DarkSite: covalent handles where structure prediction goes dark

A biologist hands you a target that matters (a transcription factor, a scaffold, an E3 ligase) and it has no pocket. Structure prediction gives a confident domain and a long, low-confidence tail, and the tail is where the biology is. DarkSite asks a different question: which cysteine on this protein is reactive and reachable, and so could anchor a covalent drug?

01What it returns

Give it a target and it returns a dossier: every cysteine on the protein, ranked, each with the evidence behind its rank.

• Evidence. Whether a reactive probe has ever been seen to label the site, and how consistently.
• Prediction. For sites never observed, a calibrated model score and the features that drive it.
• Structure. Fold confidence at the residue, how exposed the thiol is, and whether a warhead-sized fragment can be placed there.
• Chemistry. Quantum-chemical descriptors of the local pocket, computed by a teammate's pipeline.
• An experiment card. How to confirm the call at the bench, and which result would prove it wrong.

02Scored on sites it could not have seen

A model that ranks cysteines can look good just by memorizing the sites everyone already knows. So the evaluation is a time-split: train on what was known up to a cutoff date, then score the model on what was discovered afterwards. Next to it runs a deliberately simple baseline that knows only whether a site is easy to detect. The model has to beat that baseline before any of its scores mean anything.

03My part

I led the data and the pitch. I built the labels and the leak-proof benchmark split, designed the time-split and negative sets, and kept the team's evaluation honest. The rest of the team covered quantum chemistry, DFT, and machine learning.

04Built with

Python, ESM-2 protein embeddings, AlphaFold-derived structure features, Boltz-2 co-folding, xTB quantum chemistry, CatBoost and XGBoost.

05Status

Pitched at the Self-Accelerating AI for Drug Development hackathon in San Francisco, October 16–22, 2026. The code and data stay private: the training labels come from unpublished research data, so this page describes the method only.