Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5886_Библиотеки_им_академика_М_И_Перельмана.pdf
Скачиваний:
0
Добавлен:
02.09.2026
Размер:
21 Мб
Скачать
440
J. Luo et al.
information and the prior knowledge via Bayes’ theorem. Bayes’ theorem describes the conditional probability of an event based on data as well as prior information or beliefs about the event or conditions related to the event.
There are increasing numbers of empirical studies that have demonstrated Bayesian methods are explicitly selected to better accommodate reduced sample sizes, ranging from life science, psychology, exercise science, cross-cultural stud­ies, and prevention science [8, 9]. Among many Bayesian methods, a framework that combines cause and effect diagrams with Bayesian belief networks (BBNs) to estimate causal relationships could be leveraged for the biomanufacturing process [10]. The framework translates qualitative insights from a cause-and- effect diagram into a closed-form relationship between inputs and outputs utilizing BBN models, simulation, and regression. The framework starts with a typical Fishbone diagram to construct a causal network, then estimate the probability of problem occurring using BBN and probabilistic elicitation, and nally identify the vulnerabilities of the net­work through simulation. This allows investigation practitioners to identify critical root causes of a given problem when data collection activities can often consume valuable resources or events occurring at a low frequency with a high risk of severity.
When a large enough data set is generated at a commercial scale, ADA is a tool that can complement traditional univariate monitoring programs to bolster knowl­edge, improve capability, and optimize bioprocess performance. Within the data science community, the use of statistical modeling to reach conclusions from data can be divided into two separate “cultures” [11]. One assumes that the data are gen­erated by a given stochastic data model. The other uses algorithmic models and treats the data mechanism as unknown. Because data collected during commercial manufacturing (where the process runs at or close to the target as designed) may involve complex systems with unknown physical, chemical, or biological mecha­nisms, signals from stochastic data models may be weak and difcult to interpret. Therefore, algorithmic models using ADA may elucidate previously undetected relationships between inputs and outputs, and hold promise to gain actionable insights and knowledge to better understand, predict, and control sources of varia­tion in a bioprocess [12].
While there are many ADA techniques available, two types of ADA models could be built based on different biomanufacturing applications, namely diagnostic (or root cause) models and predictive models. Both models provide information to gain process understanding. Specically, the root cause model (explanatory model) focuses on actionable inputs for process troubleshooting and improvement, and makes sure the insights from the “black box” models can be explained. When the model focuses on actionable process inputs, the newly identied cause could either be veried through small-scale lab studies or large-scale process changes and sub­sequently monitoring. The predictive model uses all available inputs for output pre­diction, regardless of the model’s interpretability. This could be very helpful in making early and/or real-time forward process decisions. It could also be extremely useful if there is an intrinsic correlation between process outputs. Using easy to
15 Advanced Data Analytics Application inBiomanufacturing Processes
detect process output to predict the hard-to-measure process output could signi­cantly improve the process efciency.
With the help from an external consultant company, an internal ADA program including a cross-functional team, IT infrastructure, data engineering technique, data visualization and modeling methodology is established. This article will describe the ADA program developed in-house and the established procedure to conduct the analysis. Specically, a case study to gain additional insights for an intractable cell culture performance problem is conducted. The model insights, the observed and predicted benet based on the model outcome, are discussed. The importance of building ADA capabilities to enable more efcient and reliable bio­processes is also discussed.
441

15.2 ADA Program Development

With the help from McKinsey/Quantum Black (McK/QB) consultant rm, an in­house ADA program was developed and established. The program includes a cross­functional team, established environmental and tool, business process to run individual projects (use cases), and a library of models. In addition, data standard­ization, digitization, and data sharing are key factors to ensure the success of the program.
