ABSTRACT Protein conformational flexibility underlies a wide range of biological functions, yet experimentally probing dynamics at atomic resolution remains costly and low-throughput. Here, we present a deep learning framework that predicts protein flexibility directly from static structural descriptors, bypassing the need for molecular dynamics (MD) simulations. Using the ATLAS database of standardized all-atom MD trajectories, we encoded 1,374 protein chains as 30-dimensional Gaussian integral (GI) vectors—global shape and topology invariants of the protein backbone. Principal component analysis of GI profiles revealed four structural clusters with distinct secondary structure compositions and flexibility distributions. We trained an attention-based one-dimensional convolutional neural network (1D-CNN) to classify proteins as flexible or non-flexible based on their root-mean-square fluctuation (RMSF) relative to the dataset-wide mean. The classifier achieved an AUC of 0.772 (95% CI: 0.712–0.826) on an independent test set, with balanced sensitivity and specificity, and identified a small subset of GI components as the most predictive. In a regression setting, a recurrent neural network outperformed other architectures, attaining an R 2 of 0.537, though high-flexibility values were systematically underestimated. Cluster-specific analyses indicated that coil-rich and β-sheet–dominated proteins were more amenable to flexibility prediction than α-helical proteins, likely due to greater structural heterogeneity. Our results demonstrate that compact GI descriptors preserve sufficient information to recover MD-derived flexibility trends, offering a computationally efficient complement to simulation-based approaches. This framework enables large-scale screening of protein dynamics from structural data alone, with potential applications in structural bioinformatics, drug design, and functional annotation.

While data augmentation (DA) has been extensively studied in computer vision, its application to Inertial Measurement Unit (IMU) signals remains largely unexplored, despite IMUs’ growing importance in human motion analysis. In this paper, we present the first systematic study of IMU-specific data augmentation, beginning with a comprehensive analysis that identifies three fundamental properties of IMU signals: their time-series nature, inherent multimodality (rotation and acceleration) and motion-consistency characteristics. Through this analysis, we demonstrate the limitations of applying conventional time-series augmentation techniques to IMU data. We then introduce Motion-Drift Augmentation (MODA), a novel technique that simulates the natural displacement of body-worn IMUs during motion. We evaluate our approach across five diverse datasets and five deep learning settings, including i) fully-supervised, ii) semi-supervised, iii) domain adaptation, iv) domain generalization and v) few-shot learning for both Human Action Recognition (HAR) and Human Pose Estimation (HPE) tasks. Experimental results show that our proposed MODA consistently outperforms existing augmentation methods, with semi-supervised learning performance approaching state-of-the-art fully-supervised methods.

Human motion capture with sparse inertial sensors has gained significant attention recently. However, existing methods almost exclusively rely on a template adult body shape to model the training data, which poses challenges when generalizing to individuals with largely different body shapes (such as a child). This is primarily due to the variation in IMU-measured acceleration caused by changes in body shape. To fill this gap, we propose Shape-aware Inertial Poser (SAIP), the first solution considering body shape differences in sparse inertial-based motion capture. Specifically, we decompose the sensor measurements related to shape and pose in order to effectively model their joint correlations. Firstly, we train a regression model to transfer the IMU-measured accelerations of a real body to match the template adult body model, compensating for the shape-related sensor measurements. Then, we can easily follow the state-of-the-art methods to estimate the full body motions of the template-shaped body. Finally, we utilize a second regression model to map the joint velocities back to the real body, combined with a shape-aware physical optimization strategy to calculate global motions on the subject. Furthermore, our method relies on body shape awareness, introducing the first inertial shape estimation scheme. This is accomplished by modeling the shape-conditioned IMU-pose correlation using an MLP-based network. To validate the effectiveness of SAIP, we also present the first IMU motion capture dataset containing individuals of different body sizes. This dataset features 10 children and 10 adults, with heights ranging from 110 cm to 190 cm, and a total of 400 minutes of paired IMU-Motion samples. Extensive experimental results demonstrate that SAIP can effectively handle motion capture tasks for diverse body shapes. The code and dataset are available at https://github.com/yinlu5942/SAIP .

Transformer models excel at capturing long-range dependencies in sequential data, but lack explicit mechanisms to leverage structural patterns inherent in fixed-length input sequences. In this paper, we propose a novel sequence structure learning and modulation approach that endows Transformers with the ability to model and utilize such fixed-sequence structural properties for improved performance on inertial pose estimation tasks. Specifically, our method introduces a Sequence Structure Module (SSM) that utilizes structural information of fixed-length inertial sensor readings to adjust the input features of transformers. Such structural information can either be acquired by learning or specified based on users' prior knowledge. To justify the prospect of our approach, we show that i) injecting spatial structural information of IMUs/joints learned from data improves accuracy, while ii) injecting temporal structural information based on smooth priors reduces jitter (i.e., improves steadiness), in a spatial-temporal transformer solution for inertial pose estimation. Extensive experiments across multiple benchmark datasets demonstrate the superiority of our approach against state-of-the-art methods and has the potential to advance the design of the transformer architecture for fixed-length sequences.

