Artificial intelligence and machine learning‐assisted digital applications for biopharmaceutical manufacturing

Abstract

Artificial intelligence and automation are no longer just buzzwords in the biopharmaceutical industry. The manufacturing of a class of biologics, comprising monoclonal antibodies, cell therapies, and gene therapies, is far more complex than that of traditional small molecule drugs. Therefore, applications based on artificial intelligence are essential for successfully manufacturing this new class of biologics more quickly and more economically. Some biologics manufacturers, academic researchers, and young entrepreneurs have already begun implementing artificial intelligence‐based applications to increase operational efficiency, enhance process understanding, improve process monitoring, and achieve better regulatory compliance. Regulatory guidance from health agencies on the use of artificial intelligence and machine learning is acting as a catalyst in the adoption process of these new technologies by the biopharmaceutical industry. Research in artificial intelligence and machine learning has also advanced significantly in the last decade. At the same time, new cloud technologies have made the development and deployment of machine learning applications much easier. Several examples of artificial intelligence and machine learning applications in monoclonal antibodies manufacturing already exist. Cell and gene therapy, which present the future of medicine, will also benefit from this new technology. Overall, advancements in this domain will essentially help better serve patients' needs.

1. NEED FOR AI/ML IN BIOPHARMACEUTICAL MANUFACTURING

The manufacturing industry has observed significant advancements in artificial intelligence (AI) and machine learning (ML) driven applications over the last decade or so, which have been popularly termed the Fourth Industrial Revolution. ¹ According to market research conducted in 2022–2023, the global market size for pharmaceutical manufacturing is $516.48 billion, ² which is roughly 3.8% of the $13.5 trillion market size of the overall global manufacturing sector. ³ The global pharmaceutical manufacturing market size is expected to grow at a compound annual growth rate (CAGR) of 7.63% from 2023 to 2030. Specifically, biopharmaceutical sales, which include products derived from living organisms such as vaccines, monoclonal antibodies, and cell and gene therapies, are predicted to surpass small molecules by $120 billion by 2027, which will account for roughly 55% of all drug sales. ⁴ Monoclonal antibodies dominate the biopharmaceutical landscape. ⁵ The global cell and gene therapy market is expected to grow from $10.7 billion in 2022 to $78 billion by 2032 at a CAGR of 22.6%. ⁶ The expected growth of biologics is also desirable, as biologics offer advantages that small molecule drugs cannot. However, biologics development and manufacturing are far more challenging. Therefore, advancement in technology is needed to ensure the quality and consistency of the manufacturing process. The availability of trained personnel, specialized equipment, robotic automation and the application of AI/ML should be some of the focus areas for the biologics sector in the coming years.

For software/technology sector companies, AI/ML‐derived software is their product, but pharmaceutical companies still have to deliver “real” physical products in the form of drugs. The pharmaceutical companies may leverage AI/ML to increase their efficiency, quality, and consistency and reduce time to market significantly, but it is still not mandatory to use AI/ML to produce quality drugs. Therefore, many pharmaceutical companies are still operating in traditional settings. Adopting technology/software industry's startup culture, many biotech startups have emerged in recent times. Globally, these new biotech startups are leveraging AI/ML to accelerate drug discovery, streamline drug development, speed up clinical trials and improve manufacturing efficiency. ⁷ , ⁸ The promising work done by successful startups is attracting funding from venture capitalists worldwide. ⁸ , ⁹ Some large pharmaceutical companies have realized the potential of AI/ML and are collaborating with biotech startups and other large companies from the technology sector. ¹⁰ , ¹¹ Government grants and programs are also providing support to universities and companies for conducting research on AI/ML usage for health and biomedical research. ¹² , ¹³

2. SUPPORTING TECHNOLOGIES FOR AI/ML IMPLEMENTATION

2.1. Overview of AI/ML algorithms

Artificial intelligence (AI) is a broad field based on the idea that machines can mimic human intelligence and perform tasks such as pattern recognition, understanding languages, and making decisions. Machine learning (ML) is a subset of AI that focuses on developing algorithms to create systems that can learn from the input data and perform tasks that typically require human‐like intelligence. Depending upon the intent of the application, ML techniques can be categorized into three categories: unsupervised, supervised, and reinforcement learning.

Unsupervised learning refers to a collection of algorithms used to find patterns and groupings in the data to improve the understanding of the data. In contrast, supervised learning focuses on training a system to establish relationships between input and output variables. Reinforcement learning is the newest of the three and involves an interplay between agent, actions, the environment, and rewards. While exploring the environment, the agent takes an action, receives rewards or punishments for that action, and then updates its knowledge to improve decision‐making.

