Machine learning (ML) is increasingly being used in food science due to its ability to extract insights from large datasets. However, the advantages of ML over traditional mechanistic knowledge-based models remain unclear, especially under the limited data conditions often encountered in food bioprocesses. This study aims to address this gap by critically evaluating supervised ML techniques-specifically decision trees, support vector machines, and neural networks-in comparison to a knowledge-based model (KB), using wine fermentation as a practical, experimental example. We evaluated these approaches in three tasks. Tasks 1 and 2 use time-series fermentation data to (1) classify industrial yeast strains based on their metabolite profiles and (2) predict fermentation dynamics. Task 3 focuses on creating a fast surrogate model using ML techniques applied to synthetic data generated by a mechanistic model. For yeast strain classification, we achieved our highest test accuracy of 74% when utilizing all available metabolite data. In predicting fermentation dynamics, the KB model outperformed the ML models, achieving an average normalized root mean squared error of approximately 6%. The ML models, when additional data was incorporated, had a prediction error of around 7.6%. Lastly, a deep learning surrogate model trained solely on synthetic, mechanistic data demonstrated very low errors (around 0.6%) on test sets, compared to the KB model, while also reducing simulation time by a factor of 30. Our findings highlight the significance of experimental design: although ML models perform well when trained on large and diverse datasets, they often struggle with limited data or when predicting outcomes beyond the conditions observed during training. In contrast, mechanistic models show better generalization and biological interpretability. The complementary nature of both approaches suggests that combining them can lead to more robust, data-informed design and control in complex fermentation systems. Leveraging these complementary strengths, we developed and validated a hybrid model that integrates knowledge-based predictions with a residual neural network to correct systematic errors, reducing overall NRMSE from 6% to 5% and improving prediction for most key compounds.

Group equivariant pyramid network for respiratory motion correction on PET image.

Dataset of microbial community evolution in synthetic black-clay-based soils during ecological reconstruction.

Insulin sensitivity prediction is crucial for model-based treatment in Intensive Care Unit patients, particularly those with hyperglycemia. However, predicting insulin sensitivity is challenging due to inter- and intra-patient variability. Different neural network models are proposed and compared for predicting insulin sensitivity, including recurrent and feedforward versions of the Classification Deep Neural Network and Mixture Density Network models. These models were trained using 1879 patient records containing 123,988 insulin sensitivity values from three intensive care patient cohorts in three different countries. Results show that using patient history in prediction models can improve the accuracy of insulin sensitivity predictions. The Mixture Density Network model provided more accurate predictions, measured by a problem-specific metric that expresses clinical requirements. We demonstrated that even using up to 12 h of historical data can improve prediction accuracy. This study highlights the potential of recurrent neural network models in predicting insulin sensitivity in Intensive Care Unit patients. Our findings suggest that using recurrent neural networks and incorporating patient history can lead to more accurate predictions. These results are generalizable due to the large and diverse dataset employed, which included patients from three different cohorts in three care settings.

This study introduces a novel multilingual dataset designed to distinguish auto-tuned musical compositions from authentic recordings, addressing a significant gap in existing resources. The dataset encompasses songs in English, Mandarin, and Japanese, ensuring a diverse representation of linguistic contexts. The data collection process began with aggregating diverse datasets from the Music Information Retrieval domain, incorporating tracks from the three specified languages to capture a wide range of musical styles and recording qualities. Each audio file was subsequently standardized into 10-second intervals with the sample rate of 16 kHz to facilitate manageable analysis. For the creation of auto-tuned samples, pitch correction was implemented using the probabilistic YIN (PYIN) algorithm for accurate pitch detection, followed by transposition via the pitch-synchronized overlap and add (PSOLA) technique. To emulate realistic auto-tuning scenarios, pitch correction was randomly applied to portions of each 10-second segment, ensuring variability and realism in the dataset, which makes it suitable for training robust detection models. Additionally, time-domain labels indicating the exact locations of pitch correction within each segment were generated, providing precise annotations crucial for developing accurate detection algorithms. The resulting multilingual dataset comprises a comprehensive collection of both auto-tuned and authentic musical segments across English, Mandarin, and Japanese languages, each annotated with detailed information about pitch correction applications. This rich annotation allows for nuanced analysis and supports various research applications, while the dataset's structure and thorough documentation of its creation process make it a valuable resource for researchers in music analysis, machine learning, and audio signal processing.

