MolClass

Predict small-molecule bioactivity from chemical structure. MolClass turns labeled screening datasets into machine-learning classifiers and makes their predictions accessible through a web interface, REST APIs, and MCP tools. Use it to explore compounds, prioritize follow-up experiments, or build models from your own structure–activity data.

Try the online server: molclass.chemgrid.org No local installation is needed to browse molecules and explore published models.

Search compounds · Prediction models · Dataset review · Help

The live MolClass compound registry, captured on 3 November 2026.

What is it useful for?

• Prioritizing screening hits. Compare predicted activity across published models to help select compounds for experimental follow-up and library design.
• Exploring activity and assay liabilities. Investigate endpoints such as mutagenicity, permeability, antifungal activity, autofluorescence, and compound aggregation using the available datasets and approved models.
• Finding and comparing molecules. Search by database ID, source identifier, name, InChIKey, or SMILES; draw a structure; or find compounds containing a substructure. Molecule cards connect structures, source records, properties, and saved predictions.
• Building classifiers from your own experiments. Import labeled SDF files, choose a target property, compare algorithms and feature profiles, and review evaluation evidence before publishing a model on your own deployment.
• Automating chemical workflows. Use the REST APIs or bundled MCP server to retrieve compounds, inspect models, and predict from existing molecules or new SMILES structures.

Predictions are model-specific estimates for the dataset's class labels. Assess holdout performance, class support, and applicability-domain information alongside the prediction; experimental validation is still needed. A high class probability does not establish biological activity or safety.

Try a prediction online

1. Open Search. Enter a name, identifier, InChIKey, or SMILES, use Draw structure, or leave the query empty and choose Browse registry. Switch between Exact match and Substructure as needed.
2. Select one or more matching molecules and choose Predict selected to open the model tab with those molecules selected. Molecule cards also expose saved predictions and links back to the prediction workflow.
3. Inspect the published models and their available holdout accuracy, AUC, and F1 metrics. Select models for which the molecule has no saved result and run the predictions. Results appear as each request completes and are retained for later visits.
4. Read the predicted class and class-probability distribution together with the scaffold-based applicability information, where available. Missing metrics or applicability values mean unavailable evidence, not a score of zero.

The home page's MolClass molecule of the month opens search with a structure already loaded. When a linked structure has no exact match, Register this structure and predict queues registration and feature calculation, displays progress, and opens model selection on completion. Public prediction and registration requests are subject to the server's quota policy.

How it works

flowchart LR
    A["Labeled molecules in SDF"] --> B["Analyze and import"]
    B --> C["CDK descriptors and fingerprints"]
    C --> D["Weka training and evaluation"]
    D --> E["Human model review"]
    E --> F["Approved prediction model"]
    G["Search or register a molecule"] --> H["Predict and save results"]
    F --> H

Loading

1. Analyze and import. The upload workflow inspects an SDF file, proposes an identifier property, and lets the operator confirm the identifier and properties to import. A durable worker records molecule identities, dataset membership, selected values, import progress, and failures. Partial imports require explicit acknowledgement before model creation.
2. Generate chemical features. The Chemistry Development Kit (CDK) computes molecular descriptors and structural fingerprints. Feature profiles combine descriptor and fingerprint components; values are stored for reuse. Supported fingerprint families include MACCS, PubChem, Extended, Substructure, Klekota–Roth, Graph-only, EState, and ECFP4.
3. Define and train a model. Choose an eligible categorical target, its positive class, a feature profile, an algorithm, and a feature-selection policy. The creation API requires at least 30 labeled records and at least five records per class. Weka builds classifiers and records training, validation, holdout, and applicable cross-validation evidence, together with excluded records and checksummed model artifacts.
4. Review and publish. Successful training stops at AWAITING_APPROVAL. An explicit, audited human decision is required to publish the build. Training alone never makes a model available for prediction.
5. Predict with the published build. The predictor loads approved artifacts, checks that the molecule's feature vector matches the model contract, and saves the class distribution, predicted class, confidence, and available applicability evidence. The web interface reuses saved results for the same molecule/build.

The current model-creation implementation supports Random Forest, J48, Naive Bayes, SMO, k-NN, LibSVM, Logistic Model Tree, LogitBoost, Ensemble, Bagged J48, and AdaBoost.M1. Feature selection can use CFS, ReliefF, or None. Available profiles and deployed models are listed by the server; retained legacy model names and algorithm labels may differ from the options for new builds.

The current web interface is branded MolClass V2; the maintained database and prediction contracts use v3. The updated GitHub wiki documents the current workflow; historical PHP/Perl and v1 REST instructions remain in its revision history. Current predictions return class probabilities and confidence rather than the legacy log-odds score.

Datasets on the online server

The live dataset catalog contained 21 datasets and 63,866 imported records on 3 November 2026. Counts below are dataset records, not unique molecules across the whole server. Names follow the catalog, including historical spellings.

