Multilingual speaker verification (SV) remains challenging due to limited cross-lingual data and language-dependent information in speaker embeddings. This paper presents a language-invariant multilingual SV system for the TidyVoice 2026 Challenge. We adopt the multilingual self-supervised w2v-BERT 2.0 model as the backbone, enhanced with Layer Adapters and Multi-scale Feature Aggregation to better exploit multi-layer representations. A language-adversarial training strategy with a Gradient Reversal Layer is applied to promote language-invariant speaker embeddings. Moreover, a multilingual zero-shot text-to-speech system is used to synthesize speech in multiple languages, improving language diversity. Experimental results demonstrate that fine-tuning the large-scale pretrained model yields competitive performance, while language-adversarial training further enhances robustness. In addition, synthetic speech augmentation provides additional gains under limited training data conditions. Source code is available at https://github.com/ZXHY-82/LI-MSV-TidyVoice2026.

Protein subcellular localization prediction is essential for understanding protein function and cellular organization. However, existing methods exhibit two major limitations: (1) they overlook the critical role of evolutionarily conserved protein domains, which are fundamental functional and structural units that significantly influence functions and subcellular localization, and (2) they rarely learn residue order and backbone coordinates simultaneously, neglecting the complementary information inherent in multi-modal representations. In this paper, we propose a novel Domain-Aware Multi-View Contrastive Representation Learning for Protein Subcellular Localization prediction, named DMVCL. Firstly, it devises domain-sequence/structure attention modules, which identify functionally significant regions in protein structures/sequences that critically determine subcellular localization. Secondly, it introduces a multi-view contrastive learning framework that unites inter-view and intra-view objectives. Inter-view contrastive learning aligns protein sequences with their corresponding structures by maximizing mutual information, thereby capturing the consistency of protein residue order and backbone coordinates. Intra-view contrastive learning enhances the representation discriminability of each modality by explicitly separating proteins with no common location and attracting those with any shared localization. Extensive experiments demonstrate that DMVCL significantly outperforms existing baselines. Ablation studies and visualizations further highlight the contributions of domain-sequence/structure attention and multi-view contrastive learning in achieving superior predictive performance.

In multimodal graph learning, graph structures that integrate information from multiple sources, such as vision and text, can more comprehensively model complex entity relationships. However, the continuous growth of their data scale poses a significant computational bottleneck for training. Graph condensation methods provide a feasible path forward by synthesizing compact and representative datasets. Nevertheless, existing condensation approaches generally suffer from performance limitations in multimodal scenarios, mainly due to two reasons: (1) semantic misalignment between different modalities leads to gradient conflicts; (2) the message-passing mechanism of graph neural networks further structurally amplifies such gradient noise. Based on this, we propose Structural Regularized Gradient Matching (SR-GM), a condensation framework for multimodal graphs. This method alleviates gradient conflicts between modalities through a gradient decoupling mechanism and introduces a structural damping regularizer to suppress the propagation of gradient noise in the topology, thereby transforming the graph structure from a noise amplifier into a training stabilizer. Extensive experiments on four multimodal graph datasets demonstrate the effectiveness of SR-GM, highlighting its state-of-the-art performance and cross-architecture generalization capabilities in multimodal graph dataset condensation.

Accurate prediction of drug-target interactions (DTIs) is foundational to drug development. Over the years, representation learning methods based on sequences and relational knowledge have shown considerable promise in this field. However, DTI prediction remains a challenging task, particularly in cold-start settings and few-shot scenarios involving novel drugs or proteins. Therefore, we propose a novel DTI prediction framework. To enhance the model's generalization in settings with scarce labels and unseen entities, we introduce a link-based contrastive learning strategy. Instead of aligning entity-level global features, this strategy aligns fine-grained local features derived from both the sequence and relational modalities. Complementing this, we introduce a link-based cross-attention mechanism. This mechanism captures contextual features specific to individual drug-protein pairs conditioned on different links, providing necessary local features for contrastive learning strategies. Our model was evaluated on both cold-start and few-shot datasets involving unseen drugs or proteins, and significantly outperformed state-of-the-art (SOTA) methods. Furthermore, when evaluated in conventional data-rich settings, our model still demonstrates superior performance over current approaches.

