Browse Articles
Discover research articles across all indexed journals
Order parameters and phase transitions of continual learning in deep neural networks
Continual learning (CL) enables animals to learn new tasks without erasing prior knowledge. CL in artificial neural networks (NNs) is challenging due to catastrophic forgetting, where new learning degrades performance on older tasks. While various techniques exist to mitigate forgetting, theoretical insights into when and why CL fails in NNs are lacking. Here, we present a statistical-mechanics theory of CL in deep, wide NNs, which characterizes the network’s input–output mapping as it learns a sequence of tasks. It gives rise to order parameters (OPs) that capture how task relations and network architecture influence forgetting and anterograde interference, as verified by numerical evaluations. For networks with a shared readout for all tasks (single-head CL), the relevant-feature and rule similarity between tasks, respectively measured by two OPs, are sufficient to predict a wide range of CL behaviors on classic benchmark tasks. In addition, the theory predicts that increasing the network depth can effectively reduce interference between tasks, thereby lowering forgetting. For networks with task-specific readouts (multihead CL), the theory identifies a phase transition where CL performance shifts dramatically as tasks become less similar, as measured by another task-similarity OP. While forgetting is relatively mild compared to single-head CL across all tasks, sufficiently low similarity leads to catastrophic anterograde interference, where the network retains old tasks and interpolates new training data perfectly but completely fails to generalize new learning. Our results delineate important factors affecting CL performance and offer theoretical insights into common heuristics for mitigation of forgetting.
Attention-guided enhanced deconvolution enables reference-free cell type estimation in spatial transcriptomics
Designing multisource blue–green cooling networks by coupling landscape pattern metrics and circuit theory
Functional and antigenic constraints on the Nipah virus fusion protein
Nipah virus is a highly pathogenic virus in the family Paramyxoviridae that utilizes two distinct surface glycoproteins to infect cells. The receptor-binding protein (RBP) binds host receptors whereas the fusion protein (F) merges viral and host membranes. Here, we use nonreplicative pseudoviruses to safely measure the effects of all F single amino acid residue mutations on its cell entry function and neutralization by monoclonal antibodies. We compare mutational tolerance in F with previous experimental measurements for RBP and show that F is much more functionally constrained than the RBP. We also identify mutationally intolerant sites on the F trimer surface and core that are critical for proper function, and describe mutations that are candidates for stabilizing F in the prefusion conformation for vaccine design. We quantify how F mutations affect neutralization by six monoclonal antibodies, and show that the magnitude of mutational effects on neutralization varies among antibodies. Our measurements of mutational effects on Nipah virus F predict the ability of the antibodies to neutralize the related Hendra virus. Overall, our work defines the functional and antigenic constraints on the F protein from an important zoonotic virus.
Climate change impacts on the global potential distribution of the human flea, Pulex irritans, and the global health risks
Abstract The human flea, Pulex irritans , is a hematophagous ectoparasite and medically significant vector of zoonotic pathogens, such as Yersinia pestis (plague), Bartonella quintana (trench fever), and Rickettsia felis (flea-borne spotted fever). Despite the public health significance of P. irritans , the potential impacts of climate change on its global distribution were unstudied before. In this study, we created an ecological niche model (ENM) through integrating 564 georeferenced records and 15 bioclimatic variables using the Maximum Entropy (MaxEnt) algorithm to project the current and future habitat suitability of P. irritans under two high-emission scenarios (SSP370 and SSP585) for 2050 and 2070 from three General Circulation Models (GCMs). DIVA- GIS was used to confirm the current predictions. Results revealed that the Model’s performance was robust with high predictive accuracy (AUC = 0.898; TSS = 0.6), identifying annual mean temperature (Bio1) with 55.9% contribution as the primary distribution variable. The models project that many species across North and South America, Europe, Asia, Australia, and Africa will expand their ranges toward higher latitudes. Regions once deemed unsuitable, including northern Europe, Canada, and Russia, are forecast to become suitable habitats as these species shift their geographical distribution. On the other hand, habitat loss was predicted in Africa and Australia due to extreme warming. Two-dimensional niche analysis revealed the broad tolerances of P. irritans (2–25 °C; 0–2200 mm), confirming its invasive potential. These shifts correlate with increased plague risk in temperate zones, as warmer temperatures accelerate flea life cycles and pathogen transmission efficiency. Our findings provide the first global assessment of climate-driven redistribution of P. irritans, highlighting the urgent need for surveillance in vulnerable regions to mitigate emerging vector-borne disease threats.
