RepMed — Interpretable Drug Repositioning with Bipartite Link Prediction

This repository contains the reproducible implementation associated with the MIWAI 2026 accepted paper:

Drug Repositioning through Interpretable Link Prediction in Drug–Disease Networks

The project formulates drug repositioning as a link prediction task over a bipartite drug–disease network derived from repoDB. The final pipeline combines an interpretable topological predictor, based on Katz paths of length 3 and 5, with a lightweight chemical similarity layer derived from PubChem/DrugBank molecular descriptors. External post-ranking pharmacological assessment is performed using ChEMBL-derived indication and mechanism annotations.

Note: Predicted drug–disease links are computational repositioning hypotheses. They are not biomedical, experimental, or clinical validation.

---

Overview

The pipeline evaluates whether known drug–disease associations can be recovered from the structure of a bipartite biomedical network and then uses the selected model to rank unknown drug–disease pairs as computational repositioning hypotheses.

The main components are:

1. repoDB parsing and cleaning
2. PubChem/DrugBank chemical enrichment
3. Bipartite drug–disease network construction
4. Topological link prediction using Katz L3+L5
5. Chemical similarity support using Morgan fingerprints and Tanimoto similarity
6. Hybrid scoring
7. Final candidate ranking
8. External post-ranking ChEMBL pharmacological assessment
9. Top 1%, low-ranking 1%, and degree-matched ChEMBL control assessment
10. End-to-end reproducible execution through RepMed.py

This repository contains the consolidated final pipeline used for the Katz L3+L5 scoring, chemical support, hybrid candidate ranking, and post-ranking ChEMBL pharmacological assessment. The broader model-selection benchmark reported in the paper, including additional baselines such as Jaccard, Adamic–Adar, node2vec, GCN, and GraphSAGE, is described in the manuscript and may be released separately.

---

Method Summary

The core graph is a bipartite network:

G = (U, V, E)

where:

• U represents drug nodes;
• V represents disease nodes;
• E represents known drug–disease associations.

The main topological predictor is Katz L3+L5:

s_L3+L5(u,v) = beta^3 * A^3(u,v) + beta^5 * A^5(u,v)

where u is a drug node and v is a disease node.

In the paper notation:

• beta is the Katz attenuation parameter;
• alpha is the topological weight in the hybrid score;
• gamma is the chemical-support weight in the hybrid score.

The hybrid score combines topological and chemical evidence:

s_hybrid(u,v) = alpha * s_Katz(u,v) + gamma * s_chem(u,v)

In the final configuration used in the manuscript:

• Katz attenuation: beta = 0.05
• topological weight: alpha = 0.95
• chemical support weight: gamma = 0.05

The chemical component is based on RDKit Morgan fingerprints and Tanimoto similarity. Drugs without valid chemical descriptors receive zero chemical support, without additional penalty.

---

Main Results

The final hybrid model improved held-out link prediction performance over the topology-only Katz L3+L5 baseline.

Method	ROC-AUC	Average Precision	Accuracy	Precision	Recall	F1
Katz L3+L5	0.8581	0.8906	0.8500	0.8941	0.7942	0.8412
Hybrid Katz + Chemical	0.8911	0.9147	0.8699	0.8863	0.8486	0.8670

The hybrid model improved ROC-AUC, average precision, recall, and F1-score while preserving the dominant contribution of the interpretable graph-based score.

---

ChEMBL External Pharmacological Assessment

ChEMBL was used only as an external post-ranking pharmacological assessment layer. It was not used during training, model selection, scoring, or candidate ranking.

To reduce the risk of cherry-picking and degree/popularity artifacts, the final assessment compares:

• Top 1% of the saved final ranking: 2,500 candidate pairs
• Low-ranking 1% control: 2,500 candidate pairs from the lower tail of the same saved ranking
• Degree-matched control: 2,500 candidate pairs selected to assess whether ChEMBL support could be explained by drug or disease popularity

Candidate set	n	Mechanism annotations	Direct/Near Matches	Direct/Related Support
Top 1%	2,500	1,325	98	262 (10.48%)
Low 1% control	2,500	1,385	23	95 (3.80%)
Degree-matched control	2,500	1,544	55	162 (6.48%)

The top 1% showed higher ChEMBL support than both controls. The degree-matched control was higher than the low-ranking control, indicating that degree/popularity contributes to external support. However, the top 1% remained enriched, suggesting that degree alone does not explain the observed pharmacological support.

