Can AI discover medicines humans would never imagine?

The Question

For a century, drug discovery has moved at the pace of human intuition. Generative models change what a chemist can imagine — and what a patient might one day receive. The question is no longer whether machines can propose molecules. It is whether they can propose the right ones, for the right reasons, and whether we can tell.

The Researcher

Elena Marchetti's laboratory sits inside one of the world's largest biomedical institutes, but her team works more like a small studio. Chemists, machine-learning researchers and clinicians share a single open floor. Every model that leaves her group is expected to survive contact with a wet lab within a month.

The Research

Her current project asks whether latent-space navigation can surface therapeutic candidates for diseases so rare that no traditional pipeline can justify them. Early results, published in Nature this year, suggest that generative models trained on structural data can propose viable leads for indications with fewer than a thousand known patients worldwide.

In Practice

Two clinical collaborators are already testing molecules that emerged from the group's pipeline. Both were suggested by the model without any human seed compound. Neither would have been designed by a working medicinal chemist.

What's Next

The next question is regulatory rather than technical: what does it mean to approve a medicine whose origin cannot be fully explained? Marchetti is working with the FDA and EMA on interpretability standards for generative discovery.

Further Reading

• The Alignment ProblemBrian Christian
• The Emperor of All MaladiesSiddhartha Mukherjee
• Ways of BeingJames Bridle

Field Notes