US scientist John Jumper to leave Google DeepMind for Anthropic

This is a HUGE win for Anthropic and a big loss for Google. People with this talent are making hundreds of millions of dollars.

The story of AlphaFold is astonishing and it's worth learning about if you haven't heard of it. That AI breakthrough solution increased the knowledge of protein structures by thousands of times and reduced the time to map a protein by a similar amount. It's not overstating it to say that this is going to be one of the greatest inventions by mankind in all history.

There's a good explanation of AlphaFold on Veritasium from a year ago: "AlphaFold - The Most Useful Thing AI Has Ever Done". It includes an interview segment with Mr. Jumper starting at 1:54. Here's a direct link to John Jumper Interview.

Here's a summary of AlphaFold by Claude:

Historical Overview: From Manual Structure Determination to AlphaFold

The Early Era (1913-1960s): Foundational Breakthroughs

Although X-ray crystallography was first demonstrated by the Bragg family in 1913, protein structure determination remained an extraordinary technical challenge. After a further 20 years of painstaking work, scientists successfully determined the first protein structure: myoglobin, solved by John Kendrew in 1958. This single achievement required years of intensive effort. The second protein, hemoglobin (solved by Max Perutz in 1960), was similarly labor-intensive. These breakthroughs earned a Nobel Prize in Chemistry in 1962 and marked the birth of structural biology, but progress remained slow. In 1965, David Phillips determined the first enzyme structure, lysozyme, representing another years-long undertaking. During this era, solving a single protein structure could require a dedicated research team working for several years at tremendous cost.

The Slow Growth Period (1970s-1990s): Incremental Expansion

The Protein Data Bank (PDB) was established in 1972 with just 7 protein structures. Growth accelerated gradually as techniques improved: the first 100 structures were deposited by 1982 (10 years), reaching 1,000 structures by 1993 (an additional 11 years), and 10,000 by 1999. Throughout the 1980s and 1990s, X-ray crystallography remained the dominant experimental method, complemented by the introduction of nuclear magnetic resonance (NMR) spectroscopy in 1978 and electron microscopy in 1990. However, each structure still represented months to years of experimental work: researchers had to obtain crystals, collect diffraction data using synchrotron radiation, solve the phase problem, and build atomic models. By 1996, technological advances—including cryogenic freezing, area detectors, and multiple-wavelength anomalous diffraction (MAD)—made structure determination more efficient, resulting in four times as many new crystal structures that year compared to 1990.

The Modern Era (2000s-2010s): Steady but Plateauing Growth

The PDB reached 100,000 structures in 2014, representing exponential growth from the early years. However, this growth began to plateau around 2007. The experimental methods, though increasingly automated, remained fundamentally labor-intensive and expensive, limiting how many structures could be solved. By 2015, the PDB contained roughly 90,000 structures determined almost entirely by X-ray crystallography. The field reached a point where most readily accessible proteins had been solved, and determining new structures required increasingly specialized expertise and resources.

The AlphaFold Revolution (2020-Present): Exponential Explosion

The landscape transformed dramatically with AlphaFold 2 in 2020, which achieved near-experimental accuracy in protein structure prediction using artificial intelligence. For the first time, researchers could predict protein structures with high confidence without the years of experimental work. AlphaFold 3, released in May 2024, extended these capabilities to predict interactions between proteins, DNA, RNA, and small molecules—solving a multi-decade puzzle.

The impact has been staggering: whereas the PDB accumulated approximately 200,000 experimentally determined structures over 50 years (1972-2022), AlphaFold alone has generated over 200 million predicted structures. About 40% of new structures deposited into the PDB between 2024 and 2025 were obtained using AlphaFold. What once required years and hundreds of thousands of dollars now takes hours to days using freely available web-based tools like AlphaFold Server, with minimal computational cost. This represents approximately a thousandfold acceleration in the rate of structure prediction and a corresponding reduction in cost and time per structure—a transformation as significant to biology as the Human Genome Project was to genomics.

Current Status of AlphaFold

AlphaFold has progressed from AF1 to AF2 (which achieved near-experimental accuracy in single-chain protein folding) to AF3, expanding predictions to protein-ligand, protein-nucleic acid, and protein–protein complexes. The technology received major recognition in 2024 when its developers won the Nobel Prize in Chemistry.

Database & Accessibility

By April 2026, AlphaFold 3 (AF3) has populated the Protein Structure Database with over 200 million predicted structures. About 40% of the new structures deposited into the Protein Data Bank from 2024 to 2025 were obtained using AlphaFold. EMBL-EBI and Google DeepMind announced a partnership renewal in May 2026 and a major update to bring the AlphaFold Database in line with UniProt Knowledgebase.

Open Source Status

AlphaFold 3 was open-sourced for academic use after receiving the Nobel Prize in Chemistry, moving from a restricted web-only "black box" to an open-source model that democratized the ability to predict molecular interactions. However, as of March 2026, fewer than 50 institutions worldwide have local deployment access to the full model.

