Dual selective gleason pattern-aware multiple instance learning with uncertainty regularization for grade group prediction in histopathology images.

DeepHybridCPI: A hybrid deep learning framework for compound-protein interaction prediction.

GaussianCAD: Robust self-supervised CAD reconstruction from three orthographic views using 3D Gaussian Splatting

Dilated memory in hierarchical reinforcement learning for long-horizontal task.

Differential gene expression (DGE) analysis enables researchers to investigate the link between gene expression and the phenotypic responses observed in organisms across time, experimental, or field conditions. Accurate quantification of gene expression is essential when performing DGE experiments, with a range of methods having been developed to enable the study of gene expression within a species. Quantifying differences in expression not just within but across multiple species can also be used to reveal the genetic mechanisms underlying phenotypic differences observed between species. Accurate quantification of gene expression across multiple species requires a suitable reference; it should include each species' own expressed transcripts to mitigate reference bias, with the orthology relationships of transcripts being used to facilitate comparison of expression at the gene level. Production of such a reference remains a challenge, despite its necessity for minimising bias during multispecies DGE analysis. Our software BINge specifically aims to address this need through use of a novel approach to modelling orthology which results in multispecies transcript clusters that accurately reflect their locus orthology. Evaluation experiments demonstrate the effectiveness of this approach over existing clustering methods which have not been designed for producing a reference suitable for multispecies DGE analysis. Source code and documentation for BINge are available from the GitHub repository at https://github.com/zkstewart/BINge.

From coarse to fine-grained decomposition: Hierarchical question generation and learned importance for automated fact-checking

Twisted convolutional networks (TCNs): Enhancing feature interactions for non-spatial data classification.

Coral reefs are among the most diverse and productive marine ecosystems, providing food security and livelihoods for millions of people, especially in coastal and rural areas where small-scale reef fisheries are a main source of fish. Rising sea temperatures and pollution are causing widespread bleaching, driven by the loss of algal symbionts, and accelerating reef degradation. Underwater imaging paired with color-based health charts provides a non-invasive method for monitoring bleaching. However, current approaches require extensive manual annotation and are limited by underwater image noise such as blur and color casts. Existing AI-based monitoring approaches are often semi-autonomous and lack fine-grained bleaching localization capabilities. In this paper, we propose Coral Color-Reference Chart Automation (Coral-CRCA), a multi-stage algorithm that replicates the visual assessment process used by marine biologists to fully automate coral bleaching evaluation with color-reference charts. Initially, the pipeline incorporates a preprocessing image denoising module to improve robustness to underwater image distortions. The method then segments the coral region, isolates chart quadrants, and assigns each coral pixel to the closest reference grade by color similarity, generating pixel-level bleaching visualizations and reporting bleaching percentage. We evaluate each stage of the pipeline through comparative analyses and assess the complete system on 3400 expert-annotated, field-collected coral images from the Arabian/Persian Gulf. The method achieves a mean absolute error (MAE) of 19.17% in bleaching percentage estimation and a binary classification accuracy (bleached/healthy) of 96.12%, matching expert-level performance on this dataset. The source codes are available on this link https://github.com/MahmoudElMezain/Coral-CRCA_Color-Reference-Chart-Automation-Algorithm.

A hybrid metaheuristics-Bayesian optimization framework with safe transfer learning for continuous spark tuning

Robust gait phase recognition is essential for gait analysis, biomechanical monitoring, and human-centered robotics. A significant gap persists between computational methods and field deployment spanning generalizability, computational budget, hardware optimization, and integration with robotic frameworks. This study presents a reproducible wearable system that bridges this gap by combining robust, real-time recognition with minimal modalities. The proposed probabilistic heuristic recognition algorithm for sequential events (PHRASE) leverages statistical modeling and artificial neural networks (ANN) to achieve robust low-latency detection. The system integrates two inertial measurement units (IMUs) and a portable microcomputer within a Robot Operating System (ROS) framework, enabling seamless integration with external systems. The method was validated against benchmarks on 31 participants with diverse biometrics, sensor models, and physical conditions during multi-scenario level-ground walking, outperforming 5 state-of-the-art deep learning benchmarks. PHRASE demonstrates strong generalization, achieving an accuracy of specifically across diverse unseen subjects, conditions, and experimental setups combined. The accuracy improvement over the best-performing benchmark has a 95% confidence interval of . The wearable interface maintains a stable average inference latency of 11.6 ms. Overall, the proposed interface addresses key limitations in terms of resilience to unseen subjects, unseen sensor configurations, varying walking speeds, and latency. The source code and implementation details are available at: https://github.com/SamMans/PHRASE/tree/main . • Real-time Bayesian gait phase recognition using only two wearable IMUs. • Portable ROS-based interface for assistive and robotic gait applications. • Combination of heuristics, artificial neural networks and Bayesian inference. • More robust and generalizable than deep learning benchmarks in cross-setup testing without retraining. • Faster than deep learning benchmarks in processing latency.

