Pub Date : 2025-12-31DOI: 10.1038/s42256-025-01161-0
Xiao Xiao, Le Zhang, Hongyu Zhao, Zuoheng Wang
Cell–cell interactions (CCI), driven by distance-dependent signalling, are important for tissue development and organ function. While imaging-based spatial transcriptomics offers unprecedented opportunities to unravel CCI at single-cell resolution, current analyses face challenges such as limited ligand–receptor pairs measured, insufficient spatial encoding and low interpretability. We present GITIII (graph inductive bias transformer for intercellular interaction investigation), a lightweight, interpretable, self-supervised graph transformer-based model that conceptualizes cells as words and their surrounding cellular neighbourhood as context that shapes the meaning or state of the central cell. GITIII infers CCI by examining the correlation between a cell’s state and its niche, enabling us to understand how sender cells influence the gene expression of receiver cells, visualize spatial CCI patterns, perform CCI-informed cell clustering and construct CCI networks. Applied to four spatial transcriptomics datasets across multiple species, organs and platforms, GITIII effectively identified and statistically interpreted CCI patterns in the brain and tumour microenvironments. Xiao et al. present GITIII, a lightweight and interpretable graph transformer for inferring spatial single-cell-level interactions and quantifying the influence of neighbouring cells on the gene expression of receiver cells in spatial transcriptomics.
{"title":"Inferring spatial single-cell-level interactions through interpreting cell state and niche correlations learned by self-supervised graph transformer","authors":"Xiao Xiao, Le Zhang, Hongyu Zhao, Zuoheng Wang","doi":"10.1038/s42256-025-01161-0","DOIUrl":"10.1038/s42256-025-01161-0","url":null,"abstract":"Cell–cell interactions (CCI), driven by distance-dependent signalling, are important for tissue development and organ function. While imaging-based spatial transcriptomics offers unprecedented opportunities to unravel CCI at single-cell resolution, current analyses face challenges such as limited ligand–receptor pairs measured, insufficient spatial encoding and low interpretability. We present GITIII (graph inductive bias transformer for intercellular interaction investigation), a lightweight, interpretable, self-supervised graph transformer-based model that conceptualizes cells as words and their surrounding cellular neighbourhood as context that shapes the meaning or state of the central cell. GITIII infers CCI by examining the correlation between a cell’s state and its niche, enabling us to understand how sender cells influence the gene expression of receiver cells, visualize spatial CCI patterns, perform CCI-informed cell clustering and construct CCI networks. Applied to four spatial transcriptomics datasets across multiple species, organs and platforms, GITIII effectively identified and statistically interpreted CCI patterns in the brain and tumour microenvironments. Xiao et al. present GITIII, a lightweight and interpretable graph transformer for inferring spatial single-cell-level interactions and quantifying the influence of neighbouring cells on the gene expression of receiver cells in spatial transcriptomics.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 1","pages":"42-58"},"PeriodicalIF":23.9,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145894631","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}
Pub Date : 2025-12-31DOI: 10.1038/s42256-025-01160-1
Alex Morehead, Nabin Giri, Jian Liu, Pawan Neupane, Jianlin Cheng
The effects of ligand binding on protein structures and their in vivo functions carry numerous implications for modern biomedical research and biotechnology development efforts such as drug discovery. Although several deep learning (DL) methods and benchmarks designed for protein–ligand docking have recently been introduced, so far no previous works have systematically studied the behaviour of the latest docking and structure prediction methods within the broadly applicable context of: (1) using predicted (apo) protein structures for docking (for example, for applicability to new proteins); (2) binding multiple (cofactor) ligands concurrently to a given target protein (for example, for enzyme design); and (3) having no previous knowledge of binding pockets (for example, for generalization to unknown pockets). To enable a deeper understanding of the real-world utility of docking methods, we introduce PoseBench, a comprehensive benchmark for broadly applicable protein–ligand docking. PoseBench enables researchers to rigorously and systematically evaluate DL methods for apo-to-holo protein–ligand docking and protein–ligand structure prediction using both primary ligand and multiligand benchmark datasets, the latter of which we introduce to the DL community. Empirically, using PoseBench, we find that: (1) DL cofolding methods generally outperform comparable conventional and DL docking baseline algorithms, but popular methods such as AlphaFold 3 are still challenged by prediction targets with new protein–ligand binding poses; (2) certain DL cofolding methods are highly sensitive to their input multiple sequence alignments, whereas others are not; and (3) DL methods struggle to strike a balance between structural accuracy and chemical specificity when predicting new or multiligand protein targets. Morehead et al. introduce the benchmark PoseBench and evaluate the strengths and limitations of current AI-based protein–ligand docking and structure prediction methods.
