About MolClass

Recent improvements

• Parallelized SMOTE oversampling for large, imbalanced training sets, cutting model build time without changing the resampled output (still fully deterministic).
• Redesigned the Ensemble classifier around RandomForest, a cross-validated auto-tuned KNN, and NaiveBayes, measurably improving holdout accuracy over the previous design.
• Fixed a worker classpath bug where CDK’s descriptor engine caused a partial-catalog corruption that had been producing systematically wrong feature vectors for CDK-based models. The worker’s classpath is now generated from the same Gradle build.
• Hardened the prediction service by deduplicating overlapping CDK/Weka dependencies.
• Refactored the backend model loader from eagerly deserializing trained models into memory at startup to an on-demand, size-capped cache (four models by default) — an estimated 20× smaller resident model memory footprint at steady state.
• Benchmarked against the Tox21 Data Challenge: a default RandomForest-on-JUMBO setup matches DeepTox, the competition’s winning method (see News).
• Consolidated structure and model search into a single interface with molecule thumbnails, multi-molecule batch prediction, and per-model AUC/F1 metrics.
• Added live progress reporting during SDF upload and import, replacing a static “queued” message with real record-by-record counts.