Abstract Background Histopathology remains a cornerstone of diagnosis and prognostication in pediatric brain tumors. However, in-depth characterization of the tumor microenvironmental (TME) elucidated by spatial proteomics offers therapy-relevant insight. We present a digital pathology AI framework offering spatial proteomic insight from hematoxylin and eosin (H&E) stained pathology slides. Method We analyzed over 800 independent tissue cores originating from 79 post-mortem patients representing various pediatric brain tumor indications, with a focus on diffuse midline glioma (DMG). In a systematic benchmark of six state-of-the-art digital pathology feature extractors, we predict spatial profiles for 22 protein markers from virtual H&E images. These markers characterize the tumor microenvironment for key brain cell type (e.g. MBP) immune (e.g. CD3, CD8, CD68, CD163, PD1), stemness (e.g. Nestin), and treatment-relevant (e.g. B7H3) expression profiles. We combined the extracted features with a light gradient-boosted machine classifier, trained exclusively on one cohort (403 specimens) enabling predictions on an independently processed, previously unseen test data cohort of 405 tissue specimens. Results We established a novel AI platform for the detection of 22 antigens in H&E-stained tissue. We observed a performance advantage for feature extractors building on vision transformers with the top pipeline achieving competitive performance for most markers (test-set ROCAUC 0.57-0.90). The detection of H3K27M+ DMG cells (ROCAUCH3K27M=0.88+\-0.01) and NESTIN+ neural stem/progenitor cells (ROCAUCNESTIN=0.86+/- 0.01) and MBP (ROCAUCMBP=0.90+/- 0.01) was highly robust. Similarly, immune suppressive signals such as PD1 and CD163 showed moderate performance (ROCAUC3PD1=0.71+/-0.01, ROCAUCCD163=0.75+/-0.01). Conclusions Our AI platform demonstrates the feasibility of extracting in-depth tissue characteristics and gaining molecular insight into the TME from H&E images through digital pathology in the absence of staining, as well as enabling the spatial prediction of key molecular markers (e.g. H3K27M) for DMG identification.
Bugajska et al. (Fri,) studied this question.