Biotherapeutics are a hot topic right now — and for good reason. But even before the COVID-19 pandemic, the biotherapeutics field, like every other manufacturing sector, was exploding with all the data that are being generated by recent innovations in equipment, systems, and processes. Advances in biomanufacturing analytics, analytical technology, and machine learning have tried to keep pace; however, such tools too often are misunderstood and applied suboptimally. Thus, many companies struggle with confusion and missed opportunities when they should be reaping the benefits of improved workflows and reliable product manufacturing.
Done right, applications of data analytics and machine learning can provide bioprocessing predictions and classifications — e.g., for assessing whether product quality meets specifications. If an analysis reveals that something is awry, the cause can be identified, whether it is in processes or raw materials. Data analytics and machine learning techniques also can bolster other activities.
Preventing Adverse Effects: Safety is a primary concern during drug development and manufacture. Analytics and machine learning can help biotherapeutic companies characterize impurities and degradants accurately to ensure drug safety and effectiveness.
Maintaining Compliance: Regulatory agencies focus much of their attention on impurities and degradants, and noncompliance involves significant reputational, financial, and legal risks. New technologies for biomanufacturing can help biotherapeutic manufacturers mitigate such risks.
Accelerating Process Development: Applying process data analytics correctly improves decision-making about experiments — and about how to use the resulting data. That enables teams to converge quickly on optimal processes for manufacture of high-quality, low-risk products.
Correct Application Is Critical
The above benefits are why data scientists, data engineers, process engineers, and process scientists, and others who seek to extract value from data must learn not only which new analytical tools are best for the processes they work with, but also how to use those tools correctly.
Here’s an example: Crossvalidation commonly is used to assess the predictive performance of models and to evaluate how they perform beyond sample data sets. But most data engineers — in my estimation, as many as 90% of them — do not use crossvalidation as they should. Their calculations of model accuracy miss the mark substantially, meaning that they think their models are more accurate than they actually are. And all too often, entire process development platforms are designed presuming a certain degree of accuracy, but because the original assumptions are wrong, the end products do not meet specifications. Then whole batches must be discarded.
| Professional Education at MIT |
| Richard D. Braatz is one of three lead instructors for the MIT professional education course “Bioprocessing Data Analytics and Machine Learning.”
This recurring three-day course is designed for biopharmaceutical scientists and engineers who want to take their skills and careers to the next level. With the guidance of academic and industry experts, participants discover innovative ways to apply data analytics and learn best practices for translating biomanufacturing data into reliable models and good process decisions. This course may be taken individually or as part of MIT’s professional certificate programs in machine learning and artificial intelligence and biotechnology and life sciences. Learn more about future sessions at https://professional.mit.edu/course-catalog/bioprocess-data-analytics-and-machine-learning. |
Reckoning with Uncertainty
Many methods can ensure proper crossvalidation, but the keys are to account for uncertainty and ensure that predictions are accurate. Correct crossvalidation requires that teams compare predictions of models built from many different choices of training data, select training data sets to account for variabilities and biases that can occur for different types of data, and set aside a significant proportion of data that have not been used during model building to produce a fair estimate of model accuracy.
In drug development and manufacture, as in all manufacturing processes, data analytics must be applied well. When that happens, biotherapeutic companies will be able to develop streamlined workflows that will define highly efficient, reliable, and low-risk manufacturing processes as quickly as possible.
Dr. Richard D. Braatz is Edwin R. Gilliland Professor of Chemical Engineering at the Massachusetts Institute of Technology (MIT); 77 Massachusetts Avenue, Cambridge, MA 02139; braatz@mit.edu. Braatz researches advanced biopharmaceutical manufacturing systems and leads process data analytics, mechanistic modeling, and control systems for several MIT projects, including those for monoclonal-antibody, viral-vaccine, and gene-therapy manufacturing. During his career, he has collaborated with 20+ companies and has been published in 200+ papers and three books.
This article will appear in BPI’s October 2020 issue.