AI scientific discoveries: the landmark results and what held up

What counts as an AI-enabled scientific discovery

Whether it is called AI-enabled discovery, AI for science or AI for scientific discovery, AI plays one of three roles in the results below. It predicts a property that would otherwise need an experiment or an expensive simulation: a protein's shape, a crystal's stability, next week's weather. It searches a space far too large to test, ranking millions of molecules or programs so that people test a handful. Or it proposes something new: a molecule, a proof, a hypothesis. In every case the model is a tool that scientists built, trained and checked.

Two of the 2024 Nobel Prizes involved AI, and both went to people. The Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for foundational work that enables machine learning with artificial neural networks. The Nobel Prize in Chemistry, announced on 9 October 2024, went half to David Baker "for computational protein design" and half jointly to Demis Hassabis and John Jumper of Google DeepMind "for protein structure prediction".

Protein structure: AlphaFold and the 2024 Nobel Prize

A protein is a chain of amino acids that folds into a three-dimensional shape, and the shape decides what it does. Predicting that shape from the sequence alone had been an open problem for more than 50 years. Experiments such as X-ray crystallography, NMR and cryo-electron microscopy can determine it, but slowly: when the AlphaFold 2 paper appeared in Nature on 15 July 2021, experiments had determined the structures of around 100,000 unique proteins, a small fraction of the billions of known sequences, and one structure could take months to years of work.

AlphaFold 2 was tested blind. At CASP14 (May to July 2020), the assessment that scores predictions against recently solved structures not yet made public, its median backbone error was 0.96 Å, against 2.8 Å for the next best method; a carbon atom is about 1.4 Å wide. The network learned from structures in the Protein Data Bank (PDB) released up to 30 April 2018, and its accuracy was confirmed again on PDB structures deposited after that date. Every prediction also carries a per-residue confidence score, pLDDT, which tells a researcher which parts to trust.

The scale followed. According to the Nobel committee, AlphaFold 2 has been used to predict the structure of virtually all of the 200 million proteins researchers have identified, by more than two million people in 190 countries. A prediction takes GPU minutes to GPU hours, depending on the protein's length. It works alongside experiments rather than instead of them: the paper notes that its predictions were already helping experimentalists interpret cryo-electron microscopy maps and solve crystal structures by molecular replacement.

AlphaFold 3, published in Nature on 8 May 2024, extends prediction from single proteins to complexes of proteins, DNA and RNA, small molecules, ions and modified residues, using a diffusion-based architecture. Its authors state its limits plainly. It predicts static structures like those in the PDB, not how molecules move in solution. Like other generative models, it can hallucinate plausible-looking structure in disordered regions. On one benchmark, 4.4% of its outputs still violated chirality. The paper was published without code, only pseudocode and a non-commercial web server; Google DeepMind has since released the inference code on GitHub, with the model parameters available only from Google under separate terms of use.

Antibiotics: halicin, abaucin and generative design

Antibiotic discovery is a search problem. Chemical libraries hold millions of molecules, few of them kill bacteria, and resistance keeps eroding the drugs we have. Virtual screening with a trained model is the AI answer: score everything, test the top few.

In February 2020, the Collins Lab at MIT published a deep learning approach to antibiotic discovery in Cell. The team trained a neural network on 2,335 molecules measured for whether they stop E. coli from growing, then ran it over the Broad Institute's Drug Repurposing Hub. Its pick, a compound previously investigated as a diabetes treatment, was renamed halicin. It killed a wide range of pathogens in the lab, including Mycobacterium tuberculosis, and treated C. difficile and pan-resistant Acinetobacter baumannii infections in mice. The model then scored more than 107 million molecules from the ZINC15 database; of 23 top predictions tested, eight were antibacterial and structurally distant from known antibiotics. According to MIT's announcement, that screen took three days.

Three later studies refined the method:

• Abaucin (Nature Chemical Biology, 25 May 2023), with the halicin paper's first author, Jonathan Stokes, as senior author. The team measured about 7,500 molecules against A. baumannii, a hospital pathogen that is often multidrug-resistant, trained a network on that data and used it to find structurally new candidates. Abaucin acts on A. baumannii specifically (narrow-spectrum), disrupts lipoprotein trafficking through a mechanism involving the protein LolE, and controlled an infection in a mouse wound model.
• An explainable structural class (Nature, 20 December 2023). The Collins Lab measured antibiotic activity and human-cell toxicity for 39,312 compounds, predicted both for 12,076,365 compounds with graph neural networks, and extracted the chemical substructures behind the predictions. Of 283 compounds tested, one structural class was selective against MRSA and vancomycin-resistant enterococci and reduced infection in mouse models.
• Generative design (MIT News, 14 August 2025, on a paper in Cell). Instead of screening existing libraries, generative models designed more than 36 million possible compounds. For gonorrhea, the screen left about 1,000; the team picked 80 and found that only two could be synthesized. One of them, NG1, was effective against Neisseria gonorrhoeae in a dish and in a mouse model of drug-resistant infection. For MRSA, 22 of about 90 shortlisted designs were made, six were strongly active, and the best, DN1, cleared a MRSA skin infection in mice.

