Federated learning (FL) has expanded the scope of distributed machine learning by training a global model through the amalgamation of knowledge from various distributed clients. However, the challenge of non-i.i.d. local datasets among clients introduces statistical heterogeneity, thereby compromising the performance of a single global model. Personalized federated learning (PFL) seeks to mitigate this issue by training individualized models tailored for each client. Despite the improved performance achieved by PFL, current approaches lack interpretability in quantifying the significance of inter-client contributions. Furthermore, they often pursue individual models at clients while performing model aggregation to obtain a single global model at the FL server, which still struggles to effectively capture the distinctions among clients. In response to these problems, we propose the personalized Federated learning with Transparent Inter-client Contribution (pFedTIC.) approach. Specifically, pFedTIC. embraces a dual-level personalized model aggregation strategy to derive an individualized model for each client, and assesses inter-client contribution with the aggregation weights for enhancing transparency. It conducts model aggregation at both the server and client sides using the similarity of model parameters and learning representations, mutual contributions between clients, and model performance. Theoretical analyses are derived to establish the non-convex convergence rate of pFedTIC.. Experimental results on benchmark datasets demonstrate its superiority over competitive methods, affirming its effectiveness in handling data heterogeneity. Moreover, the experiments also showcase its capability to interpret inter-client contributions, fostering more transparent FL collaborations.

Alternatively spliced isoforms from the same gene can perform distinct functions; however, their cell-type-specific roles remain largely uncharacterized, limiting our ability to understand cellular diversity and development beyond traditional gene-level analyses. We present cIsoFun, a multi-modal fusion framework for cell-type-specific isoform function prediction from single-cell transcriptomics data. cIsoFun leverages pre-trained ESM-2 and BERT models to extract initial sequence features, constructs a multiplex heterogeneous network over genes, isoforms, GO terms, and cell types to represent their complex relationships, and applies relation-aware attention to integrate multi-modal information and refine node embeddings. It then optimizes a multi-component loss on the updated embeddings to predict isoform functions, enabling biological interpretability via sequence-importance and cell-type-specific analyses. Experiments demonstrate that cIsoFun outperforms existing methods, particularly for sparse GO terms, and reveal distinct functional programs across contexts: kidney tumor cells are enriched for metabolism and growth regulation, skin tumor cells emphasize immune surveillance and migration, and cell lines prioritize DNA repair and telomere maintenance. Sequence-importance analysis highlights critical amino-acid regions and shows that domains annotated with the same function can exhibit distinct importance profiles across spliced isoforms. Together, these results provide new insights into cell-type-specific isoform functionality and establish cIsoFun as a practical tool for single-cell isoform analysis. Code and datasets are available at www.sdu-idea.cn/codes.php?name=cIsoFun.

New Intent Discovery (NID) plays a pivotal role in dialogue systems by identifying known intents and discovering novel ones from unlabeled data. While recent methods leveraging Large Language Models (LLMs) have improved performance, they often suffer from prohibitive computational costs due to redundant, exhaustive queries. To address this, we propose Adaptive Spatio-Semantic Calibration (ASSC). Unlike previous approaches relying on indiscriminate verification, ASSC introduces a Spatial Uncertainty Identification (SUID) mechanism acting as a reliability-aware filter to detect ambiguous samples at decision boundaries. We then selectively apply Adaptive Semantic Calibration (ASC), leveraging LLMs to semantically refine the assignments of only these high-uncertainty instances. Extensive experiments on multiple challenging benchmarks demonstrate ASSC achieves consistent performance improvements over competitive baselines. Furthermore, our method reduces token consumption by approximately 80

