Nature Machine Intelligence最新文献

{"title":"Author Correction: End-to-end cryo-EM complex structure determination with high accuracy and ultra-fast speed","authors":"Jue Wang, Cheng Tan, Zhangyang Gao, Guijun Zhang, Yang Zhang, Stan Z. Li","doi":"10.1038/s42256-025-01094-8","DOIUrl":"10.1038/s42256-025-01094-8","url":null,"abstract":"","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 7","pages":"1187-1187"},"PeriodicalIF":23.9,"publicationDate":"2025-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01094-8.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145123753","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 predictive framework for liquid electrolytes takes root with BAMBOO","authors":"Ioan-Bogdan Magdău, Gábor Csányi","doi":"10.1038/s42256-025-01071-1","DOIUrl":"10.1038/s42256-025-01071-1","url":null,"abstract":"Predicting the macroscopic properties of molecular liquids from first principles is a major challenge owing to the disordered nature of liquids and the weak link between microscopic forces and thermodynamic observables. A new workflow called BAMBOO produces accurate and transferable machine learning interatomic potential simulations of liquid electrolytes.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 7","pages":"983-984"},"PeriodicalIF":23.9,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144594133","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":"Self-iterative multiple-instance learning enables the prediction of CD4+ T cell immunogenic epitopes","authors":"Long-Chen Shen, Yumeng Zhang, Zhikang Wang, Dene R. Littler, Yan Liu, Jinhui Tang, Jamie Rossjohn, Dong-Jun Yu, Jiangning Song","doi":"10.1038/s42256-025-01073-z","DOIUrl":"10.1038/s42256-025-01073-z","url":null,"abstract":"Accurate prediction of antigen presentation to CD4+ T cells and subsequent induction of immune response are fundamentally important for vaccine development, autoimmune disease treatment and cancer neoepitope discovery. In immunopeptidomics, single-allelic data offer high specificity but limited allele coverage, whereas multi-allelic data provide broader representation at the expense of weak labelling. Current computational approaches either overlook the abundance of multi-allelic data or suffer from label ambiguity due to inadequate modelling strategies. To address these limitations, we present ImmuScope, a weakly supervised deep learning framework that integrates major histocompatibility complex class II (MHC-II) antigen presentation, CD4+ T cell epitopes and immunogenicity assessment. ImmuScope leverages self-iterative multiple-instance learning with positive-anchor triplet loss to decipher peptide-MHC-II binding from weakly labelled multi-allelic data and high-confidence single-allelic data. The training dataset comprises over 600,000 ligands across 142 alleles. Additionally, ImmuScope enables the interpretation of MHC-II binding specificity and motif deconvolution of immunopeptidomics data. We successfully applied ImmuScope to identify melanoma neoantigens, uncovering mutation-driven variations in peptide-MHC-II binding and immunogenicity. Furthermore, we employed ImmuScope to evaluate the effects of SARS-CoV-2 epitope mutations associated with immune escape, with predictions well aligned with experimentally observed immune escape dynamics. Overall, by offering a unified solution for CD4+ T cell antigen recognition and immunogenicity assessment, ImmuScope holds substantial promise for accelerating vaccine design and advancing personalized immunotherapy. ImmuScope, a weakly supervised deep learning model capable of analysing multi- and single-allelic data, is introduced, facilitating interpretable neoantigen discovery and immune escape analysis.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 8","pages":"1250-1265"},"PeriodicalIF":23.9,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01073-z.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144594135","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":"Enabling large language models for real-world materials discovery","authors":"Santiago Miret, N. M. Anoop Krishnan","doi":"10.1038/s42256-025-01058-y","DOIUrl":"10.1038/s42256-025-01058-y","url":null,"abstract":"Large language models (LLMs) create exciting possibilities to accelerate scientific discovery and knowledge dissemination in materials science. While LLMs have been successfully applied to select scientific problems and rudimentary challenges, they currently fall short of being practical materials science tools. In this Perspective, we show relevant failure cases of LLMs in materials science that reveal the current limitations of LLMs related to comprehending and reasoning over complex, interconnected materials science knowledge. Given these shortcomings, we outline a framework for developing