MolClass Overview

What is MolClass?

MolClass is a small-molecule classification and bioactivity-prediction platform. It learns relationships between chemical features and categorical experimental labels, then uses approved models to classify other molecules. A web interface, REST APIs, and an MCP server expose the same registry and prediction workflows.

What it is useful for

• Prioritizing compounds for follow-up screens and library design.
• Comparing activity estimates across assay-specific models.
• Investigating assay liabilities such as autofluorescence and aggregation.
• Exploring permeability, mutagenicity, fungal activity, and other hosted endpoints.
• Training classifiers on your own labeled SDF data in a configured deployment.
• Integrating chemical search and predictions into scripts or assistant workflows.

From structure to prediction

An SDF import creates dataset memberships and preserves selected experimental properties. CDK calculates molecular descriptors and fingerprints. Weka trains classifiers using those features and categorical targets. Training records evaluation evidence and artifacts; a human must approve a build before the prediction service serves it. Predictions and their model-build identities are saved for later retrieval.

A molecule can belong to several datasets. A dataset can have several targets. A target can have several model definitions and builds. An imported dataset is therefore not the same thing as an available prediction model.

Interpreting results

Read the class labels and probability distribution for the particular model. Confidence is the probability assigned to its predicted class, not an experimentally measured probability of success. Review holdout accuracy, AUC, F1, class support, and scaffold-based applicability evidence together. Unknown applicability remains unknown. Predictions need experimental validation and do not establish safety or clinical effectiveness.

The old wiki described a log-odds output; the current prediction API returns class probabilities and confidence. Historical performance claims and original-paper dataset sizes must not be substituted for the current model's evaluation and imported membership counts.

Next steps

Use the online tool · Explore current datasets · Train your own models

Citation and contact

Wildenhain J, Fitzgerald N, Tyers M. Bioinformatics. 2012 Aug 15;28(16):2200–2201.

Contact: Jan Wildenhain.