Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    How AI designs new proteins, drugs and materials

    Generative models now propose new proteins, drug candidates and inorganic crystals instead of only analysing known ones.[1][2] The bottleneck has moved to the lab: designs must be synthesised and tested, and only a few AI-designed drugs have reached human trials.[3]

    Editor reviewedUpdated AI for scienceArtificial intelligenceScienceLife sciences

    Prediction asks “what does this molecule look like?” Design asks “what molecule should exist?” Since 2023, generative models have started answering the second question for proteins, small-molecule drugs and inorganic crystals.[1][4]

    Designing proteins

    Image generators start from random noise and clean it up into a picture. Protein design models do the same with 3D shapes: they start from a jumble of atoms and refine it into a new protein that should fold and grab a chosen target.[1]

    rfdiffusion fine-tuned the RoseTTAFold structure network on denoising, yielding a diffusion model for backbones that handles binders, symmetric assemblies and enzyme active-site scaffolding.[1][5] RFdiffusion3, released in December 2025, designs at the level of individual atoms and can target DNA and small molecules.[6][7]

    The key metric is the hit rate: how many designs actually work. The same metric applies to antibodies. In a 2025 preprint, Chai Discovery reported a 16% hit rate for fully de novo antibody design with its Chai-2 model, finding at least one binder for half of 52 targets that had no known antibody binder in the Protein Data Bank.[8] Anthropic reported in August 2026 that Claude, driving open-source design tools, reached 22–35% for small binders, against roughly 10–15% it describes as typical.[9] Results varied widely by target, from 90% to 0%.[10]

    Designing drugs

    Most medicines are small molecules that fit into a pocket on a disease protein. AI can suggest new disease targets and draw candidate molecules for them. But a molecule only becomes a medicine after years of tests in people.[11][3]

    Rentosertib is described as an AI-generated TNIK inhibitor against a target also nominated by generative AI; its 71-patient phase 2a trial was published in 2025.[11][12] In July 2026 Insilico started a 320-patient, 52-week phase III trial of rentosertib in China.[13] A 2024 pipeline analysis found 80–90% phase I success for AI-discovered molecules but about 40% in phase II, close to industry norms.[14][15]

    Predicting how tightly a molecule binds its target is a core step. The open Boltz-2 model, released in 2025, predicts structure and binding affinity together. Its authors report accuracy approaching physics-based free-energy perturbation while being at least 1,000 times more computationally efficient.[16][17] Isomorphic Labs makes a similar, company-reported claim for its proprietary engine.[18]

    Designing materials

    New battery, solar and chip materials are usually found by slow trial and error.[19] AI can screen millions of possible crystals on a computer and flag the ones likely to be stable.[4]

    gnome-materials used graph networks trained at scale, checked with first-principles calculations, to propose 2.2 million structures below the known convex hull.[4] Generative models such as MatterGen go further and generate crystals conditioned on target properties; MatterGen’s structures were more than twice as likely to be new and stable as earlier generators.[2] “Stable in a calculation” is not the same as “made and useful”: critics argue many predicted compounds are trivial variants of known ones.[20]

    Closing the loop

    Designs only count once made and measured. That is why the field is building self-driving-labs, where robots synthesise and test AI proposals and feed results back to the model.[21] Their early claims have been contested too.[22] In January 2026 the A-Lab authors corrected their paper to say its “novel” compounds were new to the prediction platform, not necessarily new to science.[23]

    Language models are now joining the loop as planners. Coscientist, described in Nature in 2023, used OpenAI‘s GPT-4 to design, plan and run chemistry experiments, including optimising a palladium-catalysed reaction.[24] Start-ups are scaling the idea: Lila Sciences reported in September 2026 that its AI-directed lab screened 2,942 green-hydrogen catalysts in three months, a company-reported result.[25] For how such agents work in general, see How AI agents work.

    Questions readers ask

    What is a "hit rate" in protein design?

    The share of designs that actually work when tested. Anthropic describes 10–15% as typical for binder design campaigns and reported 22–35% for its Claude-led campaign.[9]

    How many new materials has AI found?

