Skip to main content
Premium Trial:

Request an Annual Quote

CNS Tumor Diagnostic Features Predicted With Spatial Transcriptomics Analysis Tool

NEW YORK – A team led by investigators in Germany has developed an artificial intelligence (AI)-based tool to analyze spatial transcriptomics data in order to inform central nervous system tumor diagnoses.

"Incorporating spatial transcriptomics into neuropathological workflows enhanced our understanding of tumor microenvironment, subclonal heterogeneity, and cellular interactions," co-corresponding authors Felix Sahm from University Hospital Heidelberg and Dieter Henrik Heiland from University Hospital Erlangen and their colleagues wrote in a paper describing the approach in Nature Cancer on Wednesday.

"Through machine learning," they explained, "we mapped molecular predictions back onto histological images, assisting in the diagnostic process, particularly with limited tissue availability, which is common in brain tumors."

In their study, the authors outlined their new spatial transcriptomic analysis framework, known as "neuropathology spatial transcriptomic analysis" (NePSTA), which combines spatial transcriptomic profiles with machine-learning algorithms and graph neural network artificial intelligence to infer tumor features ranging from copy number alterations or tumor genotype to immunohistochemistry.

"This framework encompasses the entirety of contemporary molecular-based neuropathological diagnostics including automated histological, transcriptomics, and genetic and epigenetic profiling," the authors explained, noting that the "ability to generate high-dimensional data from minimal tissue holds potential as an invaluable addition to the neuropathological diagnostic arsenal, especially for challenging tissues."

Starting with a set of tissue samples from 117 individuals with neuroepithelial CNS tumors who were enrolled at four medical centers in Germany, along with 23 controls, the team trained NePSTA to predict tumor histology, DNA methylation features, and other diagnostic features using 10x Genomics Visium spatial transcriptomic data on fresh frozen or formalin-fixed paraffin-embedded tumor tissue slices.

The samples spanned IDH-wild type or IDH1/2-mutant glioblastomas, pediatric CNS tumors, and (glio)neuronal tumors collected in Heidelberg, Freiburg, Mannheim, and Memmingen, the team explained, noting that the approach was aimed at getting as much information as possible from small tissue biopsy samples.

"We exemplified NePSTA with specifically challenging neuro-oncology samples that can readily be applied to increase diagnostic accuracy and precision," the authors wrote, noting that "alignment of morphology and molecular data unites the recently diverging fields of 'traditional' and 'modern' pathology, which both have their genuine and unique advantages, to jointly optimize patient care that ultimately depends on optimal diagnostic outcomes."

After training the NePSTA tool, the researchers reported that they could predict tumor features such as methylation subclass classifications or tissue histology with roughly 89 percent accuracy for individual patients.

"Collectively, we established and validated NePSTA, a framework for comprehensive analysis of spatial transcriptomic analysis with a broad spectrum of diagnostic applications such as inferred IHC, CNAs, and in-depth genotyping," the authors wrote. "We demonstrated that integration of proximal signals of spatial transcriptomics, along with AI-based algorithms, can further predict epigenetic subgroups, perform automated segmentation, and characterize prognostic genomic alterations such as CDKN2A/B loss at spatial resolution."