NeXtMarker: Contrastive Learning for Marker-Level Interpretability in Single-Cell Multiplex Imaging

Simon Gutwein, Daria Lazic, Thomas Walter, Sabine Taschner-Mandl, Roxane Licandro

Keywords: Multiplex Imaging, Spatial Biology, Single-Cell Analysis, Deep Learning, Interpretability

TL;DR: NeXtMarker is a self-supervised, interpretable deep learning framework that enables segmentation-free single-cell phenotyping in multiplex imaging by learning marker-specific normalizations and co-expression patterns directly from raw image data.

Abstract: Understanding cell phenotypes and their spatial organization is crucial in multiplex imaging for spatial biology. Conventional analysis pipelines rely on extensive preprocessing, including background correction and segmentation, introducing biases and information loss. We present NeXtMarker, an interpretable deep learning framework for end-to-end single-cell analysis of multiplex images, eliminating the need for manual preprocessing or segmentation. NeXtMarker employs learned marker-specific normalization and interpretable feature extraction to generate biologically meaningful embeddings in a fully self-supervised manner. It directly processes raw images of cells while preserving spatial and morphological information. We demonstrate NeXtMarker’s ability to (i) resolve intercellular expression patterns and cell morphology, (ii) enable accurate cell phenotyping in a large neuroblastoma tumor dataset, and (iii) generalize to independent osteosarcoma images. NeXtMarker maintains high agreement with conventional pipelines while eliminating the need for preprocessing and segmentation and enhancing interpretability. By enabling unbiased, scalable single-cell analysis, NeXtMarker establishes a foundation for improved phenotyping in multiplex imaging. Code and pretrained models available at: https://github.com/SimonBon/NeXtMarker.

Submission Number: 39