III: Small: Learning Multi-scale Sequence Features for Predicting Gene to Microbiome Function

微生物群落在健康和环境中起着至关重要的作用。 在人类健康中，它们被称为我们的微生物组;例如，健康的肠道微生物组可以帮助消化并有效地将食物转化为我们肠道中的营养物质。 然而，什么构成“不健康”（生态失调）微生物组以及它们如何影响身体（或环境）或受其影响尚不清楚。如果我们能够理解微生物之间以及与身体之间的相互作用，那么我们就可以设计出更好的治疗方法、疗法和药物（例如益生菌和益生菌）来操纵微生物组。 为了理解微生物生态系统的“规则”，我们必须首先解决基因型到表型的问题，即确定与微生物组功能/性状变化相关的微生物遗传变化。大多数研究人员只是专注于预测环境或疾病表型，仅使用微生物群落结构（即：群落中物种的人口普查），而不考虑详细的DNA/RNA差异。毫不奇怪，大多数研究都产生了适度的预测准确性，对微生物组如何发挥作用的了解很少。将生物体和/或基因的哪些“配置”归因于特定的“微生物组状态”可以帮助我们预测疾病，了解环境如何改变微生物生态系统，并能够预测这些系统的未来变化（例如，由于化学品，温度等引起的扰动）。可以同时在多个尺度（基因组，生物体和社区水平）学习相关特征的方法，需要解释“物种普查”和微生物遗传变化（可能导致物种形成和/或功能进化的突变），影响社区结构。我们的教育活动将为生物信息学相关课程的本科和研究生教育带来前沿研究和主题，这些课程是机器学习和生物信息学硕士课程的一部分，也是德雷克塞尔大学生物信息学学士学位的一部分。此外，我们计划组织工程的德雷克塞尔学院范围内的高中课外计划，为服务不足的公立学校的科学项目的指导。需要一个统一的算法来学习多个层面的微生物组特征，以便能够预测微生物组功能，从而识别导致重要“状态”（例如疾病或健康）的生物过程（也称为利用数据来理解生命规则，NSF Big 10目标）。这样做将改变我们对大规模和小规模变化如何影响微生物组表型的理解。目前的方法非常有限。基于16 S rRNA调查的表型预测通常仅在微生物操作分类单位（OTU）上进行，其很少捕获表示总体表型变化的突变。使用宏基因组的表型预测可能比16 S调查更好，但需要许多下游分析（特征选择，统计测试）来解释（例如推断与表型相关的亚群落）这种分类。 因此，我们建议开发一种递归神经网络（RNN），它可以学习微生物组的社区水平变化和与微生物组表型相关的遗传变化。虽然大多数神经网络可以“学习”特征，但通常很难从网络中获取这些信息（即： 解释）。我们还将使用基于注意力的RNN的最新进展，这将有助于我们解释哪些多尺度特征对表型预测最重要。 我们将把我们的算法和软件提供给微生物群落，其潜在应用包括改善农业、环境监测、个性化医疗等。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Microbial communities play vital roles in health and the environment. In human health, they are referred to as our microbiomes; for example, healthy gut microbiomes can help digest and efficiently convert food to nutrients to be taken up in our gut. However, what constitutes “unhealthy” (dysbiotic) microbiomes and how they can affect or be affected by the body (or environment) is unknown. If we can understand microbes’ interactions with each other and the body, then we can design better treatments, therapies, and medicines (e.g. pre- and pro-biotics) to manipulate microbiomes. To understand the ``rules'' for microbial ecosystems, we must first solve the genotype-to-phenotype problem, i.e. identify microbial genetic changes which correlate to changes in microbiome functioning/traits. Most researchers have simply focused on predicting environmental or disease phenotypes by solely using microbiome community structure (ie: a population census of species in a community) and do not consider detailed DNA/RNA differences. It is not surprising that most studies have yielded modest prediction accuracy and little understanding of how microbiomes function. Attributing which “configurations” of organisms and/or genes contribute to a particular “microbiome state” can help us predict disease, understand how the environment may change microbial ecosystems, and be able to predict future changes of these systems (e.g. perturbations due to a chemical, temperature, etc.). Methods that can learn pertinent features at multiple scales (genome-, organism-, and community-level) simultaneously, are needed to interpret both the “species census” and microbial genetic changes (mutations that may lead to speciation and/or functional evolution) that influence community structure. Our educational activities will bring cutting edge research and topics to undergraduate and graduate education in Bioinformatics-related courses, which are part of Machine Learning and Bioinformatics Masters programs and a Bachelor’s bioinformatics minor at Drexel University. In addition, we plan to organize a Drexel College of Engineering-wide high school extracurricular program for mentoring of science projects for underserved public schools. A unified algorithm is needed to learn microbiome features on multiple levels to be able to predict microbiome functioning, thereby identifying biological processes (a.k.a harnessing data to understand the rules of life, NSF Big10 goals) that result in important “states” (e.g. disease or healthy). Doing so will transform our understanding of how large- and small-scale changes influence microbiome phenotypes. Current approaches are highly limited. Phenotype prediction based on 16S rRNA surveys is usually conducted solely on microbial operational taxonomic units (OTUs), which rarely capture the mutations that signify overall phenotypic changes. Phenotype prediction using metagenomes may perform better than 16S surveys, but many downstream analyses (feature selection, statistical tests) are needed to interpret (e.g. infer subcommunities relevant to phenotype) this classification. Therefore, we propose to develop a recurrent neural network (RNN) that can learn both community-level changes in the microbiome and genetic changes that relate to microbiome phenotypes. While most neural networks can ``learn'' features, it is usually difficult to get this information back out of the network (i.e.: interpretation). We will also use the recent advances in attention-based RNNs that will help us interpret which multi-scale features are most important to phenotype prediction. We will make our algorithms and software available to the microbiome community, whose potential applications include improving agriculture, environmental monitoring, personalized medicine, among others.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.