A deep-transfer-learning framework to transfer clinical information to single cells and spatial locations in cancer tissues

摘要 在过去的10年里，出现了新的高分辨率分子数据的爆炸性增长，这些数据彻底改变了 了解和治疗癌症的方式。它们包括单细胞转录学、空间转录学和 计算图像分析。然而，对这些数据与临床结果的相关性的研究 患者的生存、复发、转移和药物反应均被抛在脑后。与此同时，深度学习领域正在 随着许多不同的应用，包括在生物数据方面，成熟得非常快。它经常使用多层 神经网络模型用于学习和提取高度非线性的数据表示。迁移学习是 机器学习的子领域，其重点是将从一组源示例学到的知识转移到 另一种类型的样品。这两种方法的结合构成了深度迁移学习，是一种很有前途的方法 调查和了解这些新癌症数据的高分辨率分量之间的关联的解决方案 以及相应的临床结果。在这里，我们建议使用深度迁移学习来转移患者 从大型患者转录组获得的结果信息到细胞、细胞类型、空间区域、 和图像功能，然后可以根据其分配的风险进一步确定优先级，并将其评估为潜在的 侵袭性癌症的靶点。具体地说，我们将开发深度迁移学习框架Degas 对胶质母细胞瘤和多发性骨髓瘤的单细胞数据进行细胞类型优先排序和测试，以验证该方法。 然后它将应用于更具侵袭性的癌症类型的单细胞数据，如三阴性乳腺癌， 胰腺导管腺癌、非小细胞肺癌和胃癌优先考虑高危细胞和 单元类型。然后，将对其进行进一步修改，以用于空间转录(ST)数据，以确定高风险的优先顺序 乳腺癌和胰腺导管腺癌肿瘤的空间区域。由于ST数据可以充当 连接单细胞到患者水平的转录学和组织学图像的桥梁。我们将进一步利用我们的 通过ST数据将组织学图像特征与患者风险联系起来，以识别高风险图像特征的框架。 最后，我们的框架将被构建到R和Python包中，这些包可以通过GitHub和BioConductor获得 被更广泛的癌症研究社区使用。

SUMMARY In the past 10 years, there has been an explosion of new high-resolution molecular data which revolutionize the way that cancer is understood and treated. They include, single cell transcriptomics, spatial transcriptomics, and computational image analysis. However, the study of the association of those data with clinical outcomes such as survival, relapse, metastasis and drug response were left behind. In the meantime, Deep learning field is maturing very fast with many diverse applications including on biological data. It frequently utilizes multi-layer neural network models to learn and extract highly non-linear representations of data. Transfer learning is the subfield of machine learning, which focuses on transferring knowledge learned from a set of source examples to another types of samples. Combining these two approaches constitutes deep transfer learning and is a promising solution to investigate and understand the association of high-resolution components of these new cancer data with the corresponding clinical outcomes. Here we propose the use of deep transfer learning to transfer patient outcome information learned from large patient transcriptomics cohorts to the cells, cell types, spatial regions, and image features, which can then be further prioritized by their assigned risks and be evaluated as potential targets in the aggressive cancers. Specifically, we will develop deep transfer learning frameworks DEGAS for cell type prioritization and test on glioblastoma and multiple myeloma single cell data to validate this approach. Then it will be applied on single cell data of more aggressive cancer types such as triple negative breast cancer, pancreatic ductal adenocarcinoma, non-small-cell lung cancer, and gastric cancer to prioritize high risk cells and cell types. Then, it will be further modified for use with spatial transcriptomic (ST) data to prioritize high risk spatial regions of breast cancer and pancreatic ductal adenocarcinoma tumors. Since ST data can act as a bridge between single cell to patient-level transcriptomics, and histology images. We will further leverage our framework to identify high risk image features by linking histology image features to patient risk via ST data. Finally, our framework will be built into R and Python packages available through GitHub and Bioconductor for use by the broader cancer research community.