A priori adaptive evolution predictions for antibiotic resistance through genome-wide network analyses and machine learning

摘要 适应性进化(AE)既是一种“善的力量”，因为它可以帮助优化工业中的生物过程，但 当传染病利用AE逃避宿主免疫系统或成为 对药物有抗药性。长期以来，人们一直认为几乎不可能对AE做出预测，因为 随机力和事件的主要影响。然而，观察到进化的重复性 跨性状和物种的现象比之前认为的要普遍得多，这表明，有了正确的数据和 方法，可能变得(部分)可预测。事实上，我们通过对细菌病原体的实验发现 肺炎链球菌对抗生素的反应及耐药性的出现，即 为了使声发射可预测，至少需要对细菌系统的两个方面有详细的了解： 1)系统的遗传限制(即生物体网络的结构)；和2.地点和方式 在系统中，压力被体验和处理。我们通过绘制出约25%的细菌的 网络，确定表型和转录抗生素反应，应用网络分析以捕获 并在网络环境中量化响应，并利用实验进化来定位自适应 基因组的突变通过机器学习，有可能发现隐藏在基因组中的模式 使声发射预测可行的数据。这意味着网络在与环境的相互作用中形成 适应性环境，它限制了可用的解决方案，并使某些解决方案比其他解决方案更有可能，从而 推动可重复性和实现可预测性。在这项提议中，我们建立在这些令人兴奋的发展之上 目标是找出肺炎链球菌整个网络的限制条件，并开发一种机器 可以在全基因组范围内先验地预测适应性进化的学习模型。要完成 为此，我们在目标1中结合了TN-Seq、DTN-Seq和Drop-Seq的几个部分，最终形成了一个新的工具TN-Seq^2(TN-Seq 平方)，能够在高通量和全基因组范围内绘制遗传交互作用图。我们使用TN-序列^2来 在20种抗生素存在的情况下重建首个肺炎链球菌全基因组遗传相互作用网络。 在目标2中，我们创建了85株HA标记的转录因子诱导(TFI)菌株，并：a)用CHIP-Seq 所有85个转录因子在肺炎链球菌中的DNA结合位点；b)通过过度表达每个TFI菌株，然后是RNA- 我们确定每个转录因子调控信号；c)使用转录调控因子诱导的表型筛选 在20种抗生素存在的情况下，解开环境中基因和转录之间的特定联系 扰动及其表型结果。最后，在目标3中，我们训练和测试了各种机器学习 设计最佳模型的方法，该模型预测基因组中哪些基因最有可能适应 存在特定的抗生素。这种预测声发射模型的发展，不仅将在预测方面有用 抗生素耐药性的出现，但这一策略应该对大多数生物学领域有价值 从生物工程到癌症，适应性变化都很重要。

SUMMARY Adaptive evolution (AE) is both a “force of good” as it can help to optimize biological processes in industry, but it is also a “force of frustration” when infectious diseases exploit AE to escape the host immune system or become resistant to drugs. It has long been assumed close to impossible to make predictions on AE due to the presumed predominating influences of random forces and events. However, the observation that evolutionary repeatability across traits and species is far more common than previously thought, suggests that AE, with the right data and approach, may become (partially) predictable. Indeed, we found through experiments with the bacterial pathogen Streptococcus pneumoniae on its response to antibiotics and the emergence of antimicrobial resistance, that in order to make AE predictable a detailed understanding of at least two aspects of the bacterial system are required: 1.) the genetic constraints of the system (i.e. the architecture of the organismal network); and 2.) where and how in the system stress is experienced and processed. We showed that by mapping out ~25% of the bacterium's network, determining phenotypic and transcriptional antibiotic responses, applying network analyses to capture and quantify the responses in a network context, and exploiting experimental evolution to pin-point adaptive mutations in the genome it becomes possible, by means of machine learning, to uncover hidden patterns in the data that make AE predictions feasible. This means that the network in interaction with the environment shapes the adaptive landscape, it limits available solutions and makes some solutions more likely than others, thereby driving repeatability and enabling predictability. In this proposal we build on these exciting developments with the goal to map out the constraints of S. pneumoniae's entire network and develop a machine learning model that can forecast adaptive evolution a priori, and on a genome-wide scale. To accomplish this, we combine in aim 1 parts of Tn-Seq, dTn-Seq and Drop-Seq to finalize a new tool Tn-Seq^2 (Tn-Seq squared) that is able to map genetic-interactions in high-throughput and genome-wide. We use Tn-Seq^2 to reconstruct the first genome-wide genetic interaction network for S. pneumoniae in the presence of 20 antibiotics. In aim 2 we create 85 HA-tagged Transcription factor induction (TFI) strains and: a) Determine with ChIP-Seq the DNA-binding sites for all 85 TFs in S. pneumoniae; b) By overexpressing each TFI strain followed by RNA- Seq we determine each TFs regulatory signature; c) Use a Transcriptional Regulator Induced Phenotype screen in the presence of 20 antibiotics to untangle environment specific links between genetic and transcriptional perturbations and their phenotypic outcomes. Lastly, in aim 3, we train and test a variety of machine learning approaches to design an optimal model that predicts which genes in the genome are most likely to adapt in the presence of a specific antibiotic. The development of this predictive AE model, will not only be useful in predicting the emergence of antibiotic resistance, but the strategy should be valuable for most any biological field for which adaptive changes are important, ranging from biological engineering to cancer.