Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    AI for science in 2026: a crash course

    AI for science uses machine learning to predict, design and test things in nature: protein structures, weather, new materials and drug molecules.[1][2][3] By October 2026 it had won a Nobel Prize, entered daily weather forecasting at ECMWF and NOAA, and drawn billions in investment, while its biggest promises in drugs and materials remained unproven.[4][5][6][7][8]

    Editor reviewedUpdated AI for scienceArtificial intelligenceScienceLife sciences

    AI for science is the use of machine learning to do science faster: predicting the shape of molecules, forecasting weather, proposing new materials and drugs, and increasingly running the experiments too. It matters now because several of these tools have left the research paper and entered daily use, from AlphaFold’s database with millions of users to AI forecasts at Europe’s and America’s weather agencies.[9][5][6]

    Why it matters

    Many scientific problems are searches through enormous spaces: all the shapes a protein could take, all the crystals that could exist, all the ways tomorrow’s weather could unfold. Experiments and physics simulations explore those spaces slowly. AI models learn shortcuts from past data and can test millions of options in a computer first. Before AI, experiments had mapped about 100,000 protein structures; AI has now predicted structures for virtually all 200 million known proteins.[10][11]

    The common pattern is a learned surrogate for an expensive process: experimental structure determination, numerical weather prediction or density functional theory. Once trained, surrogates are orders of magnitude cheaper. GraphCast makes a 10-day forecast in under a minute on one TPU, versus hours on a supercomputer for HRES, and ECMWF reports a roughly 1,000-fold energy saving with its AIFS.[12][13] The catch is that surrogates inherit the coverage and biases of their training data, such as the protein-data-bank or ERA5 reanalysis.[14][15]

    The map of the field

    1. Biomolecular structure. Predicting 3D shapes from sequence. alphafold made this routine, and AlphaFold 3 extended it to complexes with DNA, RNA and drugs.[1][16] Start with How AI predicts protein structures.
    2. Protein design. Generating new proteins. rfdiffusion turned a structure predictor into a generator; its third version designs at atomic level.[17][18]
    3. Weather and climate. ML models trained on decades of reanalysis now rival physics models: see weathernext and ecmwf-aifs.[19][5]
    4. Materials and drugs. Screening and generating candidate crystals and molecules, from gnome-materials to companies such as isomorphic-labs.[3][20]
    5. Autonomous labs. Robots that test AI proposals and close the loop: self-driving-labs.[21]

    Key ideas

    • Training data is everything. Protein AI learned from decades of shared lab results; weather AI from 40 years of reconstructed weather.[14][15]
    • Prediction is not proof. A predicted structure or material is a hypothesis until a lab confirms it.[22]
    • Blind tests keep the field honest. The CASP competition tests predictions against structures nobody has seen yet.[1]
    • Diffusion and generative models are central to both prediction (AlphaFold 3) and design (RFdiffusion), and probabilistic ensembles in weather (GenCast, WeatherNext 2).[16][17][23]
    • Hit rate is the design metric that matters: Anthropic reported 22–35% for Claude-led binder design against a typical 10–15%, with large variation by target.[24][25]
    • Evaluation independence is fragile: CASP needed emergency funding from Google DeepMind in 2025.[26]

    Who the main players are

    Google DeepMind built AlphaFold, GNoME and the WeatherNext models.[27][3][28] The University of Washington’s Institute for Protein Design develops the open RFdiffusion design models.[18] Public weather centres ECMWF and NOAA run AI models operationally.[5][6] Isomorphic Labs, led by Nobel laureate Demis Hassabis, raised $2.1 billion in 2026; start-ups Lila Sciences, which had raised $550 million by October 2025, and Periodic Labs are building autonomous labs; and the US government launched the genesis-mission in November 2025.[29][7][30][31][32] Academic and start-up teams publish open alternatives to AlphaFold 3, such as Boltz and OpenFold3.[33][34] Frontier AI labs are also entering: Anthropic reported in August 2026 that Claude ran a protein design campaign with externally tested results.[35]

    Where the frontier is (October 2026)