Angiogenesis after ischemic brain injury contributes to the restoration of blood supply in the ischemic zone. Strategies to improve angiogenesis may facilitate the function recovery after stroke. Recent researches have demonstrated that dysfunction of long non-coding RNAs are associated with angiogenesis. We have previously reported that long non-coding RNAs (lncRNAs) are aberrantly expressed in ischemic stroke. However, little is known about long non-coding RNAs and theirs role in angiogenesis after stroke. In this study, we identified a rat lncRNAs, Meg3, and found that Meg3 was significantly decreased after ischemic stroke. Overexpression of Meg3 suppressed functional recovery and decreased capillary density after ischemic stroke. Downregulation of Meg3 ameliorated brain lesion and increased angiogenesis after ischemic stroke. Silencing of Meg3 resulted in a proangiogenic effect evidenced by increased endothelial cell migration, proliferation, sprouting, and tube formation. Mechanistically, we showed that Meg3 negatively regulated notch pathway both in vivo and in vitro. Inhibition of notch signaling in endothelial cells reversed the proangiogenic effect induced by Meg3 downregulation. This study revealed the function of Meg3 in ischemic stroke and elucidated its mechanism in angiogenesis after ischemic stroke.

Multi-speaker automatic speech recognition (ASR) aims to transcribe conversational speech involving multiple speakers, requiring the model to capture not only what was said, but also who said it and sometimes when it was spoken. Recent Speech-LLM approaches have shown the potential of unified modeling for this task, but jointly learning speaker attribution, temporal structure, and lexical recognition remains difficult and data-intensive. At the current stage, leveraging reliable speaker diarization as an explicit structural prior provides a practical and efficient way to simplify this task. To effectively exploit such priors, we propose DM-ASR, a diarization-aware multi-speaker ASR framework that reformulates the task as a multi-turn dialogue generation process. Given an audio chunk and diarization results, DM-ASR decomposes transcription into a sequence of speaker- and time-conditioned queries, each corresponding to one speaker in one time segment. This formulation converts multi-speaker recognition into a series of structured sub-tasks, explicitly decoupling speaker-temporal structure from linguistic content and enabling effective integration of diarization cues with the reasoning capability of large language models. We further introduce an optional word-level timestamp prediction mechanism that interleaves word and timestamp tokens, yielding richer structured outputs and better transcription quality. Our analysis shows that diarization systems provide more reliable speaker identities and segment-level boundaries, while LLMs excel at modeling linguistic content and long-range dependencies, demonstrating their complementary strengths. Experiments on Mandarin and English benchmarks show that the proposed approach achieves strong performance with relatively small models and training data, while remaining competitive with or outperforming existing unified approaches.

Audio-Visual Speech Recognition (AVSR) integrates acoustic and visual information to enhance robustness in adverse acoustic conditions. Recent advances in Large Language Models (LLMs) have yielded competitive automatic speech recognition performance and shown effectiveness for AVSR. However, prior approaches project audio and visual features independently or apply shallow fusion, limiting cross-modal alignment and complementary exchange while increasing the LLM's computational load. To address this, we propose AVUR-LLM, an LLM-based Audio-Visual Speech Recognition via Sparse Modality Alignment and Visual Unit-Guided Refinement. Experiments on LRS3 demonstrate state-of-the-art results for AVSR. Under additive-noise conditions at 0 dB SNR, it achieves 37

Predicting biomolecular interactions is fundamental to understanding cellular mechanisms and advancing drug discovery. However, biomolecular interactions exhibit immense diversity across multiple dimensions. Most existing computational methods are designed to handle one specific task or data modality, which limits their applicability and generalization capability in broader scenarios. To address this methodological rigidity, we propose a flexible framework for multi-modal feature fusion in biomolecular interaction prediction (FlexBIP). The core of FlexBIP lies in its modular architecture, which decouples intrinsic molecular features from complex graph topologies, enabling the adaptive integration of node attributes, edge properties, and auxiliary graph information. The flexible fusion methodology breaks through the limitations of task-specific models. This design enables FlexBIP to adaptively process and integrate biological data of different types and from various sources, including homogeneous interactions between molecules of the same type, heterogeneous interactions between different molecular classes, as well as qualitative binary, multi-class, and quantitative regression prediction tasks. Our research has yielded exciting results. In extensive testing across 15 benchmark datasets, covering 8 major categories of biomolecular associations, FlexBIP's performance comprehensively surpasses that of 25 state-of-the-art specialized models. Crucially, in data-scarce "cold-start" scenarios that simulate the discovery of new molecules, FlexBIP continues to demonstrate remarkable robustness and predictive accuracy. Furthermore, FlexBIP provides robust and reliable interpretability for various downstream analysis tasks.

