Pub Date : 2024-06-27DOI: 10.1038/s42256-024-00848-0
Alpha Renner, Lazar Supic, Andreea Danielescu, Giacomo Indiveri, Bruno A. Olshausen, Yulia Sandamirskaya, Friedrich T. Sommer, E. Paxon Frady
Analysing a visual scene by inferring the configuration of a generative model is widely considered the most flexible and generalizable approach to scene understanding. Yet, one major problem is the computational challenge of the inference procedure, involving a combinatorial search across object identities and poses. Here we propose a neuromorphic solution exploiting three key concepts: (1) a computational framework based on vector symbolic architectures (VSAs) with complex-valued vectors, (2) the design of hierarchical resonator networks to factorize the non-commutative transforms translation and rotation in visual scenes and (3) the design of a multi-compartment spiking phasor neuron model for implementing complex-valued resonator networks on neuromorphic hardware. The VSA framework uses vector binding operations to form a generative image model in which binding acts as the equivariant operation for geometric transformations. A scene can therefore be described as a sum of vector products, which can then be efficiently factorized by a resonator network to infer objects and their poses. The hierarchical resonator network features a partitioned architecture in which vector binding is equivariant for horizontal and vertical translation within one partition and for rotation and scaling within the other partition. The spiking neuron model allows mapping the resonator network onto efficient and low-power neuromorphic hardware. Our approach is demonstrated on synthetic scenes composed of simple two-dimensional shapes undergoing rigid geometric transformations and colour changes. A companion paper demonstrates the same approach in real-world application scenarios for machine vision and robotics. The inference procedure for analysing a visual scene presents a computational challenge. Renner, Supic and colleagues develop a neural network model, the hierarchical resonator, to determine the generative factors of variation of objects in simple scenes. The resonator was implemented on neuromorphic hardware, using a spike-timing code for complex numbers.
{"title":"Neuromorphic visual scene understanding with resonator networks","authors":"Alpha Renner, Lazar Supic, Andreea Danielescu, Giacomo Indiveri, Bruno A. Olshausen, Yulia Sandamirskaya, Friedrich T. Sommer, E. Paxon Frady","doi":"10.1038/s42256-024-00848-0","DOIUrl":"10.1038/s42256-024-00848-0","url":null,"abstract":"Analysing a visual scene by inferring the configuration of a generative model is widely considered the most flexible and generalizable approach to scene understanding. Yet, one major problem is the computational challenge of the inference procedure, involving a combinatorial search across object identities and poses. Here we propose a neuromorphic solution exploiting three key concepts: (1) a computational framework based on vector symbolic architectures (VSAs) with complex-valued vectors, (2) the design of hierarchical resonator networks to factorize the non-commutative transforms translation and rotation in visual scenes and (3) the design of a multi-compartment spiking phasor neuron model for implementing complex-valued resonator networks on neuromorphic hardware. The VSA framework uses vector binding operations to form a generative image model in which binding acts as the equivariant operation for geometric transformations. A scene can therefore be described as a sum of vector products, which can then be efficiently factorized by a resonator network to infer objects and their poses. The hierarchical resonator network features a partitioned architecture in which vector binding is equivariant for horizontal and vertical translation within one partition and for rotation and scaling within the other partition. The spiking neuron model allows mapping the resonator network onto efficient and low-power neuromorphic hardware. Our approach is demonstrated on synthetic scenes composed of simple two-dimensional shapes undergoing rigid geometric transformations and colour changes. A companion paper demonstrates the same approach in real-world application scenarios for machine vision and robotics. The inference procedure for analysing a visual scene presents a computational challenge. Renner, Supic and colleagues develop a neural network model, the hierarchical resonator, to determine the generative factors of variation of objects in simple scenes. The resonator was implemented on neuromorphic hardware, using a spike-timing code for complex numbers.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141462588","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 : 2024-06-24DOI: 10.1038/s42256-024-00853-3
Iksung Kang, Qinrong Zhang, Stella X. Yu, Na Ji
Widefield microscopy is widely used for non-invasive imaging of biological structures at subcellular resolution. When applied to a complex specimen, its image quality is degraded by sample-induced optical aberration. Adaptive optics can correct wavefront distortion and restore diffraction-limited resolution but require wavefront sensing and corrective devices, increasing system complexity and cost. Here we describe a self-supervised machine learning algorithm, CoCoA, that performs joint wavefront estimation and three-dimensional structural information extraction from a single-input three-dimensional image stack without the need for external training datasets. We implemented CoCoA for widefield imaging of mouse brain tissues and validated its performance with direct-wavefront-sensing-based adaptive optics. Importantly, we systematically explored and quantitatively characterized the limiting factors of CoCoA’s performance. Using CoCoA, we demonstrated in vivo widefield mouse brain imaging using machine learning-based adaptive optics. Incorporating coordinate-based neural representations and a forward physics model, the self-supervised scheme of CoCoA should be applicable to microscopy modalities in general. Adaptive optics (AO) corrects aberrations and restores resolution but requires specialized hardware. Kang et al. introduce a self-supervised AO method (CoCoA) for widefield microscopy, achieving in vivo mouse brain imaging without wavefront sensors.
