Derek Lowe on AI in Drug Discovery: Between Hype and Hope

By Allison Proffitt 

April 15, 2025 | At last week’s Venture, Innovation, and Partnering Summit at the Bio-IT World Conference & Expo, Derek Lowe, Director, Chemical Biology & Therapeutics at Novartis BioMedical Research, and John Keilty, venture partner at Third Rock Ventures, sat down for a candid conversation on the pharmaceutical industry and the current state of artificial intelligence and machine learning (AI/ML) in drug discovery. 

Lowe, known for his long-running industry commentary at the In the Pipeline blog, reflected on the cycles of hype that have accompanied new technologies over the decades—from the early days of computer-aided drug design in the late ’80s to sequencing to antisense to the current AI boom. “There have been times that I thought, ‘Ok, the hype has peaked,’ but it hasn’t peaked,” Lowe said.    

Every time a new technology emerges, there’s a wave of excitement, but “pragmatism eventually takes over,” Keilty agreed. Together they discussed how to arrive at pragmatism in drug discovery.  

AI/ML has shown promise in areas like protein structure prediction and antibody design—thanks to well-structured datasets like those found in the Protein Data Bank (PDB)—but Lowe and Keilty argued that its utility in more complex fields like small molecule discovery and clinical translation remains limited. 

Current models struggle because of the lack of consistent, high-quality datasets, especially when translating results from lab assays to human trials. Lowe pointed out that the central failure of drug development—an 85% failure rate in clinical trials—is not due to a lack of ambition, but a lack of predictive insight. However, better target selection and human toxicity prediction, the two biggest hurdles, are not yet within AI’s reach. “Those are the two things that kill most of the [drug discovery] programs,” he said.   

Despite this, both agreed that AI has meaningful roles to play—especially in discovery-stage tasks with better-defined parameters. “Proteins have a much more limited vocabulary. There’s only 20-odd words in the language, and the grammar is also pretty repetitious, too,” Lowe explained. The repetition is why models like AlphaFold work.  

But when it comes to small molecules, the chemistry is vast and the data are often inconsistent or flat-out wrong. “The problem is, the synthetic organic chemistry literature is a [mess],” Lowe said, colorfully. “It has really got a lot of problems.” A machine learning model assumes the published literature reflects careful choice in decades of experiments. But the truth is, Lowe said, scientists’ choices have always been influenced by the products that are on the shelf, the techniques he or she already knows, and other biases and inaccuracies reflected in decades of published data. 

Pushing past the hype, Lowe asserted, will not be pretty. “I think we’re going to have to get to the point where a lot of people get really pissed off and abandon the field, because I think that’s the natural end of the hype cycle,” he said. “That’s a cheerful prospect, I know.”  

These growing pains are natural and necessary, though. Lowe praised the current machine learning algorithms and compute power, but called for high quality and large amounts of data to feed into the models. “I think if you see people really getting serious on that end of it, as opposed to issuing press releases about how amazing they are, then we’re actually heading toward a better spot,” Lowe said.    

Keilty brought up the challenge of biological complexity and the need to rethink scientific reductionism. Lowe agreed, illustrating with an image analogy: just as one cannot reconstruct a Seurat painting from a close-up of colored blobs of paint, understanding disease biology from isolated molecular pathways is fundamentally limited. 

Keilty asked what low-hanging fruit is available and Lowe turned his attention to improving models. “It is kind of disgraceful that we’re at this point where our animal and small, organoid-type models for things like CNS are so terrible,” he said. “Could you stick a machine learning algorithm onto that and start looking at all these possible models and see where tiny improvements are made and build on that?”  

Lowe also expressed interest in ways to generate more data, perhaps pulling more information from existing tests results and medical images. “I’m all for collecting the data, because yeah. It may turn into gold. That’s the story of science.”  

Lowe encouraged the audience to find the balance between being “a grumpy, crusty, cynical SOB” and falling for the marketing claims. “I tell people I’m a short-term pessimist and a long-term optimist. I see no reason why these things can’t work—no mathematical, no logical, no biological or scientific reason why these things can’t work. It’s just that it’s going to be very hard to get them to work… Stick with the long-term optimism part, and don’t get too swept up in something that happens in the moment.”