📄 Research AI News
Feed for best AI research releases, updates and latest AI research news.
- Towards the explainability of protein language models
Nature Machine Intelligence, Published online: 11 May 2026; doi:10.1038/s42256-026-01232-w Hunklinger and Ferruz provide an overview of explainable artificial intelligence methods for protein language models.
- Mitigating Implicit Bias in Chinese Toxic Speech Detection via Unbiased Contrastive Learning
Mitigating human-like biases and social stereotypes in pre-trained language models (PLMs) has become a crucial task in Chinese toxic speech detection. While PLMs have achieved state-of-the-art results in mitigating explicit bias (e.g., bias on sensitive
- Uncovering Hidden Connections: Iterative Search and Reasoning for Video-grounded Dialog
Unlike conventional visual question answering, video-grounded dialog requires a deep understanding of both the dialog history and the video content to generate accurate responses. Although existing methods have achieved promising results, they still ...
- OpenEP: Open-Ended Future Event Prediction
Future event prediction (FEP) is a long-standing and crucial task in the world, as understanding the evolution of events enables early risk identification, informed decision-making, and strategic planning. Existing work typically treats event prediction
- De-Decay: Defusing Computer Vision Model Degradation through Scalable and Actionable Human-Data Alignment
Computer Vision (CV) models can become outdated after deployment as real-world data evolves, requiring intensive attention from AI engineers to address degraded performance through tasks like data relabeling to update models with new human perceptions. ...
- Novel View Synthesis and Inverse Rendering of Glossy Objects: Efficient Representations, Ambiguity Mitigation, and Editable Assets
A:Georgios Kouros; TT:PhD defense; RL:Language, Speech & Vision;
- Lense: Optimizing data preprocessing in single-cell omics using LLMs
Data preprocessing is critical for single-cell omics analyses, but default pipelines often underperform on diverse datasets, especially from emerging platforms like spatial transcriptomics. We introduce Lense, a language-model-guided method that automatically selects optimal preprocessing by comparing plots that visualize low-dimensional representations across pipeline variants. Integrated with Seurat, Lense streamlines analysis and improves preprocessing robustness without requiring manual tuning.
- Cosine Similarity Conflates Clinically Distinct Cancer Variants: A Case for Typed-Graph Retrieval in Precision Oncology Decision Support
Cancer variant interpretation increasingly relies on retrieval from biomedical knowledge bases, with cosine similarity over neural text embeddings now the dominant retrieval substrate. Whether these embeddings preserve the entity-level distinctions that variant interpretation requires that BRAF V600E and V600K are distinct alleles, that EGFR L858R is a sensitizing mutation while T790M is a resistance mutation has not been systematically measured. We hypothesize that cosine-similarity retrieval over biomedical embeddings conflates clinically distinct cancer variants at high rates, while a typed-graph approach in which each variant is a discrete node preserves variant identity by construction. We constructed a benchmark of 9 cancer variant pairs known to have differential FDA-approved therapy indications or distinct molecular biology, curated from theCIViC clinical evidence database and primary clinical literature. Pairs included BRAF V600E vs V600K (melanoma), EGFR L858R vs T790M (NSCLC
- ParSeek: Accurate cryo-EM particle picking with a deep learning model trained on synthetic data
Accurate particle picking from noisy cryo-EM micrographs is essential for high-resolution reconstruction. Current deep learning methods rely on manually annotated data, which is labor-intensive, subjective, and limits particle recall under low signal-to-noise ratio (SNR). Here we introduce ParSeek, an automated picker trained entirely on synthetic data without human annotation. Synthetic micrographs are generated by projecting known 3D structures into realistic background patches that reproduce experimental noise. On seven public cryo-EM datasets, ParSeek outperformed Topaz and CryoSegNet on four datasets, achieving the highest F1-score (up to 0.82) and reaching 0.63 on a challenging membrane protein dataset. Density maps from ParSeek-picked particles showed cross-correlation coefficients up to 0.995 with the reference and a minimal resolution difference of 0.1 [A]. ParSeek also overcame severe orientation bias on an influenza dataset, yielding a reasonable reconstruction. Applied to t
