Kangyu Wang, Zhiyun Jiang, Haibo Feng, Weijia Zhao, Lin Liu, Jianguo Li, Zhenzhong Lan, Weiyao Lin. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

While Transformer models have achieved remarkable success in various domains, the effectiveness of information propagation through deep networks remains a critical challenge.Standard hidden state residuals often fail to adequately preserve initial token-level information in deeper layers.This paper introduces ResFormer, a novel architecture that enhances information flow by incorporating value residual connections in addition to hidden state residuals.And a variant is SVFormer, where all layers share the first layer's value embedding.Comprehensive empirical evidence demonstrates ResFormer achieves equivalent validation loss with 16.11% fewer model parameters and 20.3% less training data compared to Transformer, while maintaining similar memory usage and computational cost.Besides, SVFormer reduces KV cache size by nearly half with only a small performance penalty and can be integrated with other KVefficient methods, yielding further reductions in KV cache, with performance influenced by sequence length and cumulative learning rate.

Harnessing the power of single-cell genomics technologies, single-cell Hi-C (scHi-C) and its derived technologies provide powerful tools to measure spatial proximity between regulatory elements and their target genes in individual cells. Using a global background model, we propose SnapHiC-G, a computational method, to identify long-range enhancer-promoter interactions from scHi-C data. We applied SnapHiC-G to scHi-C datasets generated from mouse embryonic stem cells and human brain cortical cells. SnapHiC-G achieved high sensitivity in identifying long-range enhancer-promoter interactions. Moreover, SnapHiC-G can identify putative target genes for noncoding genome-wide association study (GWAS) variants, and the genetic heritability of neuropsychiatric diseases is enriched for single-nucleotide polymorphisms (SNPs) within SnapHiC-G-identified interactions in a cell-type-specific manner. In sum, SnapHiC-G is a powerful tool for characterizing cell-type-specific enhancer-promoter interactions from complex tissues and can facilitate the discovery of chromatin interactions important for gene regulation in biologically relevant cell types.

The use of a mega-constellation comprised of thousands of Low Earth Orbit (LEO) satellites for global internet service has garnered significant attention. In the network layer of this system, geographical routing has been found to outperform centralized strategies due to the lower complexity and overhead. However, geographical routing can still result in “dead-ends” due to network gaps. To address this issue, we propose the use of encountering inter-satellite links (eISLs) to improve network connectivity and routing reachability. We further present a system model and analysis of eISLs, as well as our Dynamic eISLs Configuration (DeC) algorithm for establishing eISLs between encountering satellites. Our experimental results demonstrate that our proposed DeC under eISLs enabling in satellite networks can significantly reduce propagation latency by 22% and path stretch by 15% in centralized routing algorithms. Moreover, in geographical routing, DeC can effectively improve the reachable ratio from 55% to 100% while maintaining a 28% increase in throughput, outperforming schemes without eISLs. Our proposed eISL-enabled satellite network architecture shows promising results in improving routing efficiency and connectivity in LEO satellite systems.

MOTIVATION: The efficiency of CRISPR/Cas9-mediated protein knockout is determined by three factors: sequence-specific sgRNA activity, frameshift probability and the characteristics of targeted amino acids. A number of computational methods have been developed for predicting sgRNA efficiency from different perspectives. However, an integrative method that combines all three factors for rational sgRNA selection is still lacking. RESULTS: We developed GuidePro, a two-layer ensemble predictor that enables the integration of multiple factors for the prioritization of sgRNAs in protein knockouts. Tested on independent datasets, GuidePro outperforms existing methods and demonstrates consistent superior performance in predicting phenotypes caused by protein loss-of-function, suggesting its robustness for prioritizing sgRNAs in various applications of CRISPR/Cas9 knockouts. AVAILABILITY AND IMPLEMENTATION: GuidePro is available at https://github.com/MDhewei/GuidePro. A web application for prioritizing sgRNAs that target protein-coding genes in human, monkey and mouse genomes is available at https://bioinformatics.mdanderson.org/apps/GuidePro. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.