Believe the hype? Mixed signals from AI’s impact on drug development

Along with artificial intelligence has come plenty of hope and hype for its ability to sort through troves of molecules and data at super speed, shortening a clinical timeline to months as opposed to years. But can AI improve the dismal success rate of traditional drug identification methods?

Early AI wins have been promising — Insilico Medicine recently announced that an AI-identified anti-fibrotic small molecule inhibitor reached phase 2. But for every victory, there have also been flops. 

Exscientia deprioritized its cancer drug EXS-21546 after early-stage trials. BenevolentAI shelved its drug candidate for atopic dermatitis after seeing mixed results in a phase 2 trial. And Sumitomo Pharma saw its AI-identified treatment for schizophrenia, ulotaront, fall short of its primary endpoints in phase 3.

With these signposts along the way, it’s not yet clear if AI can improve upon traditional drug success rates, said Rahul Das, VP of AI and Life Science Solutions at Norstella. Between 2000 and 2015, only 14% of traditional drug candidates have ultimately hit their mark.

“We definitely know you can get a compound much faster using AI,” Das said, noting that, in some instances, the technology has shaved years off the process. “But then how they will perform in actual human clinical trials remains to be seen. Probably in two to three years, we will start seeing some of the results.”

While the future of AI-identified compounds remains murky, the technology is making a difference in other areas of drug development, Das said. 

“In clinical development we are already seeing the actual effect of AI, facilitating patient identification, faster recruitment, removing operational bottlenecks,” he said.” 

Drug companies use AI not only to flag promising compounds, but to identify and recruit patients who meet inclusion criteria for clinical trials, find high-performing sites to run trials and automate trial data. AI is also improving data collection and interpretation. 

“So now you can actually use that real-world data to identify which patients are responding and which patients are not responding to existing drugs,” Das said. Combined with biomarker data, this may make it easier to find the right drug for the right patient.

AI could also overcome one of the primary reasons people decline to participate in trials — the fear of receiving the placebo, Das said. 

“There is a concept called an external control arm, which is using real-world data to actually build a predictive model,” he said, noting that it allows investigators to treat all patients with the study drug while using computer models to predict how the control group would fare. 

As the industry awaits greater clarity as to whether AI-identified compounds are more successful, AI technology will continue to improve. 

“I believe we will continue to see a lot of benefits coming out of AI,” Das said. Even so, the industry should modulate its expectations. 

“Just because something has been designed by AI doesn’t mean it’s going to take away all the uncertainties,” Das said. “I don’t think these failures are going to be just failures — they are actually going to feed future success.”