A large number of algorithms are available to handle various types of data. The lack of data storage and computational resources is no longer a limitation, thanks to relatively inexpensive and flexible cloud services. In the last decade, biologics manufacturers either had to hire people with expertise in computer programming to write code for ML algorithms from scratch or buy expensive black‐box commercial software for data analysis. With the easy availability of ML algorithms through open‐source libraries such as, Scikit‐learn, ¹⁴ Keras, ¹⁵ TensorFlow, ¹⁶ PyTorch, ¹⁷ XGBoost, ¹⁸ the additional cost of hiring or purchasing software can be reduced. Despite this ease of availability, the selection of an ML algorithm to solve the problem at hand is not straightforward as each algorithm has its own strengths and limitations. Therefore, a careful and detailed study of these algorithms prior to application is essential. Tables 1 and 2 list the commonly used classical and advanced ML algorithms in the biopharmaceutical industry with their applications, strengths and limitations. As indicated in Tables 1 and 2, the selection of an algorithm also depends on the use case and dataset limitations. More advanced ML algorithms typically require a large amount of data. Therefore, their applications in the early phase of development may not be realistic due to the unavailability of large amounts of data unless high‐throughput methods are used to generate a large amount of data. Although most of the advanced ML algorithms can also be interpreted using explainable AI ¹⁹ concepts to some extent, process scientists still prefer classical algorithms if the difference in performance between the two types of algorithms is not practically significant. This preference aligns with regulatory expectations that highly emphasize the importance of interpretability in model performance.

TABLE 1.

List of commonly used classical ML algorithms with their intended use, strengths, and limitations.

Algorithm	Application area in Biopharmaceuticals	Strength	Limitation
Linear Regression	Biological assays, Drug stability data analysis 20 , 21	Simple to implement, interpretable, works well with linearly separable data	Assumes linearity, sensitive to outliers, Not suitable for system with multicollinearity
Logistic Regression	Pathogen safety evaluation, 22 biological assays 23	Easy to implement, interpretable, good for binary outcomes	Assumes linearity, not suitable for system with multicollinearity
Principal Component Analysis (PCA)	Process analytical technology data analysis 24 , 25 , 26 , 27	Reduces dimensionality while preserving variance	Assumes linearity, sensitive to outliers
Partial Least Squares (PLS) and variants	Process analytical technology data analysis, 28 , 29 , 30 , 31 Cell culture 32 , 33 , 34	Handles multicollinearity well, good for predictive modeling	Assumes linearity, sensitive to noise
Decision Trees	Biomanufacturing facility fit prediction 35	Easy to interpret, handles both numerical and categorical data	Prone to overfitting, sensitive to data variations
Random Forest	Chemometric model and product quality forecasting for cell‐culture, 36 , 37 , 38 chromatography for mAbs 39	Reduces overfitting, handles large datasets well	Less interpretable than decision trees, can be slow
Gradient Boosting	Process optimization 40	High predictive accuracy, handles various data types	Can be prone to overfitting, requires careful tuning
Support Vector Machines	Process performance prediction, 41 process development optimization for mAbs 42	Effective in high‐dimensional spaces, robust to overfitting	Slow with large datasets, requires careful tuning
K‐Means Clustering	Supply chain planning for pharmaceuticals, 43 online process monitoring, 44 Raman spectroscopy analysis for protein drug development 45	Simple and efficient, works well with large datasets	Not suitable for system with multicollinearity, sensitive to outliers

TABLE 2.

List of advanced ML algorithms with their intended use, strength and limitation.

Algorithm	Application area in Biopharmaceuticals	Strength	Limitation
Feedforward Neural Networks (FNN)	Drug repositing and repurposing, 46 Process intensification 47	Simple architecture, effective for basic tasks	Limited to simple relationships, prone to overfitting
Convolutional Neural Networks (CNN)	Flow imaging microscopy, 48 , 49 , 50 Process analytical technology data analysis, 51 , 52 Single‐cell cloning 53	Excellent for spatial data, reduces parameters through shared weights	Requires large datasets, computationally intensive
Recurrent Neural Networks (RNN)	Biopharmaceutical upstream process 54	Effective for sequential data, can capture temporal dependencies	Prone to vanishing gradient problem, less effective for long sequences
Long Short‐Term Memory Networks (LSTM)	Digital twin for cell culture, 55 Predictive maintenance, 56 Raman spectroscopy 57	Overcomes vanishing gradient issue, better for long sequences	More complex architecture, requires more data
Gated Recurrent Units (GRU)	Predictive maintenance, 56 Digital twin for cell culture forecasting 58	Simpler than LSTMs, effective for many sequential tasks	May not capture long‐term dependencies as well as LSTMs
Autoencoders	Fault detection 59	Effective for unsupervised learning, can learn efficient representations	Can overfit, requires careful architecture design, high sensitivity to hyperparameters
Reinforcement Learning	Chromatography optimization, 60 cell therapy manufacturing process control 61 , 62	Learns optimal actions through trial and error	Requires a lot of data, can be complex to implement

2.2. Data pre‐processing techniques

Owing to technological advancements in the computer science domain, tabular data comprising numeric values is not the only type of data that can be analyzed using ML algorithms. ML algorithms are available to analyze any type of data, including text, images, audio, video, time‐series in addition to traditional tabular data. It is true that ML algorithms only understand numerical values; therefore, all types of data need to be pre‐processed before an ML‐based model can be developed. Missing value imputation, categorical variable encoding to convert categorical to numerical, standardization, dimensionality reduction, outlier detection, and data transformation are some of the commonly used pre‐processing methods for any kind of data analysis. However, depending on the data type, additional pre‐processing steps are required. Figure 1 lists the commonly used pre‐processing steps required for different types of data analyses.

FIGURE 1.