The collaborative organization of production scheduling and machine maintenance is crucial for achieving effective manufacturing. However, since these two activities are typically managed by different self-interested managers with distinct optimization needs in practice, their collaboration faces the challenge of simultaneous consideration and balanced optimization of both parties' interests. To this end, this study investigates a novel game theory-based production-maintenance collaborative scheduling problem, wherein the typical objectives individually concerned by two activities are explicitly considered and the decision interactions between activities are modeled as a non-cooperative stochastic game. Nash equilibrium within the game model provides the ideal solution for balancing the interests of both parties. To solve this problem, an imitation-enhanced alternating-training reinforcement learning method is presented. In this method, an imitation learning-based initialization mechanism is designed to accelerate the training of two game agents, and an alternating training mechanism are developed to facilitate two agents efficiently learning the optimal responses to each other's behaviors and ultimately achieving the decisions that converge to Nash equilibrium. The superiority of the proposed method is verified through comprehensive experiments. Note to Practitioners-In real-world manufacturing, coupled production scheduling and machine maintenance usually need to be collaborative organization to ensure effective production. However, the autonomous and self-interested management of these two activities makes their collaboration extremely difficult. This study is dedicated to providing innovative solutions for this collaboration dilemma. Specifically, this study innovatively models the production and maintenance collaborative scheduling problem as a two-person non-cooperative stochastic game, and designs an imitation-enhanced alternating-training reinforcement learning method to obtain the Nash equilibrium of this game effectively. Based on the fact that Nash equilibrium represents a decision combination that neither party can change to get a higher payoff, this study is naturally able to provide a mutually acceptable benefit-optimal schedule for two self-interested activities, thereby effectively solving this collaborative problem. From a practical perspective, this study offers practitioners a practical game theory-based collaborative idea that enables them to effectively achieve the collaboration between production scheduling and machine maintenance without changing the existing autonomous management organizational structure. Moreover, the proposed collaboration idea is generalizable in that it is not limited to the flexible job shop with waiting time constraints considered in this study, but can be effectively extended to other shop environments, such as flow shop and open shop.

To achieve the extreme high data rate and extreme coverage extension requirements of sixth generation (6G) wireless communication, high-frequency bands such as millimeter wave (mmW) and sub-terahertz (sub-THz) frequencies, and non-terrestrial networks (NTN), are two of the macro trends of 6G candidate technologies, respectively. However, non linearity of power amplifiers (PA) is a critical challenge for both high-frequency bands and NTN. Therefore, low peak to average power ratio (PAPR) waveform and modulation design becomes one of the most significant research topics for 6G. To maintain fifth generation (5G) New Radio (NR) backward compatibility, many waveform and modulation enhancement schemes based on DFT-s-OFDM have been proposed, including frequency domain spectrum shaping with spectrum extension (FDSS + SE), pi/4-QPSK, symbol level interleaving, and unified non-orthogonal waveform (uNOW). This paper firstly provides a general description of the waveform and modulation enhancement schemes based on DFT-s-OFDM. To analyze the factors affecting PAPR, the time domain equivalent shaped-pluses of these enhancement schemes are derived and compared. Based on the analysis, three influencing factors that reduce the PAPR of DFT-s-OFDM are analyzed and summarized. Secondly, a low-complexity scheme, asymmetric DFT-s-OFDM, achieving time-domain compression and expansion (TD-CE) based on asymmetric DFT spreading, is proposed and discussed. Thirdly, the optimal constellation rotations for BPSK and QPSK to further improve the PAPR performance of asymmetric DFT-s-OFDM are proposed. Finally, the PAPR and block error rate (BLER) performance of these 6G candidate enhancement schemes for NTN and mmW scenarios are evaluated and compared based on link-level simulation with non-ideal factors (nonlinear PA and phase noise (PN)), which reveals that asymmetric DFT-s-OFDM with QPSK rotation can achieve the best PAPR and BLER performance among all the 6G candidate enhancement schemes for both NTN and mmW scenarios. In addition to its low-complexity and backward compatibility, asymmetric DFT-s-OFDM with QPSK rotation can be considered a promising candidate low-PAPR waveform and modulation scheme for 6G.

