Example prompt: the core MCP workflow
elio is intended to be used in three stages:
- Upload your compound data.
- Build and review models for each assay column.
- Generate new ideas that solve your liabilities, using goals set on those models. [Experimental]
You can run the whole flow from a single prompt, or drive each stage separately. The example below runs it end to end.
Before you start
You need the MCP connected (see Connecting to MCP manually) and a CSV to attach to the conversation.
Your CSV works best with:
- One SMILES column.
- A compound ID column containing numbers (for example
DM-1042). This triggers temporal validation, where the newest compounds are held out as the test set. - Between 5 and 20 assay columns. This is the range we have found works best.
- Each assay column named after the assay it holds, such as
logD_pH7_4,kinetic_solubility_uMorhERG_inhibition. Clear names give better property type inference.
The prompt
Attach your CSV to a new conversation, then paste this in alongside it. Replace the dataset name and goal thresholds with your own.
I'd like to run the full deepmirror workflow on the compound dataset I've attached.
Please work through these steps, and stop and show me the result before moving to the next one:
1. Upload the file and create a dataset called "My series - ADME". Then show me every column with the property type, unit and role deepmirror inferred, so I can check them before anything is trained. Flag anything that came back as unknown.
2. Once I've confirmed the columns, train a model for each assay column. Tell me which validation strategy each model is using, then poll until they finish and show me the metrics side by side.
3. Set the completed models as goals for the dataset: logD between 1 and 3, kinetic solubility above 50 uM. Confirm the goals registered and tell me how many of my existing compounds already meet them.
4. Run an Explore generation using those goals, starting from my five best-scoring compounds. When it finishes, show me the top 10 generated molecules by MPO score, with their predicted values and uncertainties.
Training and generation are slow — if something is still queued, wait and check again rather than giving up.