{"title":"Assessing the potential of deep learning for protein–ligand docking","authors":"Alex Morehead, Nabin Giri, Jian Liu, Pawan Neupane, Jianlin Cheng","doi":"10.1038/s42256-025-01160-1","DOIUrl":"10.1038/s42256-025-01160-1","url":null,"abstract":"The effects of ligand binding on protein structures and their in vivo functions carry numerous implications for modern biomedical research and biotechnology development efforts such as drug discovery. Although several deep learning (DL) methods and benchmarks designed for protein–ligand docking have recently been introduced, so far no previous works have systematically studied the behaviour of the latest docking and structure prediction methods within the broadly applicable context of: (1) using predicted (apo) protein structures for docking (for example, for applicability to new proteins); (2) binding multiple (cofactor) ligands concurrently to a given target protein (for example, for enzyme design); and (3) having no previous knowledge of binding pockets (for example, for generalization to unknown pockets). To enable a deeper understanding of the real-world utility of docking methods, we introduce PoseBench, a comprehensive benchmark for broadly applicable protein–ligand docking. PoseBench enables researchers to rigorously and systematically evaluate DL methods for apo-to-holo protein–ligand docking and protein–ligand structure prediction using both primary ligand and multiligand benchmark datasets, the latter of which we introduce to the DL community. Empirically, using PoseBench, we find that: (1) DL cofolding methods generally outperform comparable conventional and DL docking baseline algorithms, but popular methods such as AlphaFold 3 are still challenged by prediction targets with new protein–ligand binding poses; (2) certain DL cofolding methods are highly sensitive to their input multiple sequence alignments, whereas others are not; and (3) DL methods struggle to strike a balance between structural accuracy and chemical specificity when predicting new or multiligand protein targets. Morehead et al. introduce the benchmark PoseBench and evaluate the strengths and limitations of current AI-based protein–ligand docking and structure prediction methods.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 1","pages":"32-41"},"PeriodicalIF":23.9,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01160-1.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145894622","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}
Pub Date : 2025-12-31DOI: 10.1038/s42256-025-01167-8
Chun-Teh Chen
Deep generative models that learn intermediate surface-current maps, rather than layouts directly, offer a more stable route to inverse design of tunable and stacked metasurfaces.
