MonoXtract: A Deep Learning Framework for Extracting Transmembrane Protein Dynamics from Single-Color Single-Molecule Imaging

Abstract

Abstract Single-color fluorescence-based single-molecule techniques enable label-efficient probing of membrane-associated biomolecular dynamics. However, extracting reliable kinetic information is challenging due to signal heterogeneity, noise and the absence of reference channels. Here, we introduce MonoXtract (Single-color fluorescence Trajectory eXtraction and Analysis via Transformer networks), an end-to-end deep learning framework integrating particle tracking, trace extraction, and classification. Combining convolutional inductive bias with a Vision Transformer backbone, MonoXtract captures both fine-scale fluctuations and long-range dependencies, enabling robust state identification across temporal and noise scales. It achieves expert-level accuracy and generalizes well to diverse systems without retraining. Applied to host-defense peptide LL-37, MonoXtract uncovered a previously unrecognized cell-entry pathway that LL-37 transiently incorporates into and escapes from LL-37 oligomers in membranes, providing mechanistic insight into LL-37’s roles in antimicrobial defense and cancer biology. By delivering fast, unbiased, reproducible analysis, MonoXtract extends the analytical reach of intensity-based single-molecule techniques and accelerates complex biomolecular dynamics discoveries.