Understanding bacterial resistance by machine learning from genetic data

动机——细菌对抗生素的耐药性仍然是医学上最大的挑战之一。尽管在过去十年中有更多的药物可用，但至关重要的是，这些药物的使用是通过新的实验室测试来优化的，这些测试可以为患者提供个性化的抗生素治疗。目前对感染患者的治疗通常从一种最佳猜测抗生素开始。同时，采集样本，实验室试图培养感染细菌。如果这是成功的，细菌将在体外暴露于各种抗生素，以确定哪种抗生素可能有效治疗感染。临床医生使用这些结果来适当地改变抗生素治疗。尽管这些测试既便宜又容易，但它们对复杂的潜在耐药性机制提供了一个狭隘的代表。全基因组测序继续变得更容易获得，现在在NHS中常规用于某些感染的管理。这项技术使实验室能够读取细菌的整个遗传密码，并用于管理疫情，在某些情况下还用于检测抗生素耐药性。然而，人们对基因组数据如何与传统细菌培养技术（表型测试）联系起来的理解很差，而传统细菌培养技术在预测抗生素治愈感染的可能性方面有着长期的记录。随着新型基因组测试产生的数据变得更容易获得，我们需要更好地了解如何利用它们来优化给病人使用的抗生素。过去的研究已经显示出有希望的结果和这种方法的可行性，然而，数据集仅限于某些生物体，如结核分枝杆菌。大多数研究以二元形式代表抗生素耐药性，而不是通过一系列敏感性。问题陈述-目的是发现细菌基因组的某些部分如何影响细菌对抗生素的耐药性。由于任何基因组都是一个长序列，而且我们已经有了许多感兴趣的细菌，因此需要一种机器学习方法来找到每个基因组相关条目与给定抗生素抑制细菌生长的最低抑制浓度之间的确切相关性。最终的目标是用更快的、数学上合理的算法预测抗生素的易感性，从而取代实验室中缓慢的实验，只使用细菌基因组。工作计划-在项目的早期阶段，候选人将与利物浦临床实验室（LCL）和利物浦大学基金会信托基金会合作，收集大约500种细菌分离株的代表性银行，这些细菌分离株是感染的重要原因。通过与基因组研究中心的合作，分离的细菌将使用最先进的下一代测序技术进行测序。将使用传统技术测试分离株对一组抗生素的敏感性，该技术将被用作金标准测试。其中一个挑战将是基因组测序产生的每个细菌分离物的高维度数据。因此，这些数据将需要使用已知的基因信息进行降维，以赋予对感兴趣的抗生素的抗性。从这个过程中产生的数据将被用作机器学习模型的输入，该模型将预测细菌对一组抗生素的敏感性——输出是抑制细菌生长所需的抗生素浓度（最小抑制浓度）。该研究将通过LCL提供的数据探索该模型与患者临床特征和结果的关系。预期成果：该项目将揭示细菌基因组与其对抗生素耐药性之间的明确关系。一个实际的结果将是一个实现的算法来可靠地预测最小抑制浓度

Motivation-Bacterial resistance to antibiotics remains one of the biggest challenges in medicine. Although more drugs have become available over the past decade, it is crucial that the use of these drugs is optimised through novel laboratory tests that allow individualised antibiotic therapy for patients.State-of-the-art-Current treatment of patients with infection usually starts with a best guess antibiotic. At the same time, samples are taken and the laboratory attempts to culture the infecting bacterium. If this is successful, the bacterium is exposed in vitro to various antibiotics to determine which antibiotics are likely to effectively treat the infection. Clinicians use these results to change antibiotic treatment as appropriate. Although these tests are cheap and easy, they provide a narrow representation of complex underlying resistance mechanisms.Whole genome sequencing continues to become more accessible and is now routinely used in the NHS in the management of certain infections. This technology allows laboratories to read the whole genetic code of bacteria and is used for the management of outbreaks and, in some cases, to detect antibiotic resistance. However, there is a poor understanding of how genomic data are linked to traditional bacterial culture techniques (phenotypic tests), which have a long track record of being able to predict the likelihood of an antibiotic to cure an infection.As data produced by novel genomic tests become more accessible, we need a better understanding of how they can be used to optimise what antibiotics to give to patients. Past research has shown promising results and plausibility of this approach, however, datasets are limited to certain organisms such as Mycobacterium tuberculosis. Most studies represented antibiotic resistance in a binary form, rather than by a range of susceptibility.Problem statement-The aim is to discover how certain parts of bacterial genomes affect the bacterial resistance to antibiotics. Since any genome is a long sequence and we already have many bacteria of interest, a machine learning approach is needed to find exact correlations between relevant entries of every genome and a minimum inhibitory concentration of given antibiotic to suppress bacterial growth. The ultimate goal is to replace slow experiments in the lab by faster and mathematically justified algorithmic predictions of susceptibility to antibiotics by using only the bacterial genome.Work plan-In the early stages of the project, the candidate will collaborate with Liverpool Clinical Laboratories (LCL) and Liverpool University Foundation Trust to collect a representative bank of approximately 500 bacterial isolates that are important causes of infection.Through collaboration with the Centre for Genomic Research, the bacterial isolates will be sequenced using state-of-the-art Next Generation Sequencing techniques.The isolates will be tested for susceptibility to a panel of antibiotics using traditional techniques that will be used as the gold-standard test.One of the challenges will be the high dimensions of data produced for each bacterial isolate from genome sequencing. Therefore, the data will need to undergo dimensionality reduction using information about genes known to confer resistance to the antibiotics of interest.The data produced from this process will be used as an input to a machine learning model that will predict susceptibility of bacteria to a panel of antibiotics - with the output being a measure of antibiotic concentration required to suppress bacterial growth (minimum inhibitory concentration).The study will explore the relationship of this model to patient clinical features and outcomes, through data available from LCL.Expected Deliverables-This project will reveal an explicit relationship between genomes of bacteria and their resistance to antibiotics. A practical outcome will be an implemented algorithm to reliably predict a minimum inhibitory concent