materials science LLMs (MatSci-LLMs) that are grounded in domain knowledge, which can enable hypothesis generation followed by hypothesis testing for impactful materials science challenges. The path to attaining performant MatSci-LLMs rests, in large part, on building high-quality, multimodal datasets sourced from scientific literature, where various information extraction challenges persist. As such, we describe key materials science information extraction challenges that need to be overcome to build large-scale, multimodal datasets that capture valuable materials science principles and broader knowledge. Miret and Krishnan discuss the promise of large language models (LLMs) to revolutionize materials discovery via automated processing of complex, interconnected, multimodal materials data. They also consider critical limitations and research opportunities needed to unblock LLMs for breakthroughs in materials science.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 7","pages":"991-998"},"PeriodicalIF":23.9,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144594134","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":"Large language models to accelerate organic chemistry synthesis","authors":"Yu Zhang, Yang Han, Shuai Chen, Ruijie Yu, Xin Zhao, Xianbin Liu, Kaipeng Zeng, Mengdi Yu, Jidong Tian, Feng Zhu, Xiaokang Yang, Yaohui Jin, Yanyan Xu","doi":"10.1038/s42256-025-01066-y","DOIUrl":"10.1038/s42256-025-01066-y","url":null,"abstract":"Chemical synthesis, as a foundational methodology in the creation of transformative molecules, exerts substantial influence across diverse sectors from life sciences to materials and energy. Current chemical synthesis practices emphasize laborious and costly trial-and-error workflows, underscoring the urgent needs for advanced AI assistants. Recently, large language models, typified by GPT-4, have been introduced as an efficient tool to facilitate scientific research. Here we present Chemma, a fully fine-tuned large language model with 1.28 million pairs of questions and answers about reactions, as an assistant to accelerate organic chemistry synthesis. Chemma surpasses the best-known results in multiple chemical tasks, for example, single-step retrosynthesis and yield prediction, which highlights the potential of general artificial intelligence for organic chemistry. By predicting yields across the experimental reaction space, Chemma significantly improves the reaction exploration capability of Bayesian optimization. More importantly, integrated in an active learning framework, Chemma exhibits advanced potentials of autonomously experimental exploration and optimization in open reaction spaces. For an unreported Suzuki–Miyaura cross-coupling reaction of cyclic aminoboronates and aryl halides for the synthesis of α-aryl N-heterocycles, the human–artificial intelligence collaboration successfully explored a suitable ligand (tri(1-adamantyl)phosphine) and solvent (1,4-dioxane) within only 15 runs, achieving an isolated yield of 67%. These results reveal that, without quantum-chemical calculations, Chemma can comprehend and extract chemical insights from reaction data, in a manner akin to human experts. This work opens avenues for accelerating organic chemistry synthesis with adapted large language models. Large language models (LLMs) can be useful tools for science, but they often lack expert understanding of complex domains that they were not trained on. Zhang and colleagues fine-tuned a LLaMA-2-7b-based LLM with questions on organic chemistry reactions.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 7","pages":"1010-1022"},"PeriodicalIF":23.9,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144520683","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":"Model-based reinforcement learning for ultrasound-driven autonomous microrobots","authors":"Mahmoud Medany, Lorenzo Piglia, Liam Achenbach, S. Karthik Mukkavilli, Daniel Ahmed","doi":"10.1038/s42256-025-01054-2","DOIUrl":"10.1038/s42256-025-01054-2","url":null,"abstract":"Reinforcement learning is emerging as a powerful tool for microrobots control, as it enables autonomous navigation in environments where classical control approaches fall short. However, applying reinforcement learning to microrobotics is difficult due to the need for large training datasets, the slow convergence in physical systems and poor generalizability across environments. These challenges are amplified in ultrasound-actuated microrobots, which require rapid, precise adjustments in high-dimensional action space, which are often too complex for human operators. Addressing these challenges requires sample-efficient algorithms that adapt from limited data while managing complex physical interactions. To meet these challenges, we implemented model-based reinforcement