After the academic researches in recent years, integrated sensing and communication (ISAC) is now under study for the standardization in the 3rd Generation Partnership Project (3GPP) towards the fifth generation (5G)-Advanced (5G-A) and sixth generation (6G) wireless communication systems. In this paper, to promote the possibility of successful ISAC commercialization, we focus on communication-oriented ISAC system design. We firstly introduce the standardization progress of ISAC in 3GPP, and propose the concept of communication-oriented ISAC (CO-ISAC), and discuss the technologies under CO-ISAC framework including sensing modes, sensing waveform, multiple-input-multiple-output (MIMO) scheme, multiplexing methods for sensing and communication, and reference signal (RS). Then, taking RS design as one example, we propose a common RS for sensing and communication. Pseudo-random (PN) and Zadoff-Chu (ZC) sequences in communication systems are enhanced for sensing performance improvement. For PN, Binary Phase Shift Keying with pi 2 rotation ( pi 2-BPSK) modulation is recommended via proving the low sidelobe in aper iodic ambiguity function (AF). For ZC, the closed-form expressions of aper iodic AF and peak-to-sidelobe ratio (PSLR) are derived, based on which the root index of ZC with low PSLR is designed. Evaluation results validate the gain of proposed PN and ZC sequences on sensing performance. Finally, our views on 6G Day-1 ISAC system design are provided, especially proposing that sensing technologies should be carefully designed by taking communication performance into consideration. Our research work shows the great potential of ISAC technologies to create new values even in CO-ISAC systems. In the future, more efforts on the researches, standardization and trials from academia and industry are needed to push ISAC into reality for 5G-A and 6G.

Motivation Proteins play a crucial role in biological systems, and accurate protein sequence prediction is essential for applications such as drug discovery. Existing inverse folding models primarily rely on protein backbone structure information, overlooking the biochemical properties embedded in protein surface data that constrain its functionality, leading to limited prediction accuracy.Results We propose a novel inverse folding framework, SurfFold, which integrates both protein backbone structure and surface information for sequence prediction. Additionally, it incorporates side-chain structural information and its interaction with surface information. Then, we introduce a Representation alignment module to better fuze structure and surface Representations. Experimental results demonstrate that SurfFold achieves state-of-the-art performance on the CATH4.2 dataset, and additional experiments validate the effectiveness of the proposed modules. Moreover, the homologous structure inverse folding experiment also demonstrates that SurfFold possesses excellent capability in homologous protein design.Availability and implementation The source code and data are available at https://github.com/jiudizhengf/SurfFold.

To advance the commercialization of integrated sensing and communication (ISAC) from 6G Day 1, sensing functionality should be supported with minimal impacts on communication performance. This requirement motivates the concept of communication-oriented ISAC (CO-ISAC) systems. As an important technical enabler, reference signal (RS) sequence should be designed for CO-ISAC systems by taking both communication and sensing into consideration. The widely-used ZadoffChu (ZC) sequence in communication system is a good baseline for enhancement. With various approaches for ZC sequence enhancement, existing analyses are all based on the periodic autocorrelation and cross-correlation of sequences in time domain. Additionally, questions regarding whether and how to enhance ZC sequences for CO-ISAC remain open. In this paper, a unified RS sequence based on ZC is proposed, which integrates multiple enhancement approaches. Leveraging this unified structure, the aperiodic auto-correlation is analyzed to derive a closed-form expression for peak-to-sidelobe ratio (PSLR) of each unified sequence. Furthermore, the periodic cross-correlation is derived to determine the length of zero-correlation zone (ZCZ) for a set of unified sequences. Simulation results show that by generating multiple sequences via cyclic shifts and modulation sequences based on small number of optimized root index of ZC sequences, the proposed unified RS sequences could achieve 13dB PSLR gain, 8dB integrated sidelobe ratio (ISLR) gain and 1dB peak-to-average-power ratio (PAPR) gain compared to the ZC sequences used in new radio (NR).

