Beyond Language: Why Drug Discovery Needs a Different Kind of AI

In wet labs and dry labs, investor meetings, and medical conferences, it’s impossible to avoid discussions about how AI is transforming the way medicines are discovered.

It is and it isn’t. 

This “AI in drug discovery” dogma gets something wrong. It assumes that “AI in drug discovery” is one thing. It is not. It’s true that AI systems help drug discovery teams work faster by organizing information, automating tasks, and, most tantalizingly, generating plausible ideas for future medicines at unprecedented speed.

But there’s still a lot missing. This is language-based AI. Faster, more organized work through large language models is not the same as better prediction of the properties of a physical structure in a live organism. The hardest question in drug discovery remains stubbornly tied to the physical world, not the digital one: which molecules will actually work in biology’s messy, dynamic environment? 

To materially improve drug discovery, we need to distinguish between systems that organize information with language and systems that predict physical reality. Until the field makes the distinction between orchestrating and predicting, capital and scientific effort will continue to flow toward the wrong problems, and patients will wait longer than they should for new medicines.

Language-Based AI Is a Powerful Start. Drug Discovery Needs More.

The most visible AI systems today — including ChatGPT, Claude, and Gemini — are built to recognize patterns in symbolic information such as alphabets, sequences, and other structured data. They are already proving useful in drug discovery, from helping teams search information, summarizing knowledge to generating hypotheses and working more efficiently.

What makes drug discovery hard is that every promising molecule must pass through a gauntlet of physical tests rooted in chemistry, biology, and pharmacology. What matters is the physical properties of these molecules in the real world, not just in how language can describe them. Language- based AI can help summarize and classify what we already know. What it cannot do, at least not on its own, is reliably predict the physical properties that determine whether a molecule will actually work in a cell assay, animal model, or human patient. That is the gap Physical AI is meant to close.

Why Physical AI Matters

An AI system trained on airplane blueprints might generate plausible aircraft designs, but engineers would never skip aerodynamics, simulation, wind tunnels, stress tests, and flight testing. Drug discovery is similar. A plausible molecule is only a starting point. The real challenge is predicting whether it will survive the experimental and physical tests imposed by biology along the path of discovery, pre-clinical studies, and human trials.

Physical AI operates on a similar logic. It is built to predict how real systems behave under chemical, biological, and time constraints. In drug discovery, that means predicting molecular behavior: binding, motion, solubility, permeability, metabolism, toxicity, exposure, pharmacology, and more.

This matters because drug discovery is fundamentally about how molecules behave inside and outside the cell. A molecule does not become a medicine because it is an interesting idea, or even because it looks plausible in a model. It becomes a medicine only if it behaves correctly in the crowded, complex, dynamic environment of cells that give rise to life. Predicting that behavior is the central challenge.

Different problems will require different forms of prediction: synthesis, binding, solubility, permeability, metabolism, toxicity, exposure, and pharmacology, each of which imposes different physical constraints and require different data. To succeed, physical AI will not use one predictive model. It will be a system of task-specific models, simulations, experimental feedback loops, and decision tools. 

Molecular representation matters, too. Many current AI systems still treat molecules as strings or flat diagrams. But in drug discovery, molecules are 3-dimensional entities that change shape and properties within the crowded and changing biological environment. Small changes can have enormous consequences. Move one atom, change one bond, or alter one geometric arrangement, and a molecule can bind differently, distribute in the body differently, cause toxic effects, or fail entirely. Molecules are not static sketches. They are dynamic physical objects defined by electron clouds, geometry, motion, and quantum-mechanical interactions with their surroundings. Better prediction requires models that treat molecules much closer to that reality.

The practical lesson is simple: molecule generation is not molecule optimization. The real value in drug discovery comes not from proposing more molecules, but from learning faster which ones are actually worth making, and why. This is not just a scientific distinction. It is a practical framework for deciding where to invest capital, talent, and attention. The relevant question is no longer whether a company uses AI. It is which decisions its AI improves, how much those decisions improve, and whether those decisions ultimately lead to better medicines.

In other words, the key distinction is not between companies that use AI and companies that do not. It is between AI that helps scientists navigate existing information and AI that helps them predict behavior of novel molecules. Both matter. The first can be helpful is speeding up existing processes. But only the second can improve the odds that a molecule will become a medicine.

The Real Test: Scientific Leverage, Not Workflow Efficiency

Real progress in AI for drug discovery should not be judged only by whether it makes teams move faster. It should be judged by whether it creates scientific leverage: better decisions about what to make, what to test, what to stop, and why.

A useful rule of thumb: be skeptical when a system is easy to demo but hard to validate. In drug discovery, the most valuable models are often not the ones that produce the flashiest outputs, but the ones that quietly help teams avoid bad bets and make better molecules.

Language-based AI will absolutely be part of the future of drug discovery. It will help search the literature, map the competitive landscape, write code, coordinate teams, and increasingly orchestrate closed-loop laboratories that connect prediction, experiment, and decision-making. It will make discovery organizations more nimble, better informed, and better coordinated.

Better prediction will also require better infrastructure. Progress in drug discovery depends not just on clever models, but on rigorous benchmarks, blind challenges, and the ability to scale. It also depends on infrastructure that is open, extensible, and designed for others to build on, much like shared platforms in the technology world have accelerated innovation far beyond what any one company could do alone. Not everything needs to be open, but not everything can be built from scratch within a single organization. 

One especially promising area is co-folding, where the field has made striking progress in predicting biomolecular structure — progress significant enough to be recognized with the 2024 Nobel Prize in Chemistry. But structure alone is not enough. The next step is to connect those models to molecular recognition, energetics, mechanism, and design — the things that determine whether a structure is merely interesting or actually useful for making medicines.

This is the premise behind work done with my academic collaborators, colleagues at PsiThera, and the OpenFold community on “Grand Challenges for Predictive Modeling in Small Molecule Drug Discovery.” The impact of AI will ultimately depend on whether it helps teams predict which molecules to make next, which risks to prioritize, and which compounds have a realistic path to becoming drugs.

For investors and managers, that distinction matters. It separates operational efficiency from scientific leverage. The most important question is not whether a company can generate plausible outputs, but whether its technology improves real molecular decisions in ways that change asset quality, cycle time, capital efficiency, and ultimately probability of success.

That is the standard the field should reward: prospective validation over retrospective storytelling; physically grounded prediction over pattern matching alone; and open, extensible infrastructure over black-box claims that are difficult to test. If drug discovery makes that shift, AI will not merely accelerate existing processes. It will change how new medicines are found.

 

Woody Sherman, Ph.D., is the chief innovation officer of PsiThera and Chair of the OpenFold Consortium Executive Committee