This study assessed a novel statistical approach using space-filling designs (SFDs) and self-validating ensemble modeling (SVEM) machine learning to efficiently identify key process factors using recombinant adeno-associated virus type 9 (rAAV9) gene therapy manufacturing as a case study. Based on risk assessment of parameters that may impact rAAV9 production, we have evaluated six process parameters using 24-run SFDs generated by the JMP statistical software. SFDs are a new class of design of experiment (DoE) created with the objective of covering the entire design space as completely as possible; this in turn allows more accurate modeling of complex response surface behavior typically found in bioprocesses.
Tag: <span>machine learning</span>
Traditionally, the Six Sigma framework has underpinned quality improvement and assurance in biopharmaceutical manufacturing process management. This paper proposes a neural network (NN) approach to vaccine yield classification and compares it to an existing multiple linear regression approach. As part of the Six Sigma process, this paper shows how a data mining framework can be used to extract further value and insight from the data gathered during the manufacturing process, and how insights into yield classification can be used in the quality improvement process.
Glycosylation is one of the most common post-translational modifications in mammalian-expressed biologics, and is considered to be a critical quality attribute of therapeutic glycoproteins. Due to its biological relevance, physiochemical assessment on the glycosylation profile is always important to the success of a drug development initiative. This article describes the combination of experimental design and machine learning techniques applied to characterize and optimize a conventional, non-derivatized glycoprofiling method on glycans derived from a human immunoglobulin using high-performance anion exchange chromatography with pulsed amperometric detection (HPAEC-PAD). Two independent experimental designs, a 16-run definitive screening design (DSD) and a 28-run central composite design (CCD), were incorporated with a machine learning technique known as “self-validating ensemble modeling (SVEM)” and used to build predictive models for four chromatographic responses. We show that the predictive models created using SVEM on the DSD data reliably predicted the behavior of the chosen responses when applied to CCD validation data. This demonstrates that the DSD is an efficient alternative to the larger, traditional CCD in which the combination of experimental design and machine learning can effectively characterize and optimize analytical methods.