Language: English | Español
Prediction of drug-protein interactions using Machine Learning on real BindingDB data.
Personal project combining knowledge from computer science, business administration, and biochemistry.
- Description
- Business Impact and Operational Efficiency
- Repository Structure
- Interactive Application
- Methodology
- Results
- Interpretability
- Molecular Visualizations
- Requirements
- How to Run
- Input CSV Format
- Limitations
- Author
- License
This project implements a complete virtual screening pipeline to predict whether a small molecule (drug candidate) is likely to bind to a target protein. The model enables prioritization of compounds from libraries of thousands of molecules before performing costly experimental assays, significantly reducing both the time and cost of the drug discovery process.
Dataset: BindingDB — public database with over 2.8M affinity measurements between molecules and proteins.
As a Business Administration student, this project was designed not only as a technical challenge but as an asset optimization and risk reduction solution in the pharmaceutical R&D value chain.
- The "Funnel Filter": In drug discovery, physical High-Throughput Screening is extremely costly. This pipeline acts as a critical filter, processing a virtual library of 100,000 compounds to prioritize only the 500 candidates with the highest probability of success.
- Cost Savings (OPEX): By identifying failures early (Fail Fast), it avoids investing millions of euros in in vitro laboratory assays destined to fail, optimizing the research budget.
- Time-to-Market Reduction: Accelerating the initial screening stage allows promising molecules to reach clinical phases sooner, maximizing the net present value (NPV) of intellectual property.
DrugProtein_ML/
├── Drug_Discovery_Training.ipynb # Complete training pipeline
├── Drug_Discovery_Inference.ipynb # Prediction on new library
├── app.py # Interactive virtual screening application
├── requeriments.txt
├── xgboost.pkl # Trained model ready to use
├── examples/
│ └── mi_libreria.csv # Example CSV for testing inference
└── images/ # Plots generated during training
The project includes an interactive web application built with Streamlit that allows direct use of the model on new molecules or complete chemical libraries.
The application is deployed and can be tested directly here:
👉 https://molecular-predictive-discovery.streamlit.app/
The application allows you to:
- Analyze an individual molecule from its SMILES string
- Obtain binding probability and Binder / Non-Binder classification
- Visualize the molecular structure in 2D and 3D
- Calculate physicochemical properties and drug-likeness
- Evaluate the multivariate molecular profile
- Perform automated screening of libraries in CSV format
- Rank candidates and interactively analyze the best hits
This reproduces the real workflow of computational virtual screening prior to experimental screening.
pip install -r requirements.txt
streamlit run app.py- Loading 500k rows from BindingDB with Ki, IC50, and Kd measurements
- Chemical cleaning with RDKit: desalting, SMILES canonicalization, sanitization
- Deduplication by canonical SMILES
- Binary labeling: binder if affinity < 1000 nM (1 µM), non-binder otherwise
- 1:1 balancing via undersampling → 109,028 final molecules
| Feature | Description | Dimensions |
|---|---|---|
| Morgan FP (ECFP4) | Circular fingerprint radius 2, gold standard in QSAR | 2048 bits |
| Physicochemical Descriptors | MolWt, LogP, HBA, HBD, TPSA, RotBonds, Aromatic, QED, HeavyAtoms, FracCSP3 | 10 |
The dataset is split by Murcko scaffold: train and test sets contain entirely distinct chemical families, simulating the real-world scenario of predicting on novel compounds.
Split audit (4 checks):
- Scaffold overlap: 0% (< 1% acceptable)
- Exact SMILES in common: 0
- Tanimoto NN >= 0.85: 0% (< 5% acceptable)
- Class drift train/test: 0.6% (< 5% acceptable)
Industry standard in pharma for QSAR (Bender et al., 2022). Gradient boosting on sparse Morgan FP features with L1/L2 regularization and early stopping.