From a One Health perspective, the gut microbiota of animals acts as a major driver of microbial exchange between animals and the environment. Animals continuously release gut microbes into their surroundings, shaping environmental and human microbial communities and potentially dispersing pathogens. Characterizing gut microbiota across diverse animal hosts is therefore critical for understanding the patterns of microbial spread through ecosystems and their impact on animal, human and environmental health. Here, we introduce a large, taxonomically diverse dataset of fecal microbiomes from 715 individual animals representing over 50 mammalian and avian species. We collected samples from both wild and domestic animals with an emphasis on capturing microbial diversity across a wide range of taxa and ecological contexts. The samples were subjected to 16S rRNA gene sequencing, targeting the V3-V4 hypervariable region. Bioinformatic analysis was performed using Usearch to generate zero-radius operational taxonomic units (ZOTUs). This dataset was generated primarily for the development of microbial source tracking (MST) assays used for identifying the sources of fecal pollution in contaminated water. However, it provides a valuable resource for broader microbiome research. It enables comparative studies across host species, trophic guilds, and environmental contexts such as domestication.

While our ability to capture seafloor imagery has expanded dramatically, extracting information remains a major bottleneck, as annotation is still largely manual, time-consuming, and costly. Machine Learning (ML) classifiers offer a promising solution, but they often lack generalisation as they are typically trained on small, homogeneous datasets from selected campaigns. Here, we address this challenge by leveraging the Squidle+ platform to assemble a heterogeneous dataset of 1.7 million annotations across 150,000 images from 325 expeditions in Australian waters. To effectively train models on this scale, we introduce a novel dataset distribution strategy, the neighbour-distribution, and implement adaptive cropping to account for varying image resolutions. To assess performance at a national scale, we excluded entire campaigns from training and used them as independent test sets, allowing evaluation on complete transects, in addition to composed datasets. Using this flexible training pipeline, we combined diverse annotation schemes into three high-level Essential Ocean Variables (EOVs): Seagrass, Canopy-Forming Macroalgae, and Hard Coral cover and one species-specific label, Ecklonia radiata. This represents the first use of such a large and diverse dataset for training an ML classifier of Ecklonia radiata. The resulting models achieved test accuracies of 85%-93% for the EOVs and 96.5% for Ecklonia radiata on campaign test sets spanning Australia's coastline, demonstrating robust performance at a national scale. To enhance accessibility, both the code and trained models are made publicly available and directly integrated into Squidle+, lowering the barrier for their use in ecological monitoring programs.

ABSTRACT Photovoltaic (PV) systems are an invaluable green power solution to the energy requirements in the world yet as they are placed outside, they are subject to numerous issues and environmental challenges that may lead to loss of efficiency in the systems and even system failures. Thus, early and accurate fault diagnosis is necessary to minimize downtime and costs of maintenance. This paper proposes an automated PV-based fault diagnosis system with the aid of a pre-trained image-based InceptionV3 convolutional neural network, which can find faults in panels. The model was trained and tested on a diverse dataset of 885 images to classify six categories of conditions, including physical degradation, electrical degradation, bird droppings, dust, snow coverage, and fault-free panels. Using transfer learning and extensive data augmentation, the model achieved a good overall test-based accuracy of 85.71%. The comprehensive analysis of the performance showed that it was highly suitable for detecting Dusty, Clean, and Bird-drop classes with high F1-scores. However, the model struggled with the lower recall Physical-Damage class, indicating that it was not fully reliable in detecting subtle physical defects such as cracks. The results, complemented by ROC curves and confusion matrices, determine the extent to which the InceptionV3 architecture is appropriate for this task and highlight some diagnostic limitations. This article is important in providing more reliable and cost-effective solar energy production by creating a clear and methodologically rigorous standard for automated PV fault detection.