With the widespread deployment of three-dimensional (3D) models in industry and daily life, protecting the security of this data becomes crucial. Additionally, three-dimensional (3D) models may be distributed to users with varying security levels, necessitating distinct visualizations for each user. Recent research proposes 3D model encryption method that facilitates distinct visualizations post-decryption through hierarchical decryption. However, this method permits the decryption of 3D models at varying visual security levels based on user privileges. It has potential security vulnerabilities concerning key management and simultaneously limits its capacity to address diverse user requirements. To address this, a multi-stage secret sharing mechanism is integrated into the existing hierarchical encryption framework to bolster the security of hierarchical keys. When combined with lattice-based cryptography techniques, it ensures that only users with adequate shares can decrypt the corresponding 3D model hierarchy, achieving distinct visual effects while maintaining secret key security under diverse user needs. Experimental results demonstrate that the scheme effectively enhances security while maintaining data integrity and availability.

Recent research works have demonstrated that deep neural networks (DNNs) are vulnerable to backdoor attacks. The existing backdoor attacks can only cause targeted misclassification on backdoor instances, which makes them can be easily detected by defense methods. In this article, we propose an untargeted backdoor attack (UBA) against DNNs, where the backdoor instances are randomly misclassified by the backdoored model to any incorrect label. To achieve the goal of UBA, we propose to utilize autoencoder as the trigger generation model and train the target model and the autoencoder simultaneously. We also propose a special loss function (Evasion Loss) to train the autoencoder and the target model, in order to make the target model predict backdoor instances as random incorrect classes. During the inference stage, the trained autoencoder is used to generate backdoor instances. For different backdoor instances, the generated triggers are different and the corresponding predicted labels are random incorrect labels. Experimental results demonstrate that the proposed UBA is effective. On the ResNet-18 model, the attack success rate (ASR) of the proposed UBA is 96.48%, 91.27%, and 90.83% on CIFAR-10, GTSRB, and ImageNet datasets, respectively. On the VGG-16 model, the ASR of the proposed UBA is 89.72% and 97.78% on CIFAR-10 and ImageNet datasets, respectively. Moreover, the proposed UBA is robust against existing backdoor defense methods, which are designed to detect targeted backdoor attacks. We hope this article can promote the research of corresponding backdoor defense works.

As training a high-performance deep neural network (DNN) model requires a large amount of data, powerful computing resources and expert knowledge, protecting well-trained DNN models from intellectual property (IP) infringement has raised serious concerns in recent years. Most existing methods using DNN watermarks to verify the ownership of the models after IP infringement occurs, which is reactive in the sense that they cannot prevent unauthorized users from using the model in the first place. Different from these methods, in this article, we propose an active authorization control and user’s fingerprint tracking method for the IP protection of DNN models by utilizing sample-specific backdoor attack. The proposed method inversely and multiplely exploits sample-specific trigger as the key to implement authorization control for DNN model, in which the generated triggers are imperceptible and sample-specific for clean images. Specifically, a U-Net model is used to generate backdoor instances. Then, the target model is trained on the clean images and backdoor instances, which are inversely labeled as wrong classes and correct classes, respectively. Only authorized users can use the target model normally by pre-processing the clean images through the U-Net model. Moreover, the images processed by the U-Net model will contain unique fingerprint that can be extracted to verify and track the corresponding user’s identity. This article is the first work that utilizes the sample-specific backdoor attack to implement active authorization control and user’s fingerprint management for DNN model under black-box scenarios. Extensive experimental results on ImageNet dataset and YouTube Aligned Face dataset demonstrate that the proposed method is effective in protecting the DNN model from unauthorized usage. Specifically, the protected model has a low inference accuracy (1.00%) for unauthorized users, while maintaining a normal inference accuracy (97.67%) for authorized users. Besides, the proposed method can achieve 100% fingerprint tracking success rates on both the ImageNet and YouTube Aligned Face datasets. Moreover, it is demonstrated that the proposed method is robust against fine-tuning attack, pruning attack, pruning attack with retraining, reverse-engineering attack, adaptive attack, and JPEG compression attack. The code is available at https://github.com/nuaaaisec/SSAT .

