Nature Machine Intelligence最新文献

{"title":"A domain-adapted large language model to support clinicians in psychiatric clinical practice","authors":"Ruoxi Wang, Shuyu Liu, Ling Zhang, Xuequan Zhu, Zhi Yang, Hu Wang, Juan Huang, Yimeng Wang, Xiaofan Yang, Fei Wu, Rui Yang, Gang Wang, Cheng Jin","doi":"10.1038/s42256-026-01224-w","DOIUrl":"https://doi.org/10.1038/s42256-026-01224-w","url":null,"abstract":"Mental disorders affect nearly one billion individuals worldwide, yet professional psychiatric care remains constrained by workforce shortages and experience-dependent decision-making. Despite recent advances in large language models (LLMs), current applications in mental health are primarily patient-oriented and lack alignment with real-world psychiatric clinical workflows. Here we present PsychFound, a domain-adapted and clinician-oriented LLM developed to support psychiatric clinical practice. Developed through a three-phase framework using expert-curated psychiatric corpora and 64,588 Chinese real-world electronic health records, PsychFound integrates psychiatric professional knowledge, clinical reasoning capabilities and adaptation to the full spectrum of psychiatric clinical tasks across diagnosis, treatment planning and longitudinal management in Chinese clinical settings. In retrospective evaluations spanning three professional knowledge assessments and five clinical task benchmarks, the 7B-parameter PsychFound delivered the top overall performance among 22 LLMs. In a real-world, two-arm prospective study, resident psychiatrists assisted by PsychFound demonstrated higher consultation quality, higher diagnostic accuracy, more appropriate medication selection and reduced documentation time (all P < 0.01). A reader study with 60 psychiatrists (20 residents, 20 attendings and 20 seniors) showed that PsychFound’s clinical reasoning performance matched that of attending psychiatrists. These findings demonstrated that PsychFound provides an interpretable, expert-level decision support tool capable of improving consistency, efficiency and standardization in psychiatric clinical care.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"21 1","pages":""},"PeriodicalIF":23.8,"publicationDate":"2026-04-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147751847","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":"Adopting a human developmental visual diet yields robust and shape-based AI vision","authors":"Zejin Lu, Sushrut Thorat, Radoslaw M. Cichy, Tim C. Kietzmann","doi":"10.1038/s42256-026-01228-6","DOIUrl":"https://doi.org/10.1038/s42256-026-01228-6","url":null,"abstract":"Despite years of research and the dramatic scaling of artificial intelligence (AI) systems, a striking misalignment between artificial and human vision persists. Contrary to humans, AI relies heavily on texture features rather than shape information, lacks robustness to image distortions, remains highly vulnerable to adversarial attacks, and struggles to recognize simple abstract shapes within complex backgrounds. Here, to close this gap, we take inspiration from how human vision develops from early infancy into adulthood. We quantified visual maturation by synthesizing decades of research into a novel developmental visual diet for AI vision. Guiding AI systems through this human-inspired curriculum, which considers the development of visual acuity, contrast sensitivity and colour, produces models that better align with human behaviour on every hallmark of robust vision tested, yielding strong reliance on shape information, abstract shape recognition beyond the state of the art, and higher resilience to image corruptions and adversarial attacks. Our results thus demonstrate that robust AI vision can be achieved by guiding how a model learns, not merely how much it learns, offering a resource-efficient route towards safer and more human-like artificial visual systems.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"117 1","pages":""},"PeriodicalIF":23.8,"publicationDate":"2026-04-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147751849","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":"Programmable RNA translation through deep learning-driven IRES discovery and de novo generation","authors":"Yanyi Chu,&nbsp;Di Yin,&nbsp;Dan Yu,&nbsp;Guangxue Xu,&nbsp;Junze Zhang,&nbsp;Xiaotong Wang,&nbsp;Yue Shen,&nbsp;Yupeng Li,&nbsp;Ning Zhao,&nbsp;Yi Zhu,&nbsp;Jason Zhang,&nbsp;Hani Goodarzi,&nbsp;Mengdi Wang,&nbsp;Le Cong","doi":"10.1038/s42256-026-01213-z","DOIUrl":"10.1038/s42256-026-01213-z","url":null,"abstract":"The precise control of protein expression is a major bottleneck in the development of RNA therapeutics. Internal ribosome entry sites (IRES) overcome traditional limitations by enabling cap-independent translation initiation, making them highly desirable tools for synthetic biology and therapeutic payload expression. However, the complex structure-function relationship of IRES elements has historically hindered their rational design. Here we show that a comprehensive, end-to-end artificial intelligence framework unifies IRES identification, evolutionary optimization and de novo generation. First, IRES-LM, an ensemble of two language models trained on 46,774 sequences, predicts linear mRNA IRESs with a 15% improvement in area under the curve and F1 score over existing methods. In addition, IRES-LM demonstrates remarkable cross-applicability to circular