KRONOS

KRONOS is a self-supervised foundation model for spatial proteomics, the family of multiplexed fluorescence imaging techniques that map dozens of proteins at single-cell resolution within intact tissue. While foundation models have reshaped histopathology and natural-image analysis, they had made little headway in spatial proteomics because the data are high-dimensional, multi-channel, and heterogeneous: different experiments use different antibody panels, different marker counts, and different imaging platforms, so a model trained on one panel does not transfer cleanly to another.

KRONOS addresses this with a panel-agnostic, marker-aware design. Rather than treating each multiplexed image as a fixed stack of channels, it processes single-marker patches and attaches a non-learnable sinusoidal marker-identity embedding to each one, alongside a shared channel-wise convolutional stem. This lets a single model ingest an arbitrary number of markers, in any order, and generalize to markers never seen during training, without retraining or panel-specific fine-tuning. The result is a backbone that produces embeddings at cellular, microenvironment, and tissue scales from the same checkpoint.

KRONOS was developed by the Mahmood Lab at Brigham and Women's Hospital and Harvard Medical School, together with the Jiang Lab and a large multi-institution consortium, and was released as an arXiv preprint in June 2025 by Muhammad Shaban, Faisal Mahmood, Sizun Jiang, and colleagues.

#Key Features

• Marker-aware, panel-agnostic architecture: Sinusoidal marker-identity embeddings and a shared channel-wise stem let KRONOS accept variable marker counts and generalize to unseen markers without retraining.
• Self-supervised pretraining at scale: Trained with the DINOv2 self-distillation and masked-image-modeling framework on 47 million single-marker patches, requiring no manual labels.
• Multi-scale representations: A single checkpoint yields cell-, tissue microenvironment-, and tissue-level features, supporting phenotyping, region classification, and patient stratification.
• Segmentation-free processing: KRONOS operates on image patches rather than segmented cells, enabling efficient, scalable analysis and cross-institutional comparison of spatial patterns.
• Data efficiency and retrieval: It matches or exceeds baselines with a fraction of the labeled data and functions as an image reverse-search engine for recurring spatial motifs.

#Technical Details

KRONOS is a ViT-S/16 vision transformer (12 transformer blocks) adapted for multi-channel input, producing image-level, marker-specific, and token-level features. Each token combines a shared image embedding, a sinusoidal marker encoding, and a learnable spatial position encoding. Pretraining used DINOv2 for 125,000 iterations at a batch size of 1,024, sampling three-channel batches (one nuclear stain plus two random protein markers) from SPM-47M, a corpus of 47 million single-marker patches derived from 3.67 million multiplexed patches spanning 175 markers, 16 tissue types, 8 fluorescence platforms (including CODEX, MxIF, COMET, and IBEX), and 30 cohorts. Evaluated across 11 independent cohorts, KRONOS reached balanced accuracies of roughly 0.74-0.80 on cell phenotyping versus about 0.55 for the UNI, DINOv2, and CA-MAE baselines, an AUC of 0.91 for prostate region classification, 0.98 precision for artifact detection, and an AUC of 0.79 for treatment-response prediction in clear-cell renal cell carcinoma. Pretrained weights are available on HuggingFace under gated access.

#Applications

KRONOS is aimed at spatial-biology and computational-pathology researchers working with multiplexed tissue images. Its embeddings support label-efficient cell phenotyping, tumor-microenvironment and region classification, and patient-level stratification for outcomes such as immunotherapy response, often with far fewer labeled examples than conventional supervised pipelines. The panel-agnostic design allows datasets acquired with different antibody panels or platforms to be compared in a shared representation space, and the retrieval capability lets researchers query tissue archives for recurring spatial patterns.

#Impact

KRONOS extends the foundation-model paradigm to spatial proteomics, a modality where panel heterogeneity had blocked general-purpose pretrained backbones. Its marker-aware embeddings and segmentation-free processing offer a scalable route to cross-cohort analysis and reverse image search over multiplexed tissue data. As an arXiv preprint, its results await peer review, and while the code and pretrained weights have been released, they are distributed under a non-commercial, academic-only CC BY-NC-ND 4.0 license with gated weight access. KRONOS has not been validated for clinical diagnostic use and requires independent evaluation before any such application.