Federated learning (FL) has expanded the scope of distributed learning by training models through the combination of knowledge from various dispersed data sources. Traditional FL methods, which rely on synchronous communication, often suffer inefficiencies due to the heterogeneous nature of client devices. To address these limitations, asynchronous federated learning (AFL) allows clients to communicate with the central server at their own pace, enhancing flexibility and scalability. However, AFL struggles with efficiency and performance in scenarios involving heterogeneous clients with non-i.i.d. data and varied model architectures, which is a pervasive but under-studied framework. We formalize this problem as a new learning topic called Asynchronous Federated Learning with Data and Model Heterogeneity (AFL-DMH), and introduce the Federated Asynchronous and Personalized Prototype aggregation (FedAPP) method to tackle it. Specifically, FedAPP leverages shared prototypes to facilitate effective knowledge transfer among clients. It introduces the asynchronous and personalized prototype aggregation strategy that performs aggregation based on participant delay, models the staleness and quality of local prototypes and adopts the attention-based aggregation to generate individualized prototypes for each client, providing them with collective knowledge and thereby boosting local model training. Theoretical analyses are conducted to establish the non-convex convergence rate of FedAPP. Experiments on benchmark datasets demonstrate the superior performance and efficiency of FedAPP compared to competitive baselines, affirming its effectiveness in addressing asynchronous communication, heterogeneous data and model challenges.

Abstract Spatial transcriptomics enables high-resolution profiling of gene expression within spatial contexts, yet its potential is often hindered by fragmented toolchains, intricate parameters, and cognitive bottlenecks of interpreting high-dimensional data. While recent Large Language Model agents have attempted to automate this process, they remain constrained by rigid execution logic, lack multimodal feedback for self-correction, and operate in epistemic isolation from established biological knowledge. Here, we present STAnalyzer, an intelligent multi-agent framework designed to automate the end-to-end analytical lifecycle—from raw data processing to biological hypothesis generation. Transcending traditional pipelines, STAnalyzer employs a collaborative intelligence architecture to achieve three core capabilities: (1) Intent-Driven Orchestration , which dynamically translates natural language queries into rigorous bioinformatics workflows; (2) Multi-Modal Self-Refinement , which autonomously ensures analytical robustness through closed-loop synthesis of evidence from visual patterns and statistical metrics; and (3) Evidence-based Cross-Validation , which bridges the gap between data-driven correlations and biological causation by anchoring findings in ground-truth literature and structured databases. By eliminating manual analytical bottlenecks and ensuring rigorous evidentiary traceability and transparency, STAnalyzer makes high-resolution spatial omics more accessible to a broader research community. It provides a robust and scalable framework for cross-platform automated analysis and accelerated biological discovery, translating massive spatial datasets into verifiable biological insights.

Fairness-aware machine learning aims to build predictive models that comply with fairness requirements, particularly concerning sensitive attributes such as race, gender, and age. Among causality-based fairness notions, counterfactual fairness is widely adopted for its individual-level guarantees, requiring that an individual’s predicted outcome remains unchanged in a counterfactual world where its sensitive attribute is altered. However, existing methods critically assume that the true causal graph is fully known, which is rarely the case in practice. Moreover, counterfactual fairness suffers from inherent identifiability limitations, as counterfactual quantities cannot always be uniquely estimated from observational data, especially under incomplete causal knowledge. To address these challenges, we propose a principled framework (CF-ICG) for counterfactual fairness under imperfectly known causal graphs, e.g., Completed Partially Directed Acyclic Graphs (CPDAGs). We first introduce a criterion to determine the identifiability, and bound the counterfactual quantities under CPDAGs. Building upon this, we develop an efficient local algorithm that avoids the exhaustive enumeration of all DAGs, ensuring robustness against worst-case fairness violations. Experimental results on synthetic and real-world datasets demonstrate the practical effectiveness and theoretical soundness of CF-ICG.

With growing concerns over user privacy, federated recommendation (FedRec) has emerged as a mainstream solution for personalized recommendation services. FedRec trains user-private parameters on local clients while collaboratively updating global parameters on a centralized server. However, despite advances in optimizing these local and global parameters, existing methods overlook two key challenges: tradeoff training and distribution discrepancy. Tradeoff training balances timely local updates with diverse global parameters, limiting the model's learning ability. Distribution discrepancy arises from the divergence between locally trained global parameters and those aggregated by the server, corrupting inference performance. To fill in the gap, we propose FedSC, a principled federated recommendation framework that boosts FedRec's training and inference processes with minimal yet nontrivial efforts. During training, FedSC employs a stochastic aggregation strategy where all users participate in every round, while only a random subset is selected for aggregation, preserving the diversity of global parameters and ensuring timely local updates. During inference, FedSC makes recommendations with a consistency inference mechanism that uses the most recent locally trained global parameters of each user to improve the model's understanding of user preferences. Extensive experiments on multiple benchmark datasets demonstrate the superiority of FedSC, achieving up to a 20% improvement in most evaluation scenarios.

