AlphaFold and Beyond: AI That Proposes, Scientists Decide

In 2024, a Nobel Prize in Chemistry went, in part, to an artificial intelligence system. AlphaFold could predict the three-dimensional shape of a protein from its amino-acid sequence in hours, a problem that had resisted biologists for half a century. But the prize, and the science behind it, carries a subtler lesson than “AI solved biology.” These systems predict structures and propose molecules at a scale no laboratory could match; they do not confirm that a structure is functional, a drug is safe, or a hypothesis is true. That remains the work of experiments and of the scientists who design them. The story of AI in protein science is a story of decision support, not decision replacement.

Key takeaways

• A 50-year problem cracked: At the CASP14 assessment in 2020, DeepMind’s AlphaFold2 predicted protein structures with accuracy often indistinguishable from experiment, scoring above 90 on the GDT metric for roughly 87% of the hardest targets.
• Structure at planetary scale: By 2022 the AlphaFold Protein Structure Database held over 200 million predicted structures, covering nearly all catalogued proteins, and is now used by millions of researchers across 190+ countries.
• Nobel recognition in 2024: The Chemistry prize honored David Baker for computational protein design and Demis Hassabis and John Jumper for protein-structure prediction.
• From shapes to interactions: AlphaFold3 (2024) extended predictions to how proteins bind small-molecule drugs, DNA, RNA and ions, the chemistry that actually matters for medicine.
• Candidates, not cures: AI has surfaced novel antibiotics (halicin, MIT 2020) and generated drug candidates now in clinical trials, but every one still passes through experimental validation and human trials before it reaches patients.

AlphaFold2: a prediction engine, validated against reality

The decisive moment came at the 14th Critical Assessment of Structure Prediction (CASP14) in 2020, a blind, community-run contest in which teams predict structures that have been experimentally solved but not yet released. DeepMind’s AlphaFold2 returned models that were frequently within experimental error, achieving a Global Distance Test score above 90 for roughly 87% of the most difficult targets, a margin that led many in the field to consider single-chain structure prediction effectively solved. The key word is “assessment”: AlphaFold’s accuracy is meaningful precisely because it is measured against painstaking experimental structures from X-ray crystallography and cryo-EM. AI made the prediction; decades of laboratory science provided the answer key. The full method was published in Nature in 2021. Source: Jumper et al., Nature, 2021. Read the paper.

The 200-million-structure database: scale that reframes the question

In July 2022, DeepMind and EMBL’s European Bioinformatics Institute expanded the open AlphaFold Protein Structure Database from about one million entries to over 200 million, covering nearly every protein catalogued in UniProt; the database now holds more than 214 million predictions. DeepMind has framed this as compressing what would have taken “hundreds of millions of years” of experimental work, a vivid but order-of-magnitude estimate rather than a precise audit, and one worth treating as illustrative. What is concrete is reach: more than two million researchers across 190-plus countries had used AlphaFold by 2024. Crucially, a predicted structure is a starting hypothesis: it tells a biologist where to look, not what is true. Researchers still validate function in the lab. Source: EMBL-EBI / Google DeepMind, 2022. See the release.

The 2024 Nobel Prize: prediction and design, both human-led

The 2024 Nobel Prize in Chemistry split along two complementary lines of work. One half went to David Baker of the University of Washington for computational protein design, building entirely new proteins, including candidate drugs, vaccines and nanomaterials, that do not exist in nature. The other half went jointly to Demis Hassabis and John Jumper of Google DeepMind for protein-structure prediction with AlphaFold. The pairing is telling: prediction (reading what nature built) and design (proposing what it did not) are both AI-accelerated, yet both depend on experimentalists to express, purify and test the molecules before any claim becomes a result. The committee honored tools that expand what scientists can attempt, not tools that replace their judgment. Source: The Royal Swedish Academy of Sciences / NobelPrize.org, 2024. Read the press release.