Commonly used pre‐processing steps for (a) text ⁶³ , ⁶⁴ , ⁶⁵ (b) image ⁶⁶ , ⁶⁷ , ⁶⁸ , ⁶⁹ , ⁷⁰ , ⁷¹ , ⁷² , ⁷³ , ⁷⁴ , ⁷⁵ (c) video ⁷⁶ , ⁷⁷ , ⁷⁸ (d) time‐series ⁷⁹ , ⁸⁰ , ⁸¹ , ⁸² , ⁸³ data analyses.

2.3. Role of data systems and cloud technologies

2.3.1. Traditional on‐premises data systems

The prerequisite for an AI/ML implementation is good quality data. Therefore, data systems become an essential part of this digital ecosystem. Given, the highly regulated nature of the pharmaceutical industry, data integrity has always been the strict requirement under Good Manufacturing Practice (GMP) adopted by pharmaceutical companies. Traditionally all types of manufacturing data are captured in the form of paper batch records and multiple people sign those paper records to ensure the data integrity. Capturing data in this format may be an easy solution for the short term. However, the life‐cycle management of these paper records, data retrieval and usage for any type of analysis is always challenging. The challenge gets worse with the increase in the amount of data. To circumvent these challenges and be compliant with the regulations such as FDA 21 CFR Part 11 ⁸⁴ and EU GMP Annex 11, ⁸⁵ many large pharmaceutical companies started investing in building on‐premises IT infrastructure and moving towards electronic batch records (EBR) in the early 2000s. The implementation of Manufacturing Execution Systems (MES), which support EBRs was a step in the right direction. ⁸⁶ MES acts as a bridge between Enterprise Resource Planning (ERP) and Distributed Control System (DCS) on the production floor. MES enables the availability of manufacturing data for various purposes such as process monitoring, production planning, quality control, etc. While MES plays a crucial role in managing the manufacturing process, the system also has some limitations. One of the main limitations is access to the data for scientists with no computer programming experience. Therefore, data integrator systems like BIOVIA Discoverant ⁸⁷ are needed to ensure that data can be accessed by laboratory and process scientists via a graphical interface. ⁸⁸ The data integrator system also enables mapping across different unit operations in the manufacturing process. While scientists can access at‐line and offline measurements such as potency measurements, metabolite concentrations, etc. through data integrator systems like BIOVIA Discoverant, it is not intended to be used for real‐time data collection from operational systems like DCS. The PI system ⁸⁹ is widely used in the pharmaceutical industry as a data management platform to collect and historize the real‐time manufacturing data generated from inline sensors and equipment. The PI system provides easy access to real‐time data from DCS through its visual interface. It also provides a web API interface for scientists with programming backgrounds to query the real‐time and historical data. The PI system can be used for all parts of the manufacturing process be it upstream cell culture process or downstream purification process. ⁹⁰ , ⁹¹ , ⁹² , ⁹³ , ⁹⁴ , ⁹⁵ Data systems like Discoverant and PI often rely on on‐premises IT infrastructure, which means not only the hardware such as application (CPU, RAM, etc.) and database servers, storage solutions, networking equipment, power systems, etc., but also the human resources to manage this hardware structure. Usually the costs associated with these initial capital investments for hardware, regular maintenance and upgrades done by personnel constitute a major part of IT expenses. In addition to these costs, there will always be an increased demand for hardware to store more data and add more servers for improving accessibility to the end users.

2.3.2. Promising cloud‐based solutions

In this competitive era, pharmaceutical companies would want to reduce the associated overall IT infrastructure cost. Cloud platforms such as Amazon Web Services (AWS), ⁹⁶ Google Cloud Platform (GCP) ⁹⁷ and Microsoft Azure ⁹⁸ can help pharmaceutical companies in this effort to reduce IT costs. These cloud service providers operate on a subscription basis, which has the potential to reduce the costs associated with initial high hardware setup. Regular security updates and maintenance are also handled by these cloud providers. Scalability and flexibility are the two biggest advantages of using cloud services. On‐demand scaling of computational and memory resources provides great flexibility to companies. In a traditional on‐premises setting, the rapid deployment of even a small prototype AI/ML application requires a significant amount of collaborative work between scientists and IT teams. With the cloud setup, this kind of rapid deployment is much easier as the cloud platform provides several services catering to the needs. The cloud platform also enhances the collaborative experience of scientists working in different geographical locations as the data and applications can be accessed through the Internet. Although the cloud services are responsible for data security, it is the company's responsibility to ensure that the cloud providers follow regulations such as FDA 21 CFR Part 11 and provide an extra layer of security for the sensitive patient data.

In summary, cloud technology is needed to modernize manufacturing to improve efficiency and innovation. ⁹⁹ Biologics manufacturers have started leveraging cloud platforms to deploy their digital applications for widely providing access to process scientists and improving the efficiency of manufacturing operations. ¹⁰⁰ , ¹⁰¹ Cloud platforms' usage is not only limited to drug manufacturing, but they are also becoming very useful for effectively managing laboratory operations during the drug development phase. ¹⁰² , ¹⁰³

As mentioned above, many cloud service providers are providing similar services for computation, storage, etc. A categorization of different services provided by three major cloud providers (AWS, GCP, Azure) is summarized in Table 3 for the reader's knowledge. As shown in Table 3, similar services are offered by different cloud providers, and one or multiple providers can be selected to fulfill the needs of a biopharmaceutical company. These major cloud platforms provide a great deal of flexibility to cater to different types of industries. However, configuring these cloud systems for a specific use case may sometimes require a substantial amount of IT/cloud experience. On one hand, the upfront cost of investment in infrastructure will come down due to cloud platforms' usage, but the hiring cost of experienced IT/cloud professionals may increase. To address this business dilemma, alternative cloud platforms ¹⁰⁴ have emerged that offer more customized and easier‐to‐use solutions, which may not require extensive IT/cloud expertise. However, these alternative platforms may be more expensive than the services provided by major cloud providers. Ultimately, cloud platform usage cost and hiring cost may drive the selection of a cloud platform.