Epistasis plays a crucial role in explaining missing heritability in complex traits, yet most detection methods fail to effectively capture local and global interaction patterns and long-range dependencies critical to complex trait architecture. We propose Epiformer that leverages genome language model Evo 2 to capture long-range dependencies from genetic background, and a dual-channel network to jointly model local and global epistasis and additive effects. Epiformer can identify key SNPs and their interactions from complex genomic data to empower phenotype prediction with interpretability, which reinforces epistasis detection. Epiformer performs robustly across species, revealing biologically meaningful patterns and offering new insights into genetic architecture.

The clinical usage of Traditional Chinese Medicines (TCMs) gains increasing international attention for their distinct therapeutic effects. Precisely understanding the herb (TCM) toxicity, and identifying toxic and effective herb ingredients are crucial for their safe use. Traditional wet-lab based pipelines for assessing herb toxicity are complex and time-consuming. Given the large volume toxicity data of ingredients, building computational models for efficient ingredient-based toxicity prediction and evaluation is promising, but often fails to effectively predict the toxicity of herbs, due to the complexity herbs and toxic mechanisms. To address these challenges, we propose a heterogeneous network based approach (HerbToxNet) for predicting herb toxicity. HerbToxNet first constructs a heterogeneous network composed with herbs, ingredients and molecular targets, and leverages a heterogeneous graph attention network to learn the representation of herbs. Meanwhile, it performs contrastive learning with dynamic coefficient to refine the representation by pulling close the herbs with shared toxic labels, while pushing away the others. Next, it uses Multilayer Perceptron (MLP) on the herb representation to predict the toxic labels of this herb, and further introduces a weighted label fusion strategy that uses toxic labels of similar herbs to augment the predicted labels of this herb. HerbToxNet outperforms competitive methods and finds out novel potential toxicities, with 96% toxicity labels confirmed for canonical herbs. It can mine related toxic ingredients and targets in an interpretable way, and dissect the molecular mechanism of herb toxicity with authenticity.

Machine learning, viewed through a causal lens, has sparked a strong interest in the pursuit of fairness. Counterfactual fairness, a popular notion within this framework, evaluates the fairness of a model by comparing its predictions for each individual in the real world with those in counterfactual scenarios, where sensitive attributes (e.g., race, gender, age) are modified. Existing methods for learning counterfactual fairness models rely on having access to complete and clean data. However, real-world data are often of low quality, frequently affected by issues such as missing values. These challenges hinder the training of accurate causal models and mislead the development of counterfactual fairness models. In this paper, we propose the Counterfactual Fair prediction over Missing Data (CFairMD) method to bridge this important gap. CFairMD proposes a causal importance-weighted autoencoder to impute missing data and perform counterfactual inference. It integrates missing data within a joint model framework that captures the relationships between covariates and outcomes. Additionally, we design a counterfactual symmetric function to generate fair representations, complemented by a causally invariant penalty. Theoretical analysis shows that CFairMD achieves counterfactual fairness guarantees in a scalable and flexible manner. Extensive experimental results on four benchmarking datasets show that CFairMD outperforms eight state-of-the-art approaches, with over 22% higher counterfactual fairness gain on average.

Graph clustering, which aims to divide nodes in the graph into several distinct clusters, is a fundamental yet challenging task. Benefiting from the powerful representation capability of deep learning, deep graph clustering (DGC) methods have achieved great success in recent years. However, the corresponding survey paper is relatively scarce, and it is imminent to make a summary of this field. From this motivation, we conduct a comprehensive survey of DGC. Firstly, we introduce formulaic definition, evaluation, and development in this field. Secondly, the taxonomy of DGC methods is presented based on four different criteria, including graph type, network architecture, learning paradigm, and clustering method. Thirdly, we carefully analyze the existing methods via extensive experiments and summarize the challenges and opportunities from five perspectives, including graph data quality, stability, scalability, discriminative capability, and unknown cluster number. Besides, the applications of DGC methods in six domains, including computer vision, natural language processing, recommendation systems, social network analysis, bioinformatics, and medical science, are presented. Last but not least, this paper provides open resource supports, including 1) a collection of state-of-the-art DGC methods (papers, codes, and datasets) and 2) a flexible and extensible Python library for DGC. We hope this work can serve as a quick guide and help researchers overcome challenges in this vibrant field.

