Convolutional neural network-derived neurofibrillary tangle classifiers: Investigative tools to identify maturation levels and explore post-translational modifications using laser capture microdissection coupled mass spectrometry.

Abstract

Post-translational modifications (PTM) of tau are implicated in Alzheimer disease (AD) progression and are established biomarkers in cerebrospinal fluid and plasma; however, the labor-intensive nature of conventional proteomics limits their investigation in histology-specific contexts. We describe the findings of an artificial intelligence-guided laser capture microdissection (LCM) pipeline for harvesting neurofibrillary tangles (NFTs) by maturation level into pretangles (pre-NFTs), mature tangles (iNFTs), and ghost tangles (eNFTs) using 38 cases obtained from 2 independent biobanks. We evaluated the performance characteristics of proprietary machine learning algorithms for the subclassification of these NFT categories in anti-pTau217-stained whole slide images of entorhinal cortex, hippocampus, frontal, and parietal cortex sections. Overall precision/recall/F1 scores were highest for Classifier A (0.6/0.46/0.5). The best performing algorithm was used to guide LCM capture and inform NFT enrichment. Targeted proteomics on tau signature peptides (pTau181, pTau217, and TauMTBR) was performed on approximately 1250 NFT collections. The results demonstrated that their abundance increased from pretangles to mature tangles (∼2-11-fold increase), and that this was followed by a sharp reduction in ghost tangles (∼3-116-fold decrease), with pTau217 showing the most drastic change. Pathologist-trained NFT classifiers represent an objective albeit imperfect means to enrich specific morphologic forms permitting coupled LCM-MS (mass spectrometry) to investigate AD-associated PTM.