Abstract 6187: Multimodal characterization of tertiary lymphoid structures identified via deep learning on digitized H&E images in NSCLC

Abstract Background: Tertiary lymphoid structures (TLS) have recently shown association with improved outcomes for patients treated with immunotherapy across several cancer types. But pathologist-led identification of TLS on routine hematoxylin and eosin (H&E) stained images is time-consuming & lacks standardization. We developed a deep learning (DL) approach to identify TLS on digitized H&E whole slide images (WSI) of non-small cell lung cancer (NSCLC). We also elucidate underlying identified TLS biology by investigating corresponding transcriptomic profiles & multiplex immunofluorescence (mIF) images. Methods: DL model was trained using mature TLS (mTLS) annotated by 5 board-certified pathologists following standardized definitions across WSI of NSCLC, breast & bladder cancers. Model was validated using blinded, exhaustive annotations on 60 WSI by the same pathologists in consensus. To identify underlying biology associated with model-identified TLS presence, phase 3 clinical trial of advanced NSCLC with H&E and corresponding bulk RNA-Seq was analyzed using differentially expressed genes(DEG). Cellular phenotypes of model-identified TLS were elucidated in procured NSCLC WSI labeled with mIF panel & then H&E stained. Results: DL model achieved a validation F1-score of 0.56 (precision=0.52). DEG analyses resulted in significant association of TLS related gene signatures with model-identified TLS presence (q<0.05). Immuno-histopathological evaluation of model-identified mTLS revealed concordance with central clusters of CD20+ B cells surrounded by a mix of CD20+ B & CD3+ T cells. Conclusions: We developed a DL model to detect TLS in H&E WSI & demonstrated their concordance in transcriptomics and mIF. With further validation, DL identified TLS could potentially serve as an imaging-based biomarker in clinical outcome datasets. Table 1. Data distribution for instance-based deep learning model training, validation and independent deployment A. Model development Cancer type Training Validation Digitized H&E Slides with mTLS by Pathologists mTLS countby Pathologists Digitized H&E Slides tested mTLS count by Pathologists (in consensus) True positive mTLS count(Model identified) NSCLC 98 283 20 99 57 Breast 107 315 20 48 35 Bladder 103 262 20 65 48 B. Model deployment in independent cohorts NSCLC Dataset H&E images Other corresponding modality Samples Total Digitized Slides Digitized H&E Slides with mTLS by Pathologists Digitized H&E Slides with model identified TLS OAK (clinical trial, NCT02008227) 320 68 55 Matched patients with RNA seq 282 Procured 23 6 4 Matched slides with mIF panel (pan-cytokeratin, CD3, CD8, CD163/68, CD20, CD138, BCMA, DC-LAMP) 23 Citation Format: Niha Beig, Evan Liu, Geoffrey Schau, Assaf Amitai, Barzin Nabet, Rajiv Jesudason, Eloisa Fuentes, Hartmut Koeppen, Sertan Kaya, Namrata S. Patil, Minu K. Srivastava, Robert Johnston, Cleopatra Kozlowski, Jennifer Giltnane, Daniel Ruderman. Multimodal characterization of tertiary lymphoid structures identified via deep learning on digitized H&E images in NSCLC [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 6187.