{"title":"Coordinate-based neural representations for computational adaptive optics in widefield microscopy","authors":"Iksung Kang, Qinrong Zhang, Stella X. Yu, Na Ji","doi":"10.1038/s42256-024-00853-3","DOIUrl":"10.1038/s42256-024-00853-3","url":null,"abstract":"Widefield microscopy is widely used for non-invasive imaging of biological structures at subcellular resolution. When applied to a complex specimen, its image quality is degraded by sample-induced optical aberration. Adaptive optics can correct wavefront distortion and restore diffraction-limited resolution but require wavefront sensing and corrective devices, increasing system complexity and cost. Here we describe a self-supervised machine learning algorithm, CoCoA, that performs joint wavefront estimation and three-dimensional structural information extraction from a single-input three-dimensional image stack without the need for external training datasets. We implemented CoCoA for widefield imaging of mouse brain tissues and validated its performance with direct-wavefront-sensing-based adaptive optics. Importantly, we systematically explored and quantitatively characterized the limiting factors of CoCoA’s performance. Using CoCoA, we demonstrated in vivo widefield mouse brain imaging using machine learning-based adaptive optics. Incorporating coordinate-based neural representations and a forward physics model, the self-supervised scheme of CoCoA should be applicable to microscopy modalities in general. Adaptive optics (AO) corrects aberrations and restores resolution but requires specialized hardware. Kang et al. introduce a self-supervised AO method (CoCoA) for widefield microscopy, achieving in vivo mouse brain imaging without wavefront sensors.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141448383","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 : 2024-06-24DOI: 10.1038/s42256-024-00844-4
Qianying Cao, Somdatta Goswami, George Em Karniadakis
Neural operators map multiple functions to different functions, possibly in different spaces, unlike standard neural networks. Hence, neural operators allow the solution of parametric ordinary differential equations (ODEs) and partial differential equations (PDEs) for a distribution of boundary or initial conditions and excitations, but can also be used for system identification as well as designing various components of digital twins. We introduce the Laplace neural operator (LNO), which incorporates the pole–residue relationship between input–output spaces, leading to better interpretability and generalization for certain classes of problems. The LNO is capable of processing non-periodic signals and transient responses resulting from simultaneously zero and non-zero initial conditions, which makes it achieve better approximation accuracy over other neural operators for extrapolation circumstances in solving several ODEs and PDEs. We also highlight the LNO’s good interpolation ability, from a low-resolution input to high-resolution outputs at arbitrary locations within the domain. To demonstrate the scalability of LNO, we conduct large-scale simulations of Rossby waves around the globe, employing millions of degrees of freedom. Taken together, our findings show that a pretrained LNO model offers an effective real-time solution for general ODEs and PDEs at scale and is an efficient alternative to existing neural operators. Neural operators are powerful neural networks that approximate nonlinear dynamical systems and their responses. Cao and colleagues introduce the Laplace neural operator, a scalable approach that can effectively deal with non-periodic signals and transient responses and can outperform existing neural operators on certain classes of ODE and PDE problems.
