Abstract Introduction: GPC3 is a glycosyl-phosphatidylinositol-anchored cell surface glycoprotein that functions to control cell growth and differentiation in early development. Membrane GPC3 is highly expressed in hepatocellular carcinoma (HCC) and moderately in squamous non-small cell lung cancer (SQ-NSCLC) with limited expression in normal tissues which makes it an attractive target for immune cell engagers. This study aims to develop an AI-powered algorithm to evaluate GPC3 expression and compare it with manual immunohistochemistry (IHC) scoring in HCC, adeno-NSCLC and SQ-NSCLC. Methods: 147 NSCLC (94 SQ-NSCLC and 53 adeno-NSCLC) and 113 HCC FFPE tumor blocks were obtained from commercial sources. GPC3 protein level was quantified by a qualified IHC assay for tumor cell membrane and cytoplasm individually. PD-L1 expression is reported as tumor cell proportion score (TPS), combined positive score (CPS) and immune cell proportion score (IPS/IC) TPS, regardless of staining pattern and intensity. FFPE slides were stained for GPC3+ (GC33, Ventana), PD-L1+ (22C3, Dako) cells and then scanned at 20x using the Aperio Versa8 scanner. Machine learning (ML) models were developed using the digitized whole slide images to identify GPC3+ tumor area. Data-driven cutoffs were applied to model-generated human-interpretable features of GPC3+ percentage of tumor area to classify samples as positive and negative. As an orthogonal approach to the GPC3+ percentage cutoff method, tissue and cell annotations were used to train a convolutional neural network (CNN) to classify GPC3 positivity. Results: The epidemiology profiling results revealed that GPC3 protein is highly expressed in HCC (57.5% cases with > 1% positive cells), followed by SQ-NSCLC (52.1%) and adeno-NSCLC (5.7%). GPC3 expression was only detected in tumor cells with two staining patterns observed: membrane + cytoplasm and cytoplasm-only. There was no significant correlation between GPC3 and PD-L1 expression across tumor types, although a trend of lower level of PD-L1 was observed in the samples with GPC3 ≥50%. ML model quantification of GPC3+ and GPC3- showed high concordance with the consensus score of 2 independent pathologists on a test set of held-out samples not used in training. This concordance was comparable to that of individual annotators to consensus. Instead, GPC3 positivity results using both cutoff and CNN methods showed moderate concordance with 2-way pathologist consensus. Conclusions: Here we profiled GPC3 expression across tumor types and showed high level of expression in HCC followed by SQ-NSCLC. We have established an AI-powered digital pathology platform that can provide a standardized, scalable, and reproducible method of characterizing GPC3 positivity to support further patient selection in clinical study. Citation Format: Ellen Meng, Robin Meng, Ying Tong, Yuchen Li, Hong Wang, Paola Fiorentini, Elham Attieh, Qi Tang, Asma Kefsi, Serena Masciari, Giovanni Abbadessa, Cecile Combeau, Lei Tang, Benoit Pasquier, Rui Wang. Artificial intelligence (AI)-powered quantification of glypican-3 (GPC3) expression facilitates patient selection for GPC3-targeted therapy in solid tumors [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 6182.
Read full abstract