深度生成模型学习中间表面电流图，而不是直接布局，为可调和堆叠元表面的反向设计提供了更稳定的途径。
{"title":"Learning intermediate physical states for inverse metasurface design","authors":"Chun-Teh Chen","doi":"10.1038/s42256-025-01167-8","DOIUrl":"10.1038/s42256-025-01167-8","url":null,"abstract":"Deep generative models that learn intermediate surface-current maps, rather than layouts directly, offer a more stable route to inverse design of tunable and stacked metasurfaces.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 1","pages":"2-3"},"PeriodicalIF":23.9,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145895501","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}
Pub Date : 2025-12-31DOI: 10.1038/s42256-025-01166-9
Remmy Zen, Maximilian Nägele, Florian Marquardt
Quantum computing has the potential to solve problems that are intractable for classical computers, with possible applications in areas such as drug discovery and high-energy physics. However, the practical implementation of quantum computation is hindered by the complexity of executing quantum circuits on hardware. In particular, minimizing the number of T gates is crucial for implementing efficient quantum algorithms. AlphaTensor-Quantum1 is a reinforcement-learning-based method designed to optimize the T count of quantum circuits by formulating the problem as a tensor decomposition task. Although it has demonstrated superior performance over existing methods on benchmark quantum arithmetic circuits, its applicability has so far been restricted to specific circuit families, requiring separate, time-intensive training for each new application. This report reproduces some of the key results of the original work and extends AlphaTensor-Quantum’s capabilities to simplify random quantum circuits with varying qubit counts, eliminating the need for retraining on new circuits. Our experiments show that a general agent trained on five- to eight-qubit circuits achieves greater T-count reduction than previous methods for a large fraction of quantum circuits. Furthermore, we demonstrate that a general agent trained on circuits with varying qubit numbers outperforms agents trained on fixed qubit numbers, highlighting the method’s generalizability and its potential for broader quantum circuit optimization tasks. The reusability of AlphaTensor-Quantum is tested and the method is extended to optimize a broad range of quantum circuits without retraining, achieving greater T-count reductions and demonstrating generalizable and efficient quantum circuit optimization.
{"title":"Reusability report: Optimizing T count in general quantum circuits with AlphaTensor-Quantum","authors":"Remmy Zen, Maximilian Nägele, Florian Marquardt","doi":"10.1038/s42256-025-01166-9","DOIUrl":"10.1038/s42256-025-01166-9","url":null,"abstract":"Quantum computing has the potential to solve problems that are intractable for classical computers, with possible applications in areas such as drug discovery and high-energy physics. However, the practical implementation of quantum computation is hindered by the complexity of executing quantum circuits on hardware. In particular, minimizing the number of T gates is crucial for implementing efficient quantum algorithms. AlphaTensor-Quantum1 is a reinforcement-learning-based method designed to optimize the T count of quantum circuits by formulating the problem as a tensor decomposition task. Although it has demonstrated superior performance over existing methods on benchmark quantum arithmetic circuits, its applicability has so far been restricted to specific circuit families, requiring separate, time-intensive training for each new application. This report reproduces some of the key results of the original work and extends AlphaTensor-Quantum’s capabilities to simplify random quantum circuits with varying qubit counts, eliminating the need for retraining on new circuits. Our experiments show that a general agent trained on five- to eight-qubit circuits achieves greater T-count reduction than previous methods for a large fraction of quantum circuits. Furthermore, we demonstrate that a general agent trained on circuits with varying qubit numbers outperforms agents trained on fixed qubit numbers, highlighting the method’s generalizability and its potential for broader quantum circuit optimization tasks. The reusability of AlphaTensor-Quantum is tested and the method is extended to optimize a broad range of quantum circuits without retraining, achieving greater T-count reductions and demonstrating generalizable and efficient quantum circuit optimization.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 1","pages":"113-117"},"PeriodicalIF":23.9,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01166-9.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145894628","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}
Pub Date : 2025-12-31DOI: 10.1038/s42256-025-01170-z
Fei He, Ruixin Fei, Jordan E. Krull, Yang Yu, Xinyu Zhang, Xianyu Wang, Hao Cheng, Mingyue Gao, Li Su, Yibo Chen, Jinpu Li, Baichuan Jin, Yuzhou Chang, Anjun Ma, Qin Ma, Dong Xu
Single-cell large language models (scLLMs) capture essential biological insights from vast single-cell atlases but struggle in out-of-context applications, where zero-shot predictions can be unreliable. To address this, here we introduce a single-cell parameter-efficient fine-tuning (scPEFT) framework that integrates learnable, low-dimensional adapters into scLLMs. By freezing the backbone model and updating only the adapter parameters, scPEFT efficiently adapts to specific tasks using limited custom data. This approach mitigates catastrophic forgetting, reduces parameter tuning by over 96% and decreases GPU memory usage by more than half, thus substantially enhancing the accessibility of scLLMs for resource-constrained researchers. When validated across diverse datasets, scPEFT outperformed zero-shot models and traditional fine-tuning in disease-specific, cross-species and undercharacterized cell population tasks. Its attention-mechanism analysis identified COVID-related genes associated with specific cell states and uncovered unique blood cell subpopulations, demonstrating the capacity of scPEFT for condition-specific interpretations. These findings position scPEFT as an efficient solution for enhancing the utility of scLLMs in general single-cell analyses. He et al. present a parameter-efficient fine-tuning method for single-cell language models that improves performance on unseen diseases, treatments and cell types.