TABLE 3.

Categorization of different services provided by major cloud service providers along with their potential use cases.

Category	Amazon Web Services (AWS)	Google Cloud Platform (GCP)	Microsoft Azure	Potential Use Case in Biopharmaceutical industry
Storage	S3 (Simple Storage Service)	Cloud Storage	Azure Blob Storage	Used for storing large datasets (in any format), e.g. genomic data, cell images, etc.
	EBS (Elastic Block Store)	Persistent Disk	Azure Disk Storage	Provides high‐performance storage for applications requiring fast access to data
	Amazon FSx (File System)	Filestore	Azure Files	Facilitates easy and secured backup, of on‐premises file storage to meet regulatory, data retention, or disaster recovery requirements.
Computation	EC2 (Elastic Compute Cloud)	Compute Engine	Azure Virtual Machines	Powers high‐throughput computing tasks such as Monte‐Carlo simulations
	AWS Lambda (Serverless Computing)	Cloud Run Functions	Azure Functions	Enables event‐driven processing for tasks like machine learning model inference
	AWS Batch	Google Kubernetes Engine (GKE)	Azure Kubernetes Service (AKS)	Manages batch processing of large‐scale data loads
Machine Learning	Amazon SageMaker	Vertex AI	Azure Machine Learning	Platform for, training, and deploying machine learning models
	Amazon Comprehend (NLP)	Natural Language AI	Azure AI Language	Analyze unstructured text data, such as paper records
	Amazon Rekognition (Computer Vision)	Cloud Vision AI	Azure AI Vision	Useful in analyzing images such as cell‐images
Data Analytics	Amazon Redshift (Data Warehouse)	BigQuery	Azure Synapse Analytics	Enables large‐scale enterprise structured data analytics for making critical business decisions
	AWS Glue (ETL)	Dataflow	Azure Data Factory	Facilitates data integration and transformation, ensuring data quality and consistency
	Amazon Athena (Query Service)	Dataproc	Azure Databricks	Enables queries on large datasets, such as genomic data
Networking	Amazon Virtual Private Cloud (VPC)	Virtual Private Cloud (VPC)	Azure Virtual Network	Provides a secure environment for sensitive data processing in compliance with regulations
	AWS Direct Connect	Cloud Interconnect	Azure ExpressRoute	Ensures reliable and secure connections between on‐premises data centers and cloud resources for sensitive data transfers
Security	AWS Identity and Access Management (IAM)	Cloud Identity	Microsoft Entra ID	Manages user access and permissions to sensitive data and applications
	AWS Key Management Service (KMS)	Cloud Key Management	Azure Key Vault	Protects sensitive data by managing encryption keys, essential for maintaining data integrity and security
Compliance	AWS Compliance Programs	Compliance Offerings	Azure Compliance	Supports adherence to regulatory standards like FDA 21 CFR Part 11, ensuring data security and audit readiness

3. KEY APPLICATION AREAS OF AI/ML IN BIOPHARMACEUTICAL MANUFACTURING

3.1. Monoclonal antibodies and protein therapeutics

As mentioned before, monoclonal antibodies (mAbs) are a significant part of the biotherapeutics market. In recent years, several researchers have begun to realize the potential benefits of using AI/ML for mAbs production. Therefore, pharmaceutical companies and academic institutions are conducting extensive research in this field to improve mAbs production efficiency and enhance the scientific understanding of mAbs.

Pathogen safety evaluation is a critical part of mAb production. To ensure that the produced mAb is free of all types of harmful viruses, scientists design viral clearance studies, which are scale‐down models of the manufacturing process. Downstream purification processes are expected to be sufficient for clearing any residual viruses from the harvested cell culture fluid. Therefore, the identification of process parameters that are correlated with viral clearance capability is of critical importance. A group of researchers from the biopharmaceutical industry demonstrated the application of ML techniques for viral clearance through low pH inactivation unit operation. ²² The authors presented the application of both unsupervised and supervised learning methods for enhancing understanding and improving the efficiency of viral clearance studies. Among the different ML techniques discussed in the paper, the Random Forest algorithm achieved the highest cross‐validation overall accuracy of 94%. This type of ML model can also potentially guide the early process development of new therapeutic proteins or antibodies. The authors also mentioned that machine learning models can generate insights several orders of magnitude (weeks to hours) faster than human reviewing and summarizing learning from past wet lab experiment results.

