Researchers have used machine learning and two experimental designs to optimize a common critical quality attribute of biological therapies.
Eliza Yeung, PhD, associate director of process characterization at contract manufacturer Cytovance Biologics, and another researcher relied on statistical software to create two experimental designs to analyze glycosylation, a critical quality attribute of therapeutic glycoproteins. The study was published in the BioProcessing Journal last October.
“The goal of our paper is to give us the capability to do more with fewer experiments. We know this type of model [developed with the software] can support validation of methodologies and processes, so our goal is to have a good predictive model to control all sources of [process] variability,” she said.
Yeung hopes the model they developed can help serve their clients better. “The paper is a good case study [showing] that we have the tools and strategy to achieve a robust process,” she added.
Glycosylation plays an important role in the function and safety of therapeutic proteins, but is regarded as hard to analyze and control because it is specific to cells, proteins, and processes. As such, it is important to generate consistent glycosylation profiles and reliable structural analyses during drug development.
The paper used high-performance anion exchange chromatography with pulsed amperometric detection, a conventional glycosylation profiling method, on glycans derived from human immunoglobulin.
Two experimental designs, definitive screening and central composite design, were used with a machine learning algorithm. Both the experimental design and subsequent analyses, according to the paper, were performed with statistical software.
To optimize the performance of the glycoprofiling, the team built a predictive model based on thousands of passes through a model-fitting algorithm. This ensemble approach is known as “self-validating ensemble modeling” and, aims to reduce the instability of model performance.
The team discovered that predictive models based on one experimental design could reliably predict the results of the other.
“We’re very happy to have an approach with machine learning and design of experiment that requires fewer experiment runs,” said Yeung. “Because [glycosylation] is so complicated, it’s brilliant to gain more understanding with fewer experiments and, by means of them, validate our model.”
The team believes this, and similar approaches, can help meet International Council for Harmonization of Technical Requirements for Pharmaceuticals for Human Use (ICH) guidelines relating to “lifecycle” approach on “process development” and forthcoming guidance on “analytical procedure development.”