15.2.1 Building aCross-Functional Team forADA Program
The team composed of global and local cross-functional experts is formed. The team members and their corresponding responsibilities are as follows:
Program leader: to set up program governance and to manage ADA engagement
Project leader/translator: to set and track progress, to provide guidance/steering
to the team, and to coordinate and facilitate interactions across disciplines
Technical subject matter experts (SMEs): to share expertise and help with
problem- solving and to help dene root causes/drivers based on insights from
analytics
Data engineer (DE): to work with source system owners to make data available
and to cleanse and link data in preparation for analysis
Data scientist (DS): to dene the analytic approach and execution of analysis
and to interpret model results
Information technology (IT)/technical liaison: to provide information on techni-
cal architecture, and to establish tools and standard, and to continuously opti-
mize and quality check for development and deployment
442
J. Luo et al.
15.2.2 IT Infrastructure andTools
New environments and tools are established for the ADA program (Table15.1). The platform is built on the Google Cloud Platform (GCP). Google Cloud Storage, BigQuery, and Dataproc are often used for data storage, data exploration, and high­performance distributed computing, respectively. Python is used as the primary pro­gramming language for data analysis. The Data Pipeline was introduced and built
Table 15.1 IT infrastructure and tools used in ADA program
Environment / tool Description Key capabilities built
Google Cloud Platform (GCP)
Dataproc Analytics engine for large-
Google Cloud Storage (GCS)
BigQuery Cloud data warehouse Load/Query data in BigQuery Python A programming language Develop and improve the codebase, follow
Jupyter Notebook
Kedro An open-source Python
GitHub Version control development
JIRA Agile sprint planning tool Manage tasks and collaborate with the virtual
Conuence Remote-friendly team
https://Cloud.google.com https://spark.apache.org/ https://Python.org https://Jupyter.org https://kedro.readthedocs.io https://github.com/ https://www.atlassian.com/software/conuence
A suite of cloud computing services that runs on the same infrastructure offered by Google
scale data processing Worldwide cloud storage that
can retrieve any amount of data at any time
Open-source web application for creating and sharing documents that contain live code, equations, visualizations, and narrative text
framework for creating reproducible, maintainable and modular data science code
platform
workspace
Learn/practice/improve on working in GCP (cluster, codebase, notebook, etc.). Run data engineering pipeline and data science analytical modules in GCP.Master basic DevOps (Spinning up a cluster, Terminal SSH, basic Linux commands)
Use Dataproc and Pyspark for distributed computing
Ingest data into the GCS environment
good coding practices, peer code review, test. Perform data analysis in CLI/Jupyter
Notebooks
Understand the Data Pipeline concept Create raw layers, intermediate layers, primary layers, feature layers, model input table of the data engineering pipeline
Understand Git workow. Update data engineering pipeline with Git
team in agile ways. Implement SCRUM and work in sprints as foundations
Use as a team workspace, knowledge sharing, and collaboration tool
15 Advanced Data Analytics Application inBiomanufacturing Processes
443
using Kedro, an open-source framework for creating reproducible, maintainable, and modular data science code. All codes are version controlled in Github. Jupyter Notebook is used to support data analysis and result sharing. The project is operated in an agile paradigm running sprints using JIRA.All the knowledge and documen­tation are stored in Conuence as a team workspace.
15.2.3 Project Approach
The program is executed through individual use cases. Analytics problem-solving requires translating a business problem (use case) into an analytics problem and solving it through an iterative process of hypotheses generation, exploratory analy­sis, and feature creation. Each project could be separated into ve phases: prepara­tion, initial analysis, model iteration, translating to actionable insights, and model outcome conrmation. Table15.2 summarized the detailed deliverables from each phase. Figure15.1 showed a map of the project approach to data transformation processes (i.e., data pipeline). Throughout the project, translators, Des, and DSs worked together with the process SMEs to dene hypotheses and select variables/ features (model inputs), and to prioritize the work based on the hypothesis and data availability. Among different use cases, the continuous improvement on the overall program platform and the model methods are important, and the reusability of the pipeline built from one use case to another should be maximized.
Table 15.2 ADA project phase deliverables
Phase Deliverable
Preparation Select use case
Align on scope and denition Develop detailed execution plan Set up the technical execution environment Dene initial hypothesis Identify/prioritize data source based on initial hypotheses Align on people/capabilities and roles
Initial analysis Dene model outputs
Model selection Data ingestion (data extract from data sources) Data clean/transformation & quality check (link data based on hypotheses) Feature creation Exploratory data analysis (EDA)
Model development and iteration
Model insights Build graphical representations of the Insights gained. Develop
Insights conrmation Conrm the insight by further lab study, additional monitoring after
Build model/Analytical engine Iterate could be based on new data sources, new hypotheses, or rened features