768-dimensional embeddings from a transformer pretrained on 77M molecules (ZINC), used as input for XGBoost. The architecture avoids overfitting from an additional MLP.
Combination of both models with fixed equal weights. Weights are not optimized on the test set to avoid data leakage.
| Model | Train ROC-AUC | Test ROC-AUC | Gap | PR-AUC | F1 | MCC | Brier |
|---|---|---|---|---|---|---|---|
| XGBoost | 0.9333 | 0.8979 | 0.0354 | 0.8993 | 0.8199 | 0.6470 | 0.1291 |
| ChemBERTa+XGB | 0.8748 | 0.8357 | 0.0391 | 0.8283 | 0.7627 | 0.5238 | 0.1651 |
| Ensemble | 0.9211 | 0.8832 | 0.0379 | 0.8830 | 0.8071 | 0.6174 | 0.1404 |
XGBoost outperforms the ensemble because ChemBERTa was pretrained on ZINC, which has a different chemical distribution than BindingDB, limiting the quality of its embeddings in this domain.
All train/test gaps are below 0.05, confirming the absence of significant overfitting.
- Scaffold split with 4 integrity checks
- Early stopping in all XGBoost models
- Ensemble with fixed weights (no optimization on test set)
- Probability calibration reported (Brier score)
- Train/test gap reported for each model
- Feature interpretability (bitInfo + substructure visualization)
Morgan Fingerprints dominate the model with 98.8% of total importance. The only physicochemical descriptor in the top 20 is NumHeavyAtoms, which makes biological sense: larger molecules have more surface contact area with the protein.
Using RDKit's bitInfo, each Morgan bit can be mapped to the chemical substructure that activates it, visualized directly on the molecule.
FPs have very weak affinities (> 10,000 nM) and are structurally similar to known kinase inhibitors. FNs are ultra-potent (< 1 nM), structurally atypical, and underrepresented in the training data.
2D projection of Morgan FPs showing how binders and errors are distributed across structural space.
| Library | Recommended Version |
|---|---|
| Python | >= 3.9 |
| RDKit | >= 2023.x |
| XGBoost | >= 1.7 |
| scikit-learn | >= 1.2 |
| transformers | >= 4.x (only for ChemBERTa) |
| PyTorch | >= 2.0 (only for ChemBERTa) |
The model is already trained in xgboost.pkl. You only need your compound library.
- Prepare a CSV with a
smilescolumn (and optionally alabelcolumn with 0/1 values) - Open
Drug_Discovery_Inference.ipynb - Edit the configuration cell with the path to your CSV
- Run all cells
Output: binding probability, BINDER/NON-BINDER classification, molecular visualizations, and a report.
- Download
BindingDB_All.tsvfrom bindingdb.org (~6 GB) - Place it in the project root alongside the notebook
- Open
Drug_Discovery_Training.ipynb - Run all cells (~45 min with GPU, ~90 min without GPU)
Recommended: run on Google Colab with GPU enabled to reduce ChemBERTa training time.
smiles,name,label
CC(=O)Oc1ccccc1C(=O)O,Aspirin,0
Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1Nc1nccc(-c2cccnc2)n1,Imatinib,1| Column | Required | Description |
|---|---|---|
smiles |
Yes | SMILES string of the molecule |
name |
No | Compound name |
label |
No | Ground truth label (0/1) — enables performance metrics |
- The model predicts binding in general, not against a specific target protein. It is useful as a structural pre-filter before experimental screening.
- Trained with the publication bias of BindingDB (overrepresentation of actives and certain targets such as kinases).
- The Enrichment Factor is limited by the 1:1 balancing of the training dataset.
Personal project developed as a demonstration of the intersection between computer science, business administration, and computational biochemistry.
- Stack: Python, RDKit, XGBoost, HuggingFace Transformers, scikit-learn
- Data: BindingDB (public domain)
- Environment: Google Colab / Jupyter Notebook
MIT License — free for academic and personal use.
María Campos Carneros, 2026