KRONOS2

KRONOS2 is a marker-aware vision foundation model for multiplex immunofluorescence spatial proteomics — the assays, including CODEX/PhenoCycler and CyCIF, that measure dozens of proteins at single-cell resolution inside intact tissue. It is the successor to KRONOS and keeps the design that made the original workable: rather than treating a multiplexed image as a fixed channel stack, it encodes single-marker patches conditioned on the marker names supplied alongside them, returning one embedding per multiplexed patch. A single frozen checkpoint therefore handles arbitrary, heterogeneous antibody panels instead of requiring one model per panel.

Panel heterogeneity is what kept general-purpose backbones out of spatial proteomics: two groups studying the same disease routinely use different antibody panels, marker counts, and instruments, and a model trained on one channel layout does not transfer. Histopathology foundation models such as UNI sidestep this by working on three-channel H&E, and channel-agnostic microscopy models such as OpenPhenom target Cell Painting rather than antibody-based tissue imaging. KRONOS2 scales the marker-aware route: the backbone moves from ViT-S/16 to ViT-B/16, embedding width from 384 to 768 dimensions, and the pretraining marker vocabulary from 175 to 268 markers, learned on a larger multi-cohort corpus the developers call SPM56M.

KRONOS2 was released in July 2026 by the Mahmood Lab at Harvard Medical School and Mass General Brigham, together with the Sizun Jiang Lab at Beth Israel Deaconess Medical Center, led by Andrew H. Song, Anurag Vaidya, and Luca Alexandru Zamfira. It shipped alongside CORAL, an open-source spatial proteomics toolkit from the same groups that is the intended way to run the model. No technical report has been released yet; the model card directs users to cite the original KRONOS preprint.

#Key Features

• Marker-conditioned, panel-agnostic encoding: Embeddings are computed from single-marker patches together with their marker names, so one checkpoint ingests arbitrary panel compositions without retraining.
• Expanded marker vocabulary: Pretraining covers 268 markers, up from 175, widening the set of panels that map directly onto learned per-marker statistics.
• Multi-scale representations: The same frozen features support cell-level phenotyping, unsupervised discovery of tissue microenvironments, and patient-level outcome modeling.
• Frozen-feature deployment: Feature extraction requires no training or per-dataset refitting, and evaluation-mode inference runs in fp32, deterministic for identical inputs, software environment, and hardware.
• Strict marker matching: Channels match the pretraining vocabulary by an exact, separator-insensitive key with no fuzzy or alias step, so distinct markers such as CD24 and CD4 are never silently conflated; markers outside the vocabulary must be registered with normalization statistics computed on the user's own data.

#Technical Details

KRONOS2 is a marker-aware DINOv2 ViT-B/16 producing 768-dimensional embeddings from a reference tissue patch of 256 x 256 pixels, trained self-supervised on SPM56M and released as the teacher checkpoint. Preprocessing is marker-specific: each channel is normalized with the mean and standard deviation learned for that marker during pretraining, and one channel is designated the nuclear stain. In the developers' benchmark across four cohorts, frozen KRONOS2 features lead every encoder tested. On 16-class phenotyping in classic Hodgkin lymphoma (18 markers) it reaches 0.720 balanced accuracy, against 0.702 for a mean-marker baseline, 0.680 for Pixie, 0.662 for UNI applied per marker, and 0.308 for CA-MAE; on two 9-class DLBCL cohorts (12 markers) it reaches 0.742 and 0.796. The margin is widest on outcome prediction: on an 80-patient head and neck squamous cell carcinoma cohort with 57 markers, it achieves a 0.694 concordance index and a 0.782 three-year AUC, versus 0.602 and 0.637 for the next-best encoder.

#Applications

KRONOS2 serves spatial-biology and computational-pathology groups working with multiplexed tissue images. The CORAL toolkit wraps it in a full pipeline: ingesting OME-TIFF into OME-Zarr, segmenting tissue and cells, grid or cell-centric patching, feature extraction, phenotype classification, Leiden clustering of patch features into unsupervised spatial neighborhoods, and slide-level aggregation for survival modeling. Because the same encoder runs across panels and platforms, cohorts acquired with different antibody sets share one representation space, and CORAL's standardized preprocessing makes encoder-to-encoder comparisons reproducible. LoRA fine-tuning is supported for labs needing more accuracy on their own labeled cells.

#Impact

KRONOS2 turns the marker-aware foundation-model approach into a practical, toolkit-backed workflow, and its gains over mean-marker and clustering baselines are largest where spatial proteomics is hardest: linking tissue architecture to patient outcomes. Several caveats are material. Without a technical report, the architecture and pretraining corpus are documented only in the model card, and the benchmark numbers come from the developers' own toolkit rather than independent evaluation; the survival cohort has 80 patients. Weights are gated behind institutional email verification and a terms-of-use agreement, and are licensed CC BY-NC-ND 4.0, which rules out commercial use and redistribution of derived models. Inference code ships with the gated HuggingFace weights and through CORAL rather than a standalone repository, and KRONOS2 has not been validated for clinical diagnostic use.