1203 Automated prediction of immunotherapy response in head and neck squamous cell carcinoma by a multimodal graph neural network
Year: 2024
Type: conference-abstract
Abstract: Background Immunotherapy has had a tremendous impact on cancer treatment in the past 10 years, but despite advances, a substantial fraction of patients still fail to respond. Matching patients to the appropriate treatments in a cancer setting represents an enormous unmet clinical need. Tools are needed that can effectively screen patients for immunotherapy treatments, both for approved indications, and also for drug developers looking to expand indications. Methods We propose an AI algorithm, specifically a graph neural network (GNN), for automated screening of patient response to immunotherapy. The novelty in our algorithm is its universal graph abstraction framework that allows it to learn morphological and structural information of the tumor from any pathology image modality, including H&E and highly-multiplexed proteomics data. By being applicable to any histopathological image, our algorithm can take advantage of the vast amount of H&E pathology image data that is publicly available by learning morphological and structural features of a disease of interest that it can then import to smaller clinical or discovery cohorts of a particular trial or image modality, such as multiplexed protein imaging. Results We demonstrate the utility of this approach in the context of stratifying head and neck squamous cell carcinoma patients (HNSCC). We first pretrained our multimodal model on a large set of H&E images of HNSCC tumors from The Cancer Genome Atlas.1 We then fine-tuned the model on a smaller cohort of patients who were treated with immunotherapy, for which CODEX highly-multiplexed protein images were available.2 By learning across modalities, from the large dataset of auxiliary H&E histopathology images, our algorithm was able to significantly improve prediction performance compared to a model trained only on the small CODEX dataset. Further, we show the ability to interpret the predictions of the model based on the cellular and morphological features deemed most relevant by the model. Conclusions Although highly-multiplexed proteomic and transcriptomic imaging is becoming more widespread, it is still difficult and expensive to acquire such data for immunotherapy studies. To achieve higher accuracy for patient screening, we expect it will be necessary for algorithms to integrate a diversity of datasets and modalities. This approach can help to integrate diverse datasets from across discovery and clinical cohorts. Although we have demonstrated its efficacy here only on HNSCC, we believe it will have an impact across cancer types going forward. References The Cancer Genome Atlas Research Network. ‘Comprehensive genomic characterization defines human glioblastoma genes and core pathways.’ Nature 2008;455:1061–1068. Uppaluri Ravindra, Campbell Katie, et al. ‘Neoadjuvant and adjuvant pembrolizumab in resectable locally advanced, human papillomavirus-unrelated head and neck cancer: a multicenter, phase II trial.’ Clinical Cancer Research 2020;26(19):5140–5152.
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4404065339