Drug services firms begin embracing AI

It’s hard nowadays for any big company to distance itself from artificial intelligence and its subset, machine learning. Drug services firms are no exception. These technologies are slowly picking up momentum among contract development and manufacturing organizations (CDMOs) to innovate drug synthesis and automate quality control processes that humans have traditionally managed.

Last year, Lonza, a Swiss CDMO, initiated an AI-enabled route scouting program to identify new methods for synthesizing novel molecules that avoid manufacturing bottlenecks and speed drug development timelines.

“Many customers are approaching us especially for this service,” says Jan Vertommen, a vice president in Lonza’s small-molecule division. The service is particularly useful for smaller biotechnology companies that are developing new drugs but don’t know much about the supply chain for raw materials and intermediates, he adds.

Lonza partnered with IBM to develop models trained on data from large chemical repositories, such as Elsevier’s Reaxys, and on Lonza’s own data about raw material inventories and supply chains.

These algorithms offer new approaches for drug synthesis—traditionally the province of bench chemists. “What comes from using these models may not always be the shortest route to synthesize but the most viable one,” Vertommen says.

Once the platform generates a road map for synthesizing a drug molecule, the method is validated in the laboratory with high-throughput experimentation, often done by robots. “The robots make us the first 500 g to a kilogram to test the method,” Vertommen says. “The robots come in handy, especially at Lonza’s site in Visp, Switzerland, where labor costs are high.”

Another CDMO, Thermo Fisher Scientific, uses AI and automation to ensure quality control during production and packaging of large molecules such as vaccines and monoclonal antibodies. For instance, the company has implemented an AI-trained automated visual inspection system that rejects products that do not meet its standards.

Christy Eatmon, Thermo Fisher’s global subject matter expert on sterile drug products, explains that the firm uses the technology to inspect vials that hold viscous liquids or tiny containers that hold just a few droplets of liquid.

In the past, most batches, each composed of about 200,000 vials, were scanned by humans, Eatmon says. The inspector would typically pick up two or three vials at a time and hold them for several seconds to look for irregularities. But the human eye isn’t infallible and is easily tricked.

"When these vials are spun for proper mixing, some droplets could stick to the surface or create small bubbles that could look like cracks," Eatmon says. Some of the vials get rejected when in reality they are perfectly fine.

Since Thermo Fisher started using the AI inspection technology, the false rejection rate has fallen from an average of 20% per batch to 3%, according to Eatmon. It also saves the company about 60 h of human labor per batch, she adds.

In a draft guidance on AI usage in the drug industry released in February, the US Food and Drug Administration notes that the Center for Drug Evaluation and Research, the FDA division that assesses drug safety and efficacy, has seen a “significant increase” in the application of number of new drugs developed with AI components. The document highlights the benefits and risks associated with AI in manufacturing and urges companies to engage with the FDA early in implementing it.

Justin Butler, a partner at Eclipse Ventures, which invests in companies developing AI and automation solutions, says the CDMO industry isn’t a leader in AI adoption or in developing innovative AI technologies. “Sometimes, it is for a good reason. Introducing AI and automation could trigger regulatory trip wires,” he says.

Still, Butler is seeing a gradual change as CDMOs use AI to manage their supply chains and simplify operational tasks. And even though AI adoption among CDMOs is currently sluggish compared with other manufacturing industries, it may be hard to find a firm that doesn’t use AI by the end of 2035, he says.

Butler says it is helpful to think of AI as software that can discover unknown correlations when fed unstructured data. “It will be impossible to avoid AI, because it is the best way to do many things. Humans can stare at data for a long time but may never be able to see those correlations,” he says.