{"title":"Laplace neural operator for solving differential equations","authors":"Qianying Cao, Somdatta Goswami, George Em Karniadakis","doi":"10.1038/s42256-024-00844-4","DOIUrl":"10.1038/s42256-024-00844-4","url":null,"abstract":"Neural operators map multiple functions to different functions, possibly in different spaces, unlike standard neural networks. Hence, neural operators allow the solution of parametric ordinary differential equations (ODEs) and partial differential equations (PDEs) for a distribution of boundary or initial conditions and excitations, but can also be used for system identification as well as designing various components of digital twins. We introduce the Laplace neural operator (LNO), which incorporates the pole–residue relationship between input–output spaces, leading to better interpretability and generalization for certain classes of problems. The LNO is capable of processing non-periodic signals and transient responses resulting from simultaneously zero and non-zero initial conditions, which makes it achieve better approximation accuracy over other neural operators for extrapolation circumstances in solving several ODEs and PDEs. We also highlight the LNO’s good interpolation ability, from a low-resolution input to high-resolution outputs at arbitrary locations within the domain. To demonstrate the scalability of LNO, we conduct large-scale simulations of Rossby waves around the globe, employing millions of degrees of freedom. Taken together, our findings show that a pretrained LNO model offers an effective real-time solution for general ODEs and PDEs at scale and is an efficient alternative to existing neural operators. Neural operators are powerful neural networks that approximate nonlinear dynamical systems and their responses. Cao and colleagues introduce the Laplace neural operator, a scalable approach that can effectively deal with non-periodic signals and transient responses and can outperform existing neural operators on certain classes of ODE and PDE problems.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141448408","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 : 2024-06-24DOI: 10.1038/s42256-024-00852-4
Mayank Kejriwal, Eric Kildebeck, Robert Steininger, Abhinav Shrivastava
Environmental changes can profoundly impact the performance of artificial intelligence systems operating in the real world, with effects ranging from overt catastrophic failures to non-robust behaviours that do not take changing context into account. Here we argue that designing machine intelligence that can operate in open worlds, including detecting, characterizing and adapting to structurally unexpected environmental changes, is a critical goal on the path to building systems that can solve complex and relatively under-determined problems. We present and distinguish between three forms of open-world learning (OWL)—weak, semi-strong and strong—and argue that a fully developed OWL system should be antifragile, rather than merely robust. An antifragile system, an example of which is the immune system, is not only robust to adverse events, but adapts to them quickly and becomes better at handling them in subsequent encounters. We also argue that, because OWL approaches must be capable of handling the unexpected, their practical evaluation can pose an interesting conceptual problem. AI systems operating in the real world unavoidably encounter unexpected environmental changes and need a built-in robustness and capability to learn fast, making use of advances such as lifelong and few-shot learning. Kejriwal et al. discuss three categories of such open-world learning and discuss applications such as self-driving cars and robotic inspection.
