Multiparameter cytometry, for example with CyTOF, has enabled the interrogation of immune phenotypes in unprecedented detail in many clinical contexts. But cytometry is incapable of answering a question of critical importance to many tissue context studies, and especially understanding how local interactions between tumor cells and immune cells correlate to clinical outcomes. This becomes especially relevant to understanding the subtleties of how different immunotherapeutic approaches operate in vivo.
We recently developed a multiparameter immunofluorescence technique, termed CODEX, which allows the capture of spatial information for protein and RNA expression in tissue sections. This spatial information enables us to establish not only cell-types according to traditional phenotypic surface marker expression, but also to potentially surmise specific tissue states driving clinical responses. To make sense of the high-dimensional data afforded by CODEX, we apply here state-of-the-art deep neural networks (DNNs). These networks, which have achieved superhuman classification accuracy in many diverse domains, automatically identify cells, cell niches and regions (at multiple scales) that are capable of distinguishing healthy and diseased samples. This is done in an unbiased way, with only ‘healthy’ vs. ‘disease’ labels as additional input alongside the imaging data.
We first train DNNs to successfully classify multiparameter tissue images from independent replicates across conditions. Having achieved a high accuracy of classification, we set the network output to highlight cells and regions deemed to be most relevant to classify each condition. Applying this methodology to healthy and mrl (lupus) spleens stained for 30 markers, our neural network is able to successfully identify a not previously observed enrichment of cell confluences (niches) consisting of CD8 T-cells and conventional dendritic cells enriched in MRL samples, as well as other novel niches completely unpredicted by prior knowledge.
Our DNN enables the systematic and unbiased discovery of specific immune interactions in any tissue type. Applying our technique to the analysis of samples from immunotherapy recipients could enable the discovery of key factors in the tumor microenvironment that distinguish positive responders as well as the subsequent identification of targets for perturbation.
Citation Format: Salil S. Bhate, Nikolay Samusik, Yury Goltsev, Garry P. Nolan. Automatic identification of cell niches and immune interactions important for clinical outcomes using multiparameter imaging and deep neural networks [abstract]. In: Proceedings of the Second CRI-CIMT-EATI-AACR International Cancer Immunotherapy Conference: Translating Science into Survival; 2016 Sept 25-28; New York, NY. Philadelphia (PA): AACR; Cancer Immunol Res 2016;4(11 Suppl):Abstract nr PR14.
- ©2016 American Association for Cancer Research.