Behind the powerful computations, AI’s role in pharma is still beset by challenges

AI's ability to bring down the costs of drug development has allowed smaller companies to do much more with fewer resources. One 2023 study predicted AI-driven R&D efforts from discovery up to preclinical could create time and cost savings of at least 25% to 50%.

Small biotechs developing drugs can benefit from AI's computing power and efficiency, says Panna Sharma, CEO of Lantern Pharma, a biotech that uses a proprietary AI platform to develop oncology drugs.

The company currently has three lead drug candidates in early-stage clinical trials and an antibody-drug conjugate program in the preclinical stage. The company also collaborates with other biotech companies that want to use its AI technology platform Response Algorithm for Drug Positioning and Rescue (RADR), which aims to predict potential patient response to drugs.

"If you look at the new chipsets that are being created, we're going to approach being able to do computations that [three years ago] probably would have cost $100,000 … and sucked up a month or two of machine time … to [doing] that in a day [for] not even $1,000 to $2,000," Sharma said. "It totally changes the ability for individual researchers and small companies to be meaningful developers. And this is going to keep changing. The curve only goes one way."

As more pharmas and biotechs leverage AI platforms to develop drugs, some risks still loom large.

AI hurdles ahead

As many have long noted, AI and machine learning models are trained on available data sets, which can have gaps in patient demographics.

"There are risks that your data sets are incomplete, and we always worry about that," Sharma said. "I think that's one of the biggest challenges - incomplete data sets that lead you to bad or incomplete conclusions."

Incomplete or biased data sets are in the purview of the FDA, which noted its concerns last year regarding ethical considerations and generalizability of findings extrapolated outside testing environments with incomplete data.

Data also needs to reflect patient populations, which can be a challenge if data sets reflect a smaller group of specific patients, Sharma said. Companies may also be training AI algorithms on the data that is available to them, which may not reflect the true patient population.

"You have to question, is that the biology that I'm going to really see and target in their real world?" Sharma said.

Disease complexity

Another challenge in computational biology is the complexity of diseases, according to Sharma. AI models are not always trained on the specifics of diseases in cancer or neuroscience, for example, which can vary widely patient by patient.

"One of the biggest challenges I see as I talk to other AI companies is that you have a new generation of drug developers that don't appreciate the full complexity of the biology of their disease," he said. "You can get a lot of junky early-stage molecules, to be honest."

Relying too heavily on AI can lose the forest for the trees, he said.

"Sometimes a lot of AI [focus] tends to be too much on the software and data side and not enough on the complexities of the disease and biology side," Sharma said. You're going to have a lot of people stub their toes in the AI drug development space as a result."

As more drugs are developed using AI technology platforms, patient skepticism can also present a challenge and a risk to companies aiming to get into clinical trials.

"There may be an era in which patients are going to ask questions about whether these compounds are from AI, and they may not feel good about taking AI-developed medicines," Sharma said.

Given that risk, biotechs may have to assess how patients perceive AI drugs when enrolling in clinical trials.

Meanwhile, the FDA is still determining how it will regulate AI in drug development and expects to release guidance on the topic. At the same time, the industry continues to increase its AI adoption, and the FDA received more than 100 submissions drug and biologic application submissions using AI or machine learning components in 2021.

"[With] the pace at which AI and software is developing, we may, in certain instances, learn about the need for regulations later than we should have," particularly for medical devices that may use AI, Sharma said.

But after development, there may be less of a need for regulations because the clinical trial and drug approval pathway still hold up as the gold standard of quality despite AI's involvement, Sharma said.

Recommended Reading

• Pharma's R&D gamble: picking pipeline winners in a risky field Meagan Parrish
• Why AlphaFold 3 is stirring up so much buzz in pharma Kelly Bilodeau