- Facility-Scale Workflows for Data Acquisition, Standardization, Machine Learning Analysis, and Reproducible Science
Scientific user facilities routinely generate large-scale microscopy datasets across diverse instruments and vendors, differing substantially in file formats, dimensionality, and resolution. Beyond these inconsistencies, datasets are frequently fragmented living across isolated instruments and constrained by security policies and uneven metadata practices. Consequently, tracking, standardizing, processing, and visualizing these datasets in a manner compatible with modern machine learning and autonomous experimentation workflows remains a major challenge. While existing initiatives address data archiving, standardization, or analysis individually, few provide integrated solutions that bridge instrument-level acquisition and scalable ML workflows within heterogeneous, security-constrained user facilities. Here, we establish a deployable, facility-scale infrastructure that bridges instrument-level data generation with cloud-based ML analytics while remaining compliant with institutional n
- Predicting Discrete Structural Transformations in Small Molecules from Tandem Mass Spectrometry
Tandem mass spectrometry (MS/MS) fragments molecules into smaller pieces, generating spectra composed of m/z values and intensities that encode structural information for molecular annotation. With increasing mass spectrometry data acquisition speeds, manual annotation from MS/MS lags far behind data generation and remains a bottleneck in metabolite annotation. Current computational methods, such as molecular networking, address this challenge by organizing similar structures into families of related compounds. However, they generally provide only similarity scores, offering weak actionable insights for structural annotation. To address this limitation, we present the Molecular Transformation Graph Edit Measure (MT-GEM), a distance metric that quantifies discrete structural transformations between molecules through graph edge removals that approximate structural modifications. Building on this metric, we developed an ensemble machine learning architecture, the Spectrum Transformation E
- A fine-tuned genomic language model adds complementary nucleotide-context information to missense variant interpretation
Missense variant interpretation remains a central challenge in clinical genomics. Missense pathogenicity predictors achieve strong performance, but many emphasize protein-level consequences or overlapping annotation priors. Whether genomic language models add non-redundant nucleotide-context signal to missense interpretation remains unclear. Here, we systematically adapted genomic language models to ClinVar missense pathogenicity prediction across backbone architectures, representation strategies, classifier heads, and adaptation regimes. In our analysis, variant-position embeddings consistently outperformed pooled sequence representations, multi-species pretraining provided the strongest backbone-level advantage, and low-rank adaptation generalized better than full fine-tuning. The resulting fine-tuned model, GLM-Missense, substantially outperformed zero-shot scoring from the same pretrained model. To test whether GLM-Missense contributes information beyond existing methods, we built
- sxRaep: A Rapid and Accurate Enzyme Predictor for high-throughput mining of enzymatic sequences
Metagenomic sequencing generates petabyte-scale sequence datasets that strain both deep learning and alignment based enzyme annotation tools. A lightweight rapid and accurate filter tool is needed to identify enzymatic sequences prior to resource-intensive functional prediction. We present sxRaep (Rapid and Accurate Enzyme Predictor), a resource-efficient framework using lightweight physicochemical features for enzyme pre-screening. sxRaep achieves 6,604-fold speedup over Diamond (0.002 seconds per inference) with 62.1% memory reduction relative to Diamond (372 MB peak), while maintaining 99.4% accuracy and the highest recall in remote homology detection. This lightweight approach identifies enzymatic candidates missed by alignment-based methods without sacrificing accuracy.
