Medici OS takes a small-molecule project from target biology to hit-to-lead — bringing together the literature, protein structures, chemical libraries, docking and molecular dynamics in one workspace, with every result traced to its source.
From target biology and structure collection to screening, benchmarking and hit-to-lead — each with its own evidence.
Protein structures, chemical libraries, bioactivity, target biology, literature and trials — each with a visible status.
The platform
From the first question about a target to a defensible, documented conclusion — every step grounded in the literature, the structures and the data you bring in.
Ask a question in plain language. Medici turns it into focused queries over PubMed and returns an answer in which every statement is cited to a specific article.
A 14-layer pipeline from target biology to hit-to-lead — each layer with its own workbench, evidence, data sources and tools.
Collect experimental protein structures for a target, with method, resolution, ligands and the source record behind every entry.
Agents read the evidence in your workspace and report only what is present in a source record — each finding keeps its citation.
Ligand preparation, docking, molecular dynamics and structure prediction, run on the structures and compounds in your workflow.
Every record is labelled predicted, measured, curated or hypothesis, with its source and date. Computed results are never presented as experiments.
Workflows
A workflow is your project — for example one protein target. Each pipeline layer has its own workbench, where you collect evidence, search the literature, bind data sources and run tools. Findings, notes and results stay with the layer they belong to, and a layer with no data is clearly marked as blocked.
target · disease · organism · structure of interest
The discovery pipeline
Structure-based small-molecule discovery, laid out layer by layer. Each layer states what it needs, what was found and what is still missing — so gaps are visible instead of hidden.
The therapeutic hypothesis and the evidence behind it.
Is the target expressed, and can its activity be measured?
Experimental structures, with quality and ligand context.
Pockets, mechanism and the site hypothesis.
Which compounds are in scope, and where they came from.
Protonation, tautomers, stereochemistry and identity.
Conformers and ligand parameters for docking.
Cleaning, protonating and defining the docking box.
Docking and rescoring of the prepared library.
Known actives and decoys; how uncertain the scores are.
Experimental confirmation of computational hits.
Off-target activity and cellular mechanism.
Potency, ADMET and developability of the series.
Every record traced to its source for reproducibility.
Data integrations
The catalogue spans protein structures, chemical spaces, bioactivity, target biology, literature and docking software — open and licensed. Each source shows its status, so you can see which are connected today and which are planned.
Compute tools
Twelve established tools for preparing molecules, docking, running dynamics and predicting structure. Each lists its inputs, outputs and requirements, and says plainly when it is not installed or needs a licence.
Simulations
Run docking, dynamics and structure prediction on the structures and compounds already in your workflow. Results are recorded as predicted — never as experimental — and linked to the inputs that produced them.
AutoDock Vina · Smina · GNINA
Dock prepared ligands into a prepared receptor and rank the poses by predicted binding affinity (kcal/mol), with RMSD against a reference pose where one exists.
OpenMM · GROMACS · AmberTools
Relax a structure, then simulate a protein or complex in solvent to test whether a docked pose stays put and how the pocket moves.
ColabFold · Boltz-2 · AlphaFold 3
Fold a sequence when no experimental structure exists, with a per-residue confidence estimate (pLDDT) to guide how far to trust it.
Questions
Every record carries a status — predicted, measured, curated or hypothesis — plus its source and date. A computed result can only ever be predicted or a hypothesis; it is never recorded as an experiment. Cited answers must point to a retrieved article, and statements that can't be traced to one are removed.
Medici says so. When you run the pipeline, each layer reports what was found and what is missing — an unconnected database, a missing input, or no measurements yet — and a layer with no data is marked blocked rather than complete. Nothing is filled in with placeholder values.
A workflow is your project, for example one target. Inside it, each pipeline layer has a workbench: a focused place to gather evidence, search literature, bind data sources, run tools and keep notes for that stage.
The catalogue tracks 87 integrations across 7 areas, from structures and compound libraries to bioactivity and trials. Each shows its status, so you can see which are connected today and which are still planned. Open resources and licensed ones are both listed, and licensed ones need your own credentials.
Twelve tools cover ligand preparation, docking, molecular dynamics and structure prediction. Open-source tools run in containers; licensed ones, such as Amber, and gated ones, such as AlphaFold 3, require a licence or weights you supply. A tool that isn't installed reports that clearly.
Yes. Each run records its inputs and parameters, and every piece of evidence links back to its source, so a reviewer can retrace any result from conclusion to origin.
Sign in to open a discovery workflow, search the literature, browse protein structures and run your first docking or dynamics study.