AlphaFold Server

Google DeepMind and Isomorphic Labs launched AlphaFold Server, a free web-based platform that provides scientists with free access to AF3 structure prediction capabilities for non-commercial research. AlphaFold Server is the most accurate tool in the world for predicting how proteins interact with other molecules throughout the cell. With just a few clicks, biologists can model structures composed of proteins, DNA, RNA and a selection of ligands, ions and chemical modifications. As of March 2026, it has helped make more than 8 million structure predictions for thousands of researchers around the world.

Key Applications

Drug Discovery & Pharmaceutical Development

Since its release in May 2024, AlphaFold 3 has fundamentally transformed molecular biology by solving how to predict interactions between proteins, DNA, RNA, and small molecules with atomic precision. In early 2026, Johnson & Johnson announced a deep-integration partnership to utilize AlphaFold 3 for designing novel protein-protein interaction inhibitors.

Binding Affinity & Drug Design: 76% of AlphaFold 3's predicted binding poses land within 2 angstroms of the experimental structure, compared to 38% for DiffDock, the previous best specialized docking tool. AlphaFold 3's ability to predict drug-target interactions with unprecedented accuracy is a cornerstone of its application in drug discovery, with its predictive prowess being 50% more accurate than traditional methods.

Protein-Nucleic Acid Complex Prediction

On protein-nucleic acid complexes, AlphaFold 3 achieves an lDDT score of 0.79 compared to RoseTTAFold's 0.65-0.70 range.

Emerging Proprietary Applications

In February 2026, Isomorphic Lab's (Google DeepMind's biopharmaceutical spin-off) announced an even more powerful AI model for drug discovery that is proprietary, advancing beyond the public AlphaFold 3. In April 2025, Isomorphic Labs raised $600 million in funding and is preparing for human trials of AI-designed drugs, with active collaborations with pharmaceutical companies including Novartis and Eli Lilly.

Other Applications

Structure-based vaccine design
Drug repurposing and repositioning
Research on intrinsically disordered regions and protein dynamics
Molecular system design and protein engineering

Current Limitations

AlphaFold 3 exhibits a ~22% hallucination rate when modeling intrinsically disordered regions (IDRs)
Restricted local access limits reproducibility and fine-tuning on proprietary data
Performance diminishes with complex dynamic protein movements

AlphaFold has transitioned from a research breakthrough to an industrial tool fundamentally reshaping how pharmaceutical companies approach drug discovery, while simultaneously completing a revolution in structural biology that has compressed decades of manual work into minutes of computational time.

Nothing short of amazing.

Physical chemistry and x-ray crystallography was the state of the art when I was in school at UCSD in the 1974 to 1976 timeframe. Purifying and crystallizing the protein often required the better part of a year. Getting the x-ray diffraction patterns and doing the detailed analysis was even more difficult. The ability to predict the physical structure using only computational methods is a huge win.

Wow, you lived it! So the AlphaFold breakthrough must be really meaningful to you with that lived experience.

This is a HUGE win for Anthropic and a big loss for Google.

Make that another big loss for Google - Noam Shazeer also left Google this week.

Yep, the speed at which these things are moving is mind boggling.

It’s hard to imagine where things will be even twenty yearrs from now.

There’ll be a lot of benefits, but I also imagine a bit of social upheaval when things move this fast.

It’s amazing how all these brilliant people are moving around so much now. There’s a huge bidding war for the top talent.

I’m surprised I haven’t gotten the call.

No worries...I have you on speed dial...

Wow, you lived it! So the AlphaFold breakthrough must be really meaningful to you with that lived experience.

I left the world of molecular biology after graduation, but put lots of effort in as a undergrad. I met with my profs for dinner on many occasions to discuss progress on their tasks and offer suggestions. The suggestions did have a positive influence, so the time was well spent.

Around 1986 I had an opportunity to pursue employment with a local company that was building protein folding models. It would have been a fine opportunity to take my existing skills in molecular biology and my current skills in software engineering into a new, hybrid future. I didn't pursue that and did just fine writing software for DoD customers. AlphaFold is certainly taking the hybrid skillset to a new level. A big win for pharmaceuticals.

But AI-based drug discovery is not quite there yet:

From AI:

Limitations and Challenges While the technology provides high-confidence maps, it is not a first-principles physics simulator. Current challenges include dealing with protein flexibility (proteins naturally shift between shapes), the massive expanse of chemical space to search through, and the need for subsequent wet-lab testing to validate the AI-designed compounds.

No worries...I have you on speed dial...

Jumpers are gonna jump.

The last section I posted:

Current Limitations
* AlphaFold 3 exhibits a ~22% hallucination rate when modeling intrinsically disordered regions (IDRs)
* Restricted local access limits reproducibility and fine-tuning on proprietary data
* Performance diminishes with complex dynamic protein movements

Yikes!! I get all sorts of AI hallucinations when I ask it to sum up numbers!

.....Yikes!! I get all sorts of AI hallucinations when I ask it to sum up numbers!.....

Yup!

Drug discovery still requires lots of human ingenuity, and lots of lab experiments!