Language: English | Deutsch

Competitive intelligence platform for diabetes & obesity drug development — from raw clinical trial data to ML-powered success predictions.

Track 43 drugs across 32,800+ clinical trials, predict phase transition probabilities with XGBoost, and explore the competitive landscape through a 9-page interactive dashboard.

Live Dashboard

Commercial pharma intelligence platforms (Cortellis, GlobalData, Evaluate) cost $50,000+/year and deliver static reports. This project builds the same analytical capability from public data sources for under $60/year total infrastructure cost.

What it delivers:

• Which drugs are competing in which indications — and where the white space is
• How likely a drug is to advance from Phase 2 to Phase 3 (or from Phase 3 to approval)
• Which sponsors have the best track record, and which trials are stalling
• Patent expiry timelines, safety signals, and pricing data across US and UK markets

Nine specialized pages, each focused on a different analytical question:

Built with Streamlit and Plotly (15+ chart types: heatmaps, treemaps, scatter, donut, stacked area, line, bar).

Four binary classification models predict the probability that a drug advances:

• Drug-level temporal GroupKFold — No drug appears in both train and test. Earlier drugs train, later drugs test. This prevents the most common source of leakage in pharma ML.
• 60 point-in-time safe features — Every feature only uses information available before the prediction point. No post-market data, no future-looking variables. Down from 105 in v1 after removing unsafe features.
• CV-based model selection — Models are selected by cross-validation performance, not test-set AUC. This avoids selection bias. Final models are retrained on all data.
• Four model types per transition — Logistic Regression baseline, XGBoost aggressive, XGBoost conservative, Ridge meta-learner. XGBoost A selected for all 4 transitions.

Eight pipeline phases, each building on the previous:

0. Schema          deploy_schema.py           → 13 tables, 15 indexes
       |
1. Ingestion       1_ingestion/               → ClinicalTrials.gov, ChEMBL, OpenFDA
       |
2. Enrichment      2_enrichment/              → RxNorm codes, trial design features
       |
3. Linking         3_linking/                 → Company resolution, drug-indication mapping
       |
4. Patent & LOE    4_patent_loe/              → Orange Book, Purple Book, LOE calendar
       |
5. Safety          5_safety/                  → FDA FAERS adverse events
       |
6. Pricing         6_pricing/                 → CMS Medicare Part D, NHS OpenPrescribing
       |
7. ML              pharma-pipeline-ml/        → Feature engineering → Model training
       |
8. Dashboard       pharma-pipeline-dashboard/ → 9-page Streamlit app

pharma-pipeline-intelligence/
├── 1_ingestion/                  # Phase 1: Load raw data from APIs
├── 2_enrichment/                 # Phase 2: RxNorm, trial design features
├── 3_linking/                    # Phase 3: Entity resolution & linking
├── 4_patent_loe/                 # Phase 4: Patent & exclusivity data
├── 5_safety/                     # Phase 5: FDA adverse events
├── 6_pricing/                    # Phase 6: US & UK drug pricing
├── pharma-pipeline-ml/           # Phase 7: Feature engineering & model training
│   ├── compute_features_v2.py
│   ├── train_models_v2.py
│   ├── run_diagnostics.py
│   └── config.py
├── pharma-pipeline-dashboard/    # Phase 8: Streamlit dashboard
│   ├── app.py
│   ├── pages/                    # 9 dashboard pages
│   └── utils/                    # DB queries, charts, helpers
├── reports/                      # Phase reports & documentation
├── deploy_schema.py              # Phase 0: Create database schema
├── db_config.py                  # Shared database configuration
└── schema_stats.sql              # Database overview

• Python 3.11+
• ODBC Driver 18 for SQL Server
• Azure SQL Database (or local SQL Server)

git clone https://github.com/leelesemann-sys/pharma-pipeline-intelligence.git
cd pharma-pipeline-intelligence
python -m venv .venv
.venv\Scripts\activate
pip install -r pharma-pipeline-dashboard/requirements.txt

Copy .env.example to .env and set your Azure SQL connection string. The dashboard reads credentials from Streamlit secrets (.streamlit/secrets.toml).

streamlit run pharma-pipeline-dashboard/app.py

Execute scripts in phase order:

python deploy_schema.py
python 1_ingestion/ingest_clinicaltrials.py
python 1_ingestion/ingest_chembl_drugs.py
# ... continue through phases

This project's methodology, feature engineering, and evaluation strategy are grounded in published pharma ML research:

This project is for portfolio and educational purposes.

Built with Azure SQL, XGBoost, Streamlit, and 8 public data APIs.