learning for autonomous control of an ultrasound-driven microrobot, which learns from recurrent imagined environments. Our non-invasive, AI-controlled microrobot offers precise propulsion and efficiently learns from images in data-scarce environments. On transitioning from a pretrained simulation environment, we achieved sample-efficient collision avoidance and channel navigation, reaching a 90% success rate in target navigation across various channels within an hour of fine-tuning. Moreover, our model initially generalized successfully in 50% of tasks in new environments, improving to over 90% with 30 min of further training. We further demonstrated real-time manipulation of microrobots in complex vasculatures under both static and flow conditions, thus underscoring the potential of AI to revolutionize microrobotics in biomedical applications. Medany et al. present AI-driven microrobots that use ultrasound propulsion to learn how to navigate complex environments. These microrobots achieve 90% success after minimal training and adapt rapidly, showing promise for biomedical applications.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 7","pages":"1076-1090"},"PeriodicalIF":23.9,"publicationDate":"2025-06-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01054-2.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144488385","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":"Universal interatomic potentials shine in finding crystal structures","authors":"Ju Li","doi":"10.1038/s42256-025-01061-3","DOIUrl":"10.1038/s42256-025-01061-3","url":null,"abstract":"Various machine learning models have been developed in recent years for the discovery of crystal structures. Matbench Discovery, a new benchmark, offers an efficient way to identify the most promising architectures.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 7","pages":"985-986"},"PeriodicalIF":23.9,"publicationDate":"2025-06-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144488432","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":"End-to-end cryo-EM complex structure determination with high accuracy and ultra-fast speed","authors":"Jue Wang, Cheng Tan, Zhangyang Gao, Guijun Zhang, Yang Zhang, Stan Z. Li","doi":"10.1038/s42256-025-01056-0","DOIUrl":"10.1038/s42256-025-01056-0","url":null,"abstract":"While cryogenic-electron microscopy yields high-resolution density maps for complex structures, accurate determination of the corresponding atomic structures still necessitates significant expertise and labour-intensive manual interpretation. Recently, artificial intelligence-based methods have emerged to streamline this process; however, several challenges persist. First, existing methods typically require multi-stage training and inference, causing inefficiencies and inconsistency. Second, these approaches often encounter bias and incur substantial computational costs in aligning predicted atomic coordinates with sequence. Last, due to the limitations of available datasets, previous studies struggle to generalize effectively to complicated and unseen test data. Here, in response to these challenges, we introduce end-to-end and efficient CryoFold (E3-CryoFold), a deep learning method that enables end-to-end training and one-shot inference. E3-CryoFold uses three-dimensional and sequence transformers to extract features from density maps and sequences, using cross-attention modules to integrate the two modalities. Additionally, it uses an SE(3) graph neural network to construct atomic structures based on extracted features. E3-CryoFold incorporates a pretraining stage, during which models are trained on simulated density maps derived from Protein Data Bank structures. Empirical results demonstrate that E3-CryoFold improves the average template modelling score of the generated structures by 400% as compared to Cryo2Struct and significantly outperforms ModelAngelo, while achieving this huge improvement using merely one-thousandth of the inference time required by these methods. Thus, E3-CryoFold represents a robust, streamlined and cohesive framework for cryogenic-electron microscopy structure determination. Wang et al. present E3-CryoFold, a deep learning method for cryo-EM structure determination that enables end-to-end training and one-shot inference This method reduces inference times while boosting template modelling scores against comparable methods.