1200 Artificial intelligence image models optimizing tumor vs immune cell profiling for accurate biomarker analysis in NSCLC precision drug development
Year: 2024
Type: conference-abstract
Abstract: Background Conventional single-pathologist biopsy assessment in clinical trials may limit accurate quantification of biomarkers expressed by tumor and immune cells. Artificial intelligence (AI) models offer rapid, quantitative, reproducible approaches to optimize patient characterization and pharmacodynamic (PD) readouts but require large training sets. Here, we developed a digital image analysis pipeline to analyze at scale non-small cell lung cancer (NSCLC) samples from a first-in-human Phase I clinical trial investigating a novel small molecule targeting aryl hydrocarbon receptor (AhR), a transcription factor triggering immunosuppression, tumor cell growth and migration when translocated to the nucleus. Our main goals: TME characterization and AhR quantification/localization of expression by immunohistochemistry (IHC) scoring challenged by 1) small data set (primary/metastatic), 2) AhR pattern of expression in both nucleus/cytoplasm of tumor/immune cells. Methods Images of 55 NSCLC biopsy pairs of H&E/AhR IHC-stained sections (NCT04069026, NCT04999202, following standard ethical guidelines) were used for training. AI models for H&E-based tissue segmentation, cell detection, classification, IHC-based cell detection, classification, and nuclear AhR expression were developed. Images were first processed in small batches to accelerate data delivery during trials, using fine-tuning per batch to optimize per-sample characterization. Later, models were trained for optimal cross-cohort generalization. Correlative analysis including AhR naked-eye IHC scores and clinical data was performed on 25 monotherapy samples (IHC/H&E). Results Models improved with batched analyses with final results as follows: 1) H&E tissue segmentation and cell classification (mean F1=0.89, 0.79, respectively), 2) IHC cell classification (F1=0.86), 3) AhR expression AhR present vs. not (F1=0.89), across all intensities (mean F1=0.71), within cell classes (maximal performance when binarized carcinoma vs. other; F1=0.85). Naked-eye pathology scoring was limited to AhR tumor carcinoma cell expression. PD effect (changes in AhR nucleus expression) was mostly evident in immune cells. Lymphocytic AhR nuclear expression had a potential association with time on treatment (Log-rank p=0.026, n=15). Conclusions AI-powered digital pathology tools developed here successfully quantified/localized nuclear AhR expression in carcinoma/immune cells, enabling more granular and quantitative understanding of target biologic effects in patients. Naked-eye pathology evaluation alone missed immune cell AhR expression, displayed difficulties differentiating nuclear/cytoplasmic AhR expression and assessing all cells in a sample. Although trained on our small trial material available alone, our tools overcame target precision assessment difficulties. Limitations in generalization could be improved with larger training sets. Similar approaches could be impactful for quantitative biomarker testing on tough to score targets, providing solutions to typical first-in-human challenges of limited yet diverse samples.
Authors: Cornelius Boehm, Laura Schlieker, Martin Szyska, Christina Embacher, Timo N Haschler, Rebecca Freudling, Anke Weispfenning, Suhas Pandhe, Nader Aldoj, Ariane Sasso, Sandip Ghosh, Katja Lingelbach, Simon Heinke, Rosemarie Krupar, Julika Ribbat-Idel, Simon Schallenberg, Lukas Ruff, Sharon Ruane, Maximilian Alber, Andrea Wagner, Gabriele Leder, Jasminka Cormarkovic-Dragovic, Xin Lin, Marion Rudolph, David Schaer, Yuko Ishii, Emmanuelle di Tomaso, Vasiliki Pelekanou
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4404064968