Adult-type gliomas are among the most prevalent and lethal primary central nervous system tumors, where prompt and accurate diagnosis is essential for maximizing survival prospects. Molecular classification, particularly the detection of isocitrate dehydrogenase (IDH) mutations and 1p/19q codeletions, has become crucial for accurate diagnosis and prognosis. Artificial intelligence (AI) has emerged as a promising adjunct in enhancing diagnostic accuracy using histopathological images. Existing reviews mostly focused on radiology rather than histopathology, and no comprehensive systematic review has specifically evaluated AI performance exclusively from histopathological images for detecting these two molecular markers. This study aims to systematically evaluate the performance of AI models in detecting and classifying IDH mutation status and 1p/19q gene codeletion in adult-type gliomas using histopathological images. A systematic review was conducted in accordance with PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses-Extension for Diagnostic Test Accuracy) guidelines. Seven databases (MEDLINE, PsycINFO, Embase, IEEE Xplore, ACM Digital Library, Scopus, and Google Scholar) were searched for studies published between 2015 and 2025. Eligible studies used AI models on histopathological images for molecular classification of adult-type gliomas and reported performance metrics. Study selection, data extraction, and risk of bias assessment using a modified QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies 2) tool were conducted independently by two reviewers. Extracted data were synthesized narratively. A total of 2453 reports were identified, with 22 studies meeting the inclusion criteria. The pooled average accuracy, sensitivity, specificity, and area under the curve (AUC) across studies were 85.46%, 84.55%, 86.03%, and 86.53%, respectively. Hybrid models demonstrated the highest diagnostic performance (accuracy 92.80% and sensitivity 89.62%). In general, AI models that used multimodal data outperformed those that used unimodal data in terms of sensitivity (90.15% vs 84.31%) and AUC (88.93% vs 86.29%). Furthermore, models had a better overall performance in identifying IDH mutations than 1p/19q codeletions, with higher accuracy (86.13% vs 81.63%), specificity (86.61% vs 78.11%), and AUC (86.74% vs 85.15%). Unexpectedly, AI models designed for binary classification exhibited lower performance than those for multiclass classification in terms of both accuracy (91.98% vs 84.02%) and sensitivity (93.41% vs 80.18%). However, these differences should be interpreted as descriptive trends rather than statistically validated superiority, as formal between-group comparisons were not feasible. AI models show strong potential as complementary tools for the molecular classification of adult-type gliomas using histopathology images, particularly for IDH mutation detection. However, these findings are constrained by the limited number of studies, the focus on adult-type gliomas, lack of meta-analysis, and restriction to English-language publications. While AI offers valuable diagnostic support, it must be integrated with expert clinical judgment. Future research should prioritize larger, more diverse datasets and multimodal AI frameworks and extend to other brain tumor types for broader applicability.

Conformational biasing (CB) is a rapid and streamlined computational method that uses contrastive scoring by inverse folding models to predict protein variants biased toward desired conformational states. We successfully validated CB across seven diverse datasets, identifying variants of K-Ras, the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) spike protein, the β2 adrenergic receptor, and Src kinase with improved conformation-specific functions such as enhanced binding or enzymatic activity. Applying CB to the enzyme lipoic acid ligase (LplA), we uncovered a previously unknown mechanism controlling its promiscuous activity. Variants biased toward an "open" conformation state became more promiscuous, whereas "closed"-biased variants were more selective, enhancing LplA's utility for site-specific protein labeling with fluorophores in living cells. The speed and simplicity of CB make it a versatile tool for engineering protein dynamics with broad applications in basic research, biotechnology, and medicine.