Underwater robots have been widely used in Marine environment monitoring, deep-sea resources exploration, underwater archaeology, and other fields. The underwater robot hovering is a very demanding technology, especially in a dynamic environment, the underwater multi-disturbance robot has a great influence, and accurate hovering of the underwater robot is the basic guarantee to complete the task. In this paper, a hybrid control strategy based on a neural network and PID is proposed to realize the perception and decision of complex environment states and realize the accurate hovering of the underwater robot. Experimental results show that the hybrid control based on neural network and PID can stably and accurately complete the hovering function, which proves the effectiveness of the algorithm. (Video: https://youtu.be/1GU4BKHeTB8t)

Polyp segmentation is of great importance for the diagnosis and treatment of colorectal cancer. However, it is difficult to segment polyps accurately due to a large number of tiny polyps and the low contrast between polyps and the surrounding mucosa. To address this issue, we design an Adaptive Tiny-object Enhanced Network (ATENet) for tiny polyp segmentation. The proposed ATENet has two advantages: First, we design an adaptive tiny-object encoder containing three parallel branches, which can effectively extract the shape and position features of tiny polyps and thus improve the segmentation accuracy of tiny polyps. Second, we design a simple enhanced feature decoder, which can not only suppress the background noise of feature maps, but also supplement the detail information to improve further the polyp segmentation accuracy. Extensive experiments on three benchmark datasets demonstrate that the proposed ATENet can achieve the state-of-the-art performance while maintaining low computational complexity.

Community search is a derivative of community detection that enables online and personalized discovery of communities and has found extensive applications in massive real-world networks. Recently, there needs to be more focus on the community search issue within directed graphs, even though substantial research has been carried out on undirected graphs. The recently proposed D-truss model has achieved good results in the quality of retrieved communities. However, existing D-truss-based work cannot perform efficient community searches on large graphs because it consumes too many computing resources to retrieve the maximal D-truss. To overcome this issue, we introduce an innovative merge relation known as D-truss-connected to capture the inherent density and cohesiveness of edges within D-truss. This relation allows us to partition all the edges in the original graph into a series of D-truss-connected classes. Then, we construct a concise and compact index, ConDTruss, based on D-truss-connected. Using ConDTruss, the efficiency of maximum D-truss retrieval will be greatly improved, making it a theoretically optimal approach. Experimental evaluations conducted on large directed graphs certificate the effectiveness of our proposed method.

Community Search (CS) aims to identify densely interconnected subgraphs corresponding to query vertices within a graph. However, existing heterogeneous graph-based community search methods need help identifying cross-group communities and suffer from efficiency issues, making them unsuitable for large graphs. This paper presents a fast community search model based on the Butterfly-Core Community (BCC) structure for heterogeneous graphs. The Random Walk with Restart (RWR) algorithm and butterfly degree comprehensively evaluate the importance of vertices within communities, allowing leader vertices to be rapidly updated to maintain cross-group cohesion. Moreover, we devised a more efficient method for updating vertex distances, which minimizes vertex visits and enhances operational efficiency. Extensive experiments on several real-world temporal graphs demonstrate the effectiveness and efficiency of this solution.

The precise positioning of underwater robots is the premise and foundation to complete other operations. Since a global positioning system (GPS) cannot be used underwater, and the positioning method of the underwater robot based on inertial navigation could cause significant errors, a multi-sensor information fusion method based on Elman neural network is proposed to solve these problems. The network is trained by taking data of doppler velocity log (DVL) and inertial measurement unit (IMU) as input and GPS as output. In the underwater area without GPS, the training network is used to predict the real-time position error of the acquired DVL and IMU data. The method can realize dynamic training and learning to improve the accuracy of the system. The experimental results show that the proposed method has lower positioning error than the traditional method, effectively inhibits the accumulation error of positioning, and improves underwater robots' positioning accuracy.

BACKGROUND: Proteins form various complexes to carry out their versatile functions in cells. The dynamic properties of protein complex formation are mainly characterized by the association rates which measures how fast these complexes can be formed. It was experimentally observed that the association rates span an extremely wide range with over ten orders of magnitudes. Identification of association rates within this spectrum for specific protein complexes is therefore essential for us to understand their functional roles. RESULTS: To tackle this problem, we integrate physics-based coarse-grained simulations into a neural-network-based classification model to estimate the range of association rates for protein complexes in a large-scale benchmark set. The cross-validation results show that, when an optimal threshold was selected, we can reach the best performance with specificity, precision, sensitivity and overall accuracy all higher than 70%. The quality of our cross-validation data has also been testified by further statistical analysis. Additionally, given an independent testing set, we can successfully predict the group of association rates for eight protein complexes out of ten. Finally, the analysis of failed cases suggests the future implementation of conformational dynamics into simulation can further improve model. CONCLUSIONS: In summary, this study demonstrated that a new modeling framework that combines biophysical simulations with bioinformatics approaches is able to identify protein-protein interactions with low association rates from those with higher association rates. This method thereby can serve as a useful addition to a collection of existing experimental approaches that measure biomolecular recognition.