Dataset ID	Dataset	Imported records
1	Compound induced Neuronal Stem Cell Differentiation	1,267
2	Caco2 Penetration	100
3	Flucanozole Synergizer	1,043
4	Compound accumulation in C. elegans	476
5	Toxicity Benchmark	5,351
6	Ames Categorization	4,261
7	Blood Brain Barrier	412
8	P.falciparum Sensitivity	1,252
9	HSP90 PPI Inhibition	1,614
10	Chembank Cell Cycle Inhibitor Screen	435
11	BetaCell Transdifferentiation	907
12	Xenopus Actin Polymerization	1,359
13	Auto-fluorescence Prediction	4,910
14	Thrombin Molecule Binding	4,876
15	Metabolite Likeness	982
16	Compound aggregation	732
17	Liver Toxcicity	921
18	Antifungal Pump Inhibitors	10,427
19	Mitochondrial Fusion	19,869
23	Yeast Bioactive Library Cross-Kingdom	1,570
31	Microsome Clearance (AZ Hackathon)	1,102

All 21 catalog entries were marked MIGRATED and model-eligible at the time of this check. Dataset 23 contains 17 cross-kingdom assay targets; dataset 31 contains categorical microsomal-clearance annotations. Dataset availability and model publication are separate: a dataset or model definition does not guarantee an approved predictor for every target. The published-model API returned 21 published models at the same check.

Use Dataset review for import outcomes and current metadata, and the model tab for models currently available for prediction. The wiki's current dataset catalog describes all 21 datasets, their target labels, class counts, and available literature references. Benchmark datasets mentioned in project news are not necessarily hosted datasets.

Installation

For using the published models, start with the online server. To run your own instance, use Docker Engine with the Docker Compose plugin, Git, and internet access for the initial dependency downloads. See INSTALL.md for the full setup guide.

git clone https://github.com/jwildenhain/molclass.git
cd molclass
cp .env.example .env

Edit .env before starting:

• Replace MOLCLASS_DB_ROOT_PASSWORD and MOLCLASS_DB_PASSWORD with separate, long random secrets. Keep .env out of version control.
• Set MOLCLASS_CODE_REVISION to the immutable source revision used for the build.
• Size the database and workers for your host. The example uses an 8 GiB InnoDB buffer pool, a 16 GiB model-worker limit, and a 6 GiB predictor limit, plus the other services. These are large-host defaults; smaller deployments need adjusted settings, and training must fit the available memory.

docker compose build
docker compose up -d
docker compose ps
curl --fail http://127.0.0.1:8000/api/v1/health/readiness

Open localhost:3000. Compose starts seven services and initializes the schema from sql/v3/ when the database volume is first created. It does not download the online server's datasets or trained models. Upload your own SDF data, create model definitions, and review the resulting builds.

Model approval needs an administrative setup. The stock Compose API image does not include the host-side approval bridge, so its web approval controls remain disabled. Follow the model approval guide and review deployment notes to publish reviewed builds through the supported administrative path.

Services bind their published ports to loopback by default. A public deployment needs a TLS reverse proxy and the access/configuration controls described in the production runbook. Data intake and configuration are controlled separately by MOLCLASS_DATA_INTAKE_ENABLED and MOLCLASS_CONFIGURATION_ENABLED; see the actual service settings in docker-compose.yml. Store local dumps and snapshots in the ignored backups/ directory.

Architecture and integrations

Service	Role
frontend	Next.js / React web interface
api	FastAPI: uploads, datasets, model definitions, review, and job status
sdf-worker	Java / CDK: SDF analysis and import
model-worker	Java / CDK / Weka: feature generation, model training, and evaluation
molecule-worker	New-molecule registration, feature calculation, and ad-hoc prediction jobs
predictor	Spring Boot: molecule search and published-model predictions
db	MariaDB: registry, dataset membership, jobs, model evidence, and predictions

Long-running imports, training, and molecule registration use persisted jobs with leases and progress reporting. The current source uses CDK 2.12 and Weka 3.8.7.

REST clients use /api/v1 for dataset/import/model-management workflows and /api/v3 for the molecule registry and approved-model predictions. Start with the prediction API and SMILES registration guide.

The bundled MCP server exposes molecule search, model listing, saved results, and prediction to compatible assistants. An authenticated integration can register a SMILES structure and request all missing published-model predictions; anonymous requests follow the documented single-model/quota policy.

Documentation

• Repository documentation — technical and operational guides.
• Installation — Docker setup and local development entry points.
• Search and prediction workflow — selection, saved results, registration progress, and linked structures.
• Model creation, feature generation, and model review — training and publication contracts.
• GitHub wiki — current user guides, dataset descriptions, model training, APIs, installation, and FAQs.
• Project news — development and benchmark reports.

Citation and contact

If you use MolClass in your work, please cite:

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

Contact: Jan Wildenhain — LinkedIn.