The immune phenotype of a tumour is a key predictor of its response to immunotherapy1–4. Patients who respond to checkpoint blockade generally present with immune-inflamed5–7 tumours that are highly infiltrated by T cells. However, not all inflamed tumours respond to therapy, and even lower response rates occur among tumours that lack T cells (immune desert) or that spatially exclude T cells to the periphery of the tumour lesion (immune excluded)8. Despite the importance of these tumour immune phenotypes in patients, little is known about their development, heterogeneity or dynamics owing to the technical difficulty of tracking these features in situ. Here we introduce skin tumour array by microporation (STAMP)—a preclinical approach that combines high-throughput time-lapse imaging with next-generation sequencing of tumour arrays. Using STAMP, we followed the development of thousands of arrayed tumours in vivo to show that tumour immune phenotypes and outcomes vary between adjacent tumours and are controlled by local factors within the tumour microenvironment. Particularly, the recruitment of T cells by fibroblasts and monocytes into the tumour core was supportive of T cell cytotoxic activity and tumour rejection. Tumour immune phenotypes were dynamic over time and an early conversion to an immune-inflamed phenotype was predictive of spontaneous or therapy-induced tumour rejection. Thus, STAMP captures the dynamic relationships of the spatial, cellular and molecular components of tumour rejection and has the potential to translate therapeutic concepts into successful clinical strategies.
Self-supervised contrastive learning approaches leverage modality-specific context or invariances to pretrain models using unlabeled data. While contrastive learning has demonstrated promising on results in the image domain, there has been limited work on determining how to exploit modality-specific invariances in biosignals such as the electrocardiogram. In this work, we propose 3KG, a method to generate positive pairs for contrastive learning using physiologically-inspired 3D augmentations of the 12-lead electrocardiogram. We evaluate representation quality by fine-tuning a linear layer for the downstream task of 24-class diagnosis on the PhysioNet 2020 challenge training data, and find that models trained with physiologically-inspired augmentations both outperform and complement standard timeseries augmentations. Our best performing strategy, which incorporates spatial rotation, spatial scaling, and time masking, achieves a performance increase of 0.16, .086, and .046 in mean AUROC over a randomly initialized baseline at 1%, 10%, and 100% label fractions respectively. Additionally, we show that the strength of spatial augmentations does not significantly affect the quality of the learned representations. Finally, we investigate the clinical relevance of how physiologically-inspired augmentations affect the performance of our classifier on different disease subgroupings. As expert annotations are often expensive and scarce for medical contexts, our approach highlights the potential of machine learning to tackle medical problems with large quantities of unlabeled biosignal data by exploiting their unique biological properties.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.