AI will only revolutionize drug discovery and development if we play to its strengths

News Drug Development Opinion: AI will only revolutionize drug discovery and development if we play to its strengths September 8, 2026 | 5 min read | Thomas M. Kaiser Twitter LinkedIn Facebook Email Print Foundation models will change the way we study the natural, applied and clinical sciences, but only under certain conditions, and not in the way that some AI leaders claim. The poor reader of today’s news is bombarded by articles hyping AI or their doom and gloom opposite.

This dichotomy exists even in biopharma. Grand proclamations of the technology’s ability to accelerate novel drugs to market or sift through the scientific literature to find undiscovered targets run alongside articles that rail about how artificial intelligence in medicine, biology and chemistry is producing useless slop and will only consume exorbitant resources without accelerating research. Pinky promise, this article is something new.As a medicinal chemist, physician and machine learning (ML) scientist, I believe that AI will shift the nature of our drug design and development work, but not in the ways widely being presented.

AI approaches will change the way we study the natural, applied and clinical sciences but only under certain conditions, and we need to talk about the trends and limitations with current model types.The array of AI techniques is dizzying. There are many exciting approaches in AI subfields like ML that are becoming ubiquitous in drug development. Deep learning techniques have attracted a great deal of investment over the past five years, with one particular technique drawing significant attention more recently: the foundation model.

Reports involving use of these models have skyrocketed in the academic literature, from a single foundation model published in Q4 of 2021 to 226 foundation models published in Q2 of 2025, according to a BCG study that reviewed publications in chemistry, biology and medicine.Foundation models are broadly trained ML algorithms that use large amounts of data from a variety of public sources; large language models powering ChatGPT and Claude are familiar examples. These models are adapted to specific tasks and can include a generative output. Recently, Anthropic announced Claude Science, a foundation model approach that promises to remove the tedium from science and enhance the pace of finding connections within the scientific literature.

The live demo of the tool declared that it is ready to directly run the scientific work of drug discovery, and that AI-enabled biology will soon produce 100 years of progress in just 10 years. There would need to be more than 470 drugs approved every year for 10 years for that proposed revolution to be realized!While that ambition for AI in science is noteworthy, there are two significant roadblocks. First, the literature is filled with irreproducible results that corrode the knowledge base of foundation models.

And second, even when there are reproducible data, hidden changes in the cause-effect relationship can distort ML approaches. Artificial intelligence With tech backing, AI biotechs force biopharma into ‘fail-fast’ drug development AI is quickly becoming a central force in drug development, from powering pharma engines to permanently rewiring the capital markets. August 12, 2026 · 6 min read · Tristan Manalac Read more Irreproducible data are common and difficult to detect The progress in the capabilities of foundational modeling seems impressive at first glance.

One of the major points of excitement is the potential for drawing new connections using existing literature. The hope is to identify novel and attractive targets linked with diseases of high unmet need.But the problem with feeding scientific publications into foundation models is that several analyses by Nature and the MD Anderson Cancer Center, among others, have found there is a reproducibility crisis plaguing the literature. To put it simply, a large block of the knowledge seemingly present in peer-reviewed journals is flawed or incomplete and cannot be replicated.

Worryingly, we only learn that otherwise impressive-looking results are irreproducible, even without any malfeasance on the part of the original research team, after extensive scientific work.Importantly, this lack of reproducibility scattered throughout the scientific literature is very different from the ML reproducibility and reasoning tracing movements in AI. In AI, engineers are working toward foundation models that produce the same result when prompted with the same query for the same version of model. While important for the further development of AI approaches, a reproducible connection based on irreproducible science is sadly not progress.

Paradoxes in probability models abound While flawed data undoubtedly limit AI’s potential, the models also face another issue. Machine learning, including foundation modeling, is best understood by using the tools of probability theory. Frustratingly for ML in drug discovery, probability distributions even in reproducible datasets can change with a shifting context.

Two great examples are Bertrand’s paradox, where probabilities of a phenomenon change when using different methods of random selection from the same pool, and Simpson’s paradox, where trends in probabilities shift depending on how different subgroups are combined.Context in the case of small molecule drug design causes problems for AI because minor chemical tweaks can cause dramatic and seemingly magical changes in experimental behavior. However, when the chemist implements these same chemical tweaks for a new and distinct chemical family, those chemical tweaks can change a molecule’s properties in different and surprising ways.It is this nuanced contextual change in the rules of binding for small molecules that presents a problem for foundation models. We at Avicenna have shown that desirable chemical changes in one chemical series may be disastrous for ML modeling on another series, even for the same property (e.g., on-target potency), because the cause-effect relationship unexpectedly shifted in the data.Avicenna and others like the Doyle group out of MIT are working to solve hidden cause-effect problems in pharmacology data, but the paradoxical behavior of probabilities makes this tricky.

The context needed to model the properties of molecules in a dataset is often not known, especially for complex data like vertebrate toxicology mediators or pharmacokinetic metrics. Nevertheless, progress is being made; new methods are showing promise in addressing these cause-effect conflicts in scientific data. Artificial intelligence AI Is Changing Pharma’s Bottom Line Now—But Not Through Splashy Drug Discovery Pfizer, Eli Lilly, Novartis, Bristol Myers Squibb and AstraZeneca are all ramping up the use of AI, but drug discovery is not the primary success story—yet.

February 11, 2026 · 8 min read · Annalee Armstrong Read more Outlook for foundation models in drug development So, where can contemporary foundation model efforts bear fruit for drug design? As long as there is an effort to focus on a key, well-phrased scientific question that is mindful of contextual nuance, foundation techniques have potential for identifying novel biological networks as drug targets. Additionally, there is an enormous opportunity for mining clinical data to aid in translational development as far as patient selection is concerned, as Yatiri Bio is demonstrating and as a collaboration between Bruker and Noetik is exploring.The potential for foundation techniques to discover translational connections between clinically meaningful endpoints, appropriate patient population selection and preclinical models that reflect human disease phenotypes could improve the success rate of drug development.

But only if experimental rigor and careful evaluation of data remain in the remit of the scientist. Twitter LinkedIn Facebook Email Print Opinion Artificial intelligence Machine learning Thomas M. Kaiser Thomas M.

Kaiser, MD, PhD, is the co-founder and chief scientific officer at Avicenna Biosciences. He founded Avicenna with Pieter Burger in 2019.