Although Vision Language Models (VLMs) have shown generalization in medical imaging, pathology presents unique challenges due to ultra-high resolution, complex tissue structures, and nuanced semantics. These factors make pathology VLMs prone to hallucinations, i.e., generating outputs inconsistent with visual evidence, which undermines clinical trust. Existing RAG approaches in this domain largely depend on text-based knowledge bases, limiting their ability to leverage diagnostic visual cues. To address this, we propose Patho-AgenticRAG, a multimodal RAG framework with a database built on page-level embeddings from authoritative pathology textbooks. Unlike traditional text-only retrieval systems, it supports joint text–image search, enabling retrieval of textbook pages that contain both the queried text and relevant visual cues, thus avoiding the loss of critical image-based information. Patho-AgenticRAG also supports reasoning, task decomposition, and multi-turn search interactions, improving accuracy in complex diagnostic scenarios. Experiments show that Patho-AgenticRAG significantly outperforms existing multimodal models in complex pathology tasks like multiple-choice diagnosis and visual question answering.

Recent advances in vision-language models (VLMs) have enabled broad progress in the general medical field. However, pathology still remains a more challenging sub-domain, with current pathology-specific VLMs exhibiting limitations in both diagnostic accuracy and reasoning plausibility. Such shortcomings are largely attributable to the nature of current pathology datasets, which are primarily composed of image–description pairs that lack the depth and structured diagnostic paradigms employed by real-world pathologists. In this study, we leverage pathology textbooks and real-world pathology experts to construct high-quality, reasoning-oriented datasets. Building on this, we introduce Patho-R1, a multimodal RL-based pathology Reasoner, trained through a three-stage pipeline: (1) continued pretraining on 3.5 million image-text pairs for knowledge infusion; (2) supervised fine-tuning on 500k high-quality Chain-of-Thought samples for reasoning incentivizing; (3) reinforcement learning using Group Relative Policy Optimization and Decoupled Clip and Dynamic sAmpling Policy Optimization strategies for multimodal reasoning quality refinement. To further assess the alignment quality of our dataset, we propose Patho-CLIP, trained on the same figure-caption corpus used for continued pretraining. Comprehensive experimental results demonstrate that both Patho-CLIP and Patho-R1 achieve robust performance across a wide range of pathology-related tasks, including zero-shot classification, cross-modal retrieval, Visual Question Answering, and Multiple Choice Question.

Spatial transcriptomics (ST) technologies have transformed our understanding of tissue biology by capturing gene expression with spatial context, enabling systematic analysis of cell-cell interactions (CCIs) and spatial domains in complex tissues. However, existing computational approaches often rely on fixed proximity graphs, curated ligand-receptor (LR) databases, or deep graph neural networks that are prone to over-smoothing and lack principled uncertainty quantification. These limitations hinder the discovery of heterogeneous, directional, and long-range CCIs essential for interpreting tissue organization and disease mechanisms. Here, we present B-HGME (Bayesian Hyperspherical Graph Mixture of Experts), a scalable, unsupervised framework that jointly delineates spatial domains and infers CCI networks from ST data with principled uncertainty estimates. B-HGME integrates spatial and gene-regulatory graphs into a dual-scale structure, encodes cell representations on a unit hypersphere via coupled message passing, and decodes edges using a Bayesian mixture-of-experts governed by a Dirichlet-regularized gating network. This design enables the model to capture multi-scale, directional, and biologically coherent interactions while avoiding the over-smoothing and posterior collapse of conventional models. The hyperspherical embedding geometry ensures angular similarity is preserved in high dimensions, and edge-level credibility is derived from the Bayesian posterior, facilitating interpretable and confident CCI inference. Across multiple datasets from six major ST platforms, B-HGME consistently achieves state-of-the-art spatial clustering accuracy and uncovers biologically coherent and diverse CCIs, including novel interactions beyond curated ligand-receptor pairs. B-HGME's hyperspherical embeddings accurately localize canonical astrocytic and laminar markers (e.g.,Gfap,Pcp4,Calb1, andCamk2a) to their expected spatial niches, confirming biochemical fidelity at single-gene resolution. Inferred ligand-receptor circuits not only recover known pathways but also reveal previously uncharacterized interactions (e.g., Astro1-eL2/3, VIP-Oligo), furnishing mechanistic hypotheses for cortical layer formation and tumor-stroma crosstalk. Together, these results demonstrate that B-HGME offers a powerful tool for spatial systems biology and hypothesis generation in development, immunity, and cancer. The source code of our model is available athttps://github.com/zxj8806/B-HGME.

Link prediction (LP) is fundamental to graph-based applications, yet existing graph autoencoders (GAEs) and variational GAEs (VGAEs) often struggle with intrinsic graph properties, particularly the presence of negative eigenvalues in adjacency matrices, which limits their adaptability and predictive performance. To address this limitation, we propose Hyperspherical Kolmogorov-Arnold Networks for LP (HKANLP), a novel framework that combines multiple graph neural network (GNN)-based representation learning strategies with Kolmogorov-Arnold networks (KANs) in a hyperspherical embedding space. Specifically, our model leverages the von Mises-Fisher (vMF) distribution to impose geometric consistency in the latent space and employs KANs as universal function approximators to reconstruct adjacency matrices, thereby mitigating the impact of negative eigenvalues and enhancing spectral diversity. Extensive experiments on homophilous, heterophilous, and large-scale graph datasets demonstrate that HKANLP achieves superior LP performance and robustness compared to state-of-the-art baselines. Furthermore, visualization analyses illustrate the model's effectiveness in capturing complex structural patterns. The source code of our model is publicly available at https://github.com/zxj8806/HKANLP/.

Deep neural networks (DNNs) have demonstrated exceptional performance across a variety of applications, yet they require substantial computing and power resources. In contrast, Spiking Neural Networks (SNNs) offer significant potential for energy-efficient computing due to their binary, event-driven properties. However, existing deep generative SNNs often struggle to produce high-quality low-dimensional representations in latent space, adversely affecting the quality of their generated samples. To address this limitation, we introduce a novel generative model that integrates the principles of variational autoencoders (VAEs) with adversarial training techniques. Our model consists of a generative module based entirely on an SNN-structured VAE and a discriminator employing artificial neural networks. This discriminator enhances the training process of our SNN-based generative model by applying adversarial principles similar to those used in Generative Adversarial Networks (GANs). Additionally, we have developed a conditional generation capability within our model, enabling the controlled production of specific images based on label inputs. Experimental evaluations on multiple datasets demonstrate that our model achieves superior image quality compared to existing top-performing SNN-based generative models. The source code of our model is accessible at https://github.com/zxj8806/SGM-VaGAN.

In this work, we propose a novel Generalized Zero-Shot Learning (GZSL) approach that integrates out-of-distribution (OOD) detection mechanisms into the architecture of generative models. At the core of our method is a dual-stream Variational Autoencoder (VAE), which creates a unified latent space to align data distributions from different modalities. This alignment helps our model better identify unseen classes by reducing bias using class-specific distribution ranges. We further improve the model by adding contrastive learning and a classification loss to the latent space, which separates features of seen and unseen classes and reduces feature confusion. To enhance the quality of generated data, we combine the dual-stream VAE with a Generative Adversarial Network (GAN), enhancing the quality of synthetic data and improving the model’s robustness against variations in data distribution. Extensive validation on popular benchmarks has confirmed the effectiveness of our approach, demonstrating notable improvements in GZSL performance.