RNA IRESs, correctly identifying all 21 experimentally validated circular RNA IRESs and outperforming benchmark methods. Next, IRES-EA integrates an evolutionary algorithm with IRES-LM to induce IRES functionality through targeted mutations. Computational evaluation of 37,293 non-IRES sequences showed 60% predicted functional conversion, with large-scale massively parallel reporter assay validation of 12,000 mutated sequences demonstrating 98.4% acquired IRES functionality, confirming both computational predictions and experimental functionality. Further, IRES-DM employs a diffusion model to de novo generate novel IRES sequences that outperform the state-of-the-art method. Massively parallel reporter assay validation using another set of 12,000 IRES-DM-generated sequences revealed 99.3% detectable IRES functionality. Notably, IRES-DM shows diverse generation capacity, producing sequences ranging from natural-like candidates to structurally conserved yet sequence-divergent designs. Motif analysis revealed both natural-prevalent and design-enriched high-activity motifs. Together, this framework establishes a robust approach for programmable RNA translation, expanding the molecular toolkit for scaling up next-generation biomedical discovery and RNA-based therapeutics. Chu et al. present a framework for programmable RNA translation, of interest for RNA therapeutics. The method enables large-scale discovery and engineering of internal ribosome entry sites, which are validated by high-throughput functional assays.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 4","pages":"559-574"},"PeriodicalIF":23.9,"publicationDate":"2026-04-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147738873","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":"From embodied intelligence to physical AI","authors":"","doi":"10.1038/s42256-026-01239-3","DOIUrl":"10.1038/s42256-026-01239-3","url":null,"abstract":"Several frameworks from different disciplines are converging on the scientific question of what it takes for a system to not just predict, simulate or reason about the world, but to act physically and intelligently within it.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 4","pages":"491-492"},"PeriodicalIF":23.9,"publicationDate":"2026-04-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01239-3.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147738935","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 multimodal large language model for materials science","authors":"Yingheng Tang,&nbsp;Wenbin Xu,&nbsp;Jie Cao,&nbsp;Weilu Gao,&nbsp;Steven Farrell,&nbsp;Benjamin Erichson,&nbsp;Michael W. Mahoney,&nbsp;Andy Nonaka,&nbsp;Zhi Jackie Yao","doi":"10.1038/s42256-026-01214-y","DOIUrl":"10.1038/s42256-026-01214-y","url":null,"abstract":"Understanding and predicting the properties of inorganic materials is crucial for accelerating advancements in materials science and driving applications in energy, electronics and beyond. Integrating material structure data with language-based information through multimodal large language models (LLMs) offers great potential to support these efforts by enhancing human–artificial intelligence interaction. However, a key challenge lies in integrating atomic structures at full resolution into LLMs. In this work, we introduce MatterChat, a versatile structure-aware multimodal LLM that unifies material structural data and textual inputs into a single cohesive model. MatterChat uses a bridging module to effectively align a pretrained universal machine learning interatomic potential with a pretrained LLM, reducing training costs and enhancing flexibility. Our results demonstrate that MatterChat greatly improves performance in material property prediction and human–artificial intelligence interaction, surpassing general-purpose LLMs such as GPT-4. We also demonstrate its usefulness in applications such as more advanced scientific reasoning and step-by-step material synthesis. Tang et al. introduce MatterChat, a multimodal framework effectively integrating material structural data with large language models. It achieves high-precision property predictions and provides interpretable reasoning to accelerate materials discovery.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 4","pages":"588-601"},"PeriodicalIF":23.9,"publicationDate":"2026-04-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01214-y.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147738868","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":"Fluid thinking about collective intelligence","authors":"Justin Werfel","doi":"10.1038/s42256-026-01211-1","DOIUrl":"10.1038/s42256-026-01211-1","url":null,"abstract":"Forms of collective intelligence range from natural and artificial neural networks to swarm robotics and social insect colonies. One key axis for comparing such systems is the mobility of their individual units: systems like neural networks and wireless sensor networks typically rely on fixed topology and consistent neighbour relationships, whereas mobile robots or ants may encounter each other once and never meet again. Consequently, the core mechanisms that these systems use to compute and learn differ fundamentally between static and fluid topologies. This divide has limited the exchange of ideas across domains. This Perspective examines how mobile units achieve collective learning—through plasticity within individuals, transient