Nature Machine Intelligence最新文献

{"title":"A strong sustainability approach to AI development","authors":"Daniel W. O’Neill, Felix Creutzig","doi":"10.1038/s42256-026-01240-w","DOIUrl":"10.1038/s42256-026-01240-w","url":null,"abstract":"","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"642-644"},"PeriodicalIF":23.9,"publicationDate":"2026-05-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012339","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A generative artificial intelligence approach for peptide antibiotic optimization","authors":"Marcelo D. T. Torres, Yimeng Zeng, Fangping Wan, Natalie Maus, Jacob Gardner, Cesar de la Fuente-Nunez","doi":"10.1038/s42256-026-01237-5","DOIUrl":"10.1038/s42256-026-01237-5","url":null,"abstract":"Antibiotic resistance is rising globally, demanding faster, more reliable routes to design antimicrobial candidates. Although artificial-intelligence-based methods have accelerated antimicrobial discovery, most are designed to screen fixed libraries or generate candidates broadly, rather than optimize existing peptide scaffolds under practical design constraints. Here, to address this challenge, we present APEX generative optimization (ApexGO). ApexGO uses a transformer variational autoencoder that embeds peptide sequences in a continuous latent space, whereas Bayesian optimization efficiently proposes sequence edits to boost antimicrobial potency. Unlike traditional approaches, ApexGO generates peptide sequences through modifications of template peptides, opening avenues for peptide design and antibiotic discovery. Using ten peptides as templates, ApexGO generated optimized derivatives with enhanced antimicrobial properties. We chemically synthesized 100 of these compounds and conducted comprehensive in vitro characterizations, including assessments of antimicrobial activity, mechanism of action, secondary structure and cytotoxicity. In particular, ApexGO achieved an 85% ground-truth experimental hit rate and a 72% success rate in enhancing antimicrobial activity against Gram-negative pathogens, outperforming previously reported methods for antibiotic discovery and optimization. In two preclinical mouse models of Acinetobacter baumannii infection, artificial-intelligence-optimized molecules exhibited potent anti-infective activity superior to their template controls and comparable with or exceeding that of last-resort antibiotic. These findings highlight the potential of ApexGO as a generative artificial intelligence approach for peptide design and antibiotic optimization, offering a powerful tool to accelerate antibiotic discovery. Torres et al. present ApexGO, a generative approach capable of redesigning peptide antibiotics to better kill drug-resistant bacteria. They validated candidates in laboratory tests and mouse infections and matching or outperforming standard antibiotics.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"841-856"},"PeriodicalIF":23.9,"publicationDate":"2026-05-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01237-5.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012338","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Peripheral control enabled by distributed sensing in an octopus-inspired soft robotic arm for autonomous underwater grasping","authors":"Emanuela Del Dottore, Romina Adhami, Ebrahim Shahabi, Emanuele Solfiti, Michele Martini, Stefano Mariani, Alberto Parmiggiani, Alessio Mondini, Edoardo Sinibaldi, Barbara Mazzolai","doi":"10.1038/s42256-026-01230-y","DOIUrl":"10.1038/s42256-026-01230-y","url":null,"abstract":"Octopuses exhibit remarkable motor dexterity through distributed sensing and control in their flexible arms. Inspired by this biological model, we present a tendon-driven soft robotic arm featuring optoelectronic mechanosensors embedded in suction cups and a hierarchical behaviour-based control architecture. Each artificial sucker integrates light-emitting diodes and phototransistors to detect contact force and direction via light reflection, achieving high sensitivity (~400 mV N−1 in the 0–2-N range) and directional accuracy (error, <18°). The sensors operate reliably in dry and wet environments with minimal drift and hysteresis. Their compact, modular design facilitates system integration. The hierarchical control architecture enables local reflexes at the suction cup level and global coordination for autonomous grasping. The system reliably detects contact events, estimates force and direction and infers object position relative to the arm, enabling purposeful interaction. This work advances sensor-integrated soft robotics, demonstrating the potential of biologically inspired designs for adaptive grasping in unstructured environments. Del Dottore et al. present an octopus-inspired soft robotic arm that uses optoelectronic mechanosensors in its suction cups to detect contact forces and infer object positions.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"708-721"},"PeriodicalIF":23.9,"publicationDate":"2026-05-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01230-y.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012341","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Towards the explainability of protein language models","authors":"Andrea Hunklinger, Noelia Ferruz","doi":"10.1038/s42256-026-01232-w","DOIUrl":"10.1038/s42256-026-01232-w","url":null,"abstract":"Artificial intelligence models are transforming protein research, enabling advances in areas ranging from structure prediction to the design of functional enzymes. However, these models operate as black boxes, and their underlying working principles remain unclear. Here we survey emerging applications of explainable artificial intelligence (XAI) to protein language models and describe the potential of XAI in protein research. We organize existing work around four points in a typical modelling pipeline: the data used for training; the user-provided inputs; the internal model architecture; and input–output relationships. Across these contexts, we highlight methods and applications of XAI. In addition, from published studies we distil five potential roles for XAI in protein research: Evaluator, Multitasker, Engineer, Coach and Teacher, with the evaluator role being the only one widely adopted so far. While our analysis focuses on protein language models, our categorization is broadly applicable to any other architecture. We conclude by highlighting critical areas of application for the future and outlining a path to advance the interpretability of protein artificial intelligence. Hunklinger and Ferruz provide an overview of explainable artificial intelligence methods for protein language models.