    Google DeepMind's GNoME counted 381,000 newly discovered stable materials among 2.2 million predicted structures, though critics question how many are novel and useful.[4][20]

    Has an AI-designed drug worked in patients?

    The furthest published example, rentosertib, completed a 71-patient phase 2a trial with similar adverse-event rates across arms. A 320-patient phase III trial in China started in July 2026.[12][26][13]

    Can AI design antibodies?

    Chai Discovery reported in a 2025 preprint a 16% hit rate for fully de novo antibodies, with at least one binder for half of 52 targets.[8]

    Sources

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

    1. [1]

      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

    2. [2]

      MatterGen, a generative model for inorganic materials published in Nature in January 2025, produced structures more than twice as likely to be new and stable as earlier generative models. confirmedas of 2025-01-16

    3. [3]

      The rentosertib paper notes that few novel AI-discovered or AI-designed drugs have reached human clinical trials. confirmedas of 2025-06-03

    4. [4]

      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

    5. [5]

      The RFdiffusion paper reported strong results on binder design, symmetric oligomer design, enzyme active-site scaffolding and motif scaffolding, with hundreds of designs characterized experimentally. confirmedas of 2023-07-11

    6. [6]

      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

    7. [7]

      The Institute for Protein Design says RFdiffusion3 treats individual atoms as the units being designed and runs ten times faster than RFdiffusion2. confirmedas of 2025-12-03

    8. [8]

      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

    9. [9]

      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

    10. [10]

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

    11. [11]

      Rentosertib is described by its developers as a first-in-class small-molecule TNIK inhibitor generated with AI, against a target that was also identified using generative AI. confirmedas of 2025-06-03

    12. [12]

      A phase 2a trial of rentosertib, published in Nature Medicine in June 2025, randomized 71 patients with idiopathic pulmonary fibrosis to one of three doses or placebo for 12 weeks. confirmedas of 2025-06-03

    13. [13]

      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

    14. [14]

      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

    15. [15]

      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

    16. [16]

      Boltz-2, described in a June 2025 preprint by MIT and Recursion researchers, predicts both complex structures and small molecule–protein binding affinity, and its authors say it is the first AI model to approach free-energy perturbation accuracy while being at least 1,000 times more computationally efficient. confirmedas of 2025-06-18

    17. [17]

      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

    18. [18]

      In February 2026 Isomorphic Labs reported that its drug design engine, IsoDDE, more than doubles AlphaFold 3's accuracy on a challenging protein–ligand generalisation benchmark and predicts binding affinities more accurately than gold-standard physics-based methods; these are company-reported benchmark results. reportedas of 2026-02-10

    19. [19]

      The GNoME authors describe the discovery of inorganic crystals as bottlenecked by expensive trial-and-error approaches. confirmedas of 2023-11-29

    20. [20]

      MIT Technology Review reported that researchers who examined GNoME's predicted compounds found scant evidence of novelty, credibility and utility, describing some as merely trivial variations of known compounds. reportedas of 2025-12-15

    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]

      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

    23. [23]

      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

    24. [24]

      Coscientist, described in Nature in December 2023, is a GPT-4-driven system that autonomously designs, plans and performs complex chemistry experiments, including successful optimisation of palladium-catalysed cross-coupling reactions. confirmedas of 2023-12-20

    25. [25]

      Lila Sciences reported in September 2026 that its AI-directed lab proposed, synthesised and screened 2,942 catalysts for green hydrogen in three months and identified six high-performing material families; the result is company-reported. reportedas of 2026-09-25

    26. [26]

      The rentosertib phase 2a trial's primary endpoint, the share of patients with at least one treatment-emergent adverse event, was similar across all arms (70.6% to 83.3%). confirmedas of 2025-06-03

    Revision history (2)
    1. Page created.
    2. Added open models for binding affinity (Boltz-2) and antibody design (Chai-2), rentosertib's phase III start, AI-agent labs such as Coscientist, and the 2026 A-Lab correction.

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

    Cite this page

    "How AI designs new proteins, drugs and materials." ContentLora, updated Oct 10, 2026. https://contentlora.com/explain/how-ai-discovers-molecules-and-materials

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

    Keep exploring

    ☰ Contents