AI for Science

AI for science is the use of artificial intelligence, particularly deep learning and large language models, to accelerate scientific discovery across biology, chemistry, physics, materials science, mathematics, and earth sciences. The field reached a watershed in 2024, when the Nobel Prize in Chemistry recognized AI-based protein structure prediction and the Nobel Prize in Physics recognized foundational work on artificial neural networks ^[1][2]. Landmark systems include AlphaFold, which has been used to predict the structures of virtually all of the roughly 200 million proteins that researchers have identified; GNoME, which predicted 2.2 million new stable crystal structures; GraphCast and GenCast for weather forecasting; AlphaProof for mathematical reasoning; and Google's AI co-scientist for hypothesis generation ^{[1][7][6][8][15]}. Since 2024 a newer class of agentic "AI scientist" systems, from Google, FutureHouse, Sakana AI and Anthropic, has chained language-model agents through literature review, hypothesis generation and data analysis, with human scientists usually running the laboratory experiments ^[32][36][45].

AI for science encompasses a broad range of applications: predicting protein structures, discovering new materials, forecasting weather, generating mathematical proofs, designing drugs, optimizing fusion energy experiments, and analyzing genomic data. What unites these applications is the use of AI not merely as a tool for automation but as a method for generating new scientific knowledge, identifying patterns in data that humans cannot perceive, proposing hypotheses, and even making discoveries that advance fundamental understanding of the natural world.

Nobel Prizes 2024

The 2024 Nobel Prizes marked a watershed moment for AI in science, with both the Chemistry and Physics prizes awarded to researchers whose work centered on artificial intelligence.

Nobel Prize in Chemistry 2024

The 2024 Nobel Prize in Chemistry was awarded in two halves. One half went to David Baker at the University of Washington "for computational protein design." The other half was awarded jointly to Demis Hassabis and John Jumper at Google DeepMind "for protein structure prediction" ^[1].

"One of the discoveries being recognised this year concerns the construction of spectacular proteins. The other is about fulfilling a 50-year-old dream: predicting protein structures from their amino acid sequences. Both of these discoveries open up vast possibilities," said Heiner Linke, Chair of the Nobel Committee for Chemistry ^[1].

Hassabis and Jumper developed AlphaFold, an AI system that addressed the problem of predicting a protein's three-dimensional structure from its amino acid sequence, which researchers had been attempting since the 1970s. AlphaFold2, presented in 2020, has been used to predict the structure of virtually all the 200 million proteins that researchers have identified, and according to the Royal Swedish Academy of Sciences it has been used by more than two million people from 190 countries ^[1].

Baker's work on computational protein design took the complementary approach: rather than predicting the structure of existing proteins, Baker's group builds entirely new kinds of proteins. In 2003 his group designed a new protein unlike any other known protein, and it has since produced proteins that can be used as pharmaceuticals, vaccines, nanomaterials and tiny sensors ^[1].

Nobel Prize in Physics 2024

The 2024 Nobel Prize in Physics was awarded jointly to John J. Hopfield and Geoffrey Hinton "for foundational discoveries and inventions that enable machine learning with artificial neural networks" ^[2].

Hopfield described an associative memory in 1982, now known as the Hopfield network, which can store patterns and reconstruct them from distorted or incomplete inputs; it is described in a manner equivalent to the energy of a spin system in physics ^[2][19]. Hinton, working with Terrence Sejnowski, used the Hopfield network as the foundation for a new network that draws on statistical physics, the Boltzmann machine, published in 1985 ^[2][19]. According to the Royal Swedish Academy of Sciences, Hinton built on this work, "helping initiate the current explosive development of machine learning" ^[2].

The Physics prize was announced on 8 October 2024 and the Chemistry prize on 9 October 2024, so both awards went to AI-related work within two days ^[1][2].

Key achievements

AI has produced significant results across multiple scientific domains. The following table summarizes major achievements.

Sources for the table rows are given in the sections below and in ^[1], ^[7], ^[8], ^[9], ^[10], ^[15], ^[26], ^[27], ^[46].

How accurate is AlphaFold? Protein structure prediction

The prediction of protein structures from amino acid sequences had been one of biology's grand challenges since the 1970s. Experimental methods like X-ray crystallography and cryo-electron microscopy could determine structures but were slow and expensive, sometimes requiring years per protein ^[16].

