CyphrQ
Menu

AI in Scientific Research: What AlphaFold and AI Discovery Have Shown

AI predicted protein structures, proposed new crystals and found faster algorithms. What the papers show, what they don't, and what to watch next.

By CyphrQ.ai

Written with AI assistance and human editorial review. Every figure linked to its source.

9 min read

AI in scientific research has moved from helper to contributor in a few fields. AlphaFold predicts protein structures from their amino-acid sequence, and its creators shared the 2024 Nobel Prize in Chemistry. Google DeepMind's GNoME (2023) proposed 2.2 million candidate crystals, and its AlphaDev (2023) found faster sorting routines now in the LLVM standard C++ sort library. Each result is real, and each has limits.

This post sets out what the peer-reviewed papers and the labs' own pages report, what they do not yet show, and what to watch next. This research could change how science gets done, so the details matter.

What did AlphaFold actually do?

Proteins are chains of amino acids that fold into 3D shapes, and the shape largely decides what a protein does. Working out one shape by experiment, with methods such as X-ray crystallography, can take months or years.

In 2021 a Google DeepMind team, with John Jumper as first author, published AlphaFold in Nature. The paper calls it "the first computational method that can regularly predict protein structures with atomic accuracy", including for proteins with no similar known structure. It was tested in CASP14, a blind community assessment in which predictions are compared with experimental structures that had not yet been released.

Then the predictions were released at scale. The AlphaFold Protein Structure Database, run by Google DeepMind and EMBL's European Bioinformatics Institute (EMBL-EBI), says it "provides open access to over 260 million protein structure predictions" (as stated on the AlphaFold DB home page, 26 October 2026) under a CC-BY-4.0 licence, which allows academic and commercial use.

Why did AlphaFold's creators win a Nobel Prize?

The Royal Swedish Academy of Sciences announced on 9 October 2024 that the Nobel Prize in Chemistry 2024 went half to David Baker "for computational protein design" and half jointly to Demis Hassabis and John Jumper "for protein structure prediction".

The pairing tells you something. Prediction asks what shape a given sequence will take. Design asks the reverse: which sequence will give you the shape and function you want. Put together, they make proteins look more like an engineering material and less like something you can only study.

Can AI discover new materials?

A 2023 Nature paper from Google DeepMind, Scaling deep learning for materials discovery (Merchant and colleagues), describes a system called GNoME. Its abstract reports three figures:

Figure in the abstract What it refers to
48,000 stable crystals already identified by earlier studies, used as the starting point
2.2 million new structures found "below the current convex hull"
736 of the stable structures, those "already independently experimentally realized"

"Below the convex hull" is a computational test. It means calculations predict the crystal is more stable than the known competing arrangements of the same elements. It does not mean anyone has made it, or that it would be useful in a battery or a chip.

That is why the third number deserves the most attention. Hundreds of structures made in labs by independent groups is real evidence that the predictions point at something physical. It is also a small share of 2.2 million. Most of the remaining work sits in the gap between "predicted stable" and "made, measured and useful".

Can AI find better algorithms than people?

Two results from Google DeepMind say yes, in narrow cases where the answer can be checked.

AlphaDev (2023). The Nature paper Faster sorting algorithms discovered using deep reinforcement learning (Mankowitz and colleagues) treated writing a short sorting routine as a single-player game. The abstract says AlphaDev "discovered small sorting algorithms from scratch that outperformed previously known human benchmarks", and that they were integrated into the LLVM standard C++ sort library. Code accepted into a widely used library is stronger evidence than a benchmark table alone.

AlphaEvolve (2025). A Google DeepMind post of 14 May 2025 describes a Gemini-powered coding agent for designing algorithms. The company says it:

  • found a way to multiply 4×4 complex-valued matrices with 48 scalar multiplications, improving on Strassen's 1969 algorithm for that case;
  • was applied to more than 50 open mathematical problems, matching the best known solutions in roughly 75% of them and improving on them in about 20%;
  • produced a data-centre scheduling heuristic that recovers, on average, 0.7% of Google's worldwide compute resources.

These figures come from the company's announcement and its AlphaEvolve white paper. As of 26 October 2026 I found no peer-reviewed journal version. Read them as "Google DeepMind says", which is a step below a peer-reviewed abstract.

What do these results have in common?

Every one of them worked where an answer can be checked quickly and cheaply:

  • A predicted protein structure can be scored against experimental ones.
  • A crystal's stability can be estimated with established physics calculations.
  • A sorting routine is either correct and faster, or it isn't.
  • A matrix method either reaches the right answer in 48 multiplications, or it doesn't.

My reading: AI does science best today where a fast, trusted judge exists. Where the judge is slow (a clinical trial), expensive (a new manufacturing process) or disputed, generating ideas is the easy half and testing them is the bottleneck.

