, source_manifest.csv, LICENSES_AND_REUSE.md, README_DATA_RELEASE.txt, release_manifest.json, checksums_sha256.txt, and supplementary_analysis_outputs.zip.
These counts describe the validated eleven-source integration deposited as Zenodo v3. Use the release_manifest.json and checksums_sha256.txt in that record to check downloaded files. This repository contains code, documentation, and the heatmap preview; the multi-gigabyte integrated datasets belong in Zenodo rather than Git. Source-file modification timestamps are not proof of download dates, and recorded audit commits are not automatically an immutable code release.
Licensing note: The dataset record is public and lists CC BY 4.0, but third-party terms may limit redistribution of derived records; see the licence section above and the deposited LICENSES_AND_REUSE.md. A disclaimer by itself does not resolve licence compatibility.
The manuscript is authored by Muhammad Muneeb and David B. Ascher:
Muneeb, M. & Ascher, D. B. A source-aware drug–disease relationship dataset integrating public biomedical resources. Manuscript in preparation (2026). Update with journal/DOI after publication.
For dataset use, cite the specific published version actually used. The validated eleven-source dataset is Zenodo v3:
@dataset{muneeb_drug_disease_v3_2026,
author = {Muneeb, Muhammad},
title = {A source-aware drug--disease relationship dataset integrating public biomedical resources},
year = {2026},
version = {v3},
publisher = {Zenodo},
doi = {10.5281/zenodo.23627171},
url = {https://doi.org/10.5281/zenodo.23627171}
}Using the integrated records also requires citing the relevant upstream data resources, their versions, and applicable licences. Key source publications include:
- AACT: Tasneem et al. (2012), PLoS ONE. DOI: 10.1371/journal.pone.0033677
- ChEMBL: Zdrazil et al. (2024), Nucleic Acids Research. DOI: 10.1093/nar/gkad1004
- CTD: Davis et al., Nucleic Acids Research, 2025 update. DOI: 10.1093/nar/gkae883
- DrugCentral: Avram et al. (2023), Nucleic Acids Research. DOI: 10.1093/nar/gkac1085
- Hetionet: Himmelstein et al. (2017), eLife. DOI: 10.7554/eLife.26726
- MEDI: Wei et al. (2013), Journal of the American Medical Informatics Association. DOI: 10.1136/amiajnl-2012-001431
- Open Targets: Buniello et al. (2025), Nucleic Acids Research. DOI: 10.1093/nar/gkae1128
- PrimeKG: Chandak, Huang & Zitnik (2023), Scientific Data. DOI: 10.1038/s41597-023-01960-3
- repoDB: Brown & Patel (2017), Scientific Data. DOI: 10.1038/sdata.2017.29
- SIDER: Kuhn et al. (2016), Nucleic Acids Research. DOI: 10.1093/nar/gkv1075
- Broad Drug Repurposing Hub: Corsello et al. (2017), Nature Medicine. DOI: 10.1038/nm.4306
Acknowledgements: The creators and maintainers of the included biomedical resources, contributing clinical-trial registries, terminology projects, and open-source processing libraries are acknowledged.
Muhammad Muneeb — research and pipeline lead
Email: m.muneeb@uq.edu.au
GitHub: MuhammadMuneeb007
Repository: drug-disease-mapping
For issues involving incorrect evidence interpretation, provenance, incompatible source licences, or potential integration errors, open a GitHub issue with the source record and script details,