Predicting spatiotemporal fields governed by partial differential equations (PDEs) from sparse sensor data is a critical and long-standing challenge in science and engineering. Recent deep learning approaches, particularly neural operators, have shown considerable promise in solving PDEs. However, their performance degrades significantly in the demanding regime of extreme sparsity, characterized by spatial sensor coverage of less than 1% and limited temporal observations. To overcome this limitation, we propose a novel framework that decouples the task into two stages: spatial reconstruction and temporal extrapolation. In the first stage, rather than reconstructing the high-dimensional physical field directly, our model learns to reconstruct the complete latent features from sparse observations—features that would otherwise be extracted from a dense field. This process is stabilized by a Vector Quantization (VQ) bottleneck, which discretizes the latent space. In the second stage, a decoder-only Transformer performs temporal extrapolation by autoregressively predicting the future sequence of these discrete latent indices. This design inherently allows the model to generalize to new initial conditions and varying forecast horizons, akin to standard autoregressive models. We validate our framework on three challenging benchmarks, achieving state-of-the-art (SOTA) performance under severe sparsity constraints. Furthermore, we introduce a challenging benchmark dataset based on fire dynamics simulations. On this benchmark, our model successfully forecasts the field's evolution 30 frames into the future from a single timeframe with less than 0.1% spatial observations—a result that pushes well beyond the capabilities of existing methods.

Whisper speech recognition is crucial not only for ensuring privacy in sensitive communications but also for providing a critical communication bridge for patients under vocal restraint and enabling discrete interaction in noise-sensitive environments. The development of Chinese mandarin audio-visual whisper speech recognition is hindered by the lack of large-scale datasets. We present AISHELL6-Whisper, a large-scale open-source audio-visual whisper speech dataset, featuring 30 hours each of whisper speech and parallel normal speech, with synchronized frontal facial videos. Moreover, we propose an audio-visual speech recognition (AVSR) baseline based on the Whisper-Flamingo framework, which integrates a parallel training strategy to align embeddings across speech types, and employs a projection layer to adapt to whisper speech’s spectral properties. The model achieves a Character Error Rate (CER) of 4.13% for whisper speech and 1.11% for normal speech in the test set of our dataset, and establishes new state-of-the-art results on the wTIMIT benchmark. The dataset and the AVSR baseline codes are open-sourced at https://zutm.github.io/AISHELL6-Whisper.

Motivation Understanding cellular states and regulatory programs requires integrative analysis of multiple omics layers. Although recent spatial sequencing technologies allow molecular profiling of cells within their tissue context, paired spatial multi-omics assays are still limited by technical complexity and cost. This creates a pressing need for diagonal integration methods that enable joint analysis of unpaired spatial omics datasets.Results We propose SIVA, a deep generative framework based on Spatially-Informed Variational Autoencoders with Anchor Guidance, for diagonal integration of spatial multi-modal data. SIVA employs modality-specific variational autoencoders (VAEs) with a hybrid latent embedding that integrates Gaussian process and standard Gaussian priors, enabling joint modeling of spatially structured variation and dominant underlying data distributions across modalities. To facilitate cross-modal alignment in the absence of one-to-one cell correspondence, SIVA adopts a dual integration strategy combining global distribution alignment via Maximum Mean Discrepancy and local correspondence guidance using mutual nearest neighbor anchors. Extensive experiments across multiple cross-slice integration scenarios demonstrate that SIVA achieves robust and accurate integration of unpaired spatial omics datasets, consistently outperforming existing methods.Availability and implementation The source codes are available at https://github.com/PelenJiang/SIVA