What has AI in scientific research not shown yet?

  • That a prediction is a discovery. AlphaFold predicts a static structure. On its own, that doesn't tell you how a protein moves, how it binds a drug inside a living cell, or how it behaves in disease. Each database entry carries a confidence score the database labels pLDDT, and low confidence is a warning sign.
  • That candidate materials become products. The GNoME abstract reports 736 independent lab realisations. It does not report how many have useful properties at industrial scale.
  • That a medicine came from these systems. I found no primary source, among the pages I read, linking an approved medicine to the specific systems in this post. I'm not saying none exists; I didn't verify one.
  • That AI replaces the lab. Every result above needed experiments, standard benchmarks or real deployment before anyone believed it.

Which AI science results should you watch next?

  1. Independent replication. Count results confirmed by groups that did not build the model, as GNoME's 736 figure does.
  2. Closed-loop labs. The next step is AI that proposes, a robotic or human lab that tests, and the results fed back. Ask how many cycles ran and how many failed.
  3. Paper or press release. AlphaFold, GNoME and AlphaDev went through Nature. Newer claims often appear first as company posts. Wait for the paper, the code or both.
  4. Licences on outputs. AlphaFold DB uses CC-BY-4.0. Check the licence before you build on any model's output.
  5. Compute cost. Training these systems needs large data-centre capacity. Our post on data-centre electricity covers the official demand figures.

Going deeper: for security and IT teams

This section is for technical readers and decision-makers: research IT, R&D leaders, and the security teams who support scientists using these tools.

Treat model outputs as sourced inputs. A structure downloaded from AlphaFold DB, a candidate crystal from a published dataset, or code from an AI search each has a source, a version and a licence. Record all three. AlphaFold DB states its CC-BY-4.0 licence on its home page, and attribution is a condition of that licence, so outputs reused in publications or products need credit. Store the database release or model version next to each downloaded file, so a result can be reproduced or withdrawn if a later release changes it.

Keep the confidence data. AlphaFold DB shows a mean pLDDT confidence score for each entry. Pipelines that ingest structures should carry the confidence fields the database provides, and flag or filter low-confidence regions before they feed docking, simulation or a decision. A pipeline that drops the confidence column turns a hedged prediction into an unhedged one, and nobody downstream will notice.

Verification is the control. The common thread in this post is an automatic judge. For AI-generated code, that judge is your test suite, benchmarks and review before merge. Treat an AI-suggested algorithm like any third-party contribution: correctness tests, performance tests on your own hardware and data, and human review. A speed-up measured on someone else's machine is a hypothesis about yours.

Protect unpublished research data. Hosted AI science tools are cloud services. Unpublished sequences, compound structures and experimental results can be an institution's most valuable intellectual property. Before researchers paste them into a hosted tool, check the provider's terms on retention and model training, the region where data is processed, and whether your institution has a contract in place. Our AI at work privacy checklist lists the questions to ask.

Ask about dual use. Protein design is valuable because it creates new functions, and new functions can cut both ways. If your organisation runs computational design work, ask your biosafety or research-integrity office whether it falls under their review.

Questions for a vendor or research partner: - Which model version produced this output, and is it pinned? - What independent validation exists, and who carried it out? - What licence covers the outputs, and does it allow commercial use? - Is our input data retained, or used to train future models? - What confidence or uncertainty measure comes with each result?

How to weigh the figures here. The AlphaFold, GNoME and AlphaDev figures come from peer-reviewed Nature abstracts. The AlphaEvolve figures are the company's own. The AlphaFold DB count comes from the database's home page and has changed between releases.

A checklist for leaders weighing AI for science

  1. Name the judge. Ask how results get checked. Without a fast, trusted check, expect slow progress.
  2. Ask for replication numbers, not just counts of candidates.
  3. Separate peer-reviewed results from announcements in board papers and business cases.
  4. Record version, licence and confidence for every model output you rely on.
  5. Check data terms before unpublished research goes into a hosted tool.

Learn more

Key takeaways

  • AlphaFold predicts protein structures with what its Nature paper calls atomic accuracy. Hassabis and Jumper shared the 2024 Chemistry Nobel with David Baker.
  • AlphaFold DB says it offers over 260 million predictions under CC-BY-4.0.
  • GNoME reported 2.2 million structures predicted stable, of which 736 had been made independently in labs.
  • AlphaDev's sorting routines were integrated into LLVM's standard C++ sort library.
  • AlphaEvolve's figures, such as 48 multiplications for 4×4 complex matrices, are Google DeepMind's own claims.
  • AI does best where results can be checked quickly. Lab confirmation is still the bottleneck.

This post describes research results. It is not investment advice.

Sources