Blending deep learning with probabilistic mechanistic models to predict and understand the evolution and function of adaptive immune receptors

项目摘要 对适应性免疫受体（即抗体和T细胞受体）的科学理解有可能 彻底改变疾病的预防、诊断和治疗。高通量DNA测序和 功能性实验将适应性免疫受体的研究带入了大数据时代。到 要实现这些数据的这种潜力，就必须与适当的强大分析技术相匹配。 现有的概率和机械模型不足以捕捉这些数据的复杂性，而 简单的机器学习应用无法利用我们对免疫系统的深刻现有知识， 系统 该项目的目标是将深度学习与机械建模相结合，以预测和 了解适应性免疫受体的进化和功能。目标1：开发生成模型 捕获真实的适应性免疫受体库的复杂性的免疫受体序列。这些 我们将联合收割机与深度学习沿着与我们对VDJ重组的知识相结合，并提供一个严格的平台 进行详细的曲目比较。目的2：建立抗体体细胞定量机理模型 包括潜在生化过程的超突变。估计棘手的可能性，使用 深度学习来推断重要的潜在变量，并使用细胞中的敲除实验来验证模型 线目标3：开发混合深度学习模型，从序列数据中预测结合特性， 将大量实验衍生的结合数据与来自人类的更大的免疫序列集相结合 免疫记忆样本通过3D卷积或基于距离的惩罚来整合结构信息。 这些工具将揭示免疫库数据在医学应用中的全部力量。我们将获得更多 通过它们在相关空间中的分布对剧目进行严格的比较。这些将揭示的影响， 免疫干扰，如疫苗接种和疾病，使我们能够挑选出受影响的序列 这些扰动。我们将对体内体细胞超突变有更深入的定量了解， 统计模型，适当地捕捉突变集合的长期影响。我们也将有 能够将联合收割机指令集数据和稀疏结合数据组合以预测结合特性的算法。 总之，这些进展将使合理的疫苗设计，治疗自身免疫性疾病， 鉴定作为癌症免疫疗法的有希望的候选者的T细胞。

Project Summary Scientific understanding of adaptive immune receptors (i.e. antibodies and T cell receptors) has the potential to revolutionize prophylaxis, diagnosis, and treatment of disease. High‐throughput DNA sequencing and functional experiments have now brought the study of adaptive immune receptors into the big‐data era. To realize this potential of these data they must be matched with appropriately powerful analytical techniques. Existing probabilistic and mechanistic models are insufficient to capture the complexities of these data, while a naïve application of machine learning cannot leverage our profound existing knowledge of the immune system. The goal of this project is to blend deep learning with mechanistic modeling in order to predict and understand the evolution and function of adaptive immune receptors. Aim 1: Develop generative models of immune receptor sequences that capture the complexity of real adaptive immune receptor repertoires. These will combine deep learning along with our knowledge of VDJ recombination, and provide a rigorous platform for detailed repertoire comparison. Aim 2: Develop quantitative mechanistic models of antibody somatic hypermutation that incorporate the underlying biochemical processes. Estimate intractable likelihoods using deep learning to infer important latent variables, and validate models using knock‐out experiments in cell lines. Aim 3: Develop hybrid deep learning models to predict binding properties from sequence data, combining large experimentally‐derived binding data with even larger sets of immune sequences from human immune memory samples. Incorporate structural information via 3D convolution or distance‐based penalties. These tools will reveal the full power of immune repertoire data for medical applications. We will obtain more rigorous comparisons of repertoires via their distribution in a relevant space. These will reveal the effects of immune perturbations such as vaccination and disease, allowing us to pick out sequences that are impacted by these perturbations. We will have a greater quantitative understanding of somatic hypermutation in vivo, and statistical models that appropriately capture long‐range effects of collections of mutations. We will also have algorithms that will be able to combine repertoire data and sparse binding data to predict binding properties. Put together, these advances will enable rational vaccine design, treatment for autoimmune disease, and identification of T cells that are promising candidates for cancer immunotherapy.