    • Complexes and dynamics. CASP17, with results due 30 November 2026, focuses on immune complexes, ligand binding and conformational ensembles, where deep learning still struggles.[36][37]
    • Agents running science. AI systems now chain design tools and outsource lab tests; government programmes aim to automate experiment design.[35][38]
    • Clinical proof. AI-derived drugs pass phase I unusually often but look average in phase II so far; an FDA decision on the computationally designed zasocitinib is due in early 2027, and rentosertib began a 320-patient phase III trial in July 2026.[39][40][41][42]
    • Antibodies and binders. Generative models now design antibodies from scratch; Chai Discovery reported a 16% hit rate in a 2025 preprint.[43]
    • Real materials. Chemists dispute whether AI has yet found genuinely new, useful materials, and the A-Lab’s authors corrected their novelty claims in January 2026.[44][8][45]

    How to use this course

    Read the three fundamentals first, then the tool and institution pages, then the two debates.

    The tracker lists dated milestones and what to watch next.

    Questions readers ask

    What is the biggest success of AI for science so far?

    Protein structure prediction. AlphaFold2 reached accuracy competitive with experiment in the CASP14 blind test, and its developers shared the 2024 Nobel Prize in Chemistry.[1][4]

    Are AI weather forecasts used in practice?

    Yes. ECMWF has run its AIFS operationally since February 2025, and NOAA deployed AI-driven global models in December 2025.[5][6]

    Has AI discovered a new drug?

    AI-derived molecules are in trials, with 80–90% phase I success but about 40% in phase II, similar to industry averages. The nearest computationally designed candidate awaits an FDA decision in early 2027, and the AI-generated rentosertib entered phase III in July 2026.[39][40][41][42]

    What is a self-driving lab?

    A lab where AI plans experiments, robots carry them out, and the results feed back into the next round.[21]

    Who are the main players?

    Google DeepMind (AlphaFold, GNoME, WeatherNext), the University of Washington's Institute for Protein Design (RFdiffusion), ECMWF and NOAA in weather, companies such as Isomorphic Labs, and the US Department of Energy's Genesis Mission.[27][18][5][46][38]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      AlphaFold2, published in Nature in July 2021, was validated in the CASP14 blind assessment and predicted structures with accuracy competitive with experiment in a majority of cases. confirmedas of 2021-07-15

    2. [2]

      Google DeepMind's GraphCast, described in Science in November 2023, gave more accurate predictions than ECMWF's HRES forecast on more than 90% of 1,380 test variables and lead times. confirmedas of 2023-11-14

    3. [3]

      Google DeepMind's GNoME work, published in Nature in November 2023, used graph networks to identify 2.2 million crystal structures, of which it counted 381,000 as newly discovered stable materials. confirmedas of 2023-11-29

    4. [4]

      The 2024 Nobel Prize in Chemistry went half to David Baker for computational protein design and half jointly to Demis Hassabis and John Jumper for protein structure prediction. confirmedas of 2024-10-09

    5. [5]

      ECMWF took its Artificial Intelligence Forecasting System (AIFS) into operations on 25 February 2025, running side by side with its physics-based IFS. confirmedas of 2025-02-25

    6. [6]

      On 17 December 2025 NOAA deployed operational AI-driven global models, AIGFS, AIGEFS and the hybrid HGEFS, which NOAA says deliver more accurate guidance using a fraction of the computing resources. confirmedas of 2025-12-17

    7. [7]

      On 12 May 2026 Isomorphic Labs announced a $2.1 billion funding round led by Thrive Capital, with Alphabet and GV among existing investors and MGX, Temasek, CapitalG and the UK Sovereign AI Fund as new investors. confirmedas of 2026-05-12

    8. [8]

      MIT Technology Review reported in December 2025 that AI materials discovery had not yet produced a convincing big win or a new miracle material. reportedas of 2025-12-15

    9. [9]

      As of March 2026 the AlphaFold Database had more than 3.4 million users in 190 countries. confirmedas of 2026-03-16

    10. [10]

      Before AlphaFold2, experiments had determined structures for around 100,000 unique proteins, a small fraction of the billions of known protein sequences, because each structure takes months to years of work. confirmedas of 2021-07-15