A study from 2022 ³⁹ on the continuous manufacturing of mAbs illustrates how ML algorithms can be used to predict quality attributes pertaining to capture and polishing chromatography. Since the predictions are in real time, these predictive models can be used for process control by tracking critical quality attributes (CQAs) of in‐process samples. If the predictive models are accurate, the need for offline measurement of those CQAs will become obsolete, reducing experimental costs and accelerating the manufacturing process. If the quality of in‐process CQAs is not found to be satisfactory, the unit operation can be terminated earlier, saving time and resources.

The researchers explored several ML algorithms, such as deep neural networks (DNN), support vector regression (SVR), and random forest regression for predictive modeling. The predictive model was trained using pH, UV absorbance, and conductivity sensors to predict multiple quality attributes, namely ProteinA elute concentration, acidic charge variant (%), CEX elute concentration and aggregate (%). The random forest regression model outperformed the other algorithms with prediction errors of less than 5%. The study also discussed why other algorithms such as DNN, and gradient boosting did not perform as expected.

In a subsequent study by the same research group mentioned above, a deep learning‐based 2D‐convolutional neural network (2D‐CNN) was used for predicting critical quality attributes (CQAs) in the downstream processing of mAbs. ¹⁹ This study differs from the previous study ³⁹ in terms of pre‐processing technique before applying the modeling algorithm. A parameterization technique was used in the 2022 study before applying deep learning, whereas this study did not require parameterization, as 2D‐CNN can handle the data without any parametrization. This study focused on five key outputs, namely mAb elute concentrations from ProteinA, CEX, the percentages of acidic, basic variant and aggregate. The 2D‐CNN model was trained using routinely collected process data, showcasing superior predictive performance compared to traditional methods. The Bayesian optimization technique was integrated with 2D‐CNN for process optimization to match the biosimilar to the reference molecule with respect to the critical quality attributes. The results of the process optimization were further validated through experiments, reporting a mean percentage deviation of less than 3% from target values, underscoring its robustness and reliability. Until a few years ago, the deep learning algorithms were not widely popular for life‐science applications due to their black‐box nature. However, due to the recent advancements in the field of interpretable AI/ML, deep learning has found several applications in the highly regulated industry such as biologics manufacturing. This study also employed the Shapley method‐based SHAP (SHapley Additive explanations) technique to interpret deep learning model predictions. The Shapley method has its roots in cooperative game theory and SHAP is a specific application of the Shapley value concept in the AI/ML field. Due to this interpretable nature, the authors could reveal the significant impact of post‐viral inactivation pH on CQAs. The added interpretability helps scientists make more informed decisions in biologics manufacturing.

Spectroscopic methods have been widely used in the biopharmaceutical industry for real‐time measurements of attributes and process conditions in several unit operations, such as cell culture, chromatography, etc. Due to the nature of these methods, they produce a large amount of information‐rich data, which can be analyzed using modern ML techniques. In a paper published on the prediction of mAbs charge variants, the authors combined Raman spectroscopy with convolutional neural networks (CNNs). ⁵¹ This innovative application of ML can predict charge variants during the chromatography process in real‐time. Such integration of AI/ML frameworks with spectroscopic methods can be instrumental in ensuring consistent product quality in mAb production. Several other researchers have also published on the use of spectroscopy and ML methods for process monitoring in biopharmaceutical manufacturing. ¹⁰⁵ , ¹⁰⁶ , ¹⁰⁷ , ¹⁰⁸ , ¹⁰⁹ , ¹¹⁰

AI/ML applications are not limited to downstream purification processes. Researchers have explored their application for upstream cell culture processes as well. The prediction of antibody glycan quality from CHO cell culture media markers is one such example of AI/ML use. ¹¹¹ The study of N‐glycosylation is critical for ensuring the quality and consistency of mAbs manufacturing as it affects product stability, immunogenicity, and biological activity. N‐glycosylation is closely related to the cell culture process. Host cell type, cell culture media composition and process parameters such as oxygen level, temperature, and pH play important roles in the glycosylation process. Using AI/ML methods, the authors analyzed the data from several cell culture studies and identified significant media markers correlated with glycan abundance. Similar to the studies discussed above, the Random Forest algorithm outperformed traditional multivariate data analysis techniques.

The development of highly concentrated formulation of mAbs is another area of AI/ML application. Highly concentrated antibody formulation is required for low‐volume, subcutaneous administration of antibodies, which enables home delivery of antibodies. Low aggregation propensity and low viscosity are desired properties for a stable drug profile. However, providing home delivery convenience while maintaining stability is a challenging task as antibodies are prone to aggregation at high concentrations. AI/ML‐based predictive modeling tools can be used to assess the developability of high‐concentration antibody formulations. ¹¹² Data from 27 commercial mAbs were utilized to develop a viscosity predictive model. The number of hydrophobic residues on the light chain variable region and net charges on the light chain variable region were identified as two important features based on the logistic regression model.

3.1.1. Newly developed AI/ML techniques

Case studies discussed above are primarily based on the application of existing ML techniques for biologics manufacturing. These standard ML techniques find a large number of applications outside the pharmaceutical industry as well. Sometimes, existing ML algorithms, which are mostly imported from the computer science field, do not fulfill the modeling needs of biologics manufacturing. To this end, researchers proposed a novel approach combining time series and batch data to improve predictive modeling accuracy for CQAs in mAb production. ¹¹³ Extending the multiway partial least squares (N‐PLS) technique, the authors proposed tensorial methods to analyze datasets of differing orders (second and third‐order tensors) and improve predictions by accounting for batch‐to‐batch correlations. A typical mAb manufacturing process encompasses several complex unit operations that output data of differing orders. Therefore, an advanced method like this can certainly help with the modeling of end‐to‐end batch manufacturing of mAbs.