CANet: ChronoAdaptive network for enhanced long-term time series forecasting under non-stationarity.

Functional magnetic resonance imaging (fMRI) is a widely leveraged signal in brain decoding, experiencing rapid advancement and demonstrating tremendous promises. The development of fMRI-based large models has piqued the interest of academics, who hope to overcome the existing limitations by taking advantage of advances in large model technology and increased accessibility of mental state datasets. This also raises the challenge of effectively transferring the pre-trained large models to new applications. To tackle this problem, this paper proposed a novel uncertainty-aware variational soft prompt approach to aid the pre-trained model in adapting to a new domain, while providing an estimation of the sample aleatoric uncertainty by using the reparameterization technique and associating the prompt vectors to uncertainty through the designed loss function. The comparative experiments showed that, the proposed approach can significantly improve model transfer performance and outperform the standard prompt method, in particular for dealing with few training samples. Moreover, the proposed approach demonstrated robustness to variations in model structure and the hyperparameter of the loss function, which also highlighted the practical usability of the proposed approach. The source code is publicly available at github.com/hzx-ntu/Uncertainty-aware-Variational-Soft-Prompt-for-Few-shot-Learning .

MoGL: A mixture of heterogeneous experts for collaborative graph learning.

Multimodal information fusion for software vulnerability detection based on both source and binary codes

Harvesting AlphaEarth: Benchmarking the geospatial foundation model for agricultural downstream tasks

Research on low-dimensional multivariate information fusion prediction based on space battlefield situation information.

TRKM: Twin restricted kernel machines for classification and regression.

MGML: A plug-and-play meta-guided multi-modal learning framework for incomplete multimodal brain tumor segmentation.

Amplicon sequencing is a targeted approach used to assess the diversity of microbial communities by amplifying and sequencing a specific genetic locus from DNA. QIIME2 is one of the most prevalent methods for metagenomics analysis due to its plugin-based design wherein distinct modules can be utilized to perform specific functions. However, QIIME2 data input, and plugin utilization is cumbersome to navigate. Previous amplicon pipelines also lack host depletion and statistical biomarker identification modules from upstream and downstream analysis. To this effect, we assembled a simple and customizable Zenity based GUI workflow for analysing amplicon data with Automating Microbial Community Analysis (AMCA). The analysis integrates key attributes of amplicon analysis: host depletion with Bowtie2 and biomarker prediction by LEfSe. The bash-based analysis guides and allows the user to select filtering parameters based on intermediate results while minimizing the need to navigate command-based plugins. The outputs from the AMCA workflow include the filtered and host-depleted raw sequencing data, taxonomic abundances, alpha and beta diversity indices, alpha rarefaction analysis, phylogenetic tree (rooted and unrooted) and significant features which explain key microbial differences between conditions/classes of the experiment. The implementation of the designed workflow has been tested on a pilot study based on amplicon sequencing in 100 samples from patients of Chronic Kidney Disease and healthy controls. The exploratory LEfSE analysis revealed key taxa Streptococcus, Bacteroides and Faecalibacterium to vary between disease and control conditions. The source code related to the analysis can be assessed from the Github repository at https://github.com/Nitika-Rana/AMCA. The study delivers an efficient, user-friendly, and customizable workflow for amplicon analysis, simplifying QIIME2 execution while enabling host depletion and biomarker characterization.

PyPIMalDet: A malicious PyPI package detection method combining code features and metadata features.