The activation and differentiation of T-cells are mainly directly by their co-regulatory receptors. T lymphocyte-associated protein-4 (CTLA-4) and programed cell death-1 (PD-1) are two of the most important co-regulatory receptors. Binding of PD-1 and CTLA-4 with their corresponding ligands programed cell death-ligand 1 (PD-L1) and B7 on the antigen presenting cells (APC) activates two central co-inhibitory signaling pathways to suppress T cell functions. Interestingly, recent experiments have identified a new cis-interaction between PD-L1 and B7, suggesting that a crosstalk exists between two co-inhibitory receptors and the two pairs of ligand-receptor complexes can undergo dynamic oligomerization. Inspired by these experimental evidences, we developed a coarse-grained model to characterize the assembling of an immune complex consisting of CLTA-4, B7, PD-L1 and PD-1. These four proteins and their interactions form a small network motif. The temporal dynamics and spatial pattern formation of this network was simulated by a diffusion-reaction algorithm. Our simulation method incorporates the membrane confinement of cell surface proteins and geometric arrangement of different binding interfaces between these proteins. A wide range of binding constants was tested for the interactions involved in the network. Interestingly, we show that the CTLA-4/B7 ligand-receptor complexes can first form linear oligomers, while these oligomers further align together into two-dimensional clusters. Similar phenomenon has also been observed in other systems of cell surface proteins. Our test results further indicate that both co-inhibitory signaling pathways activated by B7 and PD-L1 can be down-regulated by the new cis-interaction between these two ligands, consistent with previous experimental evidences. Finally, the simulations also suggest that the dynamic and the spatial properties of the immune complex assembly are highly determined by the energetics of molecular interactions in the network. Our study, therefore, brings new insights to the co-regulatory mechanisms of T cell activation.

The interactions between membrane receptors and extracellular ligands control cell-cell and cell-substrate adhesion, and environmental responsiveness by representing the initial steps of cell signaling pathways. These interactions can be spatial-temporally regulated when different extracellular ligands are tethered. The detailed mechanisms of this spatial-temporal regulation, including the competition between distinct ligands with overlapping binding sites and the conformational flexibility in multi-specific ligand assemblies have not been quantitatively evaluated. We present a new coarse-grained model to realistically simulate the binding process between multi-specific ligands and membrane receptors on cell surfaces. The model simplifies each receptor and each binding site in a multi-specific ligand as a rigid body. Different numbers or types of ligands are spatially organized together in the simulation. These designs were used to test the relation between the overall binding of a multi-specific ligand and the affinity of its cognate binding site. When a variety of ligands are exposed to cells expressing different densities of surface receptors, we demonstrated that ligands with reduced affinities have higher specificity to distinguish cells based on the relative concentrations of their receptors. Finally, modification of intramolecular flexibility was shown to play a role in optimizing the binding between receptors and ligands. In summary, our studies bring new insights to the general principles of ligand-receptor interactions. Future applications of our method will pave the way for new strategies to generate next-generation biologics.

BACKGROUND: The physical interactions between proteins constitute the basis of protein quaternary structures. They dominate many biological processes in living cells. Deciphering the structural features of interacting proteins is essential to understand their cellular functions. Similar to the space of protein tertiary structures in which discrete patterns are clearly observed on fold or sub-fold motif levels, it has been found that the space of protein quaternary structures is highly degenerate due to the packing of compact secondary structure elements at interfaces. Therefore, it is necessary to further decompose the protein quaternary structural space into a more local representation. RESULTS: Here we constructed an interface fragment pair library from the current structure database of protein complexes. After structural-based clustering, we found that more than 90% of these interface fragment pairs can be represented by a limited number of highly abundant motifs. These motifs were further used to guide complex assembly. A large-scale benchmark test shows that the native-like binding is highly likely in the structural ensemble of modeled protein complexes that were built through the library. CONCLUSIONS: Our study therefore presents supportive evidences that the space of protein quaternary structures can be represented by the combination of a small set of secondary-structure-based packing at binding interfaces. Finally, after future improvements such as adding sequence profiles, we expect this new library will be useful to predict structures of unknown protein-protein interactions.