#15 · Epigenomics

Compact CpG selection and memory-limited active learning for predictive biochemistry

The dominant thread is learned models that retain predictive performance while reducing the features or training data they require. MOSurvivor jointly selects CpG sites and tunes XGBoost through multi-objective search, reducing features by 79% at near-baseline epigenetic age-prediction error. SMAL removes training data during active learning and is benchmarked on six ADMET datasets and virtual-screening tasks. FastRet predicts LC-MS retention times from molecular descriptors using LASSO and boosted regression trees, packaged as an R tool. The set is heterogeneous: only MOSurvivor addresses omics data, while the other two address small-molecule property and chromatographic prediction. What is new is explicit parsimony as a design goal, whether in feature sets, training sets, or descriptor-based models, assessed against benchmark accuracy.