A major limitation of advanced ML algorithms, such as neural networks or deep learning, is that these algorithms require a large amount of data for model training. However, when training a ML model to support biopharmaceutical manufacturing, there are often only a small number of historical batches available for training. To circumvent these types of small data problems, industry researchers proposed a novel framework that can work efficiently even with a limited number of batches. ¹¹⁴ They used data from bulk drug substance manufacturing to demonstrate the efficacy of this novel approach.

As discussed in Section 2, each ML algorithm has its own strengths and limitations. Sometimes, it is possible to derive a hybrid algorithm that combines the strengths of multiple algorithms. Decision Tree‐PLS (DT‐PLS) is one such algorithm, developed for improving predictive capability. The DT‐PLS algorithm was the outcome of a collaboration between startup and academia. ¹¹⁵ Datasets from therapeutic fusion protein production (84 fed‐batch runs at a 3.5 L scale) and mAb production (106 fed‐batch runs at 15 mL scale) were utilized to demonstrate the benefit of the DT‐PLS algorithm. The objective of this study was to model titer, VCD, LMW and neutral charge variant (C3) using cell culture data. The study also compared the DT‐PLS results against classical algorithms.

3.1.2. Application of Image‐based AI/ML techniques

Advanced ML algorithms are not only limited to tabular data analysis; they can analyze non‐tabular data, such as images, very efficiently. Protein aggregation in the final drug product is a major concern for the immunogenicity of the biopharmaceutical product, which needs to be detected and quantified in order to mitigate the issue. Flow image microscopy is a powerful tool for recording images of aggregated particles of different sizes, shapes, and compositions. With the help of advanced ML techniques, such as deep convolutional neural networks (CNNs), image data can be analyzed, and subvisible particles in protein formulations can be classified into different categories based on the stress sources such as friability, freeze‐thawing, heating, and agitation. ⁴⁸ , ¹¹⁶ Understanding the effect of different types of stresses on protein aggregation will certainly help mitigate immunogenicity problems and enhance quality control and process monitoring in biopharmaceutical manufacturing.

3.2. Potential applications of AI/ML for cell and gene therapy products

Cell and Gene therapies (CGT) have great potential to serve as alternative treatments for the unmet needs of the biopharmaceutical industry, which is why AI/ML is expected to play an even bigger role in CGT. Moreover, CGT is still a new field for exploration, and many established biopharmaceutical companies and startups are ready to invest resources into this area due to its promising future.

3.2.1. Automation and IoT for real‐time data collection

As discussed in Section 3, the most important prerequisite for using AI/ML is the data and for that, we need hardware and software infrastructure to collect the data from the instruments in almost real‐time. Therefore, instead of relying on existing infrastructure that may not be suitable for real‐time data access, companies should invest in automation (e.g. Internet of Things (IoT)) in manufacturing and process development. Due to the complex nature of these new therapies, companies either invest significantly in the human resources for manual work, which increases production costs, or explore the AI with automation pathway to enhance efficiency and reduce costs. ¹¹⁷ , ¹¹⁸

3.2.2. Flow cytometry data analysis

In contrast to mAbs, where the cells are only intermediates, in cell therapy manufacturing, the cells themselves are the product. Therefore, the characterization and measurement of cells become extremely important. To this end, flow cytometry has become a widely used tool in the manufacturing of cell therapies. During the initial years of cell therapy research, flow cytometry data analysis was mostly manual, which led to variability due to operator judgment. In the recent year, many automated software tools have emerged as alternatives to reduce this operator‐introduced variability and improve the quality, repeatability, and robustness of the manufacturing process. Flock2, flowMeans, FlowSOM, PhenoGraph, SPADE3, and SWIFT are some of the software tools used for flow cytometry data analysis. Due to the differences in the underlying ML methods employed by these software tools, their outputs may vary. Therefore, the use of these software tools for automated cell identification could be challenged by regulatory authorities. It is of paramount importance that these software tools are validated first to enhance the confidence in their outputs. Appropriately designed synthetic datasets can be used to validate the software outputs. ¹¹⁹

Analyzing cytometry data often presents challenges due to its high dimensionality, large cell numbers, and heterogeneity between datasets. Machine learning algorithms are particularly well suited for managing complex cytometry data, providing solutions for dimensionality reduction, cell population identification, data classification, and automated gating. ¹²⁰ The flow cytometry industry is increasingly adopting automated gating solutions to streamline data analysis, enhance reproducibility, and reduce the time and labor associated with manual gating. Software platforms such as Cytobank, FCS Express, and FlowJo offer automated gating capabilities. ¹²¹ Despite advancements in algorithms and software, the adoption of these solutions in the bioindustry, especially in clinical settings, remains slow. This is primarily due to biological complexity, lack of standardization, and regulatory requirements. ¹²² , ¹²³