The safety and modernization of traditional Chinese medicine (TCM) are significant concerns for human beings. Recently, the adverse effects caused by the use of certain TCMs have been frequently reported. Although TCMs may cause toxic reactions, they also play critical roles in treating multiple complex diseases. Therefore, toxicity research is urgently needed for the safe usage of TCMs. However, existing databases for TCMs primarily focus on the pharmacological effects of TCMs, with limited attention to the toxicity. They neither distinguish the toxic effects of formulas, herbs, and ingredients, nor classify and summarize targets for specific toxic manifestations, or assemble evidence from previous studies. We developed TCMToxDB, a comprehensive database that focuses on the toxicity and safe usage of TCMs. TCMToxDB systematically integrates and analyses the research results of toxic TCMs, offering users diverse information acquisition and analysis services. In addition, it assembles five canonical herb-target and ingredient-target interaction prediction algorithms with different advantages, which support the prediction of toxic targets of herbs and ingredients to empower the toxicity research of TCMs and to meet users' personal needs. TCMToxDB is accessible at https://www.sdu-idea.cn/TCMToxDB.

Recently, communication learning has emerged as a successful tool for tackling complex tasks in multi-agent reinforcement learning (MARL). Many MARL methods use graph neural networks (GNNs) to build a communication learning framework where agents and communication channels can be represented as nodes and edges in a graph. However, most GNN-based MARL methods simply aggregate features of neighboring agents to obtain message representations, which may not extract enough useful information. To tackle this problem, this paper investigates how to extract expressive information from neighboring agents to obtain high-quality message representations. Inspired by the recent success of contrastive learning methods, in this paper, we propose a multi-agent communication protocol via graph contrastive learning (MAGE), which utilizes contrast objectives to learn optimal message representations, considering feature and topological level. At the feature level, we corrupt agent features by adding more noise to insignificant neighboring agent features, to encourage the agent to recognize significant information. At the topological level, we adaptively remove edges by assigning larger removal probabilities to insignificant edges to highlight significant communication structures. Experiments across diverse benchmarks confirm that MAGE outperforms existing GNN-based MARL methods.

Effective communication is crucial for multi-agent cooperation in partially observable environments. However, a fundamental trilemma exists among task performance, communication efficiency, and human interpretability. To resolve this, we propose a multi-agent communication framework via $\textbf{G}$rounding $\textbf{L}$anguage and $\textbf{C}$ontrastive learning (GLC) to learns efficient and interpretable communication protocols. Specifically, GLC employs an autoencoder to learn discretized and compressed communication symbols, ensuring high communication efficiency. These symbols are then semantically aligned with human concepts using data generated by a Large Language Model (LLM), making them human-interpretable. Furthermore, a contrastive learning objective is introduced to ensure consistency and mutual intelligibility among all agents, thereby securing high task utility. GLC dynamically balances these objectives by the Information Bottleneck principle. Extensive experiments show that GLC outperforms state-of-the-art methods across multiple benchmarks, delivering superior task performance, higher communication efficiency, and enhanced human interpretability.

Proteoforms translated from alternatively spliced transcripts contribute to the functional repertoire of the cell by performing diverse biological functions, contributing to the functional diversity of genomics and proteomics. However, the lack of existing databases that integrate functional annotations of proteoforms, and analyse the drivers of their functional differences significantly hinders in-depth research into proteoform functions. We introduce ProteoformDB, a new web resource with integrated in-platform analytical capabilities, organizes transcript-level functional annotations of proteoforms across multiple species, and provides services for prediction of proteoform functions and analysis of functional regulatory mechanisms. ProteoformDB develops user-friendly interfaces for information search, visualization, function supplement, differential analysis, and data download services. Particularly, it enables users to investigate the impact of molecular events on the function of proteoforms at multiple levels, including sequences, domains, and post-translational modifications, among others, thereby uncovering the functional differences between protein variants. The current version includes processed data (154.83 GB) for 214 animal and 28 plant species, and will become a valuable and expandable proteoform functional resource for studying genome and transcriptome functions, disease mechanisms, and other related research.