{"title":"Challenges, evaluation and opportunities for open-world learning","authors":"Mayank Kejriwal, Eric Kildebeck, Robert Steininger, Abhinav Shrivastava","doi":"10.1038/s42256-024-00852-4","DOIUrl":"10.1038/s42256-024-00852-4","url":null,"abstract":"Environmental changes can profoundly impact the performance of artificial intelligence systems operating in the real world, with effects ranging from overt catastrophic failures to non-robust behaviours that do not take changing context into account. Here we argue that designing machine intelligence that can operate in open worlds, including detecting, characterizing and adapting to structurally unexpected environmental changes, is a critical goal on the path to building systems that can solve complex and relatively under-determined problems. We present and distinguish between three forms of open-world learning (OWL)—weak, semi-strong and strong—and argue that a fully developed OWL system should be antifragile, rather than merely robust. An antifragile system, an example of which is the immune system, is not only robust to adverse events, but adapts to them quickly and becomes better at handling them in subsequent encounters. We also argue that, because OWL approaches must be capable of handling the unexpected, their practical evaluation can pose an interesting conceptual problem. AI systems operating in the real world unavoidably encounter unexpected environmental changes and need a built-in robustness and capability to learn fast, making use of advances such as lifelong and few-shot learning. Kejriwal et al. discuss three categories of such open-world learning and discuss applications such as self-driving cars and robotic inspection.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141448341","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 : 2024-06-21DOI: 10.1038/s42256-024-00851-5
Evan E. Seitz, David M. McCandlish, Justin B. Kinney, Peter K. Koo
Deep neural networks (DNNs) have greatly advanced the ability to predict genome function from sequence. However, elucidating underlying biological mechanisms from genomic DNNs remains challenging. Existing interpretability methods, such as attribution maps, have their origins in non-biological machine learning applications and therefore have the potential to be improved by incorporating domain-specific interpretation strategies. Here we introduce SQUID (Surrogate Quantitative Interpretability for Deepnets), a genomic DNN interpretability framework based on domain-specific surrogate modelling. SQUID approximates genomic DNNs in user-specified regions of sequence space using surrogate models—simpler quantitative models that have inherently interpretable mathematical forms. SQUID leverages domain knowledge to model cis-regulatory mechanisms in genomic DNNs, in particular by removing the confounding effects that nonlinearities and heteroscedastic noise in functional genomics data can have on model interpretation. Benchmarking analysis on multiple genomic DNNs shows that SQUID, when compared to established interpretability methods, identifies motifs that are more consistent across genomic loci and yields improved single-nucleotide variant-effect predictions. SQUID also supports surrogate models that quantify epistatic interactions within and between cis-regulatory elements, as well as global explanations of cis-regulatory mechanisms across sequence contexts. SQUID thus advances the ability to mechanistically interpret genomic DNNs. The intersection of genomics and deep learning shows promise for real impact on healthcare and biological research, but the lack of interpretability in terms of biological mechanisms is limiting utility and further development. As a potential solution, Koo et al. present SQUID, an interpretability framework built using domain-specific genomic surrogate models.
{"title":"Interpreting cis-regulatory mechanisms from genomic deep neural networks using surrogate models","authors":"Evan E. Seitz, David M. McCandlish, Justin B. Kinney, Peter K. Koo","doi":"10.1038/s42256-024-00851-5","DOIUrl":"10.1038/s42256-024-00851-5","url":null,"abstract":"Deep neural networks (DNNs) have greatly advanced the ability to predict genome function from sequence. However, elucidating underlying biological mechanisms from genomic DNNs remains challenging. Existing interpretability methods, such as attribution maps, have their origins in non-biological machine learning applications and therefore have the potential to be improved by incorporating domain-specific interpretation strategies. Here we introduce SQUID (Surrogate Quantitative Interpretability for Deepnets), a genomic DNN interpretability framework based on domain-specific surrogate modelling. SQUID approximates genomic DNNs in user-specified regions of sequence space using surrogate models—simpler quantitative models that have inherently interpretable mathematical forms. SQUID leverages domain knowledge to model cis-regulatory mechanisms in genomic DNNs, in particular by removing the confounding effects that nonlinearities and heteroscedastic noise in functional genomics data can have on model interpretation. Benchmarking analysis on multiple genomic DNNs shows that SQUID, when compared to established interpretability methods, identifies motifs that are more consistent across genomic loci and yields improved single-nucleotide variant-effect predictions. SQUID also supports surrogate models that quantify epistatic interactions within and between cis-regulatory elements, as well as global explanations of cis-regulatory mechanisms across sequence contexts. SQUID thus advances the ability to mechanistically interpret genomic DNNs. The intersection of genomics and deep learning shows promise for real impact on healthcare and biological research, but the lack of interpretability in terms of biological mechanisms is limiting utility and further development. As a potential solution, Koo et al. present SQUID, an interpretability framework built using domain-specific genomic surrogate models.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141436053","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 : 2024-06-21DOI: 10.1038/s42256-024-00858-y