- Dimensionally traceable 3D microstructures for multimodal microscope calibration
High-resolution microscopy techniques are used across research and industry to analyse biological systems, from biomolecules to subcellular organelles, multicellular models and tissues. As multimodal imaging workflows and quantitative analysis of bioimaging data become increasingly widespread, there is a growing need for materials and methods to calibrate imaging systems and evaluate the fidelity of generated image data. Here, we present three-dimensional microscopy phantoms fabricated using two-photon photolithography from transparent resins that exhibit both broadband visible autofluorescence and Raman scattering across the fingerprint and C-H stretching regions. Suitable for analysis using optical profilometry, the phantoms were dimensionally calibrated with SI traceability using a metrological confocal microscope. Immersible in air and common aqueous imaging media, the phantoms are compatible with a wide variety of optical microscopy techniques, including one and two-photon excited
- Goals as dynamical attractors: a momentum-based account of stable and flexible goal commitment
Human goal pursuit is often marked by persistent activity toward achieving an objective, as well as flexibility in switching objectives based on environmental demands. How humans balance the stability and flexibility necessary for goal pursuit is the key question of this study. We propose that goal pursuit generates dynamic attractor modes in policy landscapes that produce stability in goal pursuit. The attractor properties are modulated through progress monitoring, allowing for the flexibility necessary to switch objectives in favor of alternative goals. Through simulations and behavioral cloning of human participants performing an extended goal selection task, we show how dynamic modes can develop in the latent spaces of recurrent neural networks trained with reinforcement learning. We develop metrics to quantitatively assess the attractor qualities of dynamic modes, validating them against synthetically generated dynamical systems, and use them to investigate the context modulation
- Inferotemporal Cortex Joins the Circuit Before the Code: Non-Serial Inter-Area Synergy in the Macaque Ventral Stream
The ventral visual stream is widely modeled as a serial feedforward hierarchy in which V1, V4, and IT population codes develop sequentially during object recognition. We ask whether a second, concurrent coding mode exists---one organized not by anatomical order but by joint population structure across areas. Using Partial Information Decomposition applied to simultaneous multielectrode spiking recordings across all three areas at millisecond resolution---the first simultaneous three-area spiking PID analysis of the primate ventral stream---in two macaque monkeys viewing 25,000+ natural images, we decompose population coding into serial (unique per area) and synergistic (joint across areas) components at 5 ms resolution across five CNN target representations spanning low-level spatial features to high-level object identity. Three findings replicate across both animals and all five representations. First, synergistic inter-area coupling emerges before IT carries any unique object-related
- PIXEL: Adaptive Steering Via Position-wise Injection with eXact Estimated Levels under Subspace Calibration
PIXEL: Adaptive Steering Via Position-wise Injection with eXact Estimated Levels under Subspace Calibration irep.mbzuai.ac.ae
- LIME: a fully automated pipeline for high-throughput quantification of leaf lesions
Accurate quantification of leaf lesion severity is essential for plant disease research and phenotyping but is often limited by subjective visual scoring and time-intensive manual image analysis. We present LIME, a fully automated, open-source image analysis pipeline for high-throughput quantification of leaf lesions from disease assay images. LIME integrates zero-shot leaf segmentation using the Segment Anything Model with a convolutional neural network for lesion area estimation. Applied to Arabidopsis thaliana leaves infected with Sclerotinia sclerotiorum, the proposed approach achieved a mean absolute percentage error of 12.9%, comparable to observed intrarater variability in manual scoring. Stratified evaluation across lesion-size groups demonstrated consistent prediction accuracy for small, intermediate, and large lesions, and comparative analysis showed that the deep learning-based model substantially outperformed color-based baseline methods. Under GPU-accelerated execution, LI
- Healthcare workers' acceptance of artificial intelligence in cardiac diagnosis: Implications for medical education and training programs
