Development of statistical machine learning algorithms for the manufacturing of mRNA vaccines

自新冠肺炎大流行爆发以来，信使核糖核酸疫苗已被广泛用于抵抗微博，具有显著的安全性。一种传统疫苗的新生产设施可能需要数年时间和巨额资金才能投入运营。相比之下，RNA疫苗是通过标准化的程序生产的，并进行了轻微的调整，以考虑到RNA序列长度和组成的变化。然而，在过去的几个时期里，辉瑞和莫德纳等公司由于制造中断而面临挫折。为了应对疫苗制造方面的挑战，自2018年以来，英国政府已向疫苗制造和创新中心投资近2.15亿GB，并进一步投资于CPI的RNA卓越中心。疫苗的制造是一个复杂的过程，需要过多的投入和质量措施。在基因疫苗中，起始材料是重组微生物细胞的质粒、宿主细菌和主细胞库。疫苗要经过质量控制，如核酸的完整性、含量、效力、产品和工艺相关的杂质、无菌、内毒素，以及pH和渗透压等物理化学测试。由于其化学成分，RNA通常是不稳定的，因此制造RNA疫苗涉及指示稳定性的参数，如RNA完整性、含量和效力，并辅之以pH、外观和微生物状态。最后，疫苗接受压力测试，包括温度变化、pH变化、光稳定性、湿度或多次冻融循环。我们假设，在如此复杂的制造过程中，数据科学方法目前被忽视了。特别是，我们希望重点了解在制造mRNA疫苗过程中的时间动力学。首先，该方案集中于开发方法，将线上和线下的时间数据结合起来，以预测过程结果并阐明过程生产率。在方法上，我们将建立在稀疏回归工具箱的基础上，以发现关键过程参数(CPP)，并确定这些参数如何影响关键质量属性(CQA)。我们的目标是将数学建模与建模/制造周期的迭代相结合。第二，工作方案建立在史蒂夫·布伦顿和内森·库茨开发的稀疏方法的基础上，以从数据中识别常微分方程和偏微分方程式。Brunton和Kutz的算法旨在使用稀疏回归工具箱从数据流中提取符号动态系统。他们的方法利用一组预定的状态空间变量的时间序列来描述未知的动力系统，并识别描述该系统的常微分方程组或偏微分方程组右侧的项的系数。这些项是从给定的候选基函数库中选择的。为此，作者将从时间序列中估计导数的数值方法与对基函数库执行变量选择的顺序阈值最小二乘算法(SINDY)相结合。他们的Sindy算法返回每个基函数的数值系数。如果数据集可以表示为候选基函数库中的稀疏动态系统，则大部分系数为零，并且Sindy识别非零系数。其结果是将输入数据集描述为在所选状态空间变量上表示的符号ODE或PDE，以及作为候选库函数的函数，其数值系数由Sindy找到。我们将分析由国家生物制品制造中心及其RNA卓越中心提供的生物反应器的数据集，该中心是工艺创新中心(CPI)的一部分。该工作计划位于EPSRC医疗保健技术、制造未来和数学科学主题范围内。

mRNA vaccines have been widely used to resist covid-19 since the outbreak of the pandemic with a remarkable safety profile. A new manufacturing facility of a traditional vaccine can take years and a huge amount of money to become operational. In contrast, RNA vaccines are manufactured by a standardised procedure with minor adaptations to account for variations in RNA sequence length and composition. However, over the last few periods, companies such as Pfizer and Moderna have faced setbacks due to manufacturing disruptions. To address the challenges in vaccine manufacturing, the UK government has invested almost £215 million in the Vaccines Manufacturing and Innovation Centre since 2018 with further investment in the CPI's RNA Centre of Excellence. The manufacturing of vaccines is a complex process with a plethora of inputs and quality measures. In mRNA vaccines, the starting materials are the plasmid, the host bacteria, and the master cell bank of the recombinant microbial cells. The vaccine undergoes quality controls such as integrity of the nucleic acids, content, potency, product and process-related impurities, sterility, endotoxin, and physicochemical tests like pH and osmolality. Due to its chemistry, RNA is generally unstable, and hence the manufacturing of mRNA vaccines involves stability-indicating parameters such as RNA integrity, content and potency, supplemented by pH, appearance, and microbiological status. Finally, the vaccine undergoes stress testing that can include temperature shifts, pH shifts, photostability, humidity, or numerous freeze-thaw cycles. We hypothesise that data science methods are currently overlooked in such a complex manufacturing process. In particular, we wish to focus on understanding the temporal dynamics of processes during the manufacturing of mRNA vaccines. First, the programme concentrates on developing methods to integrate online and offline temporal data to predict process outcome and shed light on process productivity. Methodologically, we will build on the toolbox of sparse regression to discover the critical process parameters (CPPs) and identify how those parameters influence critical quality attributes (CQA). Our goal is to integrate the mathematical modelling with iterations of the modelling/manufacturing cycles. Second, the work programme builds on sparse methods developed by Steve Brunton and Nathan Kutz to identify ordinary and partial differential equations from data. Brunton and Kutz's algorithm aims to extract symbolic dynamical systems from a data stream using the toolbox of sparse regression. Their method takes a time series of a predetermined set of state-space variables supposed to describe an unknown dynamical system and identifies the coefficients of the terms on the right-hand side of the ODEs or PDEs that describe the system. Those terms are selected from a given library of candidate basis functions. To do this, the authors combine numerical methods to estimate the derivatives from the time series with a sequential thresholded least squares algorithm (SINDy) that performs variable selection on the library of basis functions. Their SINDy algorithm returns numerical coefficients for each basis function. If the dataset can be represented as a sparse dynamical system in the library of candidate basis functions, most of the coefficients are zero, and SINDy identifies the non-zero coefficients. The outcome is a description of the input dataset as a symbolic ODE or PDE expressed on the chosen state-space variables and as a function of the candidate library functions, whose numerical coefficients are found by SINDy. We will analyse datasets from bioreactors provided by the National Biologics Manufacturing Centre and its RNA Centre of Excellence, part of the Centre for Process Innovation (CPI). The work plan resides within the EPSRC Healthcare Technologies, Manufacturing the Future and Mathematical Sciences themes.