Optical calcium imaging is a powerful tool for recording neural activity across a wide range of spatial scales, from dendrites and spines to whole-brain imaging through two-photon and widefield microscopy. Traditional methods for analyzing functional calcium imaging data rely heavily on spatial features, such as the compact shapes of somas, to extract regions of interest and their associated temporal traces. This spatial dependency can introduce biases in time trace estimation and limit the applicability of these methods across different neuronal morphologies and imaging scales. To address these limitations, the Graph Filtered Temporal Dictionary Learning (GraFT) uses a graph-based approach to identify neural components based on shared temporal activity rather than spatial proximity, enhancing generalizability across diverse datasets. Here we present significant advancements to the GraFT algorithm, including the integration of a more efficient solver for the L1 least absolute shrinkage and selection operator (LASSO) problem and the application of compressive sensing techniques to reduce computational complexity. By employing random projections to reduce data dimensionality, we achieve substantial speedups while maintaining analytical accuracy. These advancements significantly accelerate the GraFT algorithm, making it more scalable for larger and more complex datasets. Moreover, to increase accessibility, we developed a graphical user interface to facilitate running and analyzing the outputs of GraFT. Finally, we demonstrate the utility of GraFT to imaging data beyond meso-scale imaging, including vascular and axonal imaging.

Characterizing the cellular origin and disease-driven dynamics of cfRNA is essential for integrating cfRNA profiling into clinical workflows and precision-medicine strategies. Most cfRNA studies are restricted to bulk-level analyses, which preclude detailed analysis of alterations in the cellular origins of cfRNA. Single-cell RNA sequencing reveals cellular heterogeneity and communication, but its application to cfRNA is limited by diverse cellular origins, leaving a critical gap in understanding functional alterations in cfRNA biomarker-originating cells. In this work, we propose CellFreeGMF, a tool designed to enable diagnosis classification of clinical samples, identify cfRNA biomarkers, and analyze the alterations in their originating cells based on graph matrix factorization. Furthermore, by utilizing cell-cell communication analysis, CellFreeGMF investigates the functional alterations occurring in the cfRNA originating cells under disease conditions. We validate CellFreeGMF on diverse cell-free RNA transcriptome clinical datasets. In the case of pancreatic ductal adenocarcinoma (PDAC), CellFreeGMF not only identified cfRNA biomarkers but also traced their cellular origins to myeloid and T-cell populations. Further analysis revealed significant transcriptomic differences in these cell populations between disease and normal groups. Our user-friendly CellFreeGMF toolkit (https://cellfreegmf.readthedocs.io/) enables identifying cfRNA biomarkers and elucidating pathophysiological changes in their originating cells.

We introduce CG-Vec, a crystal graph-to-vector framework that replaces iterative message passing with compact, interpretable descriptors coupled to conventional machine learning. Across diverse datasets, CG-Vec matches the accuracy of deep graph networks on large datasets but substantially outperforms them in data-scarce regimes. The advantage is most pronounced for magnetic properties, where existing approaches have struggled: CG-Vec delivers reliable predictions of magnetization in both ferromagnetic and ferrimagnetic systems and of Curie temperature. Beyond magnetism, CG-Vec performs competitively for formation energy and band gap, demonstrating broad applicability. These results establish vectorized representations as a practical and scalable alternative to deep architectures, enabling efficient and interpretable modeling of complex material properties.

The effective integration of artificial intelligence (AI) systems into clinical medicine depends on comprehensive and transparent performance evaluation; however, the lack of standardized and widely accepted metrics poses challenges for reproducibility and model adoption. A comprehensive, machine-interpretable framework is presented to formalize the nomenclature and descriptions of 207 graphical, matrix, and scalar metrics used to measure AI model performance. The metrics taxonomy, developed as part of the Radiology Ontology of AI Datasets, Models and Projects (ROADMAP), provides a logically structured representation that captures the semantics of AI evaluation metrics, supports reasoning over metric classes, and enables automated completeness checks for AI model reporting. For each metric, the taxonomy incorporates a definition and citations to authoritative reference sources; where applicable, the taxonomy also includes synonyms, abbreviations, alternate language forms, mathematical formulae, and numerical bounds. The taxonomy supports evaluation of models operating on structured data, medical images, audio signals, and/or unstructured text. Logical axioms link each metric to one or more of 18 AI model performance criteria, including classification, calibration, image segmentation, and text analysis. By harmonizing terminology and enabling structured queries, ROADMAP's taxonomy of AI performance metrics facilitates model comparison, bias detection, and selection of appropriate evaluation methods across diverse datasets and clinical tasks. © RSNA, 2026 See also accompanying Special Report on ROADMAP ontology.

