Help

Start at Upload. Drop in an SDF file; MolClass analyzes every record, lets you pick which fields become model targets, and then imports it durably — one molecule at a time, so a single bad record never rolls back the rest. A live progress bar shows records loaded, succeeded, and skipped as the import runs.

In Model Creation, pick from datasets that already have a usable target — a property with 2 to 100 distinct classes. Click a dataset’s ID to choose the target property, feature profile (CDK descriptors alone, or JUMBO with added fingerprints), and algorithm (RandomForest, SMO, KNN, Ensemble, and more), then start the build.

Once the worker finishes training, the build lands in Model Review as AWAITING_APPROVAL— nothing is ever auto-published. Open a build to see its full holdout evidence (accuracy, AUC, F1, confusion matrix) and either approve it into the published registry or reject it.

Search has two tabs. Structure searchlooks up the registry by ID, name, InChIKey, SMILES, or a drawn structure, and shows a thumbnail per match — select any number of them and hit “Predict selected”. Model & molecule search is where that selection lands: pick one or more published models and run every molecule against every model at once.

Every prediction reports a predicted class, a confidence score, and an applicability-domain flag (whether the molecule’s scaffold actually resembles the training set). Click any molecule to open its own page, which keeps a running history of every prediction ever made against it, and lets you queue up new ones against any other published model.