Materials: GNoME's 2.2 million crystals, and what was made

New inorganic crystals underpin batteries, chips and solar cells, and a crystal is only useful if it is stable enough to exist. Databases such as the Materials Project compute stability with density functional theory (DFT), a quantum-mechanical approximation that is accurate but expensive. By the GNoME authors' own recalculation, those efforts had identified about 48,000 computationally stable crystals.

GNoME, published by Google DeepMind in Nature on 29 November 2023, used graph neural networks to filter candidate structures, ran DFT on the survivors, and fed every result back into training: active learning, which the paper calls a data flywheel. The result was 2.2 million structures below the previous convex hull, the boundary of lowest energies that defines which compounds are stable. Because new compounds compete with each other, 381,000 new entries remain on the updated hull, for 421,000 stable crystals in total. The paper also reports that 736 of its stable structures had already been made independently by other labs.

"Discovered" here has a precise meaning: predicted stable by DFT, not made in a lab or shown to be useful. In April 2024, chemists Anthony Cheetham and Ram Seshadri examined a small random subset of the predictions in Chemistry of Materials and found scant evidence of compounds that are at once novel, credible and useful, though they called the methods promising.

A companion paper, published the same day, described the A-Lab at Lawrence Berkeley National Laboratory: a self-driving laboratory that plans solid-state syntheses with models trained on the literature, runs them with robots and reads the X-ray diffraction patterns with machine learning. Berkeley Lab announced 41 new compounds from 58 attempted targets in 17 days. In March 2024, chemists at Princeton and University College London argued in PRX Energy that errors in the analysis meant no new materials had been discovered in that work. In January 2026, the authors published a correction: a manual re-analysis confirmed 36 of the 40 reported successes and found four inconclusive; one compound was dropped because it had been in the training data; and "novel" was clarified to mean new to the prediction platform, not necessarily new to science. The paper now reports 36 compounds from 57 targets. The methods behind these results, from property-predicting networks to MatterGen's generative design, are covered in AI in materials science.

Weather forecasting: GraphCast and ECMWF's AIFS

Weather is the clearest case of AI as a surrogate model for a simulation. Traditional numerical weather prediction solves the physics of the atmosphere on supercomputers; according to Google DeepMind, a 10-day forecast from ECMWF's HRES can take hours on a supercomputer with hundreds of machines. GraphCast learned from the past instead: it was trained on four decades of ECMWF's ERA5 reanalysis and steps the global weather forward six hours at a time. Published in Science on 14 November 2023, it predicts hundreds of variables 10 days ahead at 0.25° resolution in under a minute, and it outperformed the most accurate operational deterministic systems on 90% of 1,380 verification targets.

The stronger signal came from a forecast agency. On 25 February 2025, ECMWF made its own AI model, AIFS, operational, running side by side with its physics-based Integrated Forecasting System (IFS). ECMWF reports gains of up to 20% for tropical cyclone tracks and about 1,000 times less energy per forecast. The AI does not start from nothing: AIFS uses the same initial conditions as IFS, built from a short-range forecast and around 60 million quality-controlled observations, and its grid spacing is 28 km against 9 km for IFS. The observing system and the physics-based analysis remain the foundation every AI forecast starts from. How far AI reaches into climate projections, and where it stops, is covered in our guide to AI for climate research.

Mathematics and algorithms: results a program can check

Mathematics has an advantage over lab science: a construction or a proof can be checked exactly, so a model can propose freely and a checker can reject its mistakes.

• FunSearch (Nature, 14 December 2023) pairs a large language model with an automated evaluator in an evolutionary loop: the model writes programs, the evaluator scores them, the best survive. It found new constructions of large cap sets, a problem in extremal combinatorics, including the largest improvement in 20 years to the asymptotic lower bound, and better heuristics for online bin packing.
• AlphaGeometry (Nature, 17 January 2024) solved 25 of 30 recent olympiad geometry problems, against 10 for the previous best method, after training on 100 million synthetic theorems and proofs.
• AlphaProof and AlphaGeometry 2 scored 28 of 42 points at the 2024 International Mathematical Olympiad, silver-medal standard. AlphaProof works by formal verification in the language Lean, so its proofs are machine-checked, but the problems were translated into Lean by hand and some took up to three days.
• Gemini Deep Think: an advanced version scored 35 of 42 at the 2025 olympiad, gold-medal standard, working in natural language within the 4.5-hour limit, and its solutions were graded and certified by IMO coordinators. Our Gemini explainer covers the model family.
• AlphaEvolve (Google DeepMind, 14 May 2025) found a way to multiply 4×4 complex-valued matrices with 48 scalar multiplications, improving on Strassen's 1969 algorithm in that setting. Applied to more than 50 open problems, it matched the best known solutions in about 75% of cases and improved on them in about 20%.

Other fields: fusion control and AI research assistants

AI has also reached physics experiments and hypothesis generation, again with a real-world check at the end.