    11. [11]

      According to the Nobel committee, AlphaFold2 has been used to predict the structure of virtually all 200 million known proteins and has been used by more than two million people in 190 countries. confirmedas of 2024-10-09

    12. [12]

      GraphCast makes a 10-day forecast in less than a minute on a single Google TPU v4 machine, while a conventional HRES forecast can take hours on a supercomputer with hundreds of machines. confirmedas of 2023-11-14

    13. [13]

      ECMWF says the AIFS outperforms physics-based models on many measures, including tropical cyclone tracks with gains of up to 20%, and uses about 1,000 times less energy per forecast. confirmedas of 2025-02-25

    14. [14]

      AlphaFold2 was trained on a copy of the Protein Data Bank downloaded in August 2019. confirmedas of 2021-07-15

    15. [15]

      GraphCast is a graph neural network trained on four decades of ECMWF's ERA5 reanalysis data, which is itself built from observations using traditional numerical weather prediction. confirmedas of 2023-11-14

    16. [16]

      AlphaFold 3, published in Nature in May 2024, uses a diffusion-based architecture to predict joint structures of complexes containing proteins, nucleic acids, small molecules, ions and modified residues. confirmedas of 2024-05-08

    17. [17]

      RFdiffusion, published in Nature in July 2023, was made by fine-tuning the RoseTTAFold structure prediction network on structure denoising, producing a generative model of protein backbones. confirmedas of 2023-07-11

    18. [18]

      In December 2025 the Institute for Protein Design released RFdiffusion3, an all-atom model that designs proteins interacting with any type of molecule commonly found in cells, with training code and weights available through the Rosetta Commons Foundry. confirmedas of 2025-12-03

    19. [19]

      GenCast, a probabilistic machine-learning weather model published in Nature in December 2024, had greater skill than ECMWF's ENS ensemble on 97.2% of 1,320 evaluated targets. confirmedas of 2024-12-04

    20. [20]

      Isomorphic Labs says its funding will develop its AI drug design engine, IsoDDE, which it describes as working across multiple therapeutic areas and drug modalities. confirmedas of 2026-05-12

    21. [21]

      The A-Lab combines computations, literature data, machine learning and active learning to plan and interpret experiments carried out by robots. confirmedas of 2023-11-29

    22. [22]

      The GNoME paper reported that 736 of its predicted stable structures had already been independently realized experimentally. confirmedas of 2023-11-29

    23. [23]

      Google DeepMind's WeatherNext 2, announced in November 2025, uses a Functional Generative Network and surpasses the previous WeatherNext model on 99.9% of variables and lead times from 0 to 15 days. confirmedas of 2025-11-17

    24. [24]

      Anthropic reported hit rates of 22% to 35% for Claude-led binder design depending on setup, against the 10–15% it describes as typical for protein design campaigns, with 354 binders from 1,320 designs. reportedas of 2026-08-18

    25. [25]

      In Anthropic's campaign, hit rates for individual targets ranged from 90% to 0%. reportedas of 2026-08-18

    26. [26]

      In July 2025 Google DeepMind provided temporary funding to CASP after its NIH grant ran out. confirmedas of 2025-07-21

    27. [27]

      AlphaFold is an AI system developed by Google DeepMind that has predicted the structure of millions of proteins since 2021. confirmedas of 2026-03-16

    28. [28]

      GenCast was announced by Google DeepMind in December 2024 as a model delivering forecasts up to 15 days ahead. confirmedas of 2024-12-04

    29. [29]

      Demis Hassabis is chief executive of Isomorphic Labs and Max Jaderberg is its president, according to the company's May 2026 funding announcement. confirmedas of 2026-05-12

    30. [30]

      In October 2025 Lila Sciences closed a $350 million Series A, bringing its total funding to $550 million, with NVIDIA's venture arm NVentures among the investors. confirmedas of 2025-10-10

    31. [31]

      Periodic Labs emerged from stealth in September 2025 with a $300 million seed round, with a first goal of inventing better superconductors. confirmedas of 2025-09-30

    32. [32]