These results provide external pharmacological context for the ranking, not clinical validation.

---

Repository Structure

A typical project structure is:

project_root/
|
├── project/
│   └── RepMed.py
|
├── data/
│   ├── full.csv
│   ├── PubChem_search_records.json
│   └── RMResults/
|
├── requirements.txt
├── CITATION.cff
├── LICENSE
└── README.md

The repository may also include an additional preproject/ folder. This folder preserves the initial exploratory implementation used during the early development stage of the project. It contains first experimental scripts and intermediate analyses that supported the design of the final pipeline.

The preproject/ materials are kept for transparency and historical reproducibility, but the main reproducible implementation associated with the final manuscript is the consolidated RepMed.py pipeline in project/.

A typical extended structure is:

project_root/
|
├── preproject/
│   └── initial exploratory scripts and early experiments
|
├── project/
│   └── RepMed.py
|
├── data/
│   ├── full.csv
│   ├── PubChem_search_records.json
│   └── RMResults/
|
├── requirements.txt
├── CITATION.cff
├── LICENSE
└── README.md

---

Input Files

The expected input files are:

File	Description
full.csv	repoDB drug–disease association file
PubChem_search_records.json	PubChem/DrugBank chemical structure records

By default, the script expects the following layout:

project_root/
|
├── project/
│   └── RepMed.py
|
└── data/
    ├── full.csv
    ├── PubChem_search_records.json
    └── RMResults/

Data files are not necessarily redistributed with this repository. Users should obtain the original data from the appropriate data providers and respect the corresponding licenses and terms of use.

---

Main Script

The consolidated pipeline is implemented in:

project/RepMed.py

It performs the full workflow from input files to final outputs.

Default execution from inside the project/ folder:

python RepMed.py

Parameterized execution:

python RepMed.py \
  --repodb "../data/full.csv" \
  --pubchem "../data/PubChem_search_records.json" \
  --out "../data/RMResults" \
  --reuse-descriptors \
  --reuse-similarity \
  --run-chembl

On Windows PowerShell:

python .\RepMed.py `
  --repodb "..\data\full.csv" `
  --pubchem "..\data\PubChem_search_records.json" `
  --out "..\data\RMResults" `
  --reuse-descriptors `
  --reuse-similarity `
  --run-chembl

Useful fast/debug run:

python RepMed.py \
  --max-api-drugs 100 \
  --candidate-save-top-n 50000

Paper-oriented run:

python RepMed.py \
  --repodb "../data/full.csv" \
  --pubchem "../data/PubChem_search_records.json" \
  --out "../data/RMResults" \
  --alpha-katz 0.95 \
  --gamma-chemical 0.05 \
  --katz-beta 0.05 \
  --candidate-save-top-n 250000 \
  --chembl-set-size 2500 \
  --reuse-descriptors \
  --reuse-similarity \
  --run-chembl

For compatibility with earlier local runs, the script may also accept --beta-chemical as an alias for --gamma-chemical. In the paper, the chemical weight is denoted by gamma, while beta is reserved for the Katz attenuation parameter.

---

Main Parameters

Parameter	Default	Description
--katz-beta	0.05	Katz attenuation factor
--alpha-katz	0.95	Hybrid weight for normalized Katz score
--gamma-chemical	0.05	Hybrid weight for normalized chemical support
--test-size	0.20	Fraction of positive edges held out for test evaluation
--seed	42	Random seed
--min-chemical-similarity	0.20	Minimum Tanimoto similarity for chemical support
--chemical-topk	20	Number of chemical neighbors retained per drug
--morgan-radius	2	Morgan fingerprint radius
--morgan-bits	2048	Morgan fingerprint bit vector size
--candidate-save-top-n	250000	Number of final ranking rows saved
--chembl-set-size	2500	Number of candidates per ChEMBL comparison set
--run-chembl	disabled	Enables ChEMBL post-ranking pharmacological assessment

---

Main Outputs

The pipeline generates several outputs, including:

Output	Description
RM1_katz_l3_l5_test_scores.csv	Katz L3+L5 test predictions
RM2_hybrid_katz_chemical_test_scores.csv	Hybrid model test predictions
RM3_final_results_table.csv	Main performance comparison
RM4_final_hybrid_candidate_ranking.csv	Final ranked candidate drug–disease pairs
RM5_paper_table_combined_candidate_examples.csv	Representative candidate examples
RM6_chembl_assessment_candidates.csv	ChEMBL assessment for representative candidates
RM7_chembl_assessment_summary.md	ChEMBL assessment summary for representative candidates
chembl_top1pct_low1pct_degree_matched_assessment.csv	Comparative ChEMBL assessment for top, low-ranking, and degree-matched sets
chembl_top1pct_low1pct_degree_matched_summary.csv	Aggregated ChEMBL comparison
chembl_top1pct_low1pct_degree_matched_summary.json	JSON summary of ChEMBL comparison
RM8_RepMedHybrid_summary.json	Full reproducibility summary
RM9_RepMedHybrid_report.md	Markdown report of the run
README_RepMed.md	Additional execution notes generated by the pipeline

Some output filenames preserve the word validation for compatibility with earlier versions of the pipeline. Conceptually, ChEMBL is used here only as post-ranking external pharmacological assessment/context and not as clinical validation.

---

Reproducibility Notes

The core pipeline reconstructs:

• 2,326 drug nodes
• 1,451 disease nodes
• 10,469 drug–disease edges
• 3,364,557 unknown candidate pairs
• 250,000 saved ranking rows

The expanded ChEMBL assessment compares three 2,500-candidate sets:

• top 1% of the saved final ranking
• low-ranking 1% control
• degree-matched control

The ChEMBL assessment depends on external API queries, which may increase runtime and may produce minor differences if external records change over time.

The full candidate-ranking step may require several GB of RAM because it scores all unobserved drug–disease pairs before saving the selected ranking rows.

---

Dependencies

Main Python dependencies include:

• pandas
• numpy
• scikit-learn
• scipy
• requests
• rdkit

Install dependencies with:

pip install -r requirements.txt

If RDKit installation through pip is not available in your environment, use Conda:

conda install -c conda-forge rdkit

A minimal requirements.txt may include:

numpy
pandas
scipy
scikit-learn
requests
rdkit
tabulate

The tabulate package is recommended because the script uses pandas Markdown table output in generated reports.

---

Data Sources

This project uses data derived from:

• repoDB — drug repositioning validation resource
• DrugBank Live Substance data via PubChem
• ChEMBL — external post-ranking pharmacological assessment only

ChEMBL is not used for training, scoring, model selection, or candidate ranking.

Users should consult the original data providers for licensing, citation, access, and redistribution conditions.

---

Scientific Scope and Limitations

This repository supports early-stage computational candidate prioritization.

Important limitations:

• The task is missing-edge recovery and candidate ranking within a known drug–disease graph.
• The evaluation is not cold-start generalization to unseen drugs or diseases.
• Sampled unobserved drug–disease pairs are operational negatives, not confirmed biomedical negatives.
• ChEMBL is used only after ranking as external pharmacological context.
• ChEMBL support does not constitute clinical validation.
• Morgan fingerprints and Tanimoto similarity do not fully capture pharmacodynamics, pharmacokinetics, targets, or disease biology.
• Hub drugs and diseases may influence path-based ranking.

Predicted candidates should be interpreted as computational hypotheses requiring independent biomedical review, literature assessment, and, where appropriate, experimental or clinical validation.

---

Citation

Citation information for the associated MIWAI 2026 paper will be added after publication.

Preliminary citation format:

Ceulin, W. L. C., Costa Pereira, J., Camacho, R., Dutra, I.
Drug Repositioning through Interpretable Link Prediction in Drug–Disease Networks.
Accepted for oral presentation at MIWAI 2026.

For repoDB, cite:

Brown, A.S., Patel, C.J. repoDB: A New Standard for Drug Repositioning Validation.
Scientific Data 4, 170029 (2017).

---

License

This project is licensed under the PolyForm Noncommercial License 1.0.0.

Noncommercial use, including research, education, personal study, and public-interest purposes, is permitted under the terms of the license. Commercial use is not permitted without separate authorization from the copyright holder.

See the LICENSE file for the full license terms.

---

Status

This repository is associated with an accepted MIWAI 2026 paper.

The code and outputs are intended for reproducibility, inspection, and future extension of the proposed drug repositioning workflow.