formations and, notably, environmental modifications—and identifies analogous mechanisms in static networks. It then explores the advantages of mobility, showing how, for certain tasks, unit mobility can allow a collective system to achieve a given level of performance using many fewer units. An analogy between robot swarms performing a consensus task and convolutional neural networks classifying images illustrates how this principle can inform the design and use of smaller static networks, yielding resource savings. Conversely, temporary immobility or predictable movement patterns can enable mobile unit networks to perform more complex computations by leveraging the benefits of static topologies. Viewing each topology through the lens of the other may inspire advances in both domains, including novel network architectures and swarm algorithms. Collectives exhibit intrinsically different operations, resources and limitations depending on whether their components are static or mobile. Werfel examines learning in mobile collectives and shows how each class can benefit from incorporating mechanisms more commonly linked to the other.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 4","pages":"506-516"},"PeriodicalIF":23.9,"publicationDate":"2026-04-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147738869","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":"Algebraic language models for inverse design of metamaterials via diffusion transformers","authors":"Li Zheng,&nbsp;Siddhant Kumar,&nbsp;Dennis M. Kochmann","doi":"10.1038/s42256-026-01218-8","DOIUrl":"10.1038/s42256-026-01218-8","url":null,"abstract":"Generative machine learning models have revolutionized material discovery by capturing complex structure–property relationships, yet extending these approaches to the inverse design of three-dimensional metamaterials remains limited by computational complexity and underexplored design spaces due to the lack of expressive representations. Here we present DiffuMeta, a generative framework integrating diffusion transformers with an algebraic language representation, encoding three-dimensional geometries as mathematical sentences. This compact, unified parameterization spans diverse topologies, enabling the direct application of transformers to structural design. DiffuMeta leverages diffusion models to generate new shell structures with precisely targeted stress–strain responses under large deformations, accounting for buckling and contact while addressing the inherent one-to-many mapping by producing diverse solutions. Uniquely, our approach enables simultaneous control over multiple mechanical objectives, including linear and nonlinear responses beyond training domains. Experimental validation of fabricated structures further confirms the efficacy of our approach for accelerated design of metamaterials and structures with tailored properties. Architected materials are ubiquitous, yet their design remains constrained by complex structure–property relationships. Zheng et al. use diffusion models and a novel algebraic language to rapidly design shell structures with precisely tailored, exceptional mechanical properties.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 4","pages":"628-640"},"PeriodicalIF":23.9,"publicationDate":"2026-04-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147734061","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":"Flow matching for generative modelling in bioinformatics and computational biology","authors":"Alex Morehead,&nbsp;Lazar Atanackovic,&nbsp;Akshata Hegde,&nbsp;Yanli Wang,&nbsp;Frimpong Boadu,&nbsp;Joel Selvaraj,&nbsp;Alexander Tong,&nbsp;Aditi Krishnapriyan,&nbsp;Jianlin Cheng","doi":"10.1038/s42256-026-01220-0","DOIUrl":"10.1038/s42256-026-01220-0","url":null,"abstract":"Numerous problems in bioinformatics and computational biology can be framed as a task of learning a mapping from one state of a biological system to another relevant state or of exploring novel data points across biologically constrained spaces. However, manually deriving such mappings—for example, to transform cells in a diseased state back into a healthy state, or extrapolating from existing datasets to create new data—is often non-trivial and can require extraordinary domain expertise and resources. Fortunately, the field of generative artificial intelligence (AI) has introduced a new training paradigm referred to as (conditional) flow matching, which has emerged as a promising solution to this problem, with broad applicability in computer vision, natural language processing, and the physical and life sciences. Flow matching is a powerful and principled, data-driven framework for efficiently learning a mapping between arbitrary pairs of high-dimensional data distributions, making it well suited for addressing problems in molecular and cell biology. In this Review, we characterize the theoretical foundations of flow matching and its applications in biomolecular modelling for small molecules, proteins, DNA/RNA, and their interactions, as well as its uses in single/multi-cellular modelling for cell phenotyping and imaging, each contributing towards the development of an AI-based virtual cell. Finally, this review highlights open-source flow-matching methods and discusses future directions in flow-based generative modelling for bioinformatics and computational biology. Flow matching has emerged as a promising solution to mapping arbitrary pairs of high-dimensional data distributions, well suited to problems in molecular and cell biology. Morehead et al. review the theoretical foundations of flow-matching-based models and applications of flow matching in computational biology, and discuss its role in developing methods towards an AI-based virtual cell.