In the results of the CASP14 assessment, released on 30 November 2020, AlphaFold2 achieved a median Global Distance Test (GDT) score of 92.4 out of 100 across all targets, which DeepMind said corresponds to an average error (RMSD) of approximately 1.6 angstroms, comparable to the width of an atom ^[16]. It produced the best prediction for 88 of the 97 CASP14 targets ^[14]. The system, described in Nature in July 2021, combines attention-based neural network modules with an iterative "recycling" process that repeatedly refines its own predictions ^[18].

In July 2022, DeepMind and EMBL's European Bioinformatics Institute (EMBL-EBI) expanded the AlphaFold Protein Structure Database from nearly 1 million to over 200 million predicted structures, covering nearly all catalogued proteins known to science. The database is free to all ^[17].

AlphaFold3, announced on 8 May 2024 and published in Nature, extended the system's capabilities beyond individual proteins to predict the structures of complexes of proteins with DNA, RNA, ligands, ions and chemical modifications. It builds on an improved version of AlphaFold 2's Evoformer module and assembles its predictions with a diffusion network that starts from a cloud of atoms and converges on a final structure. DeepMind reported at least a 50% improvement in accuracy for protein interactions with other molecule types compared with existing prediction methods, and made most of its capabilities available free for non-commercial research through AlphaFold Server ^[3].

RoseTTAFold, developed by Minkyung Baek and David Baker's lab at the University of Washington and published in Science in 2021, provided an alternative approach using a "three-track" neural network that simultaneously considers protein sequence patterns, how amino acids interact with one another, and possible three-dimensional structures. RoseTTAFold can compute a protein structure in as little as ten minutes on a single gaming computer, and its code was released on GitHub alongside a public web server ^[4].

How accurate is AI weather forecasting? GraphCast and GenCast

Traditional numerical weather prediction relies on solving complex differential equations describing atmospheric physics, a process that requires enormous computational resources and hours of supercomputer time. AI-based approaches have demonstrated that weather forecasting can be dramatically faster and, in many cases, more accurate.

GraphCast, published in Science in November 2023, uses graph neural networks to make medium-range weather forecasts (up to 10 days) and was more accurate than the European Centre for Medium-Range Weather Forecasts' (ECMWF) HRES system, the industry gold standard, on more than 90% of 1,380 test variables and lead times. A 10-day GraphCast forecast takes less than a minute on a single Google TPU v4 machine, compared with hours of computation on a supercomputer with hundreds of machines for HRES. DeepMind open-sourced the model code ^[5].

GenCast, published in Nature in December 2024, advanced the approach further by using a diffusion model adapted to the spherical geometry of Earth. GenCast generates probabilistic forecasts, ensembles of 50 or more possible future weather trajectories, rather than a single deterministic prediction. Tested across 1,320 combinations of variables and lead times, it outperformed the ECMWF's ENS ensemble system on 97.2% of targets, and on 99.8% of targets at lead times greater than 36 hours. A single Google Cloud TPU v5 produces one 15-day GenCast forecast in about 8 minutes. DeepMind released the model's code and weights ^[6].

In November 2025, Google DeepMind and Google Research introduced WeatherNext 2, which Google says generates forecasts 8x faster, with resolution up to one hour, and can produce hundreds of possible weather scenarios from a single starting point, each in under a minute on a single TPU ^[20].

How does GNoME accelerate materials discovery?

In November 2023, Google DeepMind announced GNoME (Graph Networks for Materials Exploration), a deep learning tool that predicted 2.2 million new crystals that are stable by current scientific standards, a figure the team described as equivalent to nearly 800 years' worth of knowledge. Of these, about 380,000 were identified as the most stable and therefore the most promising candidates for experimental synthesis, and DeepMind said it would contribute them to the Materials Project database ^[7].

GNoME uses graph neural networks to predict the stability of hypothetical crystal structures. DeepMind highlighted 52,000 new layered compounds similar to graphene and 528 potential lithium-ion conductors among the predictions, pointing to possible uses in superconductors and next-generation batteries. By the time of the announcement, a literature search had found that external researchers had independently synthesized 736 of GNoME's predicted materials in concurrent work ^[7]. In a companion Nature paper, researchers at Lawrence Berkeley National Laboratory showed that an autonomous robotic laboratory, the A-Lab, could use Materials Project data and GNoME stability insights to synthesize more than 41 new materials ^[7].

