Artificial intelligence (AI) could help biopharmaceutical manufacturers detect contamination in cell cultures sooner than conventional approaches, thereby helping to reduce waste and batch losses.
The idea of using machine learning (ML), a type of AI that finds patterns in data without instruction, to identify contamination comes from researchers at Virginia Polytechnic Institute and State University, who say current detection methods take too long.
Lead author, Xuan Dung James Nguyen, PhD, told GEN, “In most bioprocesses, contamination is still detected relatively late. We rely heavily on intermittent off-line assays that can take hours to days to return a result. During that time, the culture often continues to run, so by the time contamination is confirmed, the batch is usually already lost.”
Current detection methods can misidentify signs of contamination, according to Nguyen, who said, “Small changes in pH, DO, metabolites, or gases are easy to miss or to attribute to normal process variability.
“As a result, a lot of low-level contamination is either detected very late or only discovered during downstream quality testing. That combination of delayed feedback, noisy signals, and rare but high-impact events makes early detection with conventional methods quite difficult.”
Machine learning
To develop an alternative, Nguyen and colleagues used data from 246 fermentations carried out by Salem, Virginia-based “biosolutions” developer Novonesis Biological, 23 of which were labeled contaminated batches, to build machine learning models.
The key advantage of the models is the ability to quickly differentiate meaningful signals from natural variations that occur during the culturing process.
Nguyen says, “Machine learning can learn the normal multivariate behavior of a healthy culture by looking at many process variables simultaneously—things like on-line sensors, off-line measurements, and calculated soft-sensor outputs.
“Once that normal baseline is learned, the model can flag subtle, coordinated deviations that are hard for humans or simple control charts to spot. In our work, we use anomaly-detection models for exactly this purpose: they provide an early warning that ’this batch is behaving differently’ long before traditional microbiology tests come back.”
Other advantages include the better use of data—models can be built using information already being collected—and a reduction in false positives because the models consider the full process context, rather than one variable at a time.
ML models can also be reused according to Nguyen, who adds, “By analyzing patterns across many contaminated and non-contaminated runs, ML can help identify which conditions or raw-material changes tend to precede contamination events.
“Overall, the goal is not to replace microbiology, but to add a fast, data-driven ‘smoke alarm’ that runs continuously in the background,” he says.
Resources
Another advantage ML models have over traditional contamination detection methods is that they are relatively straightforward to set up, according to Nguyen.
“You don’t need exotic hardware: modern ML models for this use case can run on standard servers or in the cloud. What matters more is having the ability to deploy models close to real time so operators can see alerts quickly.
“Good data governance and validation are needed to ensure models remain reliable as processes and equipment change,” he says.
Companies interested in the approach should consider establishing cross-functional teams—people from data science, process development, manufacturing, and quality—that can interpret the alerts and translate them into practical operating procedures.
“If those pieces are in place,” Nguyen continues, “ML-based contamination surveillance is often more of an organizational and data-engineering project than a purely algorithmic one.”
The post AI “Smoke Alarm” Could Detect Culture Contamination Earlier appeared first on GEN – Genetic Engineering and Biotechnology News.
