Kolmogorov complexity and algorithms for immunopeptidomics

我希望提供一个关于什么是信息的终极理解，并继续发明工具来推动免疫药剂学和免疫疗法的革命。因此，该建议分为两个部分：1)Kolmogorov复杂性2)免疫药剂学算法。第一部分，柯尔莫戈洛夫复杂性。我在科尔莫戈罗夫复杂性方面做了35年的研究。我们以《柯尔莫戈洛夫复杂性及其应用导论》为题撰写了一本权威的获奖著作。我将继续我之前在算法和信息距离的平均案例分析方面的工作。特别是，最简单的零射击学习类型可以用信息距离来表示。我还对使用柯尔莫戈洛夫复杂性分析深度神经网络很感兴趣。第二部分：免疫学的算法。我将更多地关注提案的这一部分。2017年，《自然·生物技术》的社论呼吁“个性化免疫疗法很流行，但新抗原的发现和验证仍然是一个令人望而生畏的问题。”在过去的4年里，我和我的学生们在这个问题上投入了大量的精力，用深度学习解决了一系列开放的问题。新抗原是由细胞表面的人类白细胞抗原分子呈递的一种短小的蛋白多肽。他们是直接的毒品目标。由于其丰度低，在个体化癌症免疫治疗的情况下，组织样本数量有限，因此在处理质谱学数据时，提高灵敏度是当务之急。这个问题要求我们在蛋白质组学方法的所有方面(数据库搜索、文库搜索和从头测序)做出重大改进，所有解决方案结合在一起将带来一个潜在的新抗原发现行业，特别是对于个性化的癌症免疫治疗，其中每个患者都需要个性化的新抗原发现。作为一个长期目标，我们将继续改进从质谱学数据中进行的肽从头测序。在过去的4年里，我们在PNAS[22]中引入了DeepNovo，并在自然方法中改进了它以用于数据独立获取(DIA)数据[21]。然后，我们引入了针对个别患者的由人类白细胞抗原多肽训练的个性化DeepNovo[21]和机器精度独立的DeepNovo[17]。沿着这一路线，我们希望继续改进DeepNovo，以处理更多的翻译后修改、具有强化学习的较长多肽和糖多肽。糖链从头测序是很困难的，因为糖链是树，而不是线性链。我们计划探索几种深度学习模型，例如图神经网络。除了从头测序，我们还将致力于以下问题：1)端到端文库/数据库搜索的高效深度学习解决方案；2)免疫原性的深度学习模型；3)改进单抗和多克隆抗体测序方法；4)应用我们的方法发现新的抗原；3)优化新冠肺炎依赖于人类白细胞抗原的T细胞疫苗的设计。

I wish to provide an ultimate understanding of what is information, and continue to invent tools to enable the revolution of immunopeptidomics and immunotherapy. Thus, this proposal is in two parts: 1) Kolmogorov complexity 2) Algorithms for immunopeptidomics. Part 1. Kolmogorov complexity. I have been doing research in Kolmogorov complexity for 35 years. We have written the authoritative and award-winning book on the subject "An introduction to Kolmogorov complexity and its applications". I will continue my previous work in average-case analysis of algorithms and information distance. In particular, the simplest type of zero-shot learning maybe formulated by information distance. I am also interested in analyzing deep neural networks using Kolmogorov complexity. Part 2. Algorithms for Immunopeptidomics. I will focus more on this part of the proposal. In 2017, Nature Biotechnology editorial appeals "Personalized immunotherapy is all the rage but the neoantigen discovery and validation remains a daunting problem." Over the past 4 years, I and my students have intensely worked on this problem, using deep learning to solve a series of open questions. Neoantigens are short protein peptides presented by HLA molecules on the surface of cells. They are direct drug targets. Due to their low abundance, and limited amount tissue samples in case of personalized cancer immunotherapy, it is imperative that we improve sensitivity when processing mass spectrometry data. The problem requires us to make significant improvements in all fronts of proteomics methods (database search, library search, and de novo sequencing) and all solutions together gives rise to a potential new industry of neoantigen discovery especially for personalized cancer immunotherapy, where each patient requires personalized neoantigen discovery. As a long-term goal, we will continue to improve peptide de novo sequencing from mass spectrometry data. Over the past 4 years, we have introduced DeepNovo in PNAS [22], and improved it for data-independent acquisition (DIA) data in Nature Methods [21]. We then introduced personalized DeepNovo trained by HLA peptides for individual patients [21] and machine precision independent DeepNovo [17]. Along this line, we wish to continue to improve DeepNovo to handle more post=translation modifications, longer peptides with reinforcement learning, and glyco peptides. Glycan de novo sequencing is hard, as glycans are trees, not linear chains. We plan to explore several deep learning models, such as graph neural networks. Beyond de novo sequencing, we will work on the following problems: 1) Efficient deep learning solutions for end-to-end library/database search; 2) Deep learning models for immunogenecity; 3) Improve monoclonal and polyclonal antibody sequencing methods; 4) Apply our methods to discover neoantigens optimize the design of an HLA dependent T cell vaccine for COVID-19.