用于空间蛋白质组学的通用AI基础模型

Standard spatial proteomics (SP) pipelines for biomarker discovery have traditionally relied on cell segmentation followed by single-marker thresholding or rule-based mechanisms to assign cell phenotypes. However, this often suffers from a laborious process of manual annotations and neglects cell states characterized by the coordinated expression of multiple markers as a function of the cellular spatial organization. Moreover, converting the preprocessed cells to actionable biological insights on the cohort level is time-consuming due to its iterative nature of hypothesis generation and testing. We sought to build an AI foundation model trained on a large body of spatial proteomics datasets to address these challenges and further accelerate crucial spatial biology tasks in a platform-agnostic manner. We present KRONOS, a foundation model for spatial proteomics that operates directly on segmentation-free multiplex image patches. KRONOS extends the self-supervised learning recipe proven to be successful for several pathology foundation models to spatial proteomics, by employing a Vision Transformer architecture that can flexibly handle the variable number of protein markers and simultaneously encode the protein expression levels and the known biological properties of the protein. The training dataset for KRONOS consists of 47 million single-marker patches spanning 175 protein markers, 16 tissue types, eight fluorescence-based imaging platforms, across five different institutions. This diverse and large-scale dataset allows KRONOS to learn rich, low-dimensional image representations that effectively capture spatial protein expressions jointly across different markers. We evaluate KRONOS across a diverse range of downstream tasks, including cell phenotyping, tissue artifact detection, and patient stratification, and across diverse disease types such as lymphoma, renal cell carcinoma, and skin cancer. Across these tasks, KRONOS consistently outperforms other pathology (UNI) and spatial proteomics foundation model (CA-MAE) baselines, underscoring the importance of a domain-specific model trained on a large corpus of data. Specifically for cell phenotyping, we show that KRONOS is data-efficient and requires only a few cell annotations, addressing a major bottleneck in spatial proteomics where cell-level annotations are expensive to produce, slow to generate, and difficult to reuse across datasets. Furthermore, KRONOS allows “reverse image search” for identifying similar spatial patterns and biological concepts within and across patient cohorts, facilitating the discovery of clinically relevant tumor microenvironments and protein biomarkers.Together, these results position KRONOS as a general-purpose, foundation model for spatial proteomics that streamlines analysis, reduces annotation burden, and unlocks new opportunities for spatial biomarker discovery.