Cell migration underlies immune surveillance, tissue repair, embryogenesis, and—when dysregulated—tumor metastasis. Yet unlike proliferation, which can be profiled at scale, migration studies remain limited by labor-intensive imaging and analysis. Existing assays often forfeit single-cell resolution, require phototoxic fluorescent labeling, or depend on tedious manual tracking, restricting the range of molecular perturbations and microenvironmental contexts that can be examined. We present Deep learning Brightfield Imaging and cell Tracking (DeepBIT), a high-throughput platform that captures live-cell behavior in multiwell plates and uses a convolutional neural network to detect and track individual cells in brightfield videos—without labels or user bias. Brightfield images are paired with nuclear fluorescence images to generate diverse ground-truth datasets, enabling automated training and eliminating manual annotation. This scalability supports a data-driven approach to systematically dissect the regulation of cell migration. Using breast cancer cells as a testbed, we tracked ~1500 cells per well across 840 conditions—including 96 FDA-approved drugs at multiple doses, a range of extracellular matrix and growth factor combinations, and CRISPR knockouts of cytoskeletal genes—yielding ~1.3 million trajectories in 30 hours (~2 minutes per condition). This dataset revealed previously unrecognized motility modulators among FDA-approved compounds and uncovered strong context dependence; for example, TNF-α and RhoA could either suppress or promote migration in the same cells depending on extracellular cues. Together, DeepBIT provides an unbiased, label-free platform for single-cell motility profiling at a scale compatible with modern drug libraries and genomic perturbation tools, enabling systematic exploration and therapeutic targeting of cell migration.

Single-cell multiomics has transformed biomedical research by enabling the study of gene expression, epigenetic modifications, and protein profiles within individual cells at an unprecedented level of detail. This approach has opened new opportunities for understanding complex biological systems, although significant challenges remain. As datasets increase in size and incorporate multiple modalities, issues such as data sparsity, integration complexity, and the need for scalable experimental methods have become prominent. In this review, recent progress combining computational tools, microfluidic technologies, and clinical applications is examined. The first section focuses on how advanced algorithms have been applied to interpret multimodal data, improve cell type identification, map developmental trajectories, and integrate diverse datasets. New techniques incorporating deep learning architectures, such as variational autoencoders, graph neural networks, and emerging foundation models, are highlighted for their role in enabling robust multimodal integration and predictive analysis. Subsequent sections address innovations in experimental workflows. Microfluidic devices integrated with smart automation and real-time monitoring have improved the reliability and efficiency of single-cell studies. These technical advances have had a tangible impact on translational research. In oncology, immunology, and infectious disease, multiomics-driven insights are informing diagnostic strategies and guiding therapeutic development. Finally, remaining challenges are considered, including regulatory requirements and the incorporation of emerging technologies such as spatial omics. Collectively, these advances point toward a future in which single-cell analysis becomes a cornerstone of precision medicine.