reports for a wide range of users
proposed change implementation
444
J. Luo et al.
Fig. 15.1 A map of the project approach and data transformation processes (i.e., data pipeline)
15 Advanced Data Analytics Application inBiomanufacturing Processes
445
15.2.4 Model Library
Machine learning is used to teach machines how to handle the data more efciently. Simply, machine learning can be better understood as “learning from data” [13]. There are two types of machine learning algorithms, unsupervised and supervised [14]. Unsupervised learning uses machine learning algorithms (e.g., K-means, PCA) to analyze and cluster unlabeled data sets. These algorithms discover hidden patterns in data without the need for human intervention (hence, they are “unsuper­vised”). It is mainly used for clustering, association, and dimensionality reduction. Supervised learning is a machine learning approach that’s dened by its use of labeled datasets. These datasets are designed to train or “supervise” algorithms into classifying data or predicting outcomes accurately. Using labeled inputs and out­puts, the model can measure its accuracy and learn over time. Currently, supervised machine learning methods are primarily used in the ADA program, which includes Linear regression, Lasso, Decision trees, Random forest, and Gradient boosting. Lasso, Random forest, and Gradient boosting are chosen because of their feature selection properties. Linear regression and Decision trees are used as the linear and nonlinear benchmarks for model comparison. Other methods such as generalized random forest have been studied and applied as appropriate.

15.3 Case Study

15.3.1 Case Study Selection forADA Modeling
The goal of the case study was to build a model to explain variation and identify actionable insights for a fed-batch Chinese hamster ovary (CHO)-based bioprocess process (Fig.15.2).
Fig. 15.2 Process owchart from seed train to large-scale cell culture process, including raw material and medium prep. HTST refers to high temperature short time
446
J. Luo et al.
Many factors were considered for the case study selection process, such as sam­ple sizes, the variability of the process output, the impact on business, the variability of data sources, data accessibility, and process knowledge. For example, to develop a meaningful ADA model, it was suggested to have a manufacturing history of at least 100 large runs. Second, if the variation in interest is observed throughout the process history, it may be a more suitable use case. Since cell culture processes typi­cally run in campaign mode, the campaign-to-campaign (raw material variability) or thaw-to-thaw variability is a common variation observed in bioprocess. Without variation observed in the historical data, there is nothing for the ADA model to optimize. In addition, where and how data were generated, captured, and stored will determine the data availability, data quality, and amount of effort for data engineering.
In this specic case, the selected process (Process A) generated more than 300 runs over the manufacturing history. As shown in Fig.15.3, 95% of all runs have normalized titer ranging from 0.77 to 1.22. Among the total observed variation, the within-campaign variation takes about 52%, and the between-campaign variation takes 48%. Table15.3 summarizes an example on process inputs and output for cell culture process. For process input, in addition to the typical primary process param­eter for characterization, the associated control strategy and secondary process parameters should also be included in the analysis.
The total number of runs available and relatively high historical variability made the process a good candidate for the analysis. In addition, most site data sources existed in digital format except for a subset of raw material release test results and paper-based executed batch records. The data availability will shorten the data prep­aration phase and enable more rapid problem solving. Based on the aforementioned
Fig. 15.3 Normalized historical titer performance of Process A by campaign (shown in different colors and separated by vertical lines) and thaw line (shown in different symbols). C# Campaign number
15 Advanced Data Analytics Application inBiomanufacturing Processes
Table 15.3 Upstream process inputs and outputs
Process input Attribute Attribute description Associated control/parameter
Thaw Time Bank storage, transfer operation Temp Setpoint, shift, shift timing Control loop, tank vs jacket temp pH Setpoint, shift, shift timing Deadband, probe calibration DO Agitation, sparge, overlay
ow rates
Medium Concentration Medium prep/transfer operation, medium
Feed Number, concentration,
volume, timing
Inoculation density
Culture duration Time Cell age Seed train maintenance, # of
Glucose Min/Max concentration Addition strategy Antifoam Addition strategy
Process output Category Attribute description
Cell state Viability, integrated growth (growth rate) Metabolite Lactate, ammonium, pCO Productivity Titer Purity Charge variants, size distribution, glycosylation Specic
modication Impurity CHOP, DNA, sequence variants Potency Potency by Anti-Proliferation, potency by Antibody-Dependent Cellular
Contaminants Bioburden, endotoxin, virus, mycoplasma
Viable cell density
nonselective passage
Deamidation, glycation, oxidation
Cytotoxicity (ADCC)
Aeration strategy
composition, raw material release specication
Medium prep/transfer operation
Solera operation, scale up strategy
2
447
factors, combined with a positive business value associated with understanding or reducing the performance variation, Process A was selected for the case study.