Multiple clustering aims to uncover diverse latent structures within the data, enabling a more comprehensive understanding of complex datasets. However, existing approaches either heavily rely on user-supplied keywords or disregard user-interested clustering types, limiting the ability to discover the full range of explainable clusterings of interests, particularly in high-dimensional settings. Furthermore, existing methods insufficiently leverage the rich textual semantics and fall short in fully integrating multi-modal information. To address these challenges, we propose MLLM enriched Multiple Clustering (MLLMMC), a novel framework that leverages multi-modal large language model (MLLM) to explore explainable non-redundant clustering. Specifically, MLLMMC first employs MLLM to generate sample descriptions, which serve as input for LLM to perform prompt-driven reasoning and infer latent clustering types, and then merges them with user-interested types to obtain diverse and explainable clustering types. For each selected type, MLLMMC utilizes MLLM to generate sample-level textual descriptions and aligns them with corresponding visual features through a cross-attention fusion module, which produces a semantically aligned and enriched representation for the target clustering type. Extensive experiments on six benchmark datasets from diverse domains demonstrate that MLLMMC achieves diverse, explainable, and high-quality clustering outcomes, outperforming state-of-the-art multiple clustering methods with a large margin.

Learning the dynamic causal structure is a difficult challenge in discovering causality from time series. Most existing studies rely on distributional or structural invariance to uncover the underlying causal dynamics, assuming stationary or partially stationary causality, which frequently conflicts with complex causal relationships in the real world. This boosts temporal causal discovery to encompass fully dynamic causality, where both instantaneous and lagged causal dependencies may change over time, bringing significant challenges to the efficiency and stability of causal discovery. To tackle these challenges, we introduce DyCausal, a dynamic causal structure learning framework that leverages convolutional networks to effectively model causal structures within coarse-grained time windows, and introduces linear interpolation to refine causal structures to each time step and recover time-varying causal graphs. In addition, we propose an acyclic constraint based on matrix norm scaling. It is more stable both theoretically and empirically, and constrains loops in dynamic causal structures with improved efficiency. Evaluations on both synthetic and real-world datasets prove that DyCausal significantly outperforms existing methods and identifies fully dynamic causal structures from coarse to fine.

Cancer survival analysis with multimodal data is crucial for precise treatments and patient benefits. However, the following challenges prohibit integrating histopathology and genomics: (i) multimodal data is not always complete, especially for the more costly genomics data; (ii) intricate interactions between different modalities are difficult to capture and understand. To response, we propose an end-to-end framework (CIMA) that coordinates Cyclic modality generation and Multi-grained multimodal Alignment. Specifically, CIMA designs a cyclic modality reconstruction module to reciprocally impute missing modalities and infer the interactions between them. Next, it introduces the multi-grained alignment module over the imputed data and interactions to mine fine-grained alignments between histopathology (slide patches) and genomics (biological pathways). CIMA then constructs the adaptive fusion module to leverage multimodal data and alignments for survival prediction. Extensive experiments on cancer benchmark datasets demonstrate that CIMA outperforms existing methods and exhibits good interpretability, providing valuable insights into intricate relationships between pathological phenotypes and biological pathways.Our code is released in the supplementary materials.

Federated prompt learning (FPL) offers a parameter-efficient way to adapt large-scale vision-language models by collaboratively optimizing lightweight prompts across distributed clients. Existing approaches typically assume that prompts learned on heterogeneous domains can be safely aggregated. We show that this assumption breaks down in cross-domain federated settings with domain and model heterogeneity: naive global averaging induces cross-domain prompt interference and can even underperform simple intra-domain collaboration. Through geometric and optimization analyses, we find that prompts from different domains are semantically separated and exhibit low gradient alignment, which explains their poor compatibility under global aggregation. Motivated by these observations, we propose DaFPL, a domain-aware FPL framework that decomposes each client prompt into two complementary parts: a domain-transferable component that captures domain-invariant semantics and a domain-specific component that preserves domain-dependent characteristics. We introduce a domain discriminator to enforce the intended functional roles of the two components. In addition, we design a domain-aware aggregation strategy that aggregates domain-transferable prompts globally through an entropy-based invariance gate, while aggregating domain-specific prompts only within each domain. We further provide a non-convex convergence analysis of DaFPL. Extensive experiments on four multi-domain benchmarks demonstrate that DaFPL consistently outperforms strong FPL baselines.