MedChem Desk - a screening hit list in, the med-chem triage out

Paste a screening hit list - identifiers, SMILES and assay values, as CSV, TSV, a pipe table or a bare .smi file - and get the medicinal-chemistry triage a project team would write, in four lanes over one sitting: the series read with scaffold clusters and the SAR each one actually supports; a progress, hold or kill call on every compound with the one liability that decided it; the ADMET risk profile and the tiered assay panel to order next, every gate falsifiable; and the make-list for the next round, each proposal from a real parent and testing exactly one hypothesis. A free in-browser scanner reads the table first and spends no credits: it sniffs the delimiter, honours quoted fields, parses every SMILES into a real molecular graph with implicit hydrogens and perceived rings, strips salts to the parent, and computes molecular formula, weight, heavy-atom count, ring and aromatic-ring counts, rotatable bonds, hydrogen-bond donors and acceptors, Ertl topological polar surface area and the fraction of sp3 carbon - all derived from the structure, never predicted. It matches 43 structural alerts across five liability families by real subgraph isomorphism rather than string search, and grades each one against the programme you declared: a covalent campaign's acrylamide is an intended warhead, an acyl halide is not excused by anything, a genotoxicity alert is never downgraded by potency, and a molecular weight of 502 is not graded like 780. No logP is predicted anywhere - the lipophilicity limbs of Lipinski and Egan are reported unassessable unless your paste carried a measured logD. Assay values are read properly: a censored bound keeps its operator so >10000 nM stays the least active compound in the set, n/a stays unmeasured, inactive stays measured-and-inactive, 1,250 is not 1.250, 0.5 uM is 500 nM, and a cell holding two measurements is refused rather than guessed. Reported masses are checked against the structure, ligand efficiency is computed only from an uncensored potency, and a list too large to send whole is sampled by a golden-ratio low-discrepancy draw so no plate layout or scaffold ordering in your file can bias it. The model must reconcile every scan flag one for one, and the page audits its reply against the scan and against itself. Derived from @k-dense-ai/medchem, @k-dense-ai/rdkit, @k-dense-ai/datamol, @k-dense-ai/deepchem and @k-dense-ai/pytdc; not affiliated with those skills' authors, with RDKit, Datamol, DeepChem, Therapeutics Data Commons, or any organisation named in the output. Not medical, regulatory or safety advice.

Every public app is built from a security-scanned skill and must pass a clean scan — skill and frontend — before it can be listed. Have a skill of your own? Turn it into an app — or read the step-by-step walkthrough.