{"title":"Harnessing the power of single-cell large language models with parameter-efficient fine-tuning using scPEFT","authors":"Fei He, Ruixin Fei, Jordan E. Krull, Yang Yu, Xinyu Zhang, Xianyu Wang, Hao Cheng, Mingyue Gao, Li Su, Yibo Chen, Jinpu Li, Baichuan Jin, Yuzhou Chang, Anjun Ma, Qin Ma, Dong Xu","doi":"10.1038/s42256-025-01170-z","DOIUrl":"10.1038/s42256-025-01170-z","url":null,"abstract":"Single-cell large language models (scLLMs) capture essential biological insights from vast single-cell atlases but struggle in out-of-context applications, where zero-shot predictions can be unreliable. To address this, here we introduce a single-cell parameter-efficient fine-tuning (scPEFT) framework that integrates learnable, low-dimensional adapters into scLLMs. By freezing the backbone model and updating only the adapter parameters, scPEFT efficiently adapts to specific tasks using limited custom data. This approach mitigates catastrophic forgetting, reduces parameter tuning by over 96% and decreases GPU memory usage by more than half, thus substantially enhancing the accessibility of scLLMs for resource-constrained researchers. When validated across diverse datasets, scPEFT outperformed zero-shot models and traditional fine-tuning in disease-specific, cross-species and undercharacterized cell population tasks. Its attention-mechanism analysis identified COVID-related genes associated with specific cell states and uncovered unique blood cell subpopulations, demonstrating the capacity of scPEFT for condition-specific interpretations. These findings position scPEFT as an efficient solution for enhancing the utility of scLLMs in general single-cell analyses. He et al. present a parameter-efficient fine-tuning method for single-cell language models that improves performance on unseen diseases, treatments and cell types.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 1","pages":"118-133"},"PeriodicalIF":23.9,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145894629","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}
Pub Date : 2025-12-31DOI: 10.1038/s42256-025-01163-y
Kevin Bijan Givechian, João Felipe Rocha, Chen Liu, Edward Yang, Sidharth Tyagi, Kerrie Greene, Rex Ying, Etienne Caron, Akiko Iwasaki, Smita Krishnaswamy
Epitope-based vaccines are promising therapeutic modalities for infectious diseases and cancer, but identifying immunogenic epitopes is challenging. Most prediction methods only use amino acid sequence information, and do not incorporate wide-scale structure data and biochemical properties across each peptide–major histocompatibility complex (MHC). We present ImmunoStruct, a deep learning model that integrates sequence, structural and biochemical information to predict multi-allele class I peptide–MHC immunogenicity. By leveraging a multimodal dataset of 26,049 peptide–MHCs, we demonstrate that ImmunoStruct improves immunogenicity prediction performance and interpretability beyond existing methods, across infectious disease epitopes and cancer neoepitopes. We further show strong alignment with in vitro assay results for a set of SARS-CoV-2 epitopes, as well as strong performance in peptide–MHC-based survival prediction for patients with cancer. Overall, this work also presents an architecture that incorporates equivariant graph processing and multimodal data integration for a long-standing challenge in immunotherapy. A multimodal deep learning model combines molecular sequence, structure and biochemical properties to predict immunogenicity in an interpretable way, providing a framework for smarter molecular prediction and hypothesis generation.
