Who Will Invent the Next Medicine—Scientists, AI, or Both?

Clarus Biotech Review

AI can widen the search for a medicine. The more interesting question is how scientists decide which possibilities deserve to become experiments.

 Picture a laboratory meeting in which the longest list of ideas comes from a machine. The most consequential contribution may still be a scientist asking, “What would convince us that this is wrong?” 
 This imagined scene captures a useful way to think about AI in medicine discovery. Generating possibilities and establishing knowledge are different activities. A promising future connects them more tightly: computational systems propose, researchers challenge, experiments reveal, and the next round begins with better questions. 
 A larger map of what might work 
 One concrete advance is the ability to model complex biological structures. The AlphaFold 3 research, published in Nature in 2024, described a system for predicting structures involving proteins, nucleic acids, small molecules, and other components. Its benchmark results demonstrated progress across several classes of molecular interactions. That is a substantial contribution to the tools researchers can use to investigate biology. [1]  
 A predicted structure can help frame an experiment. It does not, by itself, establish that a compound will reach the right tissue, produce the desired effect, or be safe in people. The distance between an attractive molecular picture and a useful treatment remains a sequence of questions that need their own evidence. 
 For a reader outside the laboratory, an architectural sketch offers a helpful analogy. A richer drawing can improve the choices a team explores. Whether the building stands, serves its occupants, and can be maintained still depends on engineering, construction, and use. In medicine, those later questions are part of the invention. 
 The experiment becomes more valuable 
 Consider a hypothetical research team with the capacity to test twenty candidates. A model produces two thousand plausible suggestions. The team's constraint has shifted: it now needs a defensible way to select twenty experiments that will teach it the most. 
 Choosing only the candidates with the highest scores could be reasonable for one purpose. Choosing a more varied group might expose a model's blind spots. Including appropriate controls could show that an apparently impressive result reflects something mundane about the assay. These are scientific decisions about what the next piece of evidence should resolve. 
 The collaboration works best when a prediction arrives with context. Which observations informed it? Where has the model been evaluated? What uncertainty matters for this particular choice? A confident answer to the wrong question can consume a laboratory's time as efficiently as a weak hypothesis. 
 Credibility depends on the decision 
 FDA's January 2025 draft guidance on AI in drug and biological product development proposed a risk-based approach to evaluating model credibility for a specified context of use. The point is useful beyond regulation: confidence in a tool should relate to the decision it is being asked to support. The cited document is draft guidance, rather than a claim that every proposed practice is a binding requirement. [2]  
 An exploratory tool used to suggest a laboratory experiment carries a different burden from a system whose output could influence a consequential assessment of safety or effectiveness. Calling both “AI” tells us very little about whether either is fit for its task. 
 For organizations evaluating these technologies, the compelling demonstration would connect a model's contribution to better decisions. Did it reveal a useful direction that the team would otherwise have missed? Did it help eliminate a weak approach before further resources were committed? Could another team reproduce the result? Those questions make the story more demanding and more interesting. 
 Shared work, clear responsibility 
 The title's question may ultimately matter less than the division of work it reveals. Scientists define meaningful problems, design tests, interpret surprises, and remain accountable for decisions. Computational tools can expand what those scientists are able to inspect and propose. Patients and clinicians help establish which improvements would matter outside the laboratory. 
 A productive future would make that chain easier to examine. An AI-generated suggestion should lead to a traceable hypothesis, a well-designed experiment, and evidence that can withstand scrutiny. Each successful handoff creates something more durable than an impressive demonstration. 
  A question to take back to the white papers  When a platform claims to accelerate discovery, ask which decision became better—and what evidence shows the improvement. Speed is most valuable when it helps a team learn something that changes its next move.  

Sources

https://www.nature.com/articles/s41586-024-07487-w

https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological

Original article: https://www.clarusreview.bio/articles/who-will-invent-the-next-medicine