A machine-learning proof of concept for drug discovery, applied to inhibition of the EGFR protein, plus a web application that makes the models usable in a browser.
Screening molecules in a laboratory is slow and expensive. This project builds a chain of models that narrows the field first, so the bench only sees the candidates worth testing.
| Folder | What it is |
|---|---|
POC_Molecular_Design/ |
Eight Jupyter notebooks, a Streamlit dashboard (app.py), and the trained models |
POC_Molecular_Design_Web/ |
Nucleus, a Next.js site that runs the chemistry in the browser through a Rust/WebAssembly engine |
Activity measurements (IC50) for target CHEMBL203 from the public ChEMBL database, pulled
automatically through the official chembl_webresource_client package — no API key. Notebook 1 fetches and
cleans the dataset that every other notebook reloads; a manual CSV export is supported as a fallback when
the network blocks the package.
| Notebook | Task | The question a researcher asks |
|---|---|---|
| 01 Acquisition & preparation | Collection and cleaning | How do raw measurements become usable data? |
| 02 Descriptor extraction | RDKit descriptors and fingerprints | How do you hand a molecule to a model? |
| 03 Exploratory analysis | EDA and chemical space | What separates an active molecule from an inactive one? |
| 04 Active / inactive | Classification | Will this molecule act on EGFR? |
| 05 pIC50 | Regression | How potent will it be? |
| 06 Chemical families | Clustering | Which families deserve to be explored first? |
| 07 Chemical space | PCA and t-SNE | Where do the active molecules sit on the map? |
| 08 Bonus | Recommendation, scaffolds, activity cliffs | What looks like my molecule, and why does a small change flip the result? |
Every notebook follows the same shape: business need, loading, processing, a simple baseline before the
advanced model, evaluation, interpretation for the researcher, joblib export, and a conclusion that states
its limits. Experiments use random_state = 42.
cd POC_Molecular_Design
pip install -r requirements.txt
# run notebooks 1, 2 and 4 first: they produce the dataset and the classification model
streamlit run app.pyThe dashboard takes a molecule as SMILES and returns its predicted activity, its estimated potency and its known analogues.
cd POC_Molecular_Design_Web
npm install
npm run dev # http://localhost:3000Molecular properties and similarity are computed by a cheminformatics engine written in Rust and compiled
to WebAssembly, running in the browser. The compiled artifact is public/molcore.wasm, lazy-loaded by
lib/wasm/molcore.ts. If WebAssembly fails to load, the site falls back to the JavaScript implementation, so
the demo never breaks. The Analytics page benchmarks JavaScript against Rust live in the browser; the numbers
are measured on the spot, not hard-coded.
npm run build:wasm # rebuild the engine after changing the Rust code
npm run verify:wasm # load the wasm in Node and check benzene, ethanol, TanimotoThe crate is compiled to raw WebAssembly, without wasm-bindgen and without a C compiler, because the
development machine had no working C linker. See rust/README.md.
- The SMILES parser and the fingerprint in the browser engine are simplified models, not RDKit. The site is a decision aid and a demonstration, not a scientific verdict.
- IC50 values carry experimental uncertainty, which bounds the accuracy any model can reach.
- A model trained on EGFR says nothing about another target, and nothing about toxicity. This is a first-pass filter.
- No external AI API is used: every model is trained locally with scikit-learn, XGBoost and RDKit.
- A generative model of molecules (GraphVAE, MolGAN, or a Transformer over SMILES) was deliberately left out of this prototype.
French versions of the documentation live in POC_Molecular_Design/README.md and
POC_Molecular_Design_Web/README.md.