{"title":"ImmunoStruct enables multimodal deep learning for immunogenicity prediction","authors":"Kevin Bijan Givechian, João Felipe Rocha, Chen Liu, Edward Yang, Sidharth Tyagi, Kerrie Greene, Rex Ying, Etienne Caron, Akiko Iwasaki, Smita Krishnaswamy","doi":"10.1038/s42256-025-01163-y","DOIUrl":"10.1038/s42256-025-01163-y","url":null,"abstract":"Epitope-based vaccines are promising therapeutic modalities for infectious diseases and cancer, but identifying immunogenic epitopes is challenging. Most prediction methods only use amino acid sequence information, and do not incorporate wide-scale structure data and biochemical properties across each peptide–major histocompatibility complex (MHC). We present ImmunoStruct, a deep learning model that integrates sequence, structural and biochemical information to predict multi-allele class I peptide–MHC immunogenicity. By leveraging a multimodal dataset of 26,049 peptide–MHCs, we demonstrate that ImmunoStruct improves immunogenicity prediction performance and interpretability beyond existing methods, across infectious disease epitopes and cancer neoepitopes. We further show strong alignment with in vitro assay results for a set of SARS-CoV-2 epitopes, as well as strong performance in peptide–MHC-based survival prediction for patients with cancer. Overall, this work also presents an architecture that incorporates equivariant graph processing and multimodal data integration for a long-standing challenge in immunotherapy. A multimodal deep learning model combines molecular sequence, structure and biochemical properties to predict immunogenicity in an interpretable way, providing a framework for smarter molecular prediction and hypothesis generation.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 1","pages":"70-83"},"PeriodicalIF":23.9,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145895500","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":"Author Correction: Scalable and robust DNA-based storage via coding theory and deep learning","authors":"Daniella Bar-Lev, Itai Orr, Omer Sabary, Tuvi Etzion, Eitan Yaakobi","doi":"10.1038/s42256-025-01175-8","DOIUrl":"10.1038/s42256-025-01175-8","url":null,"abstract":"","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 1","pages":"134-134"},"PeriodicalIF":23.9,"publicationDate":"2025-12-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01175-8.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145894634","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}
Pub Date : 2025-12-29DOI: 10.1038/s42256-025-01147-y
Zehua Chen, Yuyang Miao, Liyuan Wang, Luyun Fan, Danilo P. Mandic, Jun Zhu
Cardiovascular signals such as photoplethysmography, electrocardiography and blood pressure are inherently correlated and complementary, together reflecting the health of the cardiovascular system. However, their joint utilization in real-time monitoring is severely limited by diverse acquisition challenges from noisy wearable recordings to burdened invasive procedures. Here we propose UniCardio, a multimodal diffusion transformer that reconstructs low-quality signals and synthesizes unrecorded signals in a unified generative framework. Its key innovations include a specialized model architecture to manage the signal modalities involved in generation tasks and a continual learning paradigm to incorporate varying modality combinations. By exploiting the complementary nature of cardiovascular signals, UniCardio clearly outperforms recent task-specific baselines in signal denoising, imputation and translation. The generated signals match the performance of ground-truth signals in detecting abnormal health conditions and estimating vital signs, even in unseen domains, as well as ensuring interpretability for human experts. These advantages establish UniCardio as a practical and robust framework for advancing artificial-intelligence-assisted healthcare. UniCardio is a unified framework for versatile multimodal cardiovascular signal generation, enabling robust signal restoration and cross-modal translation to detect abnormal conditions and estimate vital signs in real-time health monitoring.