Alexander Ziller, Tamara T. Mueller, Simon Stieger, Leonhard F. Feiner, Johannes Brandt, Rickmer Braren, Daniel Rueckert, Georgios Kaissis
Artificial intelligence (AI) models are vulnerable to information leakage of their training data, which can be highly sensitive, for example, in medical imaging. Privacy-enhancing technologies, such as differential privacy (DP), aim to circumvent these susceptibilities. DP is the strongest possible protection for training models while bounding the risks of inferring the inclusion of training samples or reconstructing the original data. DP achieves this by setting a quantifiable privacy budget. Although a lower budget decreases the risk of information leakage, it typically also reduces the performance of such models. This imposes a trade-off between robust performance and stringent privacy. Additionally, the interpretation of a privacy budget remains abstract and challenging to contextualize. Here we contrast the performance of artificial intelligence models at various privacy budgets against both theoretical risk bounds and empirical success of reconstruction attacks. We show that using very large privacy budgets can render reconstruction attacks impossible, while drops in performance are negligible. We thus conclude that not using DP at all is negligent when applying artificial intelligence models to sensitive data. We deem our results to lay a foundation for further debates on striking a balance between privacy risks and model performance. Ziller and colleagues present a balanced investigation of the trade-off between privacy and performance when training artificially intelligent models for medical imaging analysis tasks. The authors evaluate the use of differential privacy in realistic threat scenarios, leading to their conclusion to promote the use of differential privacy, but implementing it in a manner that also retains performance.
{"title":"Reconciling privacy and accuracy in AI for medical imaging","authors":"Alexander Ziller, Tamara T. Mueller, Simon Stieger, Leonhard F. Feiner, Johannes Brandt, Rickmer Braren, Daniel Rueckert, Georgios Kaissis","doi":"10.1038/s42256-024-00858-y","DOIUrl":"10.1038/s42256-024-00858-y","url":null,"abstract":"Artificial intelligence (AI) models are vulnerable to information leakage of their training data, which can be highly sensitive, for example, in medical imaging. Privacy-enhancing technologies, such as differential privacy (DP), aim to circumvent these susceptibilities. DP is the strongest possible protection for training models while bounding the risks of inferring the inclusion of training samples or reconstructing the original data. DP achieves this by setting a quantifiable privacy budget. Although a lower budget decreases the risk of information leakage, it typically also reduces the performance of such models. This imposes a trade-off between robust performance and stringent privacy. Additionally, the interpretation of a privacy budget remains abstract and challenging to contextualize. Here we contrast the performance of artificial intelligence models at various privacy budgets against both theoretical risk bounds and empirical success of reconstruction attacks. We show that using very large privacy budgets can render reconstruction attacks impossible, while drops in performance are negligible. We thus conclude that not using DP at all is negligent when applying artificial intelligence models to sensitive data. We deem our results to lay a foundation for further debates on striking a balance between privacy risks and model performance. Ziller and colleagues present a balanced investigation of the trade-off between privacy and performance when training artificially intelligent models for medical imaging analysis tasks. The authors evaluate the use of differential privacy in realistic threat scenarios, leading to their conclusion to promote the use of differential privacy, but implementing it in a manner that also retains performance.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.com/articles/s42256-024-00858-y.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141436195","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 : 2024-06-21DOI: 10.1038/s42256-024-00857-z
Weixin Liang, Nazneen Rajani, Xinyu Yang, Ezinwanne Ozoani, Eric Wu, Yiqun Chen, Daniel Scott Smith, James Zou
The rapid proliferation of AI models has underscored the importance of thorough documentation, which enables users to understand, trust and effectively use these models in various applications. Although developers are encouraged to produce model cards, it’s not clear how much or what information these cards contain. In this study we conduct a comprehensive analysis of 32,111 AI model documentations on Hugging Face, a leading platform for distributing and deploying AI models. Our investigation sheds light on the prevailing model card documentation practices. Most AI models with a substantial number of downloads provide model cards, although with uneven informativeness. We find that sections addressing environmental impact, limitations and evaluation exhibit the lowest filled-out rates, whereas the training section is the one most consistently filled-out. We analyse the content of each section to characterize practitioners’ priorities. Interestingly, there are considerable discussions of data, sometimes with equal or even greater emphasis than the model itself. Our study provides a systematic assessment of community norms and practices surroinding model documentation through large-scale data science and linguistic analysis. As the number of AI models has rapidly grown, there is an increased focus on improving the documentation through model cards. Liang et al. explore questions around adoption practices and the type of information provided in model cards through a large-scale analysis of 32,111 model card documentation from 74,970 models.
