Dan Hill, Associate Director, Digital Development & Analytical, Biogen
In his talk, Dan Hill discussed a real-time amidite identification platform Biogen has been developing. He shared that this platform is part of their broader initiative in order to modernize their manufacturing procedures. He explained three important resources for manufacturing modernization, including the concept of an adaptive plant, digital plant maturity model (DPMM), and process analytical technology (PAT). One key aspect of an adaptive plant is the use of the inline technologies to provide the right insights and controls not only ensure a reliable robust process but also supports the overall integrated data flow which constitutes a pillar which other components of Industry 4.0 rely on.
While discussing the gains of data flow, Hill explained a modeling and analytics maturity model with a biology-focused approach. As we understand the data and build process monitoring capabilities, we can go further by developing predictive capabilities. Some PAT principles can be applied to a soft sensing stage. These are followed by having an understanding of the underlying causes of certain faults or outputs leading to fault classification. When one can understand the way the process is responding and the inputs to those responses, it is possible to become more adaptive and begin using model-based predictive controls.
Following explaining modeling and the analytics maturity model, Hill briefly introduced the antisense oligonucleotides (ASO), and their use and production in Biogen. Biogen has established a very active ASO program to address serious neuropathological diseases. ASOs are molecules that are engineered to bind RNA and exhibit their effect by either downregulating their translation into proteins or blocking an disease causing RNA via direct binding.
ASOs are short strands, 16-20 nucleotide in length, containing the same RNA bases and linkages with a few modifications:
At Biogen they use both 2’ modified (Methoxyethyl, MOE) and 2’ unmodified (deoxy) ASOs which make up eight distinct “flavors” of amidites based on the presence of different bases and two different 2’ profiles.
There are several opportunities at ASO manufacturing procedures to implement PAT applications. Early process development is good time to apply control strategies, monitor the control requirements, introduce new technology, and make major design changes with comparably low costs. Hill explained that the PAT system used at Biogen is specifically focused on the sequence ID.
The Biogen synthesis device has ports specific to each amidite. One essential function of their inline PAT system is to track the synthesis and ports to monitor incoming amidites and any modifications. Hill explained that amidites are responsive to infra-red (IR) and share data indicating that each DNA base (dA, dC, dG, dT) and sugar-ring modification have different IR profiles like a fingerprint in certain spectral regions.
The combination of base and modification graphs makes it possible to distinguish each of eight different amidites. Hi explained how the online PAT provides more insight and control in both the prime mode and production mode. In prime mode, it is possible to identify errors regarding amidites, solutions preparation, and connections. This allows them to prevent errors before production mode and eliminate a possible batch failure. In production, amidite additions are monitored by the PAT system to perform sequence identification.
They built a calibration set at Biogen that reflects the variability types seen in manufacturing such as concentration of amidite solutions, flow rate of the solutions through the flow path and instrument-to-instrument variations. When they manipulated each variable, the data they obtain indicated not only the robustness of the method, but was also helpful in overall technical transfer from the laboratory to manufacturing. Hill described the use of the partial least squares discriminant (PLSD) analysis model which is a multivariate model discriminating components and is helpful in setting thresholds and performing classifications based on scores. The model they used has been validated by numerous tests for the specificity of the raw spectra, model, and results. One strategy to assess the model beyond its ongoing diagnostic assessment in a comparably stable process with stable materials is to introduce new amidites. To do so, they have looked at cEt materials and obtained effective responses in the IR to distinguish different families of amidites. Instead of all-in-one model approach, they also proposed a hierarchical modeling approach which classifies based on the family of the amidites such as deoxy, MOE, cEt and so on and then identifies each base within a class. Lastly, Mr. Hill mentioned the use of Python to accelerate the model optimization which also increases the throughput of analyses.