{"title":"Versatile cardiovascular signal generation with a unified diffusion transformer","authors":"Zehua Chen, Yuyang Miao, Liyuan Wang, Luyun Fan, Danilo P. Mandic, Jun Zhu","doi":"10.1038/s42256-025-01147-y","DOIUrl":"10.1038/s42256-025-01147-y","url":null,"abstract":"Cardiovascular signals such as photoplethysmography, electrocardiography and blood pressure are inherently correlated and complementary, together reflecting the health of the cardiovascular system. However, their joint utilization in real-time monitoring is severely limited by diverse acquisition challenges from noisy wearable recordings to burdened invasive procedures. Here we propose UniCardio, a multimodal diffusion transformer that reconstructs low-quality signals and synthesizes unrecorded signals in a unified generative framework. Its key innovations include a specialized model architecture to manage the signal modalities involved in generation tasks and a continual learning paradigm to incorporate varying modality combinations. By exploiting the complementary nature of cardiovascular signals, UniCardio clearly outperforms recent task-specific baselines in signal denoising, imputation and translation. The generated signals match the performance of ground-truth signals in detecting abnormal health conditions and estimating vital signs, even in unseen domains, as well as ensuring interpretability for human experts. These advantages establish UniCardio as a practical and robust framework for advancing artificial-intelligence-assisted healthcare. UniCardio is a unified framework for versatile multimodal cardiovascular signal generation, enabling robust signal restoration and cross-modal translation to detect abnormal conditions and estimate vital signs in real-time health monitoring.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"8 1","pages":"6-19"},"PeriodicalIF":23.9,"publicationDate":"2025-12-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145895542","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}
Pub Date : 2025-12-18DOI: 10.1038/s42256-025-01150-3
Michael E. Vinyard, Anders W. Rasmussen, Ruitong Li, Allon M. Klein, Gad Getz, Luca Pinello
Single-cell sequencing measurements facilitate the reconstruction of dynamic biology by capturing snapshot molecular profiles of individual cells. Cell fate decisions in development and disease are orchestrated through an intricate balance of deterministic and stochastic regulatory events. Drift-diffusion equations are effective in modelling single-cell dynamics from high-dimensional single-cell measurements. While existing solutions describe the deterministic dynamics associated with the drift term of these equations at the level of cell state, diffusion is modelled as a constant across cell states. To fully understand the dynamic regulatory logic in development and disease, models explicitly attuned to the balance between deterministic and stochastic biology are required. To address these limitations, we introduce scDiffEq, a generative framework for learning neural stochastic differential equations that approximate biology’s deterministic and stochastic dynamics. Using lineage-traced single-cell data, we demonstrate that scDiffEq offers an improved reconstruction of cell trajectories and prediction of cell fate from multipotent progenitors during haematopoiesis. By imparting in silico perturbations to multipotent progenitor cells, we find that scDiffEq accurately recapitulates the dynamics of CRISPR-perturbed haematopoiesis. We generalize this approach beyond lineage-traced or multi-time-point datasets to model the dynamics of single-cell data from a single time point. Using scDiffEq, we simulate high-resolution developmental cell trajectories, which can model their drift and diffusion, enabling us to study their time-dependent gene-level dynamics. Vinyard et al. present a generative method to model cell dynamics using neural stochastic differential equations that learn state-dependent drift and diffusion, outperforming existing approaches and enabling perturbation studies of development and disease.
{"title":"Learning cell dynamics with neural differential equations","authors":"Michael E. Vinyard, Anders W. Rasmussen, Ruitong Li, Allon M. Klein, Gad Getz, Luca Pinello","doi":"10.1038/s42256-025-01150-3","DOIUrl":"10.1038/s42256-025-01150-3","url":null,"abstract":"Single-cell sequencing measurements facilitate the reconstruction of dynamic biology by capturing snapshot molecular profiles of individual cells. Cell fate decisions in development and disease are orchestrated through an intricate balance of deterministic and stochastic regulatory events. Drift-diffusion equations are effective in modelling single-cell dynamics from high-dimensional single-cell measurements. While existing solutions describe the deterministic dynamics associated with the drift term of these equations at the level of cell state, diffusion is modelled as a constant across cell states. To fully understand the dynamic regulatory logic in development and disease, models explicitly attuned to the balance between deterministic and stochastic biology are required. To address these limitations, we introduce scDiffEq, a generative framework for learning neural stochastic differential equations that approximate biology’s deterministic and stochastic dynamics. Using lineage-traced