# KRONOS Brings Foundation Models to Spatial Proteomics

#programming #tech

A new open-source Python package called KRONOS is changing the landscape of spatial proteomics with a foundation model trained on over 47 million image patches from 16 tissue types, 175 protein markers, 8 imaging platforms, and 5 institutions. Released just last week by researchers at Harvard's Mahmood Lab, KRONOS is designed to bring label-efficient AI to multiplexed tissue imaging with real-world scalability and generalization.

At its core, KRONOS enables powerful capabilities like unsupervised region classification, biomarker discovery, patient stratification, and even automated artifact detection — all without requiring expert annotations. Researchers can simply drop in a new image patch and retrieve biologically or visually similar regions from massive cohorts. This spatial search functionality alone opens new pathways for hypothesis generation and cohort-wide comparisons in cancer, immunology, and tissue pathology.

What Makes KRONOS Unique?

Instead of treating each imaging dataset as a separate problem, KRONOS takes a foundation model approach. It combines channel-wise image stems with marker-identity embeddings, allowing it to generalize across highly diverse marker panels. Pretrained weights and all code are released for academic use, making it possible for labs anywhere to build on the work.

It supports patch-level embeddings, nearest-neighbor queries, and easy integration with clustering or classification workflows. This makes it ideal for labs working with high-dimensional tissue datasets from imaging mass cytometry (IMC), CODEX, MIBI, and similar platforms.

Example Usage

from kronos import KronosModel
model = KronosModel.load_pretrained('kronos_spatial')
embedding = model.embed_patch(patch_image)
# embedding can now be used for UMAP, clustering, or kNN search

The KRONOS release also includes interactive notebooks demonstrating region-based querying, visualization tools, and batch processing capabilities — all in a plug-and-play format compatible with standard bioimaging workflows.

Why This Matters

Spatial proteomics is a rapidly growing field, but most methods still rely heavily on manual labeling or single-task models. KRONOS demonstrates that a single, well-trained foundation model can enable broad biological discovery across different tissues, diseases, and imaging platforms. This is especially valuable for rare disease studies and hard-to-annotate samples where expert pathologists are not available.

By lowering the barrier to entry, KRONOS helps democratize spatial tissue analysis — enabling more researchers to explore their own datasets in new ways.