A fundamental question in immunology is how the adaptive immune system encodes antigen specificity while maintaining repertoire diversity. B cell receptor (BCR) or antibody clonal families, defined by groups of B cells descending from a common ancestor, are key to deciphering this encoding. Although paired heavy and light chains jointly determine antibody specificity, most repertoire analyses have historically relied on heavy-chain-only data due to the loss of native pairing information in bulk BCR sequencing. This reliance introduces potential biases in computational clonal cluster inference, which may complicate efforts to resolve disease-associated immune signatures. Here, we leverage large-scale paired-chain BCR sequencing data to demonstrate that heavy-chain-based clustering may misrepresent true clonal architecture, and identify two major artifacts: chain-mixed clusters, in which similar heavy chains are paired with distinct light chains, and naive-like pseudo-clonal clusters, which are detected in an individual's naive B cell repertoire and exhibit highly similar heavy and light chains without reflecting true clonal expansion. To address these limitations, we present fastBCR-p, an optimized framework that integrates light-chain-informed subclustering, with public sequence aware refinement to improve clonal family inference. By resolving both technical artifacts and biological convergence, fastBCR-p improves the chain concordance and overall clustering quality of clonal inference in real-world datasets. This enables more accurate tracking of immune dynamics in health and disease and facilitates the identification of clinically relevant antibody lineages.

Biological circuits have evolved to incorporate multiple modules that perform similar functions. In the fly olfactory circuit, both lateral inhibition (LI) and neuronal spike frequency adaptation (SFA) are thought to enhance pattern separation for odor learning. However, it remains unclear whether these mechanisms play redundant or distinct roles in this process. In this study, we present a computational model of the fly olfactory circuit to investigate odor discrimination under varying noise conditions that simulate complex environments. Our results show that LI primarily enhances odor discrimination in low‑ and medium‑noise scenarios, but this benefit diminishes and may reverse under higher‑noise conditions. In contrast, SFA consistently improves discrimination across all noise levels. LI is preferentially engaged in low‑ and medium‑noise environments, whereas SFA dominates in high‑noise settings. When combined, these two sparsification mechanisms enable optimal discrimination performance. This work demonstrates that seemingly redundant modules in biological circuits can, in fact, be essential for achieving optimal learning in complex contexts.

Low-Rank Adaptation (LoRA) is a widely used parameter-efficient fine-tuning method for foundation models, but it suffers from parameter interference, resulting in suboptimal performance. Although Mixture-of-Experts (MoE)-based LoRA variants show promise in mitigating intra-task correlations in single-task instruction tuning, they introduce additional router parameters and remain ineffective in multi-task model merging where inter-task interference arises. Inspired by the fly olfactory circuit, we propose FlyLoRA, an implicit MoE-based LoRA variant that introduces: (1) rank-wise expert activation in the up-projection matrix, and (2) an implicit router that unifies expert routing and down-projection, where a frozen sparse random projection matrix replaces the traditional dense trainable version. This design resolves the trade-off between intra-task decorrelation and computational efficiency by eliminating the need for an explicit router, while inherently mitigating inter-task interference due to the orthogonality property of random matrices. Extensive experiments across four domains---general knowledge understanding, scientific question answering, mathematical reasoning, and code generation---demonstrate consistent performance improvements over existing methods. Beyond empirical gains, FlyLoRA highlights how biological structures can inspire innovations in AI technologies. Code is available at https://github.com/gfyddha/FlyLoRA.

ABSTRACT Image data is often degraded during transmission due to hardware limitations or human error, which may hinder subsequent image analysis tasks. Therefore, research on image restoration has significant practical value. Traditional matrix‐based algorithms struggle with high‐dimensional data, often failing to preserve spatial structures and risking overfitting. In this paper, we investigate tensor recovery problems under the tensor singular value decomposition framework. We introduce a non‐convex surrogate for the tensor rank—the tensor truncated Schatten‐ norm—and propose two recovery models based on this theory: a tensor completion model and a tensor robust principal component analysis model. Efficient solutions based on the alternating direction method of multipliers are developed for both models. Moreover, we provide a thorough analysis of the computational complexity and convergence behavior of our algorithms. At last, extensive experiments on synthetic data, color images, video sequences, multispectral images, and medical images demonstrate the effectiveness and robustness of the proposed methods.

Brain-computer interface (BCI) technology for emotion recognition holds significant potential for future applications in the treatment of refractory emotional disorders. Stereo-electroencephalography (SEEG), being less invasive, can precisely record neural activities originating from the cortex and the deep structures of the brain. Thus, it has broad application prospects in constructing emotion recognition BCI. In this study, SEEG data from nine subjects were collected to construct an emotion dataset, and a Spatial Transformer-based Hybrid Network (STHN) was proposed for SEEG emotion recognition. The triple-classification accuracy of STHN reached 83.56%, outperforming the baseline methods such as EEGNet, TSception, and the deep convolution neural network. Moreover, STHN can assign weights to each SEEG channel and select those channels that contribute more significantly to emotion recognition. It was found that when using the top 30% weighted SEEG channels as model inputs, the accuracy did not decrease significantly. Most of the channels with higher weights were located in brain regions strongly associated with emotions, such as the frontal lobe, the temporal lobe, and the hippocampus. This indicates that STHN is not merely a "black-box" model but possesses a degree of explainability. To the best of our knowledge, this is the first study to develop an SEEG emotion recognition algorithm, which is expected to play a crucial role in the monitoring and treatment of patients with refractory emotional disorders in the future.