How does AI prove math theorems? AlphaProof and FunSearch

AlphaProof, a system combining a pre-trained language model with the AlphaZero reinforcement learning algorithm, achieved silver-medal performance at the 2024 International Mathematical Olympiad (IMO). AlphaProof solved problems 1, 2 and 6 (two algebra problems and one number theory problem), including Problem 6, an algebra problem that was the hardest of the competition and was solved by only five human contestants, while AlphaGeometry 2 solved Problem 4 ^[8][12]. The problems were first manually translated into formal mathematical language; the systems solved one problem within minutes and took up to three days on the others ^[8]. AlphaProof generates formal proofs in the Lean proof language, allowing its solutions to be verified automatically. The methodology was published in Nature on 12 November 2025 ^[8][12].

Together, the two systems achieved a score of 28 out of 42 points, comparable to a silver medalist ^[8].

FunSearch, also from Google DeepMind and published in Nature in December 2023, took a different approach to mathematical discovery. Rather than proving existing conjectures, FunSearch pairs a pre-trained LLM with an automated evaluator to search the space of computer programs for novel mathematical constructions. It made new contributions to extremal combinatorics, finding a program that generates a cap set of size 512 in eight dimensions, larger than previously known ^[9][49]. DeepMind described the work as "the first time a new discovery has been made for challenging open problems in science or mathematics using LLMs" ^[9].

How does the AI co-scientist generate hypotheses?

The AI co-scientist is a multi-agent system, built on Gemini 2.0, that helps researchers generate, debate, and rank novel research hypotheses. Introduced by Google in February 2025 as a joint effort of Google Research, Google DeepMind and Google Cloud AI teams (arXiv preprint "Towards an AI co-scientist"), it uses self-play scientific debate, ranking tournaments and an "evolution" process, scaled with additional test-time compute ^[15]. The work was published in Nature on 19 May 2026 as "Accelerating scientific discovery with Co-Scientist" ^[30].

The system was validated across three biomedical areas, all with expert-in-the-loop guidance. For drug repurposing, experiments confirmed that drugs it suggested inhibited tumor viability at clinically relevant concentrations in multiple acute myeloid leukemia cell lines. For liver fibrosis, it identified epigenetic targets with anti-fibrotic activity in human hepatic organoids. In a test on antimicrobial resistance, researchers at Imperial College London asked it to explain a finding their group had made but not yet published; it independently proposed that capsid-forming phage-inducible chromosomal islands (cf-PICIs) interact with diverse phage tails to expand their host range, matching the unpublished result ^[15]. Nature later reported that the tool reached essentially the same conclusion in about two days ^[32].

On 19 May 2026, Google said it would make Co-Scientist available to individual researchers through Hypothesis Generation, an experimental tool in its Gemini for Science program on Google Labs, with access opening gradually. An enterprise version was in preview with organizations including Daiichi Sankyo, Bayer Crop Science and US national laboratories taking part in the Department of Energy's Genesis Mission ^[29][31].

Drug discovery

AI is being applied across the drug discovery pipeline, from target identification to molecule design to clinical trial optimization. See AI drug discovery for a fuller treatment.

A widely cited milestone is rentosertib (formerly ISM001-055), a TNIK inhibitor for idiopathic pulmonary fibrosis whose target and molecule were both found with generative AI by Insilico Medicine. The target was identified with the company's PandaOmics engine and the molecule designed with its Chemistry42 platform; the peer-reviewed paper describing the program says it took roughly 18 months from target discovery to preclinical candidate nomination ^[28]. Results of a randomized, placebo-controlled Phase 2a trial published in Nature Medicine on 3 June 2025 showed that the 60 mg once-daily group gained a mean of 98.4 mL in forced vital capacity over 12 weeks, versus a change of -20.3 mL in the placebo group ^[10]. Insilico describes rentosertib as a first for AI-discovered drugs; that framing is the company's own.

The overall impact of AI on drug discovery success rates remains debated. A 2024 analysis by Boston Consulting Group authors, published in Drug Discovery Today, found that AI-discovered molecules from AI-native biotech companies had an 80-90% success rate in Phase I trials, substantially higher than historic industry averages, but a Phase II success rate of about 40%, comparable to historic averages, on a limited sample ^[24].