Audio-visual target speaker extraction should return the speaker indicated by the video, yet a separator can ignore the visual cue and repeatedly output the acoustically dominant voice. We introduce REAL-2MIX, a Mandarin AV-TSE benchmark of jointly recorded real two-speaker mixtures with synchronized multi-view video. Each scene also contains preceding A-only and B-only stages that provide in-scene speaker references. It contains 77 scenes and 7,598 clips (11.84 hours), including 6,042 dual-annotated mixtures. After processing there leave 6,038 evaluable mixtures and 12,076 target-speaker rows. We additionally curate VOXBLINK2-AVSE from VoxBlink2, comprising 250,828 synchronized audio–lip-ROI pairs from 28,421 identities and 766.17 hours of speech. Our extractor uses frozen, 1,280-dimensional projected AV-HuBERT features, target-conditioned training, and layer-wise feature modulation. We jointly evaluate content with Qwen3-ASR-1.7B CER and target identity with WeSpeaker ResNet34 plus Overlapped Speech Detection (OSD). On the complete manifest, the best archived checkpoint obtains 0.2261 CER, 82.22

Current research in Electrocardiogram (ECG) biometrics mainly emphasizes resting-state conditions, leaving the performance decline in rest-exercise scenarios largely unresolved. This paper introduces CrossStateECG, a robust ECG-based authentication model explicitly tailored for cross-state (rest-exercise) conditions. The proposed model creatively combines multi-scale deep convolutional feature extraction with attention mechanisms to ensure strong identification across different physiological states. Experimental results on the exercise-ECGID dataset validate the effectiveness of CrossStateECG, achieving an identification accuracy of 92.50% in the Rest-to-Exercise scenario (training on resting ECG and testing on post-exercise ECG) and 94.72% in the Exercise-to-Rest scenario (training on post-exercise ECG and testing on resting ECG). Furthermore, CrossStateECG demonstrates exceptional performance across both state combinations, reaching an accuracy of 99.94% in Rest-to-Rest scenarios and 97.85% in Mixed-to-Mixed scenarios. Additional validations on the ECG-ID and MIT-BIH datasets further confirmed the generalization abilities of CrossStateECG, underscoring its potential as a practical solution for post-exercise ECG-based authentication in dynamic real-world settings.

The color appearance of pathological images varies due to differences in staining or imaging conditions such as protocols, dye ratios, and scanners, which limits the practicality of AI-aided pathological diagnosis. Based on the fact that pathological images are composed of color and structural information, we propose a template-based staining normalization method called StainPIDR, which first decouples the structural and color information contained in the pathological image and then reconstructs it using the template color mapping. In order to eliminate the influence of noise on the decoupled color information, we construct a color codebook in StainPIDR to ensure consistent color mapping under the same staining conditions. In order to select the most suitable template image from a set of reference images, we further propose a simple but effective template selection strategy. The comparison and ablation experimental results demonstrate the effectiveness of StainPIDR and our template selection strategy. The source code will be available on GitHub after acceptance.

Abstract Motivation Enzyme-reaction retrieval is increasingly used to prioritize candidate biocatalysts for experimental follow-up, where useful recommendations should indicate not only whether an enzyme can catalyze a target reaction but also under which pH and temperature conditions it should be tested. Existing retrieval models optimize catalytic matching scores, whereas catalytic optima predictors are typically developed as enzyme-level regressors because public pH and temperature annotations are sparse and often available only at the enzyme or EC-associated record level. This separation leaves a practical gap: high-ranking enzyme-reaction pairs are not evaluated for condition suitability, and enzyme-level optima predictions do not use the reaction context being retrieved. Results We present GERO, a multimodal fusion framework that uses feature-gated cross-modal fusion to integrate global enzyme sequence semantics, sequence-derived pocket geometry, and molecular reaction representations for condition-aware enzyme-reaction retrieval and catalytic optima estimation with reaction context. To evaluate this setting, we define the tolerance-restricted hit rate (Hit@k-TR), which requires both top-k retrieval of the correct candidate and condition prediction within predefined tolerances. Across enzyme- and reaction-similarity splits, GERO improves Hit@k-TR over two-stage retrieval-then-prediction baselines. Representative benchmark examples and an iodinin biosynthesis case study further illustrate GERO’s ability to provide candidate rankings together with plausible assay-condition estimates for downstream experimental prioritization. Availability and implementation Source code is available at https://github.com/ykxhs/GERO . Contact liujuan@whu.edu.cn Supplementary information Supplementary data are available at XXXX online.