3.2.3. Single‐cell RNA sequencing for cell characterization

With advancements in analytical techniques, a more sophisticated method for cell characterization, single‐cell RNA sequencing (scRNA‐seq), has emerged as an effective tool for determining cellular phenotypes through unbiased assessment of transcriptomic data. This method allows for a comprehensive evaluation of the quality and developmental or differential stages of cell products. A computational pipeline known as “scCompare” has been recently introduced to analyze scRNA‐seq data, enabling scientists in the cell therapy industry to efficiently evaluate cell products ¹²⁴ This pipeline employs several underlying algorithms, such as batch data correction and harmonization, Leiden clustering, ¹²⁵ and data projection into uniform manifold approximation and projection (UMAP) space, ¹²⁶ to generate cell type‐specific prototype signatures for comparison. In benchmark evaluations, scCompare outperforms single‐cell variational inference (scVI) in precision and sensitivity for most cell types. Developed in 2018, scVI is another state‐of‐the‐art computational framework for general scRNA‐seq analyses including data visualization, clustering, and differential expression analysis of gene expression in single cells. ¹²⁷

3.2.4. Role of process analytical technology

Similar to mAbs, process analytical technology (PAT) could play a crucial role in CGT manufacturing as well, ¹²⁸ enabling enhanced process control and continuous manufacturing of complex biological products, such as viral vectors, and autologous cell therapy. Some commonly used PAT systems in mAbs manufacturing, such as fluorescence spectroscopy, Raman spectroscopy, near‐infrared spectroscopy, can also find applications in CGT manufacturing. ¹²⁹ Another category of refractometric devices as analytical technology is gaining interest in biopharmaceuticals for their potential use for adaptive process control. ¹³⁰ , ¹³¹ , ¹³² A group of researchers used a novel refractometry‐based PAT system (Ranger system) to monitor the metabolic activity of cell cultures during lentiviral vector (LVV) production processes in real time. ¹³³ The Ranger system combined the measurement of the refractive index of cell culture with process parameters such as cumulative air addition, dissolved oxygen levels, and pH to provide a tool for assessing and controlling cell culture activity. ¹³⁴ By using this tool and advanced AI/ML techniques, such as artificial neural network (ANN) and long short‐term memory (LSTM) networks, the researchers were able to predict metabolic activity.

PAT can also support CGT manufacturing via conventional imaging technologies such as flow cytometry, microscopy, etc. Routine cell culture imaging of adherent culture can be used to automatically capture images throughout cell expansion and differentiation. When used in concert with machine learning models, real‐time estimates of parameters like confluency and cell differentiation can support process decision‐making and impact product quality. ¹³⁵ , ¹³⁶ , ¹³⁷ , ¹³⁸

4. GENERATIVE AI DRIVEN BIOPHARMACEUTICAL INNOVATION

4.1. Overview of generative AI

Generative artificial intelligence is a product of the latest generation of artificial intelligence development. Generative AI has the ability to create new content or data by learning patterns from existing information. It is a result of technical advancements in AI/ML methods over the years. Large language models (LLMs) have emerged as the most popular type of Generative AI, capable of generating human‐like content (text) based on training data. A generative pre‐trained transformer (GPT) is a type of LLM, which was used in the popular chatbot platform ChatGPT by OpenAI organization. ¹³⁹ Over the years, several versions of GPT have evolved, and with each version, the underlying model has become more complex, gaining the ability to analyze and generate larger amounts of text. In addition to OpenAI's ChatGPT, other examples of LLMs include Gemini ¹⁴⁰ by Google, Claude ¹⁴¹ by Anthropic, LLaMA ¹⁴² by Meta, and Mistral ¹⁴³ by MistralAI.

4.2. Potential applications in documentation, deviation management, and research

The impact of generative AI on the pharmaceutical industry could be enormous. According to McKinsey Global Institute (MGI) assessment, this technology has the potential to generate economic value worth $60 billion to $110 billion annually. ¹⁴⁴ This increase in the economic value is primarily due to enhanced productivity, as generative AI can speed up drug discovery and make drug development, manufacturing, and marketing more efficient.

Biopharmaceutical manufacturing is governed by standard operating procedures (SOPs). LLM‐based tools can help operators and their supervisors quickly find relevant SOPs, summarize them, and prepare a checklist comprising the steps to do a certain task. A generative AI tool can also generate a report from the bullet points written by operators on the production floor. Such help from generative AI‐based tools in day‐to‐day business has the potential to increase operational efficiency.

Deviation management is another area with potential for increased operational efficiency. Biopharmaceutical manufacturers must follow internal good manufacturing practices (GMP) and other regulatory requirements. Despite adhering to GMP, the manufacturer cannot completely avoid deviations. Typically, these deviations need to be investigated in a timely manner, requiring a significant amount of time and resources for the documentation‐related work. LLM‐based tools can assist scientists in determining the investigation strategy, which includes severity classification of deviations, mining historical deviation reports, and identifying potential root causes and corrective actions. In addition to mining internal SOPs and historical deviation reports, LLMs can assist manufacturing scientists in searching scientific literature and regulatory guidelines for process development, process optimization, and assay optimization.

Although LLMs are advanced tools that can provide human‐like services for documentation, these models may also suffer from hallucination. Therefore, expert oversight is necessary before finalizing any document for regulatory submission purposes.

