Crucible

Your mission --

--

How is it going right now

How well it holds on

nothing built

●No part of it can grab on permanently.

You versus the real medicine

you --drug --

--

What this does to the disease right now

0%of the machines jammed

Nothing built yet.

Worked out from your grip score using ordinary binding chemistry, assuming a drug level of 10 micromolar. It describes the protein, not a person, and it is only as good as the grip estimate.

Your steps do them in order

What just happened

Pick a piece below to make your first move.

Score vs the real pocket

Predicted affinity, T790M -- kcal/mol - more negative binds harder --

Ligand efficiency--per heavy atom

Selectivity--vs wild type

wild type -- best run -- record --

Property gates real rules, real ceilings

Breaches0 / 3

Your handPieces you can add pick one, then pick where

The ladder five rungs to glass

1Fast score

Every move you commit. Monte Carlo pose and torsion search against both variants, scored with AutoDock Vina's published empirical function. Free and unlimited.

2Deep search

The same function, far more sampling: 14 restarts of 900 steps instead of 4 of 220. Slow, so it is a currency rather than a move.

tokens left 2

3Robustness

Re-runs under four perturbed seeds and perturbed weights to see whether the gap survives. This is a stability check on one scoring function, not an ensemble of independent models.

4Review

A medicinal-chemistry read on the numbers, including the rejections. In production this rung is a named human chemist whose reasoning is published.

5Glass

Made and assayed, with the designer named. Locked in the prototype, because it cannot be faked.

What rung 5 actually requires:

A synthesis partner with a costed route to the exact structure
An orthogonal assay, so the readout is not one number from one method
A chemist signing their name to the order
Publication of the failures alongside the hits

Your tools

Leaderboard this disease

Your score

0

AI lab partner ask anything

It can see your molecule and all its numbers. It helps with the puzzle only. Nothing it says is medical advice.

Contacts real residues

Move log

Notebook reproducible by design

Every run carries its seed, its exact move sequence and every score it produced. Anyone can replay it bit for bit without having been here. Crowd output that cannot be reproduced is not evidence.

What is real here

Experimental: The pocket. Every wild-type atom is a measured coordinate from Protein Data Bank entry 4ZAU, the EGFR kinase domain at 2.80 angstrom resolution, solved with osimertinib bound. 264 atoms across 58 residues, the whole canonical ATP site.
Real: The scoring function. AutoDock Vina's empirical function with its published weights, over surface distances and X-Score radii, searched over rigid-body placement and rotatable torsions.
Real: The property gates. Crippen-style cLogP, Ertl polar surface area, Lipinski and Veber ceilings. The structural alerts are genuine rejection criteria.
Modelled: One residue. The T790M mutant is built by replacing the Thr790 side chain with methionine on the identical backbone, placed by rotamer search for lowest clash. The two receptors differ by nothing else, so the selectivity gap is attributable to that one change.
Not modelled: Covalency. Real osimertinib forms a covalent bond to Cys797, and an empirical scoring function cannot represent that. You are playing the non-covalent half of a problem the real drug solved with both halves.
Not a cure: A Vina-class score correlates with measured affinity at roughly r = 0.4 to 0.6. This ranks moves against each other. It does not predict what happens in a person, and nothing here is medical advice.