With increasingly complex therapies, bioprocessors face a challenge in avoiding errors. “Human error is frequent in biomanufacturing, accounting for 80% of deviations,” wrote Wei Xie, PhD, assistant professor of mechanical and industrial engineering at Northeastern University, and Giulia Pedrielli, PhD, associate professor of computing and augmented intelligence at Arizona State University. “It is increasingly realized by the biopharma industry that optimization, machine learning, and simulation approaches, which can incorporate physics-based and experiments supported knowledge, are a key to the next generation of products and processes.”

To learn more about this, GEN talked with Xie. When asked about the main benefits of understanding the bioprocessing mechanisms behind a drug, Xie described several. First, she said that understanding the mechanisms of a bioprocess “supports robust and interpretable decision making, especially for integrated and/or end-to-end biomanufacturing processes.”

As an example, she noted that “understanding how the inputs—such as media components and critical process parameters, or CPPs—propagate through mechanism pathways and influence the outputs can facilitate forward interpretable predictions and backward reasoning for root-cause analysis.”

She added that understanding the mechanisms in a bioprocess can “guide the design of experiments, data collection, and sensor-network development to improve sample efficiency and production consistency.” For instance, she said: “In a cell-culture bioreactor, we may allocate more data-collection efforts at critical times, such as the transition period from the exponential growth phase to the production phase, and this can ensure reliable bioprocess monitoring and avoid saving too much unnecessary data.”

Graph Hybrid Model

For an end-to-end representation, a bioprocessor could use the graph hybrid model, which combines mechanistic and statistical modeling.

“Since the bioprocess is the product, a hybrid model—characterizing the risk- and science-based understanding of bioprocess mechanisms—can facilitate the integration of heterogeneous data and information collected from drug discovery and production to support learning,” Xie explained.

To do that, the knowledge graph from this type of modeling can be coupled with advanced sensor monitoring and assay technologies. “Leveraging the emerging sensor technologies—such as single-molecule RNA-fluorescence in situ hybridization, two-photon excitation fluorescence, and Raman probes—can provide real-time measures on multi-omics of single molecules and single cells,” Xie said. “The multi-level bioprocessing knowledge graph hybrid model allows us to advance the understanding of complex interactions at molecular, cellular, and macro-system levels, support optimal learning and reliable monitoring, and enhance design and control decision making.”

Xie and her colleagues are well underway in developing this modeling for bioprocessors.

“Our Northeastern research team, along with multiple university and industry partners, is creating knowledge graph hybrid model based digital-twins and advanced process analytical technologies, or PATs, software for integrated bioprocessing so that these interpretable ML/AI technologies can be easily adopted by bioprocessors,” she told GEN. “New calibration methodology will be used to ensure the high fidelity of the digital twins, such as reducing the discrepancy between the digital and real commercial systems, and support the development of a reliable decision-making support platform.

“This study can facilitate drug discovery, root-cause analysis, predictive analysis, bioprocess development, and automation for monoclonal antibodies, cell/gene therapies, regenerative medicines, and RNA vaccines.”