AI-Enabled Spatial Proteomics Improves Prediction of Progression to Cancer

Spatial proteomics allows us to map protein expression within cells and tissue structures, giving insight into how tumor, stromal and immune cells interact in their native environment. Traditional approaches rely heavily on human interpretation – pathologists reviewing slides, identifying patterns and making judgment calls. That subjectivity can be limiting, especially when the biological signals are subtle or multifactorial.

AI adds scale and objectivity to the analysis of proteins in tissues and also enables complex feature data extracted by spatial analyses to be converted into clinically actionable results that physicians can discuss with their patients to make more risk-aligned clinical management decisions.

In the case of TissueCypher, a DenseNet model is used to identify Barrett’s tissue on digitized slides and to filter out common artifacts that can interfere with analysis. Additional computer vision methods then segment key tissue and cellular structures and calculate features, which are quantitative measurements of biomarkers. Those features are integrated into a locked, deterministic risk algorithm that produces a personalized score and risk classification for each patient. The result is a more consistent and clinically actionable readout than standard spatial proteomics alone can provide.

The challenge with BE is that while most patients never progress to cancer, a small subset – about one percent per year – do. Distinguishing who falls into that high-risk group can be very difficult with conventional tools. Standard practice relies on pathologists manually reviewing biopsy slides, identifying patterns and making judgment calls. That process is inherently subjective, and the differences between tissue that will progress and tissue that won’t can be very subtle, focal and multifactorial. These changes often occur in a background of chronic inflammation, which can mimic dysplasia and complicate interpretation.

BE is also genetically heterogeneous, so two patients with the same histopathology grade can carry very different levels of risk. Conventional histology often can’t capture this complexity. This is where spatial proteomics has proven particularly valuable: by quantifying multiple proteins and nuclear morphology in the spatial context, it allows us to pick up the subtle, multifactorial signals of progression that may otherwise be invisible.

TissueCypher software performs automated, high-dimensional analysis of 9 protein biomarkers and nuclear morphology in the spatial context of cells and tissue structures to objectively quantify 15 features from routine endoscopic biopsies.

The test does not require a new biopsy or special collection method. It begins with formalin-fixed paraffin-embedded tissue samples taken during endoscopy. The samples are sectioned, a panel of nine protein biomarkers and nuclei are fluorescently labeled and then the slides digitized using whole-slide fluorescence imaging.

Computational pathology software then identifies Barrett’s tissue, filters out irrelevant material and segments various tissue and cell structures. From these structures, the software extracts 15 quantitative features that capture biomarker expression, morphology and spatial relationships, signals often too complex and subtle for manual interpretation.

Those data are integrated into a locked algorithm that generates a score from 0 to 10, assigns a risk class of low, intermediate or high, and provides a 5-year probability of progression to high-grade dysplasia or EAC. The test can also predict the presence of “missed prevalent” cases of high-grade dysplasia or EAC that have been overlooked at initial endoscopy.

The result is not just a pathology report, but a personalized forecast to help guide risk-aligned patient care.

The nine proteins (p53, p16, AMACR, HER-2/neu, K20, CD68, COX-2, CD45RO and HIF1alpha) were chosen due to their known involvement in potential neoplastic progression. Some mark the loss of tumor suppression and breakdown of cell-cycle control, like p53 and p16. Others allow assessment of amplification of an oncogene (HER2/neu) and changes in lipid biometabolism (AMACR). Others, such as CD68, CD45RO and COX-2, assess the inflammatory and immune microenvironment since chronic inflammation can be an early driver of malignant transformation, and HIF1alpha assesses hypoxia response and angiogenesis.

When you put them together, you’re not just looking at one signal, you’re getting a panoramic view of the cellular system that drives BE toward cancer.

It’s this integrative, quantitative approach that has been shown to outperform any single biomarker, including p53, in predicting progression in patients with BE.

The software underlying TissueCypher was not designed specifically for BE – it’s fundamentally a tissue analysis platform. At its core, it identifies tissue specimens, filters out artifacts and segments nuclei, cells and larger tissue structures. From there, it calculates an unbiased set of features, including biomarker intensities, co-expression patterns, spatial relationships, texture and morphology. Those types of features are present in any tissue specimen and can be applied to virtually any protein biomarker that can be labeled on a slide, along with nuclear morphology.

That’s why the potential extends well beyond BE. What we’ve shown in BE is a proof point for how this framework, combining spatial insights with machine learning, can deliver more precise risk stratification and help to redirect patient care before disease advances.

More broadly, this shift is part of a new era of “spatial medicine,” where high-dimensional tissue data can not only inform cancer risk prediction but also help guide treatment response, accelerate drug development and may eventually support digital models that simulate disease over time.

Spatial medicine is an important part of the shift to more personalized strategies across medicine, not just in oncology.

There’s strong momentum, but hurdles remain. One of those is education, helping clinicians to know when to order TissueCypher and how to use the results in shared decision-making with their patients. Clinical societies are beginning to address this.

The American Gastroenterological Association’s (AGA) 2024 guideline on endoscopic eradication therapy (EET) – a set of minimally invasive procedures such as radiofrequency ablation or cryotherapy that remove or destroy Barrett’s tissue – recognized that not all non-dysplastic Barrett’s (NDBE) patients are low risk, and highlighted TissueCypher’s potential to identify those who could benefit from earlier intervention. This matters because EAC, the cancer linked to BE, carries a poor overall 5-year survival rate of ~22%, with outcomes far worse when it is detected late. The 2022 AGA Clinical Practice Update also dedicated a best practice advice statement to TissueCypher in NDBE patients, and the 2022 American College of Gastroenterology guidelines noted that molecular testing, including TissueCypher, may outperform histology and provide value for patients without dysplasia.

Supporting this, an independent pooled analysis led by Mayo Clinic investigators found that high-risk NDBE patients identified by TissueCypher were 18 times more likely to progress than those classified as low risk, and that the test identified over half of progressors missed by standard histology. Importantly, their progression risk matched patients already diagnosed with low-grade dysplasia, a group for whom guidelines recommend EET to prevent EAC.