人工智能模型的迅速扩散凸显了详尽文档的重要性,它能让用户理解、信任并在各种应用中有效地使用这些模型。尽管我们鼓励开发者制作模型卡片,但这些卡片包含多少信息或包含哪些信息却并不清楚。在本研究中,我们对 Hugging Face 上的 32111 份人工智能模型文档进行了全面分析,Hugging Face 是分发和部署人工智能模型的领先平台。我们的调查揭示了模型卡片文档的普遍做法。大多数有大量下载的人工智能模型都提供了模型卡,但信息量参差不齐。我们发现,涉及环境影响、局限性和评估的部分填写率最低,而训练部分则是填写率最高的部分。我们分析了每个部分的内容,以确定实践者的优先事项。有趣的是,我们对数据进行了大量讨论,有时对数据的重视程度甚至超过了模型本身。我们的研究通过大规模的数据科学和语言学分析,系统地评估了模型文档之外的社区规范和实践。
{"title":"Systematic analysis of 32,111 AI model cards characterizes documentation practice in AI","authors":"Weixin Liang, Nazneen Rajani, Xinyu Yang, Ezinwanne Ozoani, Eric Wu, Yiqun Chen, Daniel Scott Smith, James Zou","doi":"10.1038/s42256-024-00857-z","DOIUrl":"10.1038/s42256-024-00857-z","url":null,"abstract":"The rapid proliferation of AI models has underscored the importance of thorough documentation, which enables users to understand, trust and effectively use these models in various applications. Although developers are encouraged to produce model cards, it’s not clear how much or what information these cards contain. In this study we conduct a comprehensive analysis of 32,111 AI model documentations on Hugging Face, a leading platform for distributing and deploying AI models. Our investigation sheds light on the prevailing model card documentation practices. Most AI models with a substantial number of downloads provide model cards, although with uneven informativeness. We find that sections addressing environmental impact, limitations and evaluation exhibit the lowest filled-out rates, whereas the training section is the one most consistently filled-out. We analyse the content of each section to characterize practitioners’ priorities. Interestingly, there are considerable discussions of data, sometimes with equal or even greater emphasis than the model itself. Our study provides a systematic assessment of community norms and practices surroinding model documentation through large-scale data science and linguistic analysis. As the number of AI models has rapidly grown, there is an increased focus on improving the documentation through model cards. Liang et al. explore questions around adoption practices and the type of information provided in model cards through a large-scale analysis of 32,111 model card documentation from 74,970 models.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141435934","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 : 2024-06-21DOI: 10.1038/s42256-024-00855-1
Dong Chen, Jian Liu, Guo-Wei Wei
Despite the success of pretrained natural language processing (NLP) models in various fields, their application in computational biology has been hindered by their reliance on biological sequences, which ignores vital three-dimensional (3D) structural information incompatible with the sequential architecture of NLP models. Here we present a topological transformer (TopoFormer), which is built by integrating NLP models and a multiscale topology technique, the persistent topological hyperdigraph Laplacian (PTHL), which systematically converts intricate 3D protein–ligand complexes at various spatial scales into an NLP-admissible sequence of topological invariants and homotopic shapes. PTHL systematically transforms intricate 3D protein–ligand complexes into NLP-compatible sequences of topological invariants and shapes, capturing essential interactions across spatial scales. TopoFormer gives rise to exemplary scoring accuracy and excellent performance in ranking, docking and screening tasks in several benchmark datasets. This approach can be utilized to convert general high-dimensional structured data into NLP-compatible sequences, paving the way for broader NLP based research. Transformers show much promise for applications in computational biology, but they rely on sequences, and a challenge is to incorporate 3D structural information. TopoFormer, proposed by Dong Chen et al., combines transformers with a mathematical multiscale topology technique to model 3D protein–ligand complexes, substantially enhancing performance in a range of prediction tasks of interest to drug discovery.