Accurate prediction of submitochondrial localization is fundamental to understanding mitochondrial biogenesis and cellular metabolic pathways. While deep representations from pre-trained protein language models (pLMs) have significantly advanced the field, traditional global average pooling methods often fail to capture critical, localized N-terminal targeting signals, particularly in long sequences where these motifs are mathematically diluted. To resolve this “signal dilution” bottleneck, we developed a multi-scale architecture that explicitly integrates high-resolution N-terminal features with global evolutionary context derived from ESM-2 embeddings. The proposed framework utilizes an orthogonal mixing strategy consisting of Token-mixing and Channel-mixing. Token-mixing is specifically designed to detect spatial rhythmic patterns across residue positions, while Channel-mixing refines the biochemical signatures within the latent feature space. Extensive benchmarking across diverse datasets demonstrates that our approach effectively maintains signal integrity. Compared to existing state-of-the-art methods, the model achieves a superior overall Generalized Correlation Coefficient (GCC) of 0.7443 on the SM424-18 dataset and 0.7878 on the SubMitoPred dataset, outperforming the latest benchmarks by 9.4% and 16.1%, respectively. Furthermore, on the independent M983 test set, our method maintained a high GCC of 0.6945, demonstrating a 9.9% improvement relative to the state-of-the-art methods. This robust and efficient framework provides a high-precision tool for large-scale mitochondrial proteomics.

Electroencephalography (EEG) signals are vital physiological indicators widely used for emotion recognition due to their high temporal resolution and direct measurement of brain activity. However, due to considerable inter-individual variability in EEG signal patterns, traditional methods have showed limited effectiveness in cross-subject emotion recognition tasks. Existing Transformer models for EEG emotion recognition often employ hybrid architectures to capture feature dependencies across multiple dimensions, however, these composite structures may result in insufficient information interaction and increased model complexity. To address these limitations, this paper proposes a multi-axis adapter transformer (MAAT) network, which leverages a unified Transformer framework to extract dependencies across frequency, channel, and temporal dimensions without relying on additional model components. Firstly, a multi-axis module is designed to replace the traditional multi-head attention module in the Transformer. This module captures dependencies across frequency, channel, and temporal dimensions, effectively modeling the complex and multi-faceted relationships within EEG signals. Secondly, adapter layers are integrated into the Transformer's feed-forward layers, which facilitate cross-subject transfer learning through fine-tuning. This design enables the model to adapt to new subjects with minimal parameter adjustments, improving generalization in cross-subject emotion recognition tasks. Experimental validations were conducted on the SEED, SEED-IV, and DEAP datasets. The MAAT model achieved high accuracy in both subject-dependent and subject-independent EEG emotion recognition tasks. These results confirm the model's effectiveness, robustness, and generalizability across diverse EEG datasets, outperforming existing methods in cross-subject recognition scenarios.

A critical challenge in agricultural automation is the precise detection of mango leaf diseases that compromise crop quality and yield. To address the limitation of existing heavy models in resource-constrained agricultural environments, this study proposes MangoLeafNet-XAI, a novel lightweight deep learning architecture. The model synergistically integrates Efficient Channel Attention (ECA) modules with a DenseNet-121 backbone to adaptively refine features and capture subtle pathological patterns with high precision. The proposed framework was rigorously evaluated using a 5-fold cross-validation and soft-voting ensemble strategy across three public datasets (MLDID, Mango Leaf Disease, and Harumanis). These datasets encompass diverse environmental conditions and distinct disease classes, including Anthracnose, Bacterial Canker, Die Back, Gall Midge, Powdery Mildew, Sooty Mould, and Cutting Weevil. MangoLeafNet-XAI achieved state-of-the-art accuracies of 98.83% on MLDID, 98.09% on the Mango Leaf Disease Dataset, and 98.76% on the Harumanis dataset. A primary contribution of this work is the optimal balance between performance and computational efficiency, utilizing only 6.9 million parameters, making it highly suitable for deployment on edge devices. Moreover, the interpretability of AI methods, such as Grad-CAM and LIME, that are used to explain the rationale behind predictions to offer pathological explanations, also validate the focus on clinically important aspects of the model. The results discuss the key limitations of existing methods, such as computational complexity, inability to interpret the findings, and dataset-dependent overfitting, and demonstrate a high level of resilience and generalizability on diverse datasets. MangoLeafNet-XAI will be a new benchmark of reliable, deployable, as well as accurate disease diagnosis systems, in smart agriculture.

MedSapiens: Taking a pose to rethink medical imaging landmark detection.