Integrative Machine Learning Models for Discovery and Validation of Biological Knowledge

下一代测序技术的出现彻底改变了基因组、转录组和外显子组的研究方式，通常会产生大量需要分析的数据。在这方面，机器学习提出了应用于知识发现、检索和生物现象预测的成功方法。特别是深度学习，在生物学、医学、数据和网络安全、文本挖掘和计算机视觉等广泛的应用中，已经成功地应用于提取知识，这只是其中的几个。然而，在现有的数据和方法中存在一些挑战和限制，以及一些有待克服的障碍，包括缺乏对不同变量的注释，缺乏大规模和标记的训练数据，变量的多样性，不同数据集之间格式的变化，以及在提取的知识中缺乏样本特异性。******本研究计划旨在开发集成的机器学习系统，用于从多变量、多样本数据集和不同的生物指标中提取相关的生物学知识，包括“组学”、文献中的文本和图形。整合模型将涉及不同类型疾病的多个数据集、多组学和多个变量。在不同的子项目中，我们计划应用集成不同类型数据的多模式、多任务和迁移（深度和浅层）机器学习方法。待开发的方法将整合不同的表征深度学习方案，利用不同形式的训练，如对手、卷积和循环网络，有或没有记忆。******综合机器学习方法的发展尚未在多组学数据中显着出现，因此，将文本数据与不同形式的分子测量相结合是开发新方法的有前途的途径，这些方法也可以用于其他领域，例如数据安全，网络和增材制造，仅举几例。缺乏可靠的算法来整合和消除不一致的数据是至关重要的，因为它们是缺失的变量，因此，综合半监督方法在这方面是至关重要的。将开发的方法将被其他研究人员使用，通过共享出版物和一个将部署在标准生物信息学和开源平台上的系统，从大型数据集中发现生物知识。此外，预计该研究项目将培养1名博士后、3名博士生、数名硕士生和本科生，掌握大数据分析和工具平台软件开发的关键技能。

The advent of next generation sequencing has revolutionized the way the genome, transcriptome and exome are studied, typically producing huge amounts of data to be analyzed. In this regard, machine learning has proposed a paramount of successful methods applied to knowledge discovery, retrieval and prediction of biological phenomena. Deep learning, in particular, has been successfully applied to extract knowledge in a wide range of applications in biology, medicine, data and network security, text mining, and computer vision, just to mention a few. There are, however, some challenges and limitations in the available data and the current approaches, as well as some obstacles yet to overcome, including lack of annotation of different variables, lack of large-scale and labelled training data, variety of variables, variations of formats across different datasets, and lack of sample-specificity in the knowledge extracted.******This research program aims to develop integrative machine learning systems used to extract relevant biological knowledge from multiple variables, multiple datasets with large numbers of samples, and different biological indicators including “-omics”, text and graphics from the literature. The integrative model will involve multiple datasets, multi-omics and multiple variables of different types of diseases. In the different sub-projects, we are planning to apply multi-modal, multi-task and transfer (deep and shallow) machine learning approaches that integrate different types of data. The approaches to be developed will integrate different schemes of representational deep leaning that utilize different forms of training such as adversary, convolutional and recurrent networks, with or without memory. ******Development of integrative machine learning approaches have not emerged significantly in multi-omics data, and hence, incorporating textual data, integrated with molecular measurements of different forms is a promising avenue for developing novel approaches, which can be then used in other fields as well, such as data security, networking and additive manufacturing, just to mention a few. The lack of reliable algorithms for integrating and disambiguating inconsistent data are crucial, as they are for missing variables thus, integrative semi-supervised approaches are crucial in this regard. The methods to be developed will be used by other researchers in discovering biological knowledge from large datasets via sharing publications and a system that will be deployed in standard bioinformatics and open source platforms. In addition, it is expected that under this research program one postdoctoral fellow, three PhD students, and several Master's and undergraduate students will be trained, gaining key skills in big data analytics and software development of tools and platforms.