Correction for Singer, Confronting the inevitable: Harnessing technology to contain systemic scientific fraud
Expanding SARS-CoV-2 antigen targets beyond spike using a baculovirus expression system
Abstract The ongoing antigenic evolution of SARS-CoV-2, particularly within the spike glycoprotein, threatens the long-term efficacy of spike-focused vaccines and serodiagnostics. While most authorized COVID-19 vaccines exclusively target the spike protein, growing evidence underscores multivalent strategies incorporating conserved viral antigens. Here, we employed a baculovirus expression vector system (BEVS) to co-express all four structural proteins, including spike, nucleocapsid, membrane, and envelope, in Sf9 insect cells. Surface-displayed antigens were used in a cell-based ELISA to profile IgG responses in convalescent sera. All antigens elicited detectable antibody binding, with nucleocapsid and membrane proteins provoking significantly stronger responses than spike ( p < 0.001), and envelope protein showing intermediate reactivity. Statistical analyses revealed distinct patterns of antigenic reactivity. These findings validate the utility of BEVS for multiplex antigen presentation and highlight the immunological and diagnostic value of conserved non-spike antigens. Combined, this study advocates multivalent vaccine strategies and refines serodiagnostics by leveraging broader antigenic targets to counter immune escape.
A daily cycle of White Collar Complex dephosphorylation sustains circadian rhythmicity in <i>Neurospora</i>
The transcription factor complex White Collar Complex (WCC) functions both as a photoreceptor and as the circadian positive element. In response to light, WCC acutely activates ~5% of all genes, whereas in the dark it influences expression of about 40% of the transcriptome. Among WCC targets is frq , which is acutely light-activated through the pLRE ( proximal Light-Response Element ) and circadian-regulated through the C-box ( Clock-box ) promoter element that is not responsible for light-driven expression. The FRQ–FRH complex (FFC), which includes CK-1a, represses WCC activity at the C-box by phosphorylating WCC at >95 sites, but FFC has no described role in the light. We validated the expectation that FFC also silences C-box promoters in constant light, thereby confirming two classes of WCC targets: C-box -like genes that are normally repressed in light and pLRE -like genes that remain light-active despite FFC-driven WCC phosphorylation. Derepression of C-box -like promoters in frq -null fungi may explain reported noncircadian phenotypes such as reduced virulence and conidiation. Reanalysis of WCC circadian regulation revealed that, while most WCC is phosphorylated and repressed at dusk, subsequent circadian activation results from transient dephosphorylation of only a small subset of the WCC pool. This small active pool drives frq expression, nucleating the FFC, which rephosphorylates WCC to repress it again, generating a phosphorylation/dephosphorylation cycle that can persist for days without new WCC synthesis. The realization that both FFC and WCC are regulated primarily through phosphorylation rather than protein turnover leaves the circadian oscillator looking much like a “phoscillator,” emphasizing the primacy of posttranslational regulation in timekeeping.
Consensus land-cover mapping improves grassland classification in European mountain landscapes
Abstract Accurate land-cover information is essential for biodiversity monitoring, yet existing 10-m global and continental land-cover datasets vary in accuracy and thematic consistency, particularly for grasslands in complex mountain environments. We assessed six state-of-the-art land-cover products (Dynamic World, ESA WorldCover, Esri Land Cover, Corine Land Cover+ Backbone, ELC10, S2GLC) across the Alps and Carpathians and developed three consensus maps using weighted voting, accuracy-confusion weighting, and an accuracy-weighted Random Forest ensemble. All datasets were validated against an independent set of expert-interpreted reference samples. Individual products showed large discrepancies in grassland extent, elevation distribution, and landscape structure. Global datasets (Dynamic World, Esri Land Cover) underestimated grassland extent, whereas ESA WorldCover and Corine Land Cover+ reported higher proportions. Consensus approaches substantially reduced these inconsistencies. The Random Forest ensemble achieved the highest accuracy (overall 90–92%), outperforming individual datasets and improving both user’s and producer’s accuracies for grassland (> 84%). Consensus datasets also better captured expected elevation and slope gradients, producing more spatially coherent and ecologically realistic grassland patterns. By integrating multiple land-cover sources, consensus approaches effectively mitigated dataset-specific biases and increased the reliability of grassland mapping in heterogeneous mountain systems. Consequently, consensus land-cover products provide a robust and ecologically meaningful alternative to single-source datasets for environmental assessments in complex mountain regions.