On 8 December 2025, the US Food and Drug Administration qualified its first AI drug development tool, AIM-NASH, a cloud-based tool that helps pathologists score liver biopsies in clinical trials for metabolic dysfunction-associated steatohepatitis (MASH). Pathologists remain responsible for the final interpretation ^[25].

Fusion energy

Controlling the superheated plasma inside a tokamak fusion reactor is one of the most complex control problems in engineering. Google DeepMind, working with the Swiss Plasma Center at EPFL, showed in a February 2022 Nature paper that deep reinforcement learning could control the magnetic coils of the Variable Configuration Tokamak (TCV) in Lausanne, using a single neural network to control all 19 coils at once. The controllers, trained in simulation, held plasmas steady and sculpted them into a range of shapes, including a "droplet" configuration with two plasmas in the vessel at once, which had never been done in TCV before ^[22][11].

On 16 October 2025, DeepMind announced a research partnership with Commonwealth Fusion Systems (CFS) to apply AI to SPARC, a compact tokamak that CFS aims to make the first magnetic fusion machine to produce net fusion energy. The collaboration covers fast, differentiable plasma simulation with TORAX (an open-source simulator written in JAX that DeepMind released in 2024), searching for the most efficient and robust paths to maximizing fusion energy, and using reinforcement learning for real-time plasma control ^[11].

Separately, a Princeton-led team reported in Nature in February 2024 that an AI controller trained on past experimental data from the DIII-D National Fusion Facility in San Diego could forecast tearing-mode instabilities up to 300 milliseconds in advance and adjust operating parameters to avoid them, using a reinforcement learning algorithm ^[21].

AlphaGenome (2025-2026)

Google DeepMind introduced AlphaGenome on 25 June 2025, a model designed to predict the regulatory function of DNA sequences up to one million base pairs long, and made it available in preview for non-commercial research through an API. The accompanying paper was published in Nature on 28 January 2026, with the model source code released on GitHub under the Apache 2.0 licence and the trained weights made available on Kaggle and Hugging Face under non-commercial model terms ^[13][50][53]. Unlike AlphaMissense, which classifies missense variants at the protein level ^[27], AlphaGenome operates at the level of gene regulation, predicting where genes start and end in different cell types and tissues, where they are spliced, how much RNA is produced, and which DNA bases are accessible, close to one another, or bound by certain proteins, at the resolution of individual DNA letters ^[13].

AlphaGenome is available through a free non-commercial API with a Python software development kit ^[50]. In its announcement, DeepMind showed the model predicting that mutations seen in patients with T-cell acute lymphoblastic leukemia would activate the nearby TAL1 gene by introducing a MYB DNA binding motif, replicating the known disease mechanism ^[13].

How AI changes the scientific method

AI is not merely accelerating existing scientific workflows; it is changing how science is done at a fundamental level.

Hypothesis generation

Traditionally, scientific hypotheses originate from human intuition informed by domain knowledge and literature review. AI systems can now scan vast bodies of scientific literature, identify patterns across disparate fields, and propose hypotheses that human researchers might not consider. Purpose-built systems such as Google's AI co-scientist formalize this by having specialized agents generate competing hypotheses and rank them in an "idea tournament" before surfacing the most promising candidates to human researchers ^[29].

Experiment design

AI can optimize experimental designs by predicting which experiments are most likely to yield informative results. In materials science, for example, AI systems can prioritize which of millions of candidate materials should be synthesized first; the A-Lab at Lawrence Berkeley National Laboratory combined AI guidance with robotic synthesis to make new materials predicted to be stable ^[7]. Bayesian optimization and active learning techniques allow AI to design sequential experiments that maximize information gain.

Data analysis at scale

Modern scientific instruments generate data at rates that exceed human analytical capacity. Particle physics experiments at the Large Hadron Collider, genomic sequencing facilities, and astronomical surveys all produce petabytes of data that require automated analysis. AI systems can identify signals in noisy data, classify objects, and detect anomalies that would be invisible to human inspection.

Simulation and surrogate modeling

Physics-based simulations are essential to many scientific fields but are often computationally expensive. AI can be trained to approximate the outputs of these simulations (a technique called surrogate modeling), enabling scientists to explore parameter spaces orders of magnitude faster. DeepMind's TORAX plasma simulator, for example, is built in JAX so that it can integrate AI-powered models, and DeepMind says it lets CFS run millions of virtual experiments before SPARC is turned on ^[11].