15.3.2 Data Sources Identication
A holistic review of the manufacturing process was conducted, and 11 data sources were identied that store manufacturing information. The data in 11 data sources were stored or extracted in different formats, such as Oracle databases, time-series data historian, scanned paper copy, and Microsoft Excel le.
448
J. Luo et al.
15.3.3 Hypothesis Generation
The generation of the hypotheses is a collaborative effort between the business, the analytical team, and domain experts. Hypotheses are at the heart of every project and inform how data is ingested and how analytics is performed. In this use case, more than 100 hypotheses were generated and prioritized among process, raw mate­rial, and test method areas.
Based on the importance of the hypothesis and data availability, data ingestion process was prioritized. For example, a key hypothesis for culture performance variation was associated with raw material variation [15]. Therefore, incorporating raw material test results into the ADA model was required. However, because raw material release test information was paper based, digitizing raw material data became a top priority early in the project. In addition, the analysis is to identify fac­tor to further improve titer/reduce titer variability; only process inputs are included in the model.
15.3.4 EDA andFeature Engineering
Once the raw data were available, data was cleaned, rearranged in smaller tables, and data from different tables were joined as a pipeline. Then data was ready for exploratory data analysis (EDA) and feature engineering. Once the feature was cre­ated, all features were joined to a master data table, and a model could be built based on the master data. Prior to running the model, Variance Ination Factor (VIF) was used for feature selection; this is to deal with multicollinearity between features and achieve more stable explanations of models. A VIF of 5 was used as a threshold for feature selection, which means the feature will be removed from model input if it can be explained 80% by other features. Based on the EDA or the initial model outcomes, additional data might be added to the pipeline or additional features could be created for model updates. As mentioned above, the overall pipeline was built on Kedro using Python [16].
15.3.5 Feature Engineering Example
For mammalian cell culture, pH setpoint and pH control are known factors to impact cell growth and productivity [17, 18]. Therefore, it is expected to build features based on pH information. Features created from pH information described below show an example of the data ingestion process. The online pH (PI data historian) for one inoculum train stage is the raw data from the manufacturing process. In addition to the culturing phase, the pH information before (medium only) and during inocu­lation (transfer cells from the previous culture to medium), and transfer out
15 Advanced Data Analytics Application inBiomanufacturing Processes
449
(inoculation for the next stage) could be extracted as different phases from data. In addition, during the culturing phase, the pH experienced a shift, which is due to the interaction between cell growth and pH control deadband. Therefore, pH during culture could be further partitioned into before, during, and after the shift. During culture transfer out, pH control was turned off, and the probe would be exposed to air at a certain level (tank-specic information). Linking other time-series data (e.g., tank volume or weight) and equipment information is, thus, required to get the transfer-out information. With this detailed process knowledge, the features were built for pH in one inoculum train culture stage as shown in Table15.4. As part of feature generation, EDA to review pH effect is performed to verify the correction of data ingestion and linkage and to understand if there is any effect.
Table 15.4 Data ingestion process for online pH in one inoculum train
Steps pH example Chart example
Raw data pH in one inoculum train stage
Intermediate data
Primary layer Based on pH and weight trend,
Feature For each pH phase, explore the
a
Might go back to raw data based on initial data extraction frequency
Associate pH with equipment data and tank weight (kg) to identify different phases of pH
separate pH trend in one culture stage into six different phases 1: medium batch 2: inoculation 3: pH at the top of deadband 4: pH transition from top to
bottom of dead band 5: pH at the bottom of deadband 6: Transfer out
pH: min, max, average, 25, 50, 75 percentile of pH For each pH phase, explore the duration, the ratio of each duration to overall duration For culturing phase, explore any pH excursion: spike, change due to online/ofine adjustment,
a
etc.
a
a
Max phase 3 pH for all runs in one inoculum train stage