Interpretable deep learning models for translational medicine

了解细胞信号系统的状态有助于深入了解细胞在生理条件下的行为 以及病态的情况。蜂窝信令系统被组织为层次结构(级联)和信号 分子通常被组成编码以控制细胞过程，如基因表达。这 该项目旨在开发高级深度学习模型(DLMS)来模拟基于 基因表达数据。在过去的3年里，该项目取得了重大进展，但面临的挑战 留下来。重要的是，当代DLM的行为就像“黑匣子”，因为它很难解释信号是如何 以及如何解释隐藏节点在DLm中表示哪个信号。这种黑箱性质 阻止研究人员使用DLMS获得生物学见解，即使这些模型可能会 在许多任务中，例如预测癌症的药物敏感性，在建模数据方面优于其他类型的模型 细胞。在这种竞争更新中，我们建议开发新的DLMS和创新的推理算法来 训练“可解释的”DLMS并将其应用于翻译研究。拟议的研究具有创新性，并且 在以下几个方面具有很高的意义：1)我们的新型DLMS和算法利用了大数据 由于细胞信号机制的系统化学/遗传扰动，所以我们可以使用 作为辅助信息的扰动条件，以揭示信号如何在DLm中编码。2)我们整合 因果推理原理和信息论与深度学习方法相结合，使DLMS具有可解释性。 因此，研究人员可以从这些模型中获得机械性的见解。3)创新应用 可解释的DLMS将促进翻译研究。例如，我们将训练可解释的DLMS来建模 单细胞水平上的细胞信号传递，并利用这些信息研究细胞间的相互作用 肿瘤微环境中的细胞之间的相互作用，以揭示癌症的免疫逃逸机制。我们还将 使用从可解释的DLMS获得的信息来预测癌细胞的药物敏感性。我们预计我们的 这项研究不仅将在深度学习方法方面带来重大进展，而且还将在精确医学方面带来重大进展。

Understanding the state of cellular signaling systems provides insights to how cells behave under physiological and pathological conditions. Cellular signaling systems are organized as hierarchy (cascade) and signals of a molecular is often compositionally encoded to control cellular processes, such as gene expression. This project aims to develop advanced deep learning models (DLMs) to simulate cellular signaling systems based on gene expression data. In last 3 years, the project has made significant progresses, but the challenges remain. Importantly, contemporary DLMs behave as “black boxes”, in that it is difficult to interpret how signals are encoded and how to interpret which signal a hidden node represent in a DLM. This black-box nature prevents researchers from gaining biological insights using DLMs, even though these models can be much superior in modeling data than other types of models in many tasks, e.g., predicting drug sensitivity of cancer cells. In this competitive renewal, we propose to develop novel DLMs and innovative inference algorithms to train “interpretable” DLMs and apply them in translational research. The proposed research is innovative and of high significance in several perspectives: 1) Our novel DLMs and algorithms take advantage of big data resulting from systematic chemical/genetic perturbations of cellular signaling machinery, so that we can use the perturbation condition as side information to reveal how signals are encoded in a DLM. 2) We integrate principles of causal inference and information theory with deep learning method to make DLMs interpretable. As results, that researchers can gain mechanistic insights from such models. 3) Innovative application of interpretable DLMs will advance translational research. For example, we will train interpretable DLMs to model cellular signaling at the level of single cells and use this information investigate inter-cellular interactions among cells in tumor microenvironment to shed light on immune evasion mechanisms of cancers. We will also use information derived from interpretable DLMs to predict cancer cell drug sensitivity. We anticipate that our study will bring forth significant advances not only in deep learning methodology but also in precision medicine.