Inverse design

Rather than analyzing what exists, AI enables scientists to specify desired properties and work backward to find or design systems that exhibit those properties. This "inverse design" paradigm is being applied in materials science (designing materials with target properties), drug discovery (designing molecules that bind specified targets), and protein engineering (designing proteins with desired functions).

Agentic AI scientists (2024-2026)

From 2024 onward, several labs built systems that string AI agents together to carry out more of the research loop: reading literature, proposing hypotheses, writing and running analysis code, and drafting reports or papers. In nearly all published cases, human scientists chose the problem and ran any wet-lab experiments. A Nature technology feature in September 2026 described such tools as increasingly common, and quoted the Berkeley genome-editing researcher Fyodor Urnov, an early user of FutureHouse's Kosmos, on his rule for working with them: "Trust, but verify" ^[32].

Sakana AI. Sakana's AI Scientist, introduced on 13 August 2024, was designed to take a code template and autonomously generate ideas, run experiments and write a full machine learning paper for about $15 per paper ^[33]. Sakana reported that a paper by the improved AI Scientist-v2 scored an average of 6.33 in blind review at the ICLR 2025 "I Can't Believe It's Not Better" workshop, above the average acceptance threshold, and was withdrawn before publication as agreed with the organizers. The work was published in Nature on 25 March 2026; Sakana acknowledges that the system is limited to computational experiments and sometimes produces naive ideas, hallucinations and inaccurate citations ^[34]. An independent evaluation of the first version by Joeran Beel and colleagues found that 42% of its experiments failed because of coding errors and that some generated papers contained hallucinated numerical results ^[35].

FutureHouse and Edison Scientific. FutureHouse, a nonprofit in San Francisco, announced Robin on 20 May 2025 as a workflow that orchestrates its literature agents Crow and Falcon and its data analysis agent Finch. Robin proposed testing compounds that enhance phagocytosis by retinal pigment epithelium cells, analyzed the experiments that human researchers ran, and identified the Rho-kinase inhibitor ripasudil as a candidate treatment for dry age-related macular degeneration ^[36]. The paper was published in Nature on 19 May 2026 ^[37]. On 5 November 2025, FutureHouse launched Edison Scientific as a commercial spinout to develop and deploy its AI Scientist for commercial applications ^[38]; Edison now markets Kosmos as "The AI Scientist for R&D" ^[52]. A preprint posted the same week described Kosmos, which runs for up to 12 hours, executing an average of 42,000 lines of code and reading 1,500 papers per run; independent scientists found 79.4% of the statements in its reports accurate, and collaborators said a single 20-cycle run was equivalent on average to six months of their own research time ^[39].

OpenAI. Kevin Weil announced OpenAI for Science on 2 September 2025, describing its goal as "the next great scientific instrument: an AI-powered platform that accelerates scientific discovery" ^[40]. The group launched Prism, a free AI-assisted workspace for scientific writing integrated with GPT-5.2, on 27 January 2026 ^[41]. In April 2026, TechCrunch reported that Weil was leaving OpenAI and that OpenAI for Science was being absorbed into other research teams ^[42].

Anthropic. Anthropic launched Claude for Life Sciences on 20 October 2025, adding connectors to scientific platforms such as Benchling and BioRender, Agent Skills and a life-sciences prompt library, with the stated goal of eventually allowing AI models to make discoveries autonomously ^[43]. On 30 June 2026 it released Claude Science in beta for Claude Pro, Max, Team and Enterprise users: a workbench in which a coordinating agent draws on more than 60 curated skills and connectors for fields such as genomics, proteomics, structural biology and cheminformatics, and a reviewer agent checks citations and calculations ^[44].

On 23 September 2026, Anthropic introduced a life sciences research group and laboratory in the Bay Area and reported the discovery of array-associated reverse transcriptases (ART) ^[45]. According to the accompanying technical report, Claude Code agents running Claude Mythos 5 surveyed reverse transcriptase loci across 1.9 billion metagenomic protein clusters in a campaign of 949 agent sessions and 215.6 million tokens over 21.5 hours without human intervention; one agent reading raw DNA noticed a tandem repeat array beside an unusual reverse transcriptase gene found in jumbo phages ^[46]. Anthropic's post gives rounded figures of roughly 950 agents, 210 million tokens and 21 hours, and says its scientists' involvement was limited to the initial prompt and the lab work, which is performed by human scientists ^[45]. The system consists of the reverse transcriptase, a partner gene and an array of roughly 200-nucleotide units, each a short repeat (15 to 49 nucleotides) followed by a longer spacer, that is expressed as distinct short RNAs during phage infection ^[45][46]. Anthropic describes Claude as appearing to be "the first to notice" the system's defining features, and the report states that the authors "have not shown that the RT is active or that the unit RNAs are its substrates"; the system's function is unknown ^[45][46].