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 4","pages":"517-534"},"PeriodicalIF":23.9,"publicationDate":"2026-04-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147738864","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":"Competing Biases underlie Overconfidence and Underconfidence in LLMs","authors":"Dharshan Kumaran,&nbsp;Stephen M. Fleming,&nbsp;Larisa Markeeva,&nbsp;Joe Heyward,&nbsp;Andrea Banino,&nbsp;Mrinal Mathur,&nbsp;Razvan Pascanu,&nbsp;Simon Osindero,&nbsp;Benedetto De Martino,&nbsp;Petar Veličković,&nbsp;Viorica Patraucean","doi":"10.1038/s42256-026-01217-9","DOIUrl":"10.1038/s42256-026-01217-9","url":null,"abstract":"Large language models (LLMs) are increasingly deployed in high-stakes applications where reliable confidence estimation is crucial for trustworthy artificial intelligence (AI). However, their confidence dynamics remain poorly understood, with users reporting paradoxical behaviours: LLMs exhibit reduced flexibility in updating initial responses while simultaneously showing excessive sensitivity to contradictory feedback. Understanding these confidence patterns is essential for developing more reliable AI systems and improving human–AI interaction. Here we show that LLM confidence is governed by two competing mechanisms that explain this paradox. First, we identify a choice-supportive bias: when LLMs view their initial answers, they exhibit inflated confidence and maintain their original responses at rates exceeding optimal decision-making, even when presented with contrary evidence. Second, we demonstrate systematic overweighting of contradictory information: LLMs update their confidence more strongly in response to opposing advice than supporting advice, deviating markedly from optimal Bayesian reasoning. These mechanisms operate across diverse models and generalize from simple factual queries to reasoning tasks. Our computational modelling reveals that these two principles—self-consistency preservation and hypersensitivity to contradiction—capture LLM behaviour across domains. These findings provide an understanding of when and why LLMs exhibit adherence to initial responses versus disproportionate updating, with implications for enhancing the robustness and transparency of LLM decision-making. Kumaran et al. show that large language model (LLM) confidence is shaped by two competing biases: a choice-supportive bias that inflates confidence in initial answers, and a systematic overweighting of contradictory advice, deviating from optimal Bayesian reasoning.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 4","pages":"614-627"},"PeriodicalIF":23.9,"publicationDate":"2026-04-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01217-9.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147734062","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":"Molecular deep learning at the edge of chemical space","authors":"Derek van Tilborg,&nbsp;Luke Rossen,&nbsp;Francesca Grisoni","doi":"10.1038/s42256-026-01216-w","DOIUrl":"10.1038/s42256-026-01216-w","url":null,"abstract":"Molecular machine learning models often fail to generalize beyond the chemical space of their training data, limiting their ability to reliably perform predictions on structurally novel bioactive molecules. Here, to advance the ability of machine learning to go beyond the ‘edge’ of their training chemical space, we introduce a joint modelling approach that combines molecular property prediction with molecular reconstruction. This approach allows the introduction of unfamiliarity, a reconstruction-based metric that enables the estimation of model generalizability. Via a systematic analysis spanning more than 30 bioactivity datasets, we demonstrate that unfamiliarity not only effectively identifies out-of-distribution molecules but also serves as a reliable predictor of classifier performance. Even when faced with the presence of strong distribution shifts on large-scale molecular libraries, unfamiliarity yields robust and meaningful molecular insights that go unnoticed by traditional methods. Finally, we experimentally validate unfamiliarity-based molecule screening in the wet lab for two clinically relevant kinases, discovering seven compounds with low micromolar potency and limited similarity to training molecules. This demonstrates that unfamiliarity can extend the reach of machine learning beyond the edge of the charted chemical space, advancing the discovery of diverse and structurally novel molecules. van Tilborg et al. introduce a deep learning model trained to simultaneously predict bioactivity and reconstruct molecules. The model provides a principled measure of distribution shifts, experimentally validated by discovering active inhibitors for two clinically relevant kinase targets.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 4","pages":"575-587"},"PeriodicalIF":23.9,"publicationDate":"2026-04-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-026-01216-w.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147734063","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}