Computational approaches for the analyses of spatial profiling technologies

点击翻译按钮获取中文摘要

Aim 1: Unsupervised training of image region encoder for highly-multiplexed spatial proteomics data. We will first identify the optimal encoder to obtain image region representations. This encoding will be learned in a completely unsupervised approach to use all the available regions in the spatial imaging data without requiring any clinical labels, or any individual cell location information from any preprocessing step. The learned encoding will be used in Aim 2. Aim 2: Interpretable differential analysis across cancer types to identify spatial topological features associated with clinical characteristics like survival outcomes. We will use interpretable machine learning methods to predict clinical characteristics from images represented in terms of learned image region encodings. The interpretability of the prediction method will help in visualizing image regions associated with different clinical categories, like good and poor prognosis. The commonality or differences of important image regions in the encoded space will identify underlying key similarities or differences across clinical outcomes, which could be used to identify tumor immune evasion mechanisms and help in diagnosis and therapy. Preliminary Data. We have collected spatial proteomics images and corresponding patient survival from external databases for colorectal cancer (432 samples), head and neck cancer (308 samples) and breast cancer (681 samples) for spatial imaging technologies like CODEX and IMC. We also have Lung adenocarcinoma CODEX images collected internally at NCI.