Multimodal Machine-Learning and High Performance Computing Strategies for Big MS Proteomics Data

项目摘要/摘要 质谱(MS)数据是用于大规模系统生物学的高维数据 蛋白质组学。目前最先进的质谱仪可以从 一个单一的有机体和实验。这些高维数据使用数据库搜索进行处理 和Denovo算法都取得了不同程度的成功。这项研究的首要目标是 开发、测试、集成和评估新的图像处理和深度学习算法， 使我们能够以确定和定量的方式推断和鉴定可靠的多肽序列。我们的 长期目标是改进基于MS的蛋白质组数据的识别，使用新的 可扩展的算法。这项建议的目标是调查、设计和实施 机器学习深度学习算法用于从MS数据中识别多肽。自.以来 深度学习非常擅长在高维数据中发现错综复杂的结构 发现暗蛋白质组学数据和更准确地推断多肽的理想解决方案。我们 可以预测，这些方法与传统数值算法的结合将导致 提出了一种基于多模式融合的优化和准确的多肽推导方法 用于大规模MS数据的系统。此外，我们还将设计和实现数据增强， 内存高效的索引和高性能计算(HPC)可实现这些结果 计算时间更短，效率更高。因此，这条新的调查路线是 由于它有可能改进长期停滞不前的提高准确性的努力， MS数据分析和搜索工具的可靠性和重复性。最接近预期的 这项工作的成果是一套新的深度学习和图像处理工具，将使 更深入地了解基于MS的蛋白质组学数据。这一结果将产生重要的积极影响 立即产生影响，因为这些拟议的研究任务将为开发 新一类算法，并将提供快速、高吞吐量、灵敏和可重复的 基于MS的蛋白质组学的可靠工具。

Project Abstract/Summary Mass spectrometry (MS) data is high-dimensional data that is used for large-scale system biology proteomics. The current state of the art mass spectrometers can generate thousands of spectra from a single organism and experiment. This high-dimensional data is processed using database searches and denovo algorithms with varying degrees of success. The overarching objective of this study is to develop, test, integrate and evaluate novel image-processing and deep-learning algorithms that will allow us to deduce and identify reliable peptide sequences in a definitive and quantitative fashion. Our long-term goal is to improve on identification of MS based proteomics data using novel and scalable algorithms. The objective of this proposal is to investigate, design and implement machine-learning deep-learning algorithms for identification of peptides from MS data. Since deep-learning is very good at discovering intricate structures in high-dimensional data it will be ideal solution for discovering dark proteomics data and more accurate deduction of peptides. We predict that the integration of these methods, along with traditional numerical algorithms, will lead to a multimodal fusion-based approach for an optimized and accurate peptide deduction system for large-scale MS data. Further, we will design and implement data augmentation, memory-efficient indexing, and high-performance computing (HPC) to achieve these outcomes more efficiently with a shorter computational time. Therefore, this new line of investigation is significant since it has the potential to improve on long-stalled effort to increase accuracy, reliability and reproducibility of MS data analysis and search tools. The proximate expected outcome of this work is a novel set of deep-learning and image-processing tools which will allow much better insight in MS based proteomics data. The results will have an important positive impact immediately because these proposed research tasks will lay the groundwork to develop a new class of algorithms and will provide rapid, high-throughput, sensitive, and reproducible and reliable tools for MS based proteomics.