single-cell data, we demonstrate that scDiffEq offers an improved reconstruction of cell trajectories and prediction of cell fate from multipotent progenitors during haematopoiesis. By imparting in silico perturbations to multipotent progenitor cells, we find that scDiffEq accurately recapitulates the dynamics of CRISPR-perturbed haematopoiesis. We generalize this approach beyond lineage-traced or multi-time-point datasets to model the dynamics of single-cell data from a single time point. Using scDiffEq, we simulate high-resolution developmental cell trajectories, which can model their drift and diffusion, enabling us to study their time-dependent gene-level dynamics. Vinyard et al. present a generative method to model cell dynamics using neural stochastic differential equations that learn state-dependent drift and diffusion, outperforming existing approaches and enabling perturbation studies of development and disease.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 12","pages":"1969-1984"},"PeriodicalIF":23.9,"publicationDate":"2025-12-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145770795","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}
Pub Date : 2025-12-18DOI: 10.1038/s42256-025-01115-6
Gregory Serapio-García, Mustafa Safdari, Clément Crepy, Luning Sun, Stephen Fitz, Peter Romero, Marwa Abdulhai, Aleksandra Faust, Maja Matarić
The advent of large language models (LLMs) has revolutionized natural language processing, enabling the generation of coherent and contextually relevant human-like text. As LLMs increasingly power conversational agents used by the general public worldwide, the synthetic personality traits embedded in these models by virtue of training on large amounts of human data are becoming increasingly important to evaluate. The style in which LLMs respond can mimic different human personality traits. Here, as these patterns can be a key factor determining the effectiveness of communication, we present a comprehensive psychometric methodology for administering and validating personality tests on widely used LLMs, as well as for shaping personality in the generated text of such LLMs. Applying this method to 18 LLMs, we found that: personality measurements in the outputs of some LLMs under specific prompting configurations are reliable and valid; evidence of reliability and validity of synthetic LLM personality is stronger for larger and instruction-fine-tuned models; and personality in LLM outputs can be shaped along desired dimensions to mimic specific human personality profiles. We discuss the application and ethical implications of the measurement and shaping method, in particular regarding responsible artificial intelligence. Serapio-García, Safdari and colleagues develop a method based on psychometric tests to measure and validate personality-like traits in LLMs. Large, instruction-tuned models give reliable personality measurement results, and specific personality profiles can be mimicked in downstream tasks.
{"title":"A psychometric framework for evaluating and shaping personality traits in large language models","authors":"Gregory Serapio-García, Mustafa Safdari, Clément Crepy, Luning Sun, Stephen Fitz, Peter Romero, Marwa Abdulhai, Aleksandra Faust, Maja Matarić","doi":"10.1038/s42256-025-01115-6","DOIUrl":"10.1038/s42256-025-01115-6","url":null,"abstract":"The advent of large language models (LLMs) has revolutionized natural language processing, enabling the generation of coherent and contextually relevant human-like text. As LLMs increasingly power conversational agents used by the general public worldwide, the synthetic personality traits embedded in these models by virtue of training on large amounts of human data are becoming increasingly important to evaluate. The style in which LLMs respond can mimic different human personality traits. Here, as these patterns can be a key factor determining the effectiveness of communication, we present a comprehensive psychometric methodology for administering and validating personality tests on widely used LLMs, as well as for shaping personality in the generated text of such LLMs. Applying this method to 18 LLMs, we found that: personality measurements in the outputs of some LLMs under specific prompting configurations are reliable and valid; evidence of reliability and validity of synthetic LLM personality is stronger for larger and instruction-fine-tuned models; and personality in LLM outputs can be shaped along desired dimensions to mimic specific human personality profiles. We discuss the application and ethical implications of the measurement and shaping method, in particular regarding responsible artificial intelligence. Serapio-García, Safdari and colleagues develop a method based on psychometric tests to measure and validate personality-like traits in LLMs. Large, instruction-tuned models give reliable personality measurement results, and specific personality profiles can be mimicked in downstream tasks.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":"7 12","pages":"1954-1968"},"PeriodicalIF":23.9,"publicationDate":"2025-12-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s42256-025-01115-6.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145770796","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}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：