In February 2022, a Nature paper from DeepMind and EPFL's Swiss Plasma Center described a deep reinforcement learning controller for the magnetic coils of the TCV research tokamak. The controller learned by interacting with a tokamak simulator, then ran directly on the hardware, holding a range of plasma shapes, including two separate plasmas in the vessel at once.

In February 2025, Google introduced an AI co-scientist, a multi-agent system built with Gemini 2.0 that generates hypotheses and research proposals, with experts in the loop. Drug repurposing candidates it proposed for acute myeloid leukemia were then shown to reduce the viability of cancer cell lines in lab experiments. Asked why a family of bacterial genetic elements turns up across many bacterial species, a question an Imperial College London group had already answered by experiment but not yet published, it independently proposed the same mechanism.

What makes AI useful in science

The same conditions recur in every result above. Where they are missing, AI rarely delivers.

Data comes first: AlphaFold stands on structures solved by experiment, GraphCast on four decades of reanalysis, the antibiotic models on compounds measured in the lab. A model learns only what someone recorded, consistently and with the right label.

Simulation makes the loop affordable: a trained model answers in minutes what took hours of supercomputing or months at the bench, so it can filter candidates in front of the expensive method, as GNoME did for DFT.

Search is where the payoff is: ranking millions of candidates and sending the best few to the lab turns a fixed budget of experiments into a better-aimed one.

The limits: lab validation, reproducibility and hype

A prediction is not a discovery until something real confirms it

Each step from model to reality loses candidates. A compound that is stable in a calculation is not necessarily one a lab can make, and one a lab can make is not necessarily useful: that is the core of the GNoME critique. Of 80 generated antibiotic designs checked with synthesis vendors, only two could be made. A-Lab's count fell from 41 compounds to 36 after chemists re-read the diffraction data. And a mouse is not a patient.

Reproducibility needs held-out data and published code

In an August 2023 survey in Patterns, Sayash Kapoor and Arvind Narayanan of Princeton found data leakage in 17 fields that had adopted machine learning, affecting 294 papers and in some cases producing wildly over-optimistic conclusions. In their own re-analysis of civil war prediction, complex models that had seemed far better than logistic regression were not, once the leakage was fixed. They propose "model info sheets", which researchers fill in to detect and prevent each type of leakage they catalogue, including temporal leakage, where test data comes from before the training data. The AlphaFold 2 evaluation on structures released after its training cutoff is the pattern to copy; a paper published without code, as AlphaFold 3 was, is harder for others to check.

Hype travels faster than peer review

A lab announcement is written to be read; the paper is written to be checked. GNoME's announcement presented 380,000 stable materials with the potential to power future technologies, while what the paper established was stability computed by DFT. Preprints have not been peer reviewed at all: in May 2025, MIT asked arXiv to withdraw a November 2024 economics preprint, "Artificial Intelligence, Scientific Discovery, and Product Innovation", after an internal review left it with no confidence in the data's provenance, reliability or validity. Peer review is not the end of checking either: A-Lab passed review at Nature and was corrected more than two years later, after outside chemists examined its data. That correction is the system working.

How to check an AI discovery claim

A short routine catches most overstatements:

1. Find the primary source. Note the journal or preprint server and the date, and read the abstract, not the headline.
2. Separate predicted from measured. Computed stable, synthesized, active in a dish, effective in mice and effective in people are five different claims.
3. Check that the test was blind. CASP14 used unreleased structures; AlphaFold 2 was re-tested on structures released after its training data. A random split of time-ordered data can leak.
4. Ask "better than what?" GraphCast was compared with HRES, AlphaGeometry with the previous best method. A result with no baseline is a demonstration, not evidence.
5. Look for independent replication and critique. GNoME's 736 structures made by other labs count for more than its 2.2 million predictions.
6. Check for corrections, code and data. The article page lists corrections; the methods section says whether code and data are available.

What AI in science means for organizations with data

The pattern that worked in these labs transfers to any organization with years of measurements: test results, quality records, sensor logs, transactions or support tickets. The lessons are the same, at a smaller scale.

1. Treat historical data as the asset. AlphaFold learned from structures experiments had already solved, and abaucin's model from about 7,500 measurements its team made. If your records are inconsistent, missing labels or scattered across systems, fix that before choosing a model.
2. Test the way the scientists did. Hold back the most recent data for testing, as AlphaFold's evaluation did, and compare the model with the method you use today.
3. Put a cheap check between the model and the decision. A simulation, a rule, a small experiment or a person should confirm what the model proposes.
4. Use AI to rank and narrow, not to decide. The halicin screen scored more than 107 million molecules and tested 23. The value came from the 23 tests.
5. Record what was used. Which data, which model version and which results, so an outcome can be reproduced or corrected later.

If your data lives in spreadsheets and disconnected systems, the first step is the platform under it: our data platform service builds ingestion and a governed warehouse or lakehouse, engineered with tests, lineage and access control. When a model step is ready for a real workflow, our AI and automation service runs an evaluation set from your real cases before every change, and nothing changes in your systems until a rule or a person approves it.