5. REGULATORY ASPECTS OF AI/ML IMPLEMENTATION

As mentioned before, the biopharmaceutical industry has begun to realize the immense benefits of AI/ML applications. However, any major positive disruption in this industry cannot materialize without the approval from regulatory agencies worldwide. It is encouraging to know that these agencies also agree that artificial intelligence can transform the industry. This transformation will not only reduce time to market and enhance efficiency but also ensure quality and improve regulatory compliance. Therefore, these agencies have started discussions with industry and the public in general.

In the United States, the Center for Drug Evaluation and Research (CDER) has recognized that excessive regulatory oversight will slow down its adoption rate of new technologies. ¹⁴⁵ Therefore, the agency is willing to provide some flexibility in manufacturing processes to keep pace with technological advancements. This is a continuation of the Emerging Technology Program (ETP) from 2014, which served as a platform to collaborate with companies to support technological advancements in manufacturing. The FDA recognizes that AI can play a critical role in areas such as process design, advanced process control, process monitoring, fault detection and deviation trend monitoring. FDA also raises an important discussion related to data safety, security and traceability for cloud applications. Per FDA discussion with industry, non‐time‐critical applications can be hosted on cloud servers. However, software applications that execute direct control actions should be on‐premise servers, close to the manufacturing unit to ensure no impact on performance and security. It is the manufacturer's responsibility to define standards for model validation and ensure no adverse impact on product quality due to the model usage. FDA also raises concerns over real‐time updates of AI models. Typically, any updates in the model or training data are governed by change control processes within the company's quality system. If the model is updated close to real time, the implementation of the change control process may not be realistic. Therefore, the manufacturer must ensure that there is ample clarity about the lifecycle management of AI models. If there is any impact from the model update on product quality, then a product comparability assessment should be performed.

In the more recent draft guidance by U.S. FDA, ¹⁴⁶ they have focused on the use of AI to support regulatory decision making for drug and biological products. However, the draft guidance does not address the use of AI models for increasing operational efficiency e.g. using LLMs for drafting/writing a regulatory submission. The risk‐based credibility assessment framework has been proposed in the draft to help biologics manufacturers establish the credibility of AI model outputs to make regulatory decisions. Figure 2 presents the FDA proposed risk‐based credibility assessment framework comprising seven steps with an example from drug manufacturing.

FIGURE 2.

Risk‐based credibility assessment framework.

Biopharmaceutical drug manufacturers can also learn from the regulatory guidance issued for AI/ML applications for software‐based medical devices. Recent FDA guidance ¹⁴⁷ prescribes the pre‐determined change control plan for AI/ML models emphasizing documenting detailed changes, validation plans, and impact assessments in a modification protocol. This type of change control process will ensure the continued safety of such AI/ML models for GMP use.

European Medicines Agency (EMA) ¹⁴⁸ also opines that AI/ML applications in biologics manufacturing are expected to increase in the coming years for process design and scale up, process optimization, quality control and real‐time batch release. The EMA also highlights the need for interpretable and explainable models instead of black‐box models to enhance the acceptance of such models within the industry and across regulatory agencies worldwide.

6. CONCLUSION

In an era dominated by artificial intelligence, biopharmaceutical companies cannot afford to miss the opportunity to overhaul their businesses by leveraging this technology. Increased operational efficiency (e.g., reduced process deviations and batch rejections, improved process monitoring, enhanced process understanding, and better regulatory compliance) are some of the direct benefits for the biopharmaceutical industry. These benefits would not only translate into higher profitability for manufacturers but also potentially reduce the cost of drugs for patients. Biopharmaceutical manufacturers can learn a great deal from the software industry.

The first and foremost requirement for developing any AI/ML assisted application is good quality data in an easily accessible digital format. As discussed before, biopharmaceutical manufacturers need to invest in building cloud‐based data systems. The GMP validation of these data systems is the bedrock for building AI/ML‐based digital applications that can assist in making commercial manufacturing decisions. Hiring human resources with a data science background and/or providing data science training to the existing workforce is also critical in this AI implementation journey.

It's not only about the application of machine learning algorithms or cloud technology; the culture and mindset of the people working in the biopharmaceutical companies need to change. New life‐saving treatments, such as cell and gene therapies are far more complex than chemically synthesized small molecules. Without automation and artificial intelligence, the development journey of these new treatments will require enormous time and resources.

As discussed earlier, the increased adoption of artificial intelligence by large companies in the biopharmaceutical industry will also motivate young researchers and entrepreneurs to start their own companies, fulfilling the unmet research needs of this sector. Essentially, this will create an ecosystem that benefits patients by providing treatments for currently uncurable diseases. Alongside industry and academic researchers, regulatory agencies will also need to play an important role as catalysts to boost the applications of artificial intelligence in biopharmaceutical manufacturing. By providing clear guidance on the life cycle management of artificial intelligence applications, regulatory agencies can help biopharmaceutical manufacturers better serve the patients.

AUTHOR CONTRIBUTIONS

Dr. Shyam Panjwani performed the roles of conceptualization, investigation, writing, visualization, supervision, and editing. Drs. John Mason and Hao Wei performed the role of writing.

FUNDING INFORMATION

No funding was received for this work.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

DATA AVAILABILITY STATEMENT

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.