      On 24 November 2025 the US president signed an executive order launching the Genesis Mission, a national initiative to accelerate scientific breakthroughs using AI. confirmedas of 2025-11-24

    33. [33]

      The Boltz team describes Boltz-1 as the first fully open-source model to approach AlphaFold 3 accuracy, and releases Boltz code and weights under the MIT licence for academic and commercial use. confirmedas of 2026-10-10

    34. [34]

      OpenFold3-preview, from Columbia University's AlQuraishi Lab and the OpenFold consortium, aims to be a bitwise reproduction of AlphaFold 3 and is available under the Apache 2.0 licence for academic and commercial use. confirmedas of 2026-10-10

    35. [35]

      In August 2026 Anthropic reported that Claude models, using open-source protein design tools, designed binders against 14 of 15 targets, with designs produced and tested by Adaptyv Bio and Twist Bioscience. reportedas of 2026-08-18

    36. [36]

      The CASP17 prediction season ended on 11 September 2026, with assessment results due on 30 November 2026, a day before the CASP17 conference in Rome. confirmedas of 2026-10-10

      • CASP17 home page · Protein Structure Prediction Center (UC Davis) (retrieved 2026-10-10)
    37. [37]

      CASP17 emphasises areas where deep learning has yet to deliver, including immune complexes, protein–ligand complexes, nucleic acids and conformational ensembles. confirmedas of 2026-10-10

      • CASP17 home page · Protein Structure Prediction Center (UC Davis) (retrieved 2026-10-10)
    38. [38]

      The Genesis Mission tasks the Department of Energy with integrating its supercomputers and datasets into a closed-loop AI platform intended to automate experiment design, accelerate simulations and generate predictive models. confirmedas of 2025-11-24

    39. [39]

      A 2024 analysis of AI-native biotech pipelines in Drug Discovery Today found an 80–90% phase I success rate for AI-discovered molecules, higher than historic industry averages. confirmedas of 2024-06-01

    40. [40]

      The same 2024 analysis found a phase II success rate of about 40% for AI-discovered molecules, on a limited sample, comparable to historic industry averages. confirmedas of 2024-06-01

    41. [41]

      On 14 September 2026 Takeda said the FDA had accepted its new drug application for zasocitinib (TAK-279), an oral TYK2 inhibitor for moderate-to-severe plaque psoriasis, under priority review with a target action date in the first quarter of 2027. confirmedas of 2026-09-14

    42. [42]

      On 7 July 2026 Insilico Medicine announced the start of a phase III trial of rentosertib, a randomised, double-blind, placebo-controlled study expected to enrol 320 patients with idiopathic pulmonary fibrosis in China over 52 weeks. confirmedas of 2026-07-07

    43. [43]

      In a preprint posted in July 2025, Chai Discovery reported that its Chai-2 model achieved a 16% hit rate in fully de novo antibody design and found at least one binder for 50% of 52 targets, testing 20 or fewer designs per target, none of which had an existing antibody binder in the Protein Data Bank. confirmedas of 2025-07-06

    44. [44]

      A March 2024 PRX Energy perspective concluded that no new materials had been discovered in the A-Lab work and that about two-thirds of claimed successes were likely known disordered versions of predicted compounds. disputedas of 2024-03-07

    45. [45]

      A January 2026 author correction to the A-Lab paper said its novelty claims meant materials new to the prediction platform, not necessarily new to science; a manual re-analysis confirmed 36 of the 40 reported successes, with four inconclusive, and one compound was removed because it had been in the training data. confirmedas of 2026-01-19

    46. [46]

      Isomorphic Labs was founded in 2021, is headquartered in London and has offices in Cambridge, Massachusetts and Lausanne, Switzerland. confirmedas of 2026-05-12

    Revision history (2)
    1. Page created.
    2. Refresh: added open AlphaFold 3 alternatives, Chai-2 antibody design, rentosertib's phase III start, the A-Lab correction and Lila Sciences' funding.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "AI for science in 2026: a crash course." ContentLora, updated Oct 10, 2026. https://contentlora.com/explain/ai-for-science

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    ☰ Contents