AI-Guided Precision Tumor Enrichment for Efficient Histo-Proteomics

Introduction: Laser Microdissection (LMD) coupled with Mass Spectrometry (MS) enhances clinical proteomics studies by preserving spatial information while leveraging unbiased MS characterization. When integrated with machine learning tools for image analysis, proteomic data can reveal biological insights linked to histological features, including cell composition and proximity, that might not be fully discerned with either approach independently. We implemented an AI-based approach to examine histopathological features of LMD resected regions and integrate cellular features and protein expression patterns. This innovative approach can be leveraged to define histo-proteomic signatures for improved therapeutic prediction. Methods/DesignIn this study, we developed an eosin-free, hematoxylin (HTX)-based tissue imaging approach and trained an advanced a deep neural network model (DNN, HALO-AI) to define and annotate cell types (e.g. tumor epithelia, immune, and stromal), histological cell subtypes, and nuclei features. The AI-driven automated pipeline facilitated high-throughput transfer of cell annotations onto regions of interest (ROIs) for tumor epithelial (TE) enrichment using the LMD7 (Leica) platform. Subsequent proteomic characterization of tissue lysates was performed using two MS approaches: data independent acquisition (DIA) on a timsTOF Pro (Bruker) and parallel reaction monitoring (PRM) on a Exploris 480 (Thermo). ResultsWe developed a stream-lined, AI-guided, HTX-based digital imaging LMD workstream for tissue segmentation and cell annotation for precise laser microdissection. The AI-trained cell classifiers achieved 87-98% tumor epithelial purity (F1 > 0.85) and displayed a high degree of concordance across three tested independent H&E cohorts (n = 211). Python-enabled automation and additional strategies significantly reduced the total time needed for LMD enrichment, relative to manual methods; boosting throughput by ~500% (p < 0.001). The AI-histology subtype model discriminated adenocarcinoma and squamous carcinoma with high accuracy (F1 > 0.85), verified by board-certified pathologist and comparative analysis of known NSCLC subtype-specific markers - KRT5, KRT6A, and NAPSA. Furthermore, cell proximity measurements delineated the degree of tumor infiltration.Conclusion/DiscussionLeveraging PRM data, we found good correlation between abundance profiles for select immune cell markers (e.g. CD8 and CD45) AI-derived image analysis outputs. Protein-based measurements and spatial analysis helped infer the degree of immune cell infiltration within the tumor environment. Taken together, our AI-guided LMD workflow for tissue annotation and enrichment can be leveraged as a high-throughput precision approach which, when coupled with MS, offers the opportunity for gaining biological insight into oncological diseases.

Learning Objectives: 

1. Comprehend the potential for automating laser microdissection procedures with AI to support high-throughput enrichment and sample characterization for proteomics analysis

2. Facilitate discussion and networking to enable collaborative process developments with AI-based digital pathology, high-precision tumor enrichment, and proteomics analysis

3. Leverage AI, digital imaging, and histological features to correlate histopathology and spatial proteomics for enhanced therapeutic understanding