Tools and platforms

Several major AI tools and platforms have been developed specifically for scientific applications.

Sources: ^{[3][4][5][6][7][11][13][17][23][27][31][44][50][51][52]}.

Challenges

Reproducibility

Scientific reproducibility requires that results can be independently verified. AI-based scientific results face reproducibility challenges at multiple levels: the training data may not be fully available, model architectures may be proprietary, random seeds affect results, and computational requirements may be prohibitive for independent replication. Some developers have responded by building provenance into their tools; Anthropic, for example, says Claude Science saves the exact code, environment and message history behind each figure it generates ^[44].

Interpretability

Many AI systems, particularly deep neural networks, operate as "black boxes" whose internal reasoning is difficult to interpret. In science, understanding why a prediction is correct is often as important as the prediction itself. A model that accurately predicts protein structures but provides no insight into the underlying physics offers less scientific value than one whose predictions can be mapped to physical principles. Explainable AI techniques are being applied to scientific AI, but interpretability remains a significant challenge, especially for the most complex models.

Data quality and bias

AI models are only as good as the data they are trained on. In scientific applications, training data may contain systematic biases, errors, or gaps. Models trained on biased data may produce predictions that are accurate for well-represented categories but unreliable for underrepresented ones. Careful curation of training data and validation against diverse test sets are essential.

Integration with domain expertise

Effective AI for science requires deep integration between AI researchers and domain scientists. AI practitioners who lack domain knowledge may build models that optimize the wrong objectives or that miss important physical constraints. Conversely, domain scientists who lack AI expertise may misapply tools or misinterpret results. The September 2026 Nature feature on AI co-scientists reported researchers spending close to an hour correcting a tool's misunderstandings of their question before it produced a useful strategy, and raised the concern that relying on such tools could leave students fewer chances to build the expertise needed to judge their output ^[32].

Computational cost

Training and running large AI models requires substantial computational resources. AlphaFold2 was trained on 128 TPU v3 cores ^[18], and AlphaProof was trained by proving or disproving millions of problems over a period of weeks before the 2024 IMO ^[8]. The computational cost of AI for science creates equity concerns: well-funded institutions and wealthy nations can leverage AI for science, while others may be left behind. Open-source models and publicly available predictions (like the AlphaFold Protein Structure Database) partially address this concern but do not eliminate it.

Validation and trust

Scientific claims require rigorous validation. AI predictions, no matter how compelling, must be verified experimentally before they can be accepted as scientific knowledge. The GNoME materials discovery project, for example, identified 2.2 million candidate materials, but their scientific value depends on experimental confirmation. As of the November 2023 announcement, 736 of the predicted materials had been independently created in labs, a meaningful validation but still a small fraction of the total predictions ^[7]. Agentic systems raise the same issue at the level of hypotheses: independent scientists judged 79.4% of the statements in Kosmos reports accurate, leaving about one in five not judged accurate ^[39], and Anthropic's ART result remains a sequence-level and RNA-expression finding until the enzyme's activity is shown ^[46].

Institutional and funding landscape

AI for science has attracted significant institutional investment. Google DeepMind has been one of the most prominent contributors, producing AlphaFold, GraphCast, GenCast, GNoME, AlphaProof, AlphaGenome and other scientific AI systems, and co-developing the AI co-scientist. On 29 October 2025, Google DeepMind and Google.org announced the AI for Math Initiative with five institutions (Imperial College London, the Institute for Advanced Study, the Institut des Hautes Etudes Scientifiques, the Simons Institute for the Theory of Computing at UC Berkeley, and the Tata Institute of Fundamental Research), offering Google.org funding and access to tools including Gemini Deep Think, AlphaEvolve and AlphaProof ^[23].

Other AI developers built dedicated science efforts in 2025 and 2026, including OpenAI for Science and Anthropic's life sciences organization and laboratory (see above) ^[40][45].

