A blind source separation approach for deconvolution of bulk transcriptional data leads to early detection of ATF cell-states in complex bacterial populations, in vitro and in vivo

摘要--项目3 包括耐受性、持久性和异质性抗性(HR)在内的暂时性细菌细胞状态是 抗生素治疗失败(ATF)和抗生素耐药性的促进者。重要的是，他们在任何 目前采用的是诊断化验或药敏试验。耐人寻味的是，在治疗不同 对于不同类型的癌症，医生经常面临类似的治疗失败问题。事实证明，这些 表观遗传细胞状态为高水平耐药突变的出现创造了更多的机会。此外， 由于表型的暂时性，它们本身可以直接驱动(易感)的重新出现。 毒品压力消退后的人口。虽然这些细胞状态越来越被认为是坐着的驱动因素 在治疗失败的根源上，正在出现新的战略，以具体识别、跟踪和瞄准它们。至 为了实现这种高度靶向的治疗，开发出了描绘出复合体的组成的方法 癌症组织，例如通过单细胞RNA-Seq(scRNA-Seq)，或计算大块的去卷积 RNA-Seq数据。虽然细菌上的scRNA-Seq在技术上仍然具有挑战性，但我们发现通过修改 现有的工具，可以在复杂的细菌种群中识别特定的细菌细胞状态。然而， 当前工具的能力是有限的，并且通过实施最先进的机器学习 算法还有很大的改进空间。此外，ATF细胞状态的特征也很差，使其 目前不可能有效地定义它们。在此，追求三个目标，以开发一种基于 在大量的RNA-Seq数据上，将复杂的细菌种群解剖成其独立的细胞状态，并计算其 频率和麦克风。在目标1中，通过跟随广泛的时间序列来生成大量且多样化的时间RNA-Seq数据集 当他们接触抗生素时，菌株和物种的多样性，以及种群的子集转换为 ATF单元状态。在目标2中，探索了一种盲源分离算法来设计一台最先进的机器 一种学习工具，可以将大量的RNA-Seq数据从复杂的细菌群体中解卷到细胞状态和 他们组成人口的频率。此外，通过重构每个细胞状态的表达简档 我们实现了转录熵计算，从而实现了细胞状态特定的MIC预测。在目标3中， 该方法通过回顾预测患者样本中ATF细胞状态的存在而得到验证。最后， 将模型的适用性扩展到来自宿主和细菌的批量双RNA-Seq数据，并在 病人血清样本。因此，该项目不仅告知ATF细胞状态是如何发展和被 维持在种群中，但也创造了一条通往可以检测到它们的诊断方法的发展道路 在活动性感染中。再加上项目2的附带敏感性，这最终可能使链接成为可能 从检测到靶向治疗决策。

SUMMARY – PROJECT 3 Transient bacterial cell-states including tolerance, persistence and hetero-resistance (HR) are harbingers of antibiotic treatment failure (ATF) and enablers of antibiotic resistance. Importantly, they are missed in any currently employed diagnostic assay or antibiotic susceptibility tests. Intriguingly, in the treatment of different types of cancer, physicians are often confronted with similar treatment failure issues. It turns out that these epigenetic cell-states create extended opportunities for high-level resistance mutations to emerge. Moreover, due to the phenotype’s transience, they themselves can directly drive the re-emergence of the (susceptible) population after drug pressure subsides. While these cell-states are increasingly recognized as drivers that sit at the root of treatment failure, new strategies are emerging to specifically identify, track and target them. To achieve such highly targeted treatment, approaches are developed that map out the composition of complex cancer tissue, for instance through single cell RNA-Seq (scRNA-Seq), or computational deconvolution of bulk RNA-Seq data. While, scRNA-Seq on bacteria remains technically challenging we found that by modifying existing tools, specific bacterial cell-states can be identified in complex bacterial populations. However, the capabilities of current tools are limited, and through the implementation of state-of-the-art machine learning algorithms there is much room for improvement. Moreover, ATF cell-states are poorly characterized, making it currently impossible to effectively define them. Herein, 3 aims are pursued to develop an approach that, based on bulk RNA-Seq data, dissects a complex bacterial population into its separate cell-states, and calculates their frequencies and MICs. In Aim 1 a large and diverse temporal RNA-Seq dataset is generated by following a wide variety of strains and species while they are exposed to antibiotics and a subset of the population switches to an ATF cell state. In Aim 2 a blind source separation algorithm is explored to design a state-of-the-art machine learning tool that deconvolves bulk RNA-Seq data from a complex bacterial population into the cell-states and their frequencies that make up the population. Moreover, by reconstituting each cell-state’s expression profile we enable transcriptional entropy calculations and thereby cell-state specific MIC predictions. In Aim 3 the approach is validated by retrospectively predicting the presence of ATF cell-states in patient samples. Finally, the model’s applicability is extended to bulk dual RNA-Seq data from host and bacterium, and validated on patient serum samples. This project therefore not only informs on how ATF cell-states develop and are maintained in a population, but also creates a path towards the development of diagnostics that can detect them in an active infection. Combined with the collateral sensitivities from Project 2 this could eventually enable linking detection to targeted treatment decisions.