SwitchClass: dissecting attenuated and escalated molecular features via a label-switch classification framework

Biological systems exhibit complex molecular trajectories in response to perturbations, ranging from changes that revert or attenuate towards homeostasis to alterations that persist or escalate. Capturing these complex patterns is essential for understanding molecular resilience and maladaptive persistence. Here, we introduce SwitchClass, a label-switch classification framework that distinguishes attenuated and escalated molecular features across biological states. By training dual classifiers with inverted outcome labels, SwitchClass computes a differential feature importance score ({delta}) that quantifies directional change in high-dimensional data. Applied to colorectal cancer proteomics spanning healthy, pre-treatment, and post-treatment samples, Switch-Class reveals proteins that normalise after therapy and those remaining dysregulated, uncovering partial molecular recovery. In phosphoproteomics of dietary perturbation and reversal, it uncovers phosphorylation sites linked to incomplete restoration of insulin signalling. In single-cell transcriptomes from COVID-19 patients with varying severities, it identifies cell-type-specific transcripts that mark either the resolution or persistence of inflammatory activity. Together, these analyses establish SwitchClass as a generalisable and interpretable framework for mapping directional molecular changes underlying adaptation, divergence, and disease severity across biological systems. SwitchClass is freely available from https://github.com/PYangLab/SwitchClass.