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"649-662"},"PeriodicalIF":23.9,"publicationDate":"2026-05-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012340","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Re-thinking human–machine interaction and the governance of AI in the military domain","authors":"Ingvild Bode, Katherine Chandler","doi":"10.1038/s42256-026-01231-x","DOIUrl":"10.1038/s42256-026-01231-x","url":null,"abstract":"Human control is a key principle for the governance of artificial intelligence (AI) in the military domain, underlining the importance of human responsibility in the use of military AI. But what human control exactly entails is unclear. This Perspective proposes that safeguarding human control requires examining the diverse forms of human–machine interaction that occur across the AI life cycle. Here we review how actions at distinct stages, including research and development as well as testing, evaluation, validation and verification, impact human control. We evaluate these dynamics through potential applications of AI-based decision support systems to assist in international humanitarian law assessments. Arguing that dynamics of human–machine interaction require more attention to uphold human control, we offer three recommendations to bolster this principle: contestation mechanisms to help humans cross-check and validate information generated by AI; continuous training to help users navigate unexpected scenarios and contexts that lack sufficient data; and documentation. Bode and Chandler discuss how human–machine interactions across the AI life cycle affect human control and decision-making in the military domain.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"663-669"},"PeriodicalIF":23.9,"publicationDate":"2026-05-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012343","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Learning the chemical language of natural products","authors":"Xu Guo, Celia M. Rava, Allison S. Walker","doi":"10.1038/s42256-026-01241-9","DOIUrl":"10.1038/s42256-026-01241-9","url":null,"abstract":"A promising foundation model is developed for a range of downstream applications in natural product mining.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"645-646"},"PeriodicalIF":23.9,"publicationDate":"2026-05-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012346","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Platonic representation of foundation machine learning interatomic potentials","authors":"Zhenzhu Li, Aron Walsh","doi":"10.1038/s42256-026-01235-7","DOIUrl":"10.1038/s42256-026-01235-7","url":null,"abstract":"Foundation machine learning interatomic potentials (MLIPs) have emerged as powerful tools for atomistic simulation, yet different models encode chemical environments in incompatible latent spaces, limiting direct comparison and interoperability. The platonic representation hypothesis suggests that sufficiently capable models converge towards a shared statistical representation of reality. Here, motivated by this hypothesis, we show that independently developed MLIPs exhibit statistically consistent geometric organization of atomic environments. By projecting embeddings relative to a set of atomic anchors, we unify the latent spaces of seven MLIPs—spanning equivariant, non-equivariant, conservative and non-conservative architectures—into a common latent space that preserves chemical periodicity and structural invariants. This unified framework enables cross-model optimal transport, interpretable embedding arithmetic and the detection of representational biases. Furthermore, we show that deviation in this space provides a ground-truth-free measure for atypical structures, and signals physical prediction failures. Our results suggest that the platonic representation offers a practical route towards interoperable, comparable and interpretable foundation models for materials science. Li and Walsh show that a unified ‘platonic’ geometry emerges across the representations of independent foundation models for learning interatomic potentials. The observation enables cross-model comparison, embedding arithmetic and ground-truth-free diagnostics.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"830-840"},"PeriodicalIF":23.9,"publicationDate":"2026-05-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01235-7.