Repository navigation
The public server exposes molecule search, cards, datasets, and published-model browsing. No local installation is required. Upload, configuration, approval, and authenticated integration capabilities depend on deployment settings; the old wiki's registration-and-email workflow is not a general description of the current app.
No. Current outputs include class probabilities and confidence for the predicted class. Labels and calibration are model-specific. Check performance and applicability evidence; a probability is not experimental confirmation.
No. Imported records, model definitions, successful builds, and human-published models are separate. See Current datasets for the catalog and the live model selector for approved predictors.
The current model-creation workflow builds classifiers from eligible categorical labels. Prepare a scientifically justified categorical target if your source measurements are continuous. It does not automatically turn a clearance or permeability value into a validated regression model.
The current creation API requires at least 30 labeled records and five per class, with 2–100 distinct labels on an eligible target. These are acceptance limits, not a recommendation that minimal datasets yield reliable models. Review class balance, chemistry coverage, exclusions, and held-out evaluation.
That old wiki limit is not the current analyzer contract. Current source defaults to a 2 GiB SDF and 16 MiB per record; proxies and deployment resources may impose tighter constraints. Training memory depends on both data and model configuration.
A successful build waits for human approval. Only published builds are served. In stock Docker Compose, web approval needs additional administrative setup. See Model training and review.
The molecule worker must be running and able to reach the database and predictor. Retain the job ID and inspect its status/progress. Refresh resumes monitoring; submitting again is not a recovery mechanism. Clearing the UI does not cancel backend work.
Feature generation can fail or exclude records, and some evaluation statistics are undefined for sparse or degenerate class support. Unknown is not zero. Read warnings and the specific build's evidence.
Public molecule cards can be cached for up to an hour. The interactive prediction flow loads saved history separately. A new build also has a different identity from an older saved result.
Use the REST API and MCP guide. New-SMILES registration is asynchronous; anonymous submissions require a model ID and obey a daily quota. A valid integration key enables all-missing-model mode. API keys do not grant model-approval rights.
Wiki home · Online MolClass · Current datasets · Repository documentation
Documentation reviewed 3 November 2026. Live catalog counts and deployment state can change.