OTRec

OTRec (Open Targets Recommender) ranks druggable genes for a disease using a two-tower recommender trained on Open Targets data. It predicts disease–target prioritization rather than compound–protein binding or drug–target interaction. This repository contains the code, released predictions, and the interactive demo for the OTRec manuscript.

Try OTRec

• Web demo: explore disease → target and target → disease rankings in the OTRec Hugging Face app.
• Released predictions: Outputs/S1-DL_novel_predictions.csv (novel candidates) and Outputs/S2-DL_novel+known_candidates.csv (novel + known).
• Simple local query, no model install needed:

pip install pandas
python examples/query_predictions.py --disease obesity
python examples/query_predictions.py --target GIPR

See examples/README.md.

Released paper data

All manuscript data files live in Outputs/:

File	Contents
S1-DL_novel_predictions.csv	214,968 novel disease–target predictions (score ≥ 0.65; 4,347 diseases)
S2-DL_novel+known_candidates.csv	S1 plus 67,532 known clinical associations (282,500 rows)
S3-feature_ablation.csv	Per-feature-group ablation (temporal setting)
S5-ottree_hyperparameter_sweep.csv	CatBoost baseline hyperparameter sweep
S7-temporal_predictions.parquet	Per-pair temporal predictions behind Tables 4, 5, and S8
S8-shortlist_stratification.csv	Shortlist metrics by therapeutic area / known targets / annotation volume
S9-gene_family_overlap.csv	Per-family gene-family overlap statistics (25.12 model)
Table 1/2/3 ... .csv	Main benchmark tables
InterFeat_reranked_candidates/	InterFeat annotations for selected candidates
CV_*, temporal_* directories	Per-fold and per-run metrics behind Tables 1–3

Versions: the manuscript benchmarks and the S1/S2 prediction files are frozen to the paper analysis (Open Targets 22.02 and 25.06). The interactive web app has since been refreshed to Open Targets 25.12 (analysis/retrain_2512/), so live app scores can differ from S1/S2.

Citation

Ofer D., Linial M. OTRec: A Deep Learning Recommender for Druggable Disease–Target Prioritization. bioRxiv 2025.12.21.695803. doi: 10.64898/2025.12.21.695803

---

Reproducing the analysis

Repository structure

OTRec/
├── 0-OT-PreProcess_Recc.ipynb  # Data preprocessing and feature engineering
├── 1-Train-DL-Retriever.ipynb  # Trains the two-tower model; writes S1/S2
├── 2-Temporal-Eval.ipynb       # Temporal split validation (2022 vs 2025)
├── dl_model_def.py             # Keras two-tower model definition
├── utils.py                    # Data loading and metric helpers
├── examples/                   # Query the released predictions with pandas
├── gradio/                     # Standalone interactive web demo
├── analysis/                   # Supplementary and rebuttal analyses
│   ├── scripts/                # Reproduction scripts (S3, S5, S7, S8, ...)
│   └── retrain_2512/           # Open Targets 25.12 deployment refresh
├── baselines/                  # Baseline reproduction (run_baselines.py, run_temporal_repeated.py)
└── Outputs/                    # Released predictions, tables, figures, metrics

Data prerequisites

The repository does not contain the raw Open Targets Platform (OTP) data. To retrain or re-evaluate, download the parquet exports from Open Targets Data Downloads:

• Core: associationByOverallDirect (rename folder to association_overall_direct), associationByDatatypeDirect (→ association_by_datatype_direct), target, disease, knownDrugsAggregated (→ known_drug)
• Features: targetPrioritisation (→ target_prioritisation), targetEssentiality (→ target_essentiality), diseasePhenotypes (→ disease_phenotype), optional mousePhenotypes (→ mouse_phenotype)
• Temporal validation additionally needs the Release 22.02 core datasets.

Expected layout (siblings of this repository; set Config.DATA_DIR in the notebooks to change it):

project_root/
├── OTRec/                 # this repository
└── data/
    ├── opentargets/       # current release (paper: 25.06)
    │   ├── association_overall_direct/ ... mouse_phenotype/
    └── historical_ot/
        └── 22_02/         # historical release for the temporal split

Environment

Python 3.10+:

pip install tensorflow pandas scikit-learn catboost gradio pyarrow fastparquet

Run order

1. 0-OT-PreProcess_Recc.ipynb — filters raw data, applies the druggable-genome constraint, builds the training frame, trains the CatBoost baseline.
2. 1-Train-DL-Retriever.ipynb — trains the two-tower model; writes weights and the S1/S2 prediction files.
3. 2-Temporal-Eval.ipynb — temporal validation: 2022 model vs 2025 clinical outcomes.

Baselines: baselines/run_baselines.py (CV) and baselines/run_temporal_repeated.py (temporal, five runs). Supplementary analyses: see analysis/README.md.

Interactive demo (local)

cd gradio
python app.py

The app downloads model weights at runtime; see gradio/README.md for details and caveats.

Key results

• Temporal validation (predict 2025 clinical-trial entries from 2022 data; mean of five runs): OTRec ROC-AUC 0.872 / PR-AUC 0.288 vs Open Targets score 0.559 / 0.082.
• Target-disjoint 5×5 CV (targets unseen in training): OTRec ROC-AUC 0.950 / PR-AUC 0.844.

License

MIT. Open Targets Platform data are CC0 1.0; other evidence sources retain their own licenses.