The monoclonal antibody industry is benefitting from machine learning (ML), according to new analysis showing that use of the approach in process development is increasing.
The study, by researchers at the University of Melbourne and CSL Innovation in Australia, looked at the extent to which supervised learning modeling (SL)—an ML technique that uses labeled data to train predictive algorithms—is employed in mAb production.
In addition, the study assessed how widely SL is being used in combination with data-driven optimization (DO)—an approach that uses data to tweak and improve process models in real-time.
And—based on a literature review—over the past ten years the combination of SL and DO has had a significant impact on the development of commercial mAb production platforms.
The authors write: “The increasing number of studies in this field in the past years strongly suggests an industry-wide motivation to develop novel SLDO techniques to improve their platform process optimization.”
They add, “While DoE-based mechanistic modeling remains an integral part of the initial data generation and process characterization, SLDO can outperform it when it comes to modeling and optimizing more complex data and experimental requirements at the later stages of platform development.”
However, the reviewers suggest hybrid approaches combining elements of SLDO and traditional experimental methods are likely to be more effective for process development than predictive models alone.
“Rather than proposing new industry best practices, the studies considered in this review have demonstrated the usefulness of incorporating SLDO into the traditional approaches to characterization, predictive modeling, monitoring, and optimization of the bioprocesses in various applications.”
Learning impetus
Despite their findings, the authors acknowledge that further development of machine learning modeling techniques is required for the mAb industry to fully realize the potential benefits of the approach.
The main challenge, they say, is the quality of the process data that is available to the biopharmaceutical industry for ML training purposes.
“The results indicated that data in the biopharmaceutical landscape is laden with inevitable industry-specific challenges, most notably heterogeneity and high dimensionality.
“Existing methods such as MVDA [multivariate data analysis] and ANN [artificial neural networks] have yet to fully overcome these complicated limitations, hence the need for new methodologies and frameworks to be developed in the future.”
However, the authors are optimistic the industry will be willing to put the effort into advancing ML methods, citing the wider move towards digitization as a positive sign.
“As models and frameworks continue to be refined as part of the digital transformation movement in biomanufacturing technologies, the realization of a holistic digital twin platform across all unit operations will eventually necessitate efforts to integrate multiple algorithms, data sources, and instruments.
“This cements the pivotal role of SLDO in the foreseeable future of the biopharmaceutical R&D sector.”
