Machine learning approaches to mine evolving 3D genomes

导致今天已知的基因组和物种多样性的进化过程可以被认为是一个长达30亿年的大规模平行选择实验。这项实验的结果，即现代物种的基因组，今天可以以低成本进行测序，数百种脊椎动物的基因组现在可以公开获得。我的研究计划的长期目标是开发和应用计算方法，以最好地利用令人难以置信的丰富数据集，以更好地了解现代基因组(如人类基因组)的调节功能。我们站在科学的三大趋势的汇合点上：(I)基因组测序能力的提高；(Ii)深度学习方法的发展，这种方法可以利用大数据集来识别模式并以前所未有的准确性预测功能。(3)认识到表观基因组学，特别是3D表观基因组学在调节基因表达方面发挥了驱动作用。在这一令人兴奋的时刻，我们的短期目标是：目标1：开发能够利用多物种比较的机器学习方法，以提高对各种类型的转录和转录后调控元件预测的准确性。我们将开发机器学习方法，将现有生物信息学工具对多个同源序列做出的预测结合起来，以提高准确性。然后，我们将开发新的半监督机器学习方法，直接在标记的基因组数据和未标记的同源区域的组合上训练深度学习方法。目的2：利用最大似然法改进哺乳动物进化史的重建。我们将使用深度学习和强化学习来提高全基因组比对、祖先基因组推断的准确性。我们还将专注于开发机器学习方法来从DNA序列预测细胞类型特定的3D基因组组织(间隔，TADS)，并将该方法应用于现代基因组和推断的祖先基因组。目的3：研究基因组环境的变化如何影响调节功能，以及如何最好地将其考虑到预测中。利用AIMS 1和2中开发的方法，我们将研究基因组重排和转座元件插入如何改变(预测)3D基因组构象和调节区的活性。然后，我们将修改在目标1中开发的机器学习模型，以使用有关3D基因组环境的信息来增强它们。总而言之，这项研究计划将开发尖端的机器学习，以帮助研究基因组进化，并从研究进化的基因组中学习识别功能。通过将进化论、基因组学、生物信息学和机器学习的方法结合在一起，它不仅提供了在生物学和计算机科学方面取得重大概念进展的机会，而且还将成为跨学科研究培训的一个很好的渠道。

The evolutionary process that led to the diversity of genomes and species known today can be thought of as a 3 Billion-year long, massively parallel selection experiment. The outcome of this experiment, the genomes of modern species, can today be sequenced at low cost, and the genomes of hundreds of vertebrates are now publicly available. The long-term objective of my research program is to develop and apply computational methodologies to best take advantage of that incredibly rich data set to better understand the regulatory function of modern genomes such as the human genome. We stand at the convergence of three major trends in science: (i) the increased ability to sequence genomes; (ii) the development of deep learning approaches that can take advantage of large data sets to identify patterns and predict function with an unprecedented accuracy. (iii) The realization that epigenomics, and in particular 3D epigenomics play a driver role in regulating gene expression. In this exciting conjuncture, our short-term objectives are: Aim 1: Develop machine-learning approaches that can take advantage of multi-species comparison to boost accuracy of prediction of various types of regulatory elements, both transcriptional and post-transcriptional. We will develop machine learning approaches that can combine predictions made by existing bioinformatics tools on multiple orthologous sequence to boost accuracy. We will then develop new semi-supervised machine learning approaches train deep learning approaches directly on a combination of labeled genomic data and unlabeled orthologous regions. Aim 2: Use ML to improve the reconstruction of mammalian evolutionary history. We will improve the accuracy of whole-genome alignment, ancestral genome inference using deep learning and reinforcement learning. We will also focus on developing machine learning approaches to predict cell-type specific 3D genome organization (compartments, TADs) from DNA sequence, and apply the approach to both modern genomes and inferred ancestral genomes. Aim 3: Study how changes in genomic context impact regulatory functions, and how best to take this into consideration for prediction. Taking advantage of the approaches developed in Aims 1 and 2, we will study how genome rearrangements and transposable element insertion alter (predicted) 3D genome conformation and the activity of regulatory regions. We will then revise the machine learning models developed in Aim 1 to augment them with information about 3D genome context. In summary, this research program will develop cutting-edge machine learning to help study genome evolution and learn to recognize function from studying evolving genomes. By bringing together approaches from evolution, genomics, bioinformatics, and machine learning, it not only offers the opportunity for significant conceptual steps forward in both biology and computer science, but will also serve as a great conduit for training in interdisciplinary research.