Binding potency prediction
Which molecules are worth making?
Testing a compound against a target costs weeks and thousands of dollars. DOXLAS reads a structure and estimates how tightly it will bind — before anyone runs the assay. Paste a SMILES string below, or start from a known drug.
—
Model performance
held-out test set · R² and RMSE on pIC50Redox Atlas
target curation in progressThe two targets above are validation work — well-studied proteins with dense public data, chosen to prove the pipeline behaves. The atlas DOXLAS is named for is the redox set below. Curation is underway; no models are trained on these yet, and nothing is claimed for them.
How this demo works
method & limitsStructures and descriptors are computed live. SMILES parsing, 2D depiction, and physicochemical properties run in your browser through RDKit compiled to WebAssembly. Nothing is sent to a server.
The potency estimate is a surrogate, not the production model. It is a similarity-weighted k-nearest-neighbour readout over a curated set of reference actives, using Morgan fingerprints (ECFP6 — radius 3, folded to 2048 bits) and Tanimoto similarity. The neighbours it used are listed so you can judge the estimate yourself. The full random-forest ensemble reported above runs offline against the complete ChEMBL extract.
Read R² as variance explained, not accuracy. An R² of 0.712 means the model accounts for roughly 71% of the variance in pIC50 across the held-out set — it does not mean 71% of predictions are correct. RMSE is the more directly interpretable figure: an error of ~0.68 log units is about a five-fold spread in predicted potency.
Confidence tracks similarity. When the closest reference compound is below ~0.35 Tanimoto, the molecule sits outside the applicability domain and the estimate should be treated as uninformative. The readout says so explicitly rather than returning a confident-looking number.