National governments have also invested. On 24 November 2025, a US executive order launched the Genesis Mission, a Department of Energy-led effort to use AI, the national laboratories, supercomputers and federal data to "double the productivity and impact of American science and engineering within a decade" ^[47].

Philanthropic programs also support the field. The Eric and Wendy Schmidt AI in Science Fellowship, run by Schmidt Sciences, supports about 160 postdoctoral fellows and 20 faculty fellows each year across nine partner universities to bring AI techniques into science and engineering research ^[48].

Current state (2026)

As of September 2026, AI for science is in a period of rapid expansion following the recognition of the 2024 Nobel Prizes. Several trends characterize the current landscape.

The success of AlphaFold has established a template: identify a hard scientific prediction problem, frame it as a machine learning task, train a large model on available data, and achieve results that rival or exceed experimental methods. This template is being applied to an expanding set of problems, from predicting the properties of quantum materials to simulating cellular processes to modeling climate dynamics.

However, the field is also confronting the limits of the current approach. Many scientific problems lack the large, well-curated datasets that enabled AlphaFold's success. Protein structures had decades of accumulated experimental data in the Protein Data Bank; other domains may not have equivalent resources. Generating high-quality training data for scientific AI remains a bottleneck.

The role of AI in science is also evolving from prediction toward generation. Rather than just predicting properties of existing systems, AI is increasingly used to design new ones: new proteins, new materials, new drug candidates, new experimental protocols. The emergence of agentic systems such as the AI co-scientist, Robin, Kosmos and Anthropic's genome-mining agents, which propose hypotheses, analyze data and in some cases flag anomalies in raw sequence data, represents a shift in how AI contributes to science, from analysis toward initiating lines of investigation that humans then test ^{[30][37][39][46]}.

Finally, questions about the long-term relationship between AI and scientific understanding remain open. AlphaFold can predict protein structures with high accuracy, but it does not fully explain the physical principles that determine protein folding. AI may accelerate discovery while leaving the deeper task of understanding to future work. Whether AI will eventually contribute to fundamental scientific understanding, not just empirical prediction, is one of the most profound open questions in the field.

See also

• AlphaFold
• Google DeepMind
• Deep learning
• Reinforcement learning
• AlphaGo
• Machine learning
• AI drug discovery
• AI co-scientist
• FutureHouse
• Claude Science
• Claude for Life Sciences
• Array-associated reverse transcriptase
• Genesis Mission