AlphaFold3 and AI drug candidates: from shapes to medicines

Structure alone rarely cures anyone; interactions do. In May 2024, DeepMind and Isomorphic Labs published AlphaFold3 in Nature, using a diffusion-based architecture to predict how proteins bind not just to each other but to small-molecule drugs, DNA, RNA and ions, the binding events that drive disease and therapy. Parallel efforts target the molecules themselves. In 2020, an MIT team led by Jonathan Stokes and James Collins used a deep-learning model to screen vast chemical libraries and surface halicin, a structurally novel antibiotic, then confirmed its activity in the lab and in mice. More recently, Insilico Medicine advanced INS018_055 (ISM001-055), a TNIK inhibitor for idiopathic pulmonary fibrosis designed using generative AI; in November 2024 the company reported positive topline Phase IIa results, with the candidate appearing safe, well tolerated and showing encouraging signals on lung function. These remain candidates inside the multi-year gauntlet of human trials, where regulators and clinicians, not algorithms, decide. Source: Abramson et al., Nature, 2024; Stokes et al., Cell, 2020; Insilico Medicine, 2024. AlphaFold3 paper.

What’s next: faster hypotheses, same burden of proof

The trajectory is clear: models are moving from static structures to dynamic interactions, from reading proteins to designing them, and from one molecule at a time to libraries screened in silico. The bottleneck is shifting too. When generating plausible candidates becomes cheap, the scarce resources become experimental capacity, careful validation and good scientific judgment about which of thousands of AI proposals is worth a wet-lab week. AlphaFold’s own caveats matter here: predictions come with per-residue confidence scores precisely so scientists know where to trust the model and where to verify. The promise is not automation of discovery but acceleration of it, with humans still steering.

The through-line: decision support, not replacement

Across every example here, the same division of labor holds. AI compresses an impossibly large search space into a ranked list of strong hypotheses: this is the likely fold, this molecule may kill these bacteria, this compound might inhibit that kinase. Scientists then do what AI cannot, run the experiment, weigh the confidence score against the assay, decide whether a promising signal survives contact with a living system, and choose what to pursue. The 2024 Nobel did not retire the laboratory; it gave researchers a faster way to decide where to point it. AI predicts and proposes; people validate and decide.

Methodology & sources

• AlphaFold2 method and CASP14 accuracy — Jumper et al., Nature (2021)
• 200M+ structures released in the AlphaFold DB — EMBL-EBI (2022)
• Database now exceeds 214M structures; usage and impact framing — Google DeepMind (2024)
• 2024 Nobel Prize in Chemistry (Baker; Hassabis & Jumper) — NobelPrize.org (2024)
• AlphaFold3 predicts protein-ligand, DNA, RNA and ion interactions — Abramson et al., Nature (2024)
• AI antibiotic discovery (halicin), 107M molecules screened — Stokes et al., Cell (2020)
• Generative-AI drug INS018_055 Phase IIa topline results — Insilico Medicine (2024)

Frequently asked questions

Did AI replace the lab in protein science?

No. AI predicts structures and proposes molecules, but scientists still validate them experimentally and decide what to pursue. AlphaFold’s accuracy was established by comparing its predictions against structures solved in the lab, and AI-generated drug candidates must still pass laboratory testing and human clinical trials before reaching patients.

How accurate is AlphaFold?

At the CASP14 assessment in 2020, AlphaFold2 scored above 90 on the Global Distance Test for roughly 87% of the hardest targets, often within experimental error for single protein chains. Accuracy varies by protein and region, which is why AlphaFold publishes per-residue confidence scores so researchers know where predictions are reliable and where to verify.

Have AI-designed drugs reached patients yet?

Not as approved medicines, but candidates are advancing through trials. Insilico Medicine’s generative-AI-designed INS018_055 for idiopathic pulmonary fibrosis reported positive topline Phase IIa results in 2024, and AI-surfaced antibiotics such as halicin have shown activity in lab and animal models. All remain in the multi-year validation process that determines whether a candidate becomes a real treatment.

Part of our Real-World AI Use Cases series — how AI supports high-stakes decisions across surprising domains.