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 7","pages":"1091-1103"},"PeriodicalIF":23.9,"publicationDate":"2025-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144371011","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":"Dimensions underlying the representational alignment of deep neural networks with humans","authors":"Florian P. Mahner, Lukas Muttenthaler, Umut Güçlü, Martin N. Hebart","doi":"10.1038/s42256-025-01041-7","DOIUrl":"10.1038/s42256-025-01041-7","url":null,"abstract":"Determining the similarities and differences between humans and artificial intelligence (AI) is an important goal in both computational cognitive neuroscience and machine learning, promising a deeper understanding of human cognition and safer, more reliable AI systems. Much previous work comparing representations in humans and AI has relied on global, scalar measures to quantify their alignment. However, without explicit hypotheses, these measures only inform us about the degree of alignment, not the factors that determine it. To address this challenge, we propose a generic framework to compare human and AI representations, based on identifying latent representational dimensions underlying the same behaviour in both domains. Applying this framework to humans and a deep neural network (DNN) model of natural images revealed a low-dimensional DNN embedding of both visual and semantic dimensions. In contrast to humans, DNNs exhibited a clear dominance of visual over semantic properties, indicating divergent strategies for representing images. Although in silico experiments showed seemingly consistent interpretability of DNN dimensions, a direct comparison between human and DNN representations revealed substantial differences in how they process images. By making representations directly comparable, our results reveal important challenges for representational alignment and offer a means for improving their comparability. An interpretability framework that compares how humans and deep neural networks process images has been presented. Their findings reveal that, unlike humans, deep neural networks focus more on visual properties than semantic ones, highlighting divergent representational strategies.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 6","pages":"848-859"},"PeriodicalIF":23.9,"publicationDate":"2025-06-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01041-7.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144340984","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":"Robust virtual staining of landmark organelles with Cytoland","authors":"Ziwen Liu, Eduardo Hirata-Miyasaki, Soorya Pradeep, Johanna V. Rahm, Christian Foley, Talon Chandler, Ivan E. Ivanov, Hunter O. Woosley, See-Chi Lee, Sudip Khadka, Tiger Lao, Akilandeswari Balasubramanian, Rita Marreiros, Chad Liu, Camille Januel, Manuel D. Leonetti, Ranen Aviner, Carolina Arias, Adrian Jacobo, Shalin B. Mehta","doi":"10.1038/s42256-025-01046-2","DOIUrl":"10.1038/s42256-025-01046-2","url":null,"abstract":"Correlative live-cell imaging of landmark organelles—such as nuclei, nucleoli, cell membranes, nuclear envelope and lipid droplets—is critical for systems cell biology and drug discovery. However, achieving this with molecular labels alone remains challenging. Virtual staining of multiple organelles and cell states from label-free images with deep neural networks is an emerging solution. Virtual staining frees the light spectrum for imaging molecular sensors, photomanipulation or other tasks. Current methods for virtual staining of landmark organelles often fail in the presence of nuisance variations in imaging, culture conditions and cell types. Here we address this with Cytoland, a collection of models for robust virtual staining of landmark organelles across diverse imaging parameters, cell states and types. These models were trained with self-supervised and supervised pre-training using a flexible convolutional architecture (UNeXt2) and augmentations inspired by image formation of light microscopes. Cytoland models enable virtual staining of nuclei and membranes across multiple cell types—including human cell lines, zebrafish neuromasts, induced pluripotent stem cells (iPSCs) and iPSC-derived neurons—under a range of imaging conditions. We assess models using intensity, segmentation and application-specific measurements obtained from virtually and experimentally stained nuclei and membranes. These models rescue missing labels, correct non-uniform labelling and mitigate photobleaching. We share multiple pre-trained models, open-source software (VisCy) for training, inference and deployment, and the datasets. Ziwen Liu et al. report Cytoland, an approach to train robust models to virtually stain landmark organelles of cells and address the generalization gap of current models. The training pipeline, models and datasets are shared under open-source permissive licences.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 6","pages":"901-915"},"PeriodicalIF":23.9,"publicationDate":"2025-06-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01046-2.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144340983","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}