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Press release: The Nobel Prize in Chemistry 2024 - NobelPrize.org
2. ^1 ^2 ^3 ^4 ^5 ^6Press release: The Nobel Prize in Physics 2024 - NobelPrize.org
3. ^1 ^2AlphaFold 3 predicts the structure and interactions of all of life's molecules - Google DeepMind / Isomorphic Labs
4. ^1 ^2RoseTTAFold: Accurate protein structure prediction accessible to all - Institute for Protein Design, UW
5. ^1 ^2GraphCast: AI model for faster and more accurate global weather forecasting - Google DeepMind
6. ^1 ^2 ^3GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy - Google DeepMind
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Millions of new materials discovered with deep learning - Google DeepMind
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7AI achieves silver-medal standard solving International Mathematical Olympiad problems - Google DeepMind
9. ^1 ^2 ^3FunSearch: Making new discoveries in mathematical sciences using Large Language Models - Google DeepMind
10. ^1 ^2Xu, Z. et al. "A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial." Nature Medicine, published 3 June 2025
11. ^1 ^2 ^3 ^4Google DeepMind is bringing AI to the next generation of fusion energy - Google DeepMind (16 October 2025)
12. ^1 ^2Olympiad-level formal mathematical reasoning with reinforcement learning - Nature (2025)
13. ^1 ^2 ^3 ^4AlphaGenome: AI for better understanding the genome - Google DeepMind
14. ^AlphaFold - Wikipedia (CASP14 results summary)
15. ^1 ^2 ^3 ^4 ^5Accelerating scientific breakthroughs with an AI co-scientist - Google Research
16. ^1 ^2AlphaFold: a solution to a 50-year-old grand challenge in biology - Google DeepMind (30 November 2020)
17. ^1 ^2AlphaFold reveals the structure of the protein universe - Google DeepMind (28 July 2022)
18. ^1 ^2Jumper, J. et al. "Highly accurate protein structure prediction with AlphaFold." Nature 596, 583-589 (2021)
19. ^1 ^2The Nobel Prize in Physics 2024 - Popular science background - NobelPrize.org
20. ^WeatherNext 2: Google DeepMind's most advanced forecasting model - Google (17 November 2025)
21. ^Engineers use AI to wrangle fusion power for the grid - Princeton Engineering (February 2024)
22. ^Accelerating fusion science through learned plasma control - Google DeepMind (16 February 2022)
23. ^1 ^2Google DeepMind and Google.org announce AI for Math Initiative - Google (29 October 2025)
24. ^Jayatunga, M. K. P. et al. "How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons." Drug Discovery Today 29(6), 104009 (2024) - PubMed
25. ^FDA Qualifies First AI Drug Development Tool, Will Be Used in 'MASH' Clinical Trials - U.S. Food and Drug Administration (8 December 2025)
26. ^Chai Discovery Team. "Zero-shot antibody design in a 24-well plate." bioRxiv (July 2025)
27. ^1 ^2 ^3A catalogue of genetic mutations to help pinpoint the cause of diseases - Google DeepMind (19 September 2023)
28. ^Ren, F. et al. "A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models." Nature Biotechnology (2024)
29. ^1 ^2Co-Scientist: A multi-agent AI partner to accelerate research - Google DeepMind (19 May 2026)
30. ^1 ^2 ^3Gottweis, J. et al. "Accelerating scientific discovery with Co-Scientist." Nature, published 19 May 2026
31. ^1 ^2Gemini for Science: AI experiments and tools for a new era of discovery - Google
32. ^1 ^2 ^3 ^4Dolgin, E. "AI co-scientists are revolutionizing how research is done." Nature 657, 1108-1110 (21 September 2026)
33. ^1 ^2The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery - Sakana AI (13 August 2024)
34. ^1 ^2The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature - Sakana AI (26 March 2026)
35. ^Beel, J., Kan, M.-Y. and Baumgart, M. "Evaluating Sakana's AI Scientist: Bold Claims, Mixed Results, and a Promising Future?" arXiv:2502.14297
36. ^1 ^2 ^3Demonstrating end-to-end scientific discovery with Robin: a multi-agent system - FutureHouse
37. ^1 ^2 ^3Ghareeb, A. E. et al. "A multi-agent system for automating scientific discovery." Nature, published 19 May 2026
38. ^Announcing Edison Scientific - FutureHouse (5 November 2025)
39. ^1 ^2 ^3 ^4Mitchener, L. et al. "Kosmos: An AI Scientist for Autonomous Discovery." arXiv:2511.02824
40. ^1 ^2Kevin Weil, "I'm starting something new inside OpenAI!" - X (2 September 2025)
41. ^1 ^2OpenAI launches Prism, a new AI workspace for scientists - TechCrunch (27 January 2026)
42. ^Kevin Weil and Bill Peebles exit OpenAI as company continues to shed 'side quests' - TechCrunch (17 April 2026)
43. ^Claude for Life Sciences - Anthropic (20 October 2025)
44. ^1 ^2 ^3 ^4Claude Science, an AI workbench for scientists, is now available - Anthropic (30 June 2026)
45. ^1 ^2 ^3 ^4 ^5 ^6 ^7Claude discovers a novel enzyme system with CRISPR-like repeats - Anthropic (23 September 2026)
46. ^1 ^2 ^3 ^4 ^5 ^6 ^7Yoon, P. H. et al. "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays." Anthropic technical report (September 2026)
47. ^Energy Department Launches 'Genesis Mission' to Transform American Science and Innovation Through the AI Computing Revolution - U.S. Department of Energy (24 November 2025)
48. ^AI in Science Fellowship - Schmidt Sciences
49. ^Romera-Paredes, B. et al. "Mathematical discoveries from program search with large language models." Nature 625, 468-475 (2024)
50. ^1 ^2 ^3Avsec, Z. et al. "Advancing regulatory variant effect prediction with AlphaGenome." Nature (28 January 2026)
51. ^synthetic-sciences/openscience: The open-source AI workbench for scientific research - GitHub
52. ^1 ^2 ^3Edison Scientific - Kosmos: The AI Scientist for R&D
53. ^google-deepmind/alphagenome_research - GitHub