A systematic map of methods for assessing societal benefits of Earth science information
Remotely sensed Earth science information (ESI) has become increasingly central to addressing global challenges, yet its societal value, i.e., the difference ESI makes in real-world decisions and outcomes, is rarely quantified. In this study, we systematically map peer-reviewed literature that explicitly assesses the societal value of ESI across instrumental, intrinsic, and relational value types, and the diversity of approaches used to assess those values. Drawing from 13,823 publications across Scopus, Web of Science, and a curated library of ESI valuation studies, we identify 171 studies that applied ESI in a decision context and used a valuation method to compare outcomes with and without ESI. The majority of these studies employed decision analysis methods (e.g., Value of Information, Cost–Benefit Analysis), focusing primarily on quantitative instrumental values (e.g., profit, crop yield, lives saved), particularly in agricultural contexts. A smaller set of studies applied preference elicitation methods (e.g., stated preference, surveys, interviews, focus groups) to capture qualitative benefits and relational values including quality of life improvements, empowerment, and procedural justice. Many excluded studies demonstrated scientific value of ESI but did not explicitly translate that into societal value, revealing the need for a more systematic approach to ESI valuation. By promoting a more inclusive, interdisciplinary, and flexible portfolio of valuation methods, we aim to expand our understanding of the societal benefits of ESI to help guide investment in future missions, enhance public support, and ensure that science and policy goals are well aligned.
LINC01857 promotes clear cell renal cell carcinoma progression by scaffolding DNMT1 to suppress WIF1 expression
Structural analysis of rhodopsin states in megabody complexes
Rhodopsin, the most intensively studied G protein–coupled receptor (GPCR), is activated by light-induced isomerization of its chromophore 11- cis -retinal. This study employed cryogenic electron microscopy (cryo-EM) to investigate rhodopsin structure using a megabody (Mb7) as a negative allosteric modulator. Three distinct cryo-EM structures were solved: ground-state rhodopsin, photoactivated rhodopsin, and apo-rhodopsin, all in complex with Mb7. Photoactivated rhodopsin and apo-rhodopsin, both in complex with Mb7, maintain a conformation remarkably similar to ground-state rhodopsin rather than adopting a Meta-II-like conformation. Structural elements, including the conserved residues of the NPxxY motif and the ionic lock, remain in positions corresponding to inactive rhodopsin. The megabody forms extensive interactions with rhodopsin’s extracellular loop 2, N terminus, and glycans. The findings demonstrate that Mb7 stabilizes photoactivated rhodopsin in a Meta-I-like conformation, preventing progression to the active Meta-II state through specific immobilization of the extracellular domain. This work establishes a foundation for cryo-EM-guided discovery of ligands modulating rhodopsin.
Deep learning enabled pseudonymization for preserving data privacy of financial identifiers in public documents in India
Abstract The increasing digitization and transmission of government-issued electronic documents have intensified the need to protect the ’Handwritten signatures’-recognized as ’critical biometric identifiers’ from identity-related data breaches. For instance, as per 2025-RSA ID IQ Report, 40% of respondents reported Identity-related data breaches and 66% emphasized the significant damages caused by these breaches to their organizations. The existing privacy-preserving anonymization research is primarily focusing on facial features and fingerprints, whereas the Pseudonymization of handwritten signatures in publicly accessible documents remains largely underexplored in the literature. This research study proposes a new Fully Convolutional Neural Network (CNN)-based Pseudonymization framework using SuperPoint architecture integrated with Differentiable output decoding, which aims to identify and pseudonymize the handwritten signatures in public-domain documents, specifically in Indian Government issued Permanent Account Number (PAN) cards. In contrast to traditional anonymization approaches, this pseudonymization technique preserves document utility by securing sensitive data and thereby enables traceable identity protection without compromising the structural integrity of input documents. Extensive evaluations are carried out on a curated dataset of 500+ real-world PAN cards, which establishes the model’s robustness and applicability in large-scale deployments. The results of comparative analysis with baseline techniques including ORB, FAST, SIFT and deep CNN, clearly demonstrate the superior performance of the proposed method in terms of various metrics such as Precision, Recall, SSIM, runtime efficiency, and spatial overhead. In addition, the research findings suggest practical implications for embedding CNN-based pseudonymization into Public-sector Document processing pipelines, which supports secure and utility-preserved digital archiving in compliance with modern privacy GDPR standards.
