The Protein Folding Revolution Is Just Getting Started

In 2020, DeepMind's AlphaFold2 achieved something the scientific community had spent five decades trying to solve: predicting the three-dimensional structure of proteins from their amino acid sequences with near-experimental accuracy. The announcement was met with something rare in science — genuine, cross-disciplinary awe.

But the announcement was also widely misunderstood. The headlines said "AI solves protein folding." The implication was that the problem was done, the chapter closed. What actually happened was more interesting: a new chapter opened, one that's being written faster than almost anyone anticipated.

What protein structure tells you (and doesn't)

A protein's structure determines its function. If you know the shape of a protein involved in a disease, you can — in principle — design a molecule that fits into it like a key into a lock, disrupting or enhancing its activity. This is the basis of most modern drug development.

The problem is "in principle." Structure is necessary but not sufficient for drug design. Proteins are dynamic — they flex and shift and change shape in context. They interact with other proteins, with membranes, with small molecules in complex ways. AlphaFold gave us high-quality static snapshots. The biology happens in the movie.

Subsequent models — including ESMFold, RoseTTAFold, and now a wave of generative protein design tools — have started to capture more of that dynamism. The field has moved from "predict structure from sequence" to "design sequences that fold into target structures" to "design proteins with specified functions." Each step compounds the previous one.

The drug discovery acceleration

Traditional drug discovery timelines run 10-15 years from target identification to approval. The bottlenecks are many, but structure-based drug design — knowing what shape you're trying to hit — has historically been one of the slower, more expensive steps.

That's changing. Several biotech companies founded in the last five years are now running what would have been years of structural biology work in weeks. Insilico Medicine put an AI-designed drug into Phase I trials in three years. Recursion, Relay Therapeutics, and others are publishing results that would have seemed implausible a decade ago.

None of this means every drug candidate will succeed. The clinical failure rate for drug development remains brutal — around 90% of Phase I compounds don't make it to approval. AI doesn't fix the biology; it just helps you navigate to the right starting point faster.

The part nobody's talking about enough

The most underreported aspect of the protein structure revolution is what it means for biology as a basic science — not just drug development.

We now have structural data for essentially every human protein. We're generating structural data for pathogen proteins, environmental proteins, synthetic proteins, at scales that weren't conceivable five years ago. This is creating a new kind of biological map — one that lets researchers ask questions they couldn't previously formulate.

How do different organisms' versions of the same protein differ? What structural changes accompany evolution? How does structure correlate with disease? These questions are now computationally tractable in ways they weren't before, and we're only beginning to develop the tools to exploit that tractability.

The protein folding problem is solved. The protein design problem is just beginning. And the protein understanding problem — what all these structures mean for life, disease, and the possibilities of designed biology — may be the defining scientific project of the next decade.