If you’ve been following the buzz around artificial intelligence and pharmaceuticals, you’ve probably heard bold claims that AI is on the verge of revolutionizing drug discovery. But digging a little deeper, it turns out AI-designed drugs haven’t yet cleared the final, toughest hurdle in drug development — the phase 3 clinical trials, where efficacy and safety are tested on a large scale.
That might soon change. I recently discovered that InSilico Medicine’s small molecule, rentosertib, could become the first AI-designed drug to officially enter phase 3 trials within the next couple of years. This drug targets idiopathic pulmonary fibrosis, a chronic lung-scarring disease. Their 71-patient phase 1/2 study in China demonstrated that rentosertib was safe and well-tolerated, a key milestone on this high-stakes journey.
The promise of AI is to be faster and a little more sensitive in detecting signals in a large ocean of noise.
How AI turbo-charges drug discovery – and where it hits limits
One of the biggest strengths of AI, especially machine learning, lies in its ability to sift through massive biological datasets efficiently, mapping out protein targets or genes worthy of deeper exploration. I came across insights from Chris Meier, formerly at pharma and now with Boston Consulting Group, who emphasized that AI can be a turbocharger for drug discovery by hunting signals that might be missed by human researchers.
Research even shows AI-discovered molecules in early clinical stages can have success rates of 80-90%, substantially above historical averages of around 66%. That’s a striking statistic, suggesting AI does pick some promising candidates more reliably — at least early on.
But there’s a catch. While AI excels at mining chemical databases and predicting which molecules might interact with known targets, experts like Andreas Bender at Khalifa University warn that much of this exploration remains within well-mapped biological territory. In other words, AI mostly suggests candidates against targets we already understand, which might partly explain why early safety signals look promising.
Medicinal chemist Derek Lowe also stresses caution. He points out that many AI-claimed breakthroughs involve targets already known to disease biology, and he worries about overselling AI’s revolutionary potential amid waves of enthusiasm for computational methods over the years. Adding to the challenge is AI’s dependence on existing data, which suffers from biases — for example, failed experiments or negative results rarely get published, skewing the information AI learns from.
The hype-versus-hope tightrope in AI-driven pipelines
Given these realities, where does AI make the biggest difference and where does it struggle? Machine learning can suggest novel molecule designs and speed up early lab testing, but it’s less adept at predicting complex human responses such as unexpected toxicity. This limitation becomes crucial because late-stage failures in clinical trials cost hundreds of millions of dollars and years of time.
Phase 1 trials focus on safety with a handful of participants, which is relatively affordable. But phase 2 and especially phase 3 trials require large patient cohorts and multi-year commitments. AI-designed drugs like rentosertib still must prove they can effectively treat disease at this scale — no guarantee yet. And many industry insiders think AI mostly helps with the initial, less expensive steps, while the costly, more uncertain phases remain a hurdle.
Still, the momentum is undeniable. Major pharma companies are investing billions into AI biotech partnerships. For instance, Isomorphic Labs, part of Alphabet, signed big deals this year with Eli Lilly and Novartis. Companies like Benevolent and Recursion also showcase how AI-driven automation and machine learning can shorten the drug development timeline substantially. Recursion’s recent 18-month journey from target initiation to new drug application submission is well below the industry average of 42 months, which is impressive.
Yet reality bites. Some AI-focused biotechs have trimmed pipelines or even shuttered clinical programs, signaling that financial and clinical challenges remain substantial. As Lina Nilsson from Recursion mentioned, strategic prioritization means doubling down on oncology and rare diseases — areas that might better fit AI’s strengths and current data availability.
Why I’m cautiously optimistic about AI’s long game in drug discovery
There is a data gap, especially around complex patient biology and toxicity prediction, that AI cannot overcome without better, more transparent clinical and experimental datasets. But the foundation in speeding up target ID and lead optimization is solid. Eventually, as datasets improve and models grow more sophisticated, I expect AI to bridge more of those gaps.
So, while rentosertib and its forthcoming phase 3 trial results may be a litmus test for AI’s true transformative impact, the pharmaceutical industry’s ongoing embrace of AI-powered discovery tools signals a shift unlikely to be reversed. It’s a fascinating moment where technology is reshaping hope for faster, smarter drug development — even if the full promise is still unfolding.
If anything, the journey of AI in drug discovery reminds me that progress in medicine never rushes. It requires measured optimism, relentless iteration, and respect for the unknowns ahead.