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012348","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Reusability report: Meta-learning for antigen-specific T cell receptor binder identification","authors":"Fei He, Xianyu Wang, Dong Xu","doi":"10.1038/s42256-026-01236-6","DOIUrl":"10.1038/s42256-026-01236-6","url":null,"abstract":"Accurate prediction of peptide–T cell receptor (TCR) binding is vital for immunotherapy, vaccine design and diagnostics. PanPep, a meta-learning framework, was developed to generalize diverse TCR binder predictions. Here we present a comprehensive and unbiased evaluation of PanPep’s reusability and practical utility. We reproduced its reported performance on original datasets and further benchmarked it against the control tools using both classification metrics and virtual screening enrichment evaluations. Leveraging a newly curated independent dataset, we demonstrated that PanPep generalizes better than other methods to unseen antigens with few or no known TCR binders under a background-drawn negative sampling strategy; however, this advantage diminished under reshuffled negatives, which present a more challenging evaluation setting by introducing harder negative examples. We further extended PanPep to peptide–TCRα and peptide–TCRαβ binding prediction, demonstrating its applicability in more biologically and physiologically relevant contexts. Despite its strengths, PanPep exhibits limitations in early binder enrichment and reduced robustness to novel TCRs, indicating sensitivity of performance to model architecture, training data composition and negative sampling strategies. This work establishes a reproducible and extensible benchmarking framework for general peptide–TCR binding prediction. The results provide practical guidance for model selection and robustness assessment in real-world TCR binder discovery. At the same time, our findings highlight that accurate, generalizable TCR binding prediction remains an open challenge. He et al. comprehensively test the reusability of PanPep, a meta-learning framework for peptide–TCR binding prediction. The authors demonstrate the utility of PanPep in extended testing scenarios, leveraging proposed independent datasets and negative sampling strategies.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"818-829"},"PeriodicalIF":23.9,"publicationDate":"2026-05-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012337","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Force-free molecular dynamics through autoregressive equivariant networks","authors":"Fabian L. Thiemann, Thiago Reschützegger, Massimiliano Esposito, Tseden Taddese, Juan D. Olarte-Plata, Fausto Martelli","doi":"10.1038/s42256-026-01227-7","DOIUrl":"10.1038/s42256-026-01227-7","url":null,"abstract":"Molecular dynamics simulations play a crucial role in scientific research. Yet their computational cost often limits the timescales and system sizes that can be explored. Most data-driven efforts have focused on reducing the computational cost of the accurate interatomic forces required to solve the equations of motion. However, despite their success, these machine-learning interatomic potentials are still bound to small time steps. In this work, we introduce TrajCast, a transferable and data-efficient framework based on autoregressive equivariant message-passing networks, that directly updates the atomic positions and velocities, lifting the constraints imposed by traditional numerical integration. We benchmark our framework across various systems, including a small molecule, crystalline material and bulk liquid, demonstrating excellent agreement with reference molecular dynamics simulations for structural, dynamical and energetic properties. Moreover, we show that TrajCast can generalize in a zero-shot manner to unseen regions of phase space, producing physically meaningful ensembles in metastable equilibrium and out-of-equilibrium regimes beyond the training data, without compromising accuracy. Depending on the system, TrajCast allows for forecast intervals up to 30× larger than traditional MD time steps, generating over 15 ns of trajectory data per day for a solid with more than 4,000 atoms. By enabling efficient, large-scale simulations over extended timescales, TrajCast can accelerate materials discovery and explore physical phenomena beyond the reach of traditional simulations and experiments. Thiemann et al. introduce TrajCast, a neural network that bypasses force calculations to directly predict atomic trajectories, enabling time steps up to 30 times longer while accurately reproducing physical properties of molecules and materials.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"764-776"},"PeriodicalIF":23.9,"publicationDate":"2026-05-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01227-7.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012333","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A collaborative constrained graph diffusion model for the generation of realistic synthetic molecules","authors":"Manuel Ruiz-Botella, Marta Sales-Pardo, Roger Guimerà","doi":"10.1038/s42256-026-01229-5","DOIUrl":"10.1038/s42256-026-01229-5","url":null,"abstract":"Developing new molecular compounds is crucial to address pressing challenges, from health to environmental sustainability. However, exploring the molecular space to discover new molecules is difficult owing to the vastness of the space. Here we introduce CoCoGraph, a collaborative and constrained graph diffusion model capable of generating molecules that are guaranteed to be chemically valid. Thanks to the constraints built into the model and to the collaborative mechanism, CoCoGraph outperforms state-of-the-art approaches on standard benchmarks while being more efficient. Analysis of 36 chemical properties also demonstrates that CoCoGraph generates molecules with distributions more closely matching real molecules than current models. To illustrate the potential of the model, we created a database of 8.2 million synthetically generated molecules, show how this database and CoCoGraph could be used for molecular discovery and conduct a Turing-like test with organic chemistry experts to further assess the plausibility of the generated molecules, and the potential biases and limitations of CoCoGraph. The collaborative constrained graph diffusion model CoCoGraph generates novel molecules that are guaranteed to be valid and more realistic than state-of-the-art outputs, while achieving faster performance with up to an order of magnitude fewer parameters.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 5","pages":"749-763"},"PeriodicalIF":23.9,"publicationDate":"2026-05-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148012326","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}