{"title":"Multiscale topology-enabled structure-to-sequence transformer for protein–ligand interaction predictions","authors":"Dong Chen, Jian Liu, Guo-Wei Wei","doi":"10.1038/s42256-024-00855-1","DOIUrl":"10.1038/s42256-024-00855-1","url":null,"abstract":"Despite the success of pretrained natural language processing (NLP) models in various fields, their application in computational biology has been hindered by their reliance on biological sequences, which ignores vital three-dimensional (3D) structural information incompatible with the sequential architecture of NLP models. Here we present a topological transformer (TopoFormer), which is built by integrating NLP models and a multiscale topology technique, the persistent topological hyperdigraph Laplacian (PTHL), which systematically converts intricate 3D protein–ligand complexes at various spatial scales into an NLP-admissible sequence of topological invariants and homotopic shapes. PTHL systematically transforms intricate 3D protein–ligand complexes into NLP-compatible sequences of topological invariants and shapes, capturing essential interactions across spatial scales. TopoFormer gives rise to exemplary scoring accuracy and excellent performance in ranking, docking and screening tasks in several benchmark datasets. This approach can be utilized to convert general high-dimensional structured data into NLP-compatible sequences, paving the way for broader NLP based research. Transformers show much promise for applications in computational biology, but they rely on sequences, and a challenge is to incorporate 3D structural information. TopoFormer, proposed by Dong Chen et al., combines transformers with a mathematical multiscale topology technique to model 3D protein–ligand complexes, substantially enhancing performance in a range of prediction tasks of interest to drug discovery.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141436150","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 : 2024-06-18DOI: 10.1038/s42256-024-00843-5
Yuanqi Du, Arian R. Jamasb, Jeff Guo, Tianfan Fu, Charles Harris, Yingheng Wang, Chenru Duan, Pietro Liò, Philippe Schwaller, Tom L. Blundell
Machine learning has provided a means to accelerate early-stage drug discovery by combining molecule generation and filtering steps in a single architecture that leverages the experience and design preferences of medicinal chemists. However, designing machine learning models that can achieve this on the fly to the satisfaction of medicinal chemists remains a challenge owing to the enormous search space. Researchers have addressed de novo design of molecules by decomposing the problem into a series of tasks determined by design criteria. Here we provide a comprehensive overview of the current state of the art in molecular design using machine learning models as well as important design decisions, such as the choice of molecular representations, generative methods and optimization strategies. Subsequently, we present a collection of practical applications in which the reviewed methodologies have been experimentally validated, encompassing both academic and industrial efforts. Finally, we draw attention to the theoretical, computational and empirical challenges in deploying generative machine learning and highlight future opportunities to better align such approaches to achieve realistic drug discovery end points. Data-driven generative methods have the potential to greatly facilitate molecular design tasks for drug design.