Investigation into the interactive feedback and rock burst mechanism under mining disturbance
Universal relation between spectral and wavefunction properties at criticality
Quantum-chaotic systems exhibit several universal properties, ranging from level repulsion in the energy spectrum to wavefunction delocalization. On the other hand, if wavefunctions are localized, the energy levels exhibit no level repulsion and their statistics is Poisson. At the boundary between quantum chaos and localization, however, one observes critical behavior, not complying with any of those characteristics. An outstanding open question is whether there exists yet another type of universality, which is genuine for the critical point. Previous work suggested that there may exist a relation between the global characteristics of the energy spectrum, such as the spectral compressibility χ , and the degree of wavefunction delocalization, expressed via the fractal dimension D 1 of the Shannon–von Neumann entropy in a preferred (e.g., real-space) basis. Here, we study physical systems subject to local and nonlocal hopping, both with and without time-reversal symmetry, with the Anderson models in dimensions three to five being representatives of the first class, and the random banded matrices as representatives of the second class. Our thorough numerical analysis supports the validity of the simple relation χ + D 1 = 1 in all systems under investigation. Hence, we conjecture that it represents a universal property of a broad class of critical models. Moreover, we test and confirm the accuracy of our surmise for a closed-form expression of the spectral compressibility in the one-parameter critical manifold of random banded matrices. Based on these findings, we derive a universal function D 1 ( r ) of the averaged level spacing ratio r , which is valid for a broad class of critical systems.
Promoting sustainable urban tourism in Historic Sari, Iran through GIS-based walkability planning
Identifying genome-by-childhood trauma interactions for depression using a forest-based approach in the UK Biobank and Adolescent Brain Cognitive Development Study
Depression is shaped by both genetic and environmental factors, but genome-wide interaction studies (GWIS) often lack power to detect complex gene–environment (G × E) interactions. We applied a forest-based machine learning approach to 38,018 UK Biobank (UKB) participants, examining interactions between 285,677 single-nucleotide polymorphisms (SNPs) and three trauma types (childhood, adult, and catastrophic trauma). While GWIS detected no significant interactions, we identified 8,225 potentially important SNP–environment pairs across 1,732 genes, with childhood trauma contributing most prominently. Stratified heritability was higher among childhood trauma–exposed individuals (13.3%) versus those unexposed (6.0%). Many identified genes overlapped with known psychiatric risk loci and accounted for most of the SNP-based heritability. Thirteen top genes were replicated in the Adolescent Brain Cognitive Development Study. Our findings highlight the polygenic G × E nature of depression and the critical role of childhood trauma in modulating genetic risk, demonstrating the value of forest-based methods in detecting complex gene–environment interactions.
Benchmarking imputation strategies for missing time-series data in critical care using real-world-inspired scenarios
Abstract Handling missing data remains a central challenge in Intensive Care Units (ICU) time-series analysis, where gaps frequently arise from non-random mechanisms such as sensor disconnections and workflow-driven interruptions. In this study, we benchmarked multiple imputation strategies on monitoring data from MIMIC-IV and designed masking scenarios that reflect ICU missingness patterns observed in the database, thereby approximating real-world conditions and clarifying how conclusions depend on both the chosen imputation method and the missingness scenario. We compared commonly used simple statistical approaches (mean, LOCF, interpolation), classical machine learning techniques (MICE, MissForest), and several deep learning architectures (Transformers, RNNs, GANs, VAEs). Transformer and GAN models achieved the best overall performance, whereas linear interpolation remained a strong baseline. Crucially, results were scenario-dependent: MCAR produced optimistic error estimates and compressed differences between methods, whereas structured gaps revealed clearer performance separations. Our findings suggest that, while deep learning methods improve overall imputation accuracy, linear interpolation is often nearly as effective and offers a lighter, more interpretable approach. This work introduces a practical framework for evaluating time-series imputation strategies under realistic constraints, with a focus on clinical relevance. Further analysis of downstream impact under clinically realistic scenarios and using tailored imputation strategies by variable type remains needed.