Abstract A066: Scalable visualization of perineuronal nets in the brain tumor microenvironment using generative deep learning

Abstract Brain tumor-related epilepsy remains a major source of morbidity, and emerging evidence implicates tumor-associated degradation of perineuronal nets (PNNs) in seizure susceptibility. However, systematic study of PNN remodeling in human brain tumors has been limited by the scalability and tissue requirements of immunohistochemistry, constraining large-cohort, clinically anchored analyses of tumor-PNN interactions. We developed VirtualPNN, a generative deep learning framework that imputes aggrecan-like PNN staining directly from routine hematoxylin and eosin (H&E) histopathology. The study cohort comprised 133 tissue microarray cores from 44 patients, including low-grade glioma, glioblastoma, brain metastases, and non-neoplastic epilepsy-associated tissue. VirtualPNN incorporates domain-informed constraints designed to suppress biologically implausible vascular-associated signal. Model performance was evaluated using five-fold cross-validation with core-level splits, quantitative similarity metrics, patient-level feature-space analyses, targeted vessel-only supervision, blinded neuropathologist review, and robustness testing in tumor tissue lacking PNN morphology. Across held-out test folds, VirtualPNN achieved the lowest Fréchet Inception Distance compared with CUT, CycleGAN, and AI-FFPE baselines, indicating closer distributional similarity to aggrecan immunohistochemistry. Vessel-targeted negative supervision substantially reduced off-target vascular staining while preserving perisomatic, lattice-like PNN morphology. When virtual staining was aggregated at the patient level, VirtualPNN preserved higher-order feature-space concordance with aggrecan staining regardless of seizure history. Robustness analysis across 84 PNN-negative tumor and brain cores demonstrated significantly lower spurious signal generation by VirtualPNN compared with baseline models. Residual false-positive staining was enriched in hypocellular regions and reflected structured, feature-dependent behavior rather than random model failure. Neuropathologist review confirmed capture of key PNN morphological hallmarks, including perisomatic localization and reticulated organization. VirtualPNN provides a scalable, tissue-sparing approach for visualizing aggrecan-like PNN patterns from routine H&E sections. By enabling retrospective, population-scale analyses of extracellular matrix remodeling across diverse brain tumor cohorts, this framework establishes a foundation for studying the brain tumor microenvironment and its relationship to epilepsy and tumor biology at scale. Generative artificial intelligence was used solely to assist with grammar and clarity editing of the abstract text. All scientific content, analyses, and conclusions were generated by the authors. Citation Format: Vishva Natarajan, Tanmay Shukla, Scott Palisoul, Madhumala Sadanandappa, Chun-Chieh Lin, George J. Zanazzi, Jennifer Hong, Saeed Hassanpour. Scalable visualization of perineuronal nets in the brain tumor microenvironment using generative deep learning [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Brain Cancer; 2026 Mar 23-25; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2026;86(6_Suppl):Abstract nr A066.