{"title":"Machine learning-aided generative molecular design","authors":"Yuanqi Du, Arian R. Jamasb, Jeff Guo, Tianfan Fu, Charles Harris, Yingheng Wang, Chenru Duan, Pietro Liò, Philippe Schwaller, Tom L. Blundell","doi":"10.1038/s42256-024-00843-5","DOIUrl":"10.1038/s42256-024-00843-5","url":null,"abstract":"Machine learning has provided a means to accelerate early-stage drug discovery by combining molecule generation and filtering steps in a single architecture that leverages the experience and design preferences of medicinal chemists. However, designing machine learning models that can achieve this on the fly to the satisfaction of medicinal chemists remains a challenge owing to the enormous search space. Researchers have addressed de novo design of molecules by decomposing the problem into a series of tasks determined by design criteria. Here we provide a comprehensive overview of the current state of the art in molecular design using machine learning models as well as important design decisions, such as the choice of molecular representations, generative methods and optimization strategies. Subsequently, we present a collection of practical applications in which the reviewed methodologies have been experimentally validated, encompassing both academic and industrial efforts. Finally, we draw attention to the theoretical, computational and empirical challenges in deploying generative machine learning and highlight future opportunities to better align such approaches to achieve realistic drug discovery end points. Data-driven generative methods have the potential to greatly facilitate molecular design tasks for drug design.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141425425","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 : 2024-06-17DOI: 10.1038/s42256-024-00847-1
Huan Yee Koh, Anh T. N. Nguyen, Shirui Pan, Lauren T. May, Geoffrey I. Webb
In drug discovery, determining the binding affinity and functional effects of small-molecule ligands on proteins is critical. Current computational methods can predict these protein–ligand interaction properties but often lose accuracy without high-resolution protein structures and falter in predicting functional effects. Here we introduce PSICHIC (PhySIcoCHemICal graph neural network), a framework incorporating physicochemical constraints to decode interaction fingerprints directly from sequence data alone. This enables PSICHIC to attain capabilities in decoding mechanisms underlying protein–ligand interactions, achieving state-of-the-art accuracy and interpretability. Trained on identical protein–ligand pairs without structural data, PSICHIC matched and even surpassed leading structure-based methods in binding-affinity prediction. In an experimental library screening for adenosine A1 receptor agonists, PSICHIC discerned functional effects effectively, ranking the sole novel agonist within the top three. PSICHIC’s interpretable fingerprints identified protein residues and ligand atoms involved in interactions, and helped in unveiling selectivity determinants of protein–ligand interaction. We foresee PSICHIC reshaping virtual screening and deepening our understanding of protein–ligand interactions. Predicting the binding affinity between small-molecule ligands and proteins is a key task in drug discovery; however, sequence-based methods are often less accurate than structure-based ones. Koh et al. develop a graph neural network using physicochemical constraints that discovers interactions between small molecules and proteins directly from sequence data and that can achieve state-of-the-art performance without the need for costly, experimental 3D structures.
{"title":"Physicochemical graph neural network for learning protein–ligand interaction fingerprints from sequence data","authors":"Huan Yee Koh, Anh T. N. Nguyen, Shirui Pan, Lauren T. May, Geoffrey I. Webb","doi":"10.1038/s42256-024-00847-1","DOIUrl":"10.1038/s42256-024-00847-1","url":null,"abstract":"In drug discovery, determining the binding affinity and functional effects of small-molecule ligands on proteins is critical. Current computational methods can predict these protein–ligand interaction properties but often lose accuracy without high-resolution protein structures and falter in predicting functional effects. Here we introduce PSICHIC (PhySIcoCHemICal graph neural network), a framework incorporating physicochemical constraints to decode interaction fingerprints directly from sequence data alone. This enables PSICHIC to attain capabilities in decoding mechanisms underlying protein–ligand interactions, achieving state-of-the-art accuracy and interpretability. Trained on identical protein–ligand pairs without structural data, PSICHIC matched and even surpassed leading structure-based methods in binding-affinity prediction. In an experimental library screening for adenosine A1 receptor agonists, PSICHIC discerned functional effects effectively, ranking the sole novel agonist within the top three. PSICHIC’s interpretable fingerprints identified protein residues and ligand atoms involved in interactions, and helped in unveiling selectivity determinants of protein–ligand interaction. We foresee PSICHIC reshaping virtual screening and deepening our understanding of protein–ligand interactions. Predicting the binding affinity between small-molecule ligands and proteins is a key task in drug discovery; however, sequence-based methods are often less accurate than structure-based ones. Koh et al. develop a graph neural network using physicochemical constraints that discovers interactions between small molecules and proteins directly from sequence data and that can achieve state-of-the-art performance without the need for costly, experimental 3D structures.","PeriodicalId":48533,"journal":{"name":"Nature Machine Intelligence","volume":null,"pages":null},"PeriodicalIF":18.8,"publicationDate":"2024-06-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141333577","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}