Background The integration of artificial intelligence (AI) in cardiology requires healthcare worker acceptance for successful implementation. Understanding attitudes and educational needs is crucial for developing effective training programs. Methods A cross-sectional survey was conducted among 408 healthcare workers treating cardiac diseases in Riyadh, Saudi Arabia. We assessed AI acceptance, knowledge levels, and training preferences using validated scales. Statistical analyses included descriptive statistics, chi-square tests, correlation analysis, reliability testing, and multiple logistic regression. Results Of 408 participants, 407 provided complete responses. The sample comprised predominantly young (87.0% aged [≤]30), female (75.7%) medical residents (89.9%) with limited AI experience (86.7% never used AI clinically). Internal consistency was excellent (Cronbach's = 0.892). Moderate acceptance was observed: 49.9% were aware of AI applications in cardiology, 46.7% were willing t
- MISP-Bench: Decomposing User-Provided False Priors into Answer, Rationale, and Guard Effects
Large language models in clinical and educational settings routinely receive user-provided context containing incorrect prior beliefs. Existing benchmarks measure aggregate susceptibility to such priors but do not disentangle which structural component (the asserted answer, the supporting rationale, or their combination) drives the damage, nor test whether safety meta-prompts such as "verify the reasoning firs" consistently mitigate it. We introduce MISP-Bench, a factorial benchmark of 1,724 audited multiple-choice items (1,430 MedMCQA medical + 294 GSM8K quantitative) evaluated under 13 prompt conditions across 10 open-weight instruction-tuned models (1B-27B) in chain-of-thought and direct modes, with approximately 1.33M audited response records across three runs per condition. Distractors were generated by GPT-5.4 and the model was excluded from the evaluated set to prevent circular evaluation. Targeted and arbitrary distractor subsets yield similar aggregate Misinformation Damage In
- The German National Cohort: Ophthalmological Assessment, Baseline Profile and Potential for AI-based Eye Research
Objective: To describe the ophthalmic examination protocol within the German National Cohort (NAKO) / NAKO Gesundheitsstudie, to report the baseline profile of participants undergoing ophthalmological assessment, and to illustrate the potential of these data as a population-based open resource for artificial intelligence (AI) research in eye health. Design: Baseline analysis of ophthalmic data within the nationwide, population-based multicenter prospective NAKO study. Participants: 48,460 adults in the ophthalmological level 2 module of 205,053 adults enrolled in NAKO, aged 19-74 years, with mean age 48.9 {+/-} 12.5 years and 52.7% male. Methods: All participants underwent standardized assessments of a wide range of biomedical examinations and detailed questionnaire-based data collection, including non-dilated color fundus imaging, visual acuity testing, recording of a brief ocular history. Ocular and systemic health measures were summarized using descriptive statistics. Fundus image q
- Artificial intelligence for detecting bipolar disorder in electronic health records of patients with affective diagnoses: a diagnostic accuracy study
Background: Bipolar disorder (BD) is frequently underdiagnosed, particularly in patients presenting with depressive disorders, leading to delays in appropriate treatment. Artificial intelligence (AI) applied to electronic health records (EHRs) may improve early detection by identifying clinically relevant symptom patterns. Objective: To evaluate the diagnostic performance of a natural language processing (NLP)-based AI model for detecting BD-related features in EHRs of patients with affective diagnoses. Methods: A retrospective diagnostic accuracy study was conducted using 500 EHRs from a psychiatric referral hospital in Bogota, Colombia (2020-2024). The model extracted 18 predefined clinical domains from unstructured text and classified patients into four risk categories. Diagnostic performance was assessed in a validation subset of 100 records using independent psychiatric evaluation as the reference standard. Sensitivity, specificity, positive and negative predictive values, F1-scor
- SEIR-IoT cyber-physical architecture with dual parametric coupling for epidemic scenario simulation using synthetic biomedical signals
This study presents a proof-of-concept cyber-physical architecture integrating a SEIR epidemiological model (Susceptible-Exposed-Infectious-Recovered), implemented in MATLAB, with a simulated Internet of Things (IoT) acquisition and transmission stage based on the ESP32 microcontroller and the ThingSpeak platform. The system generates synthetic biomedical signals of body temperature and peripheral oxygen saturation (SpO2), structured across three levels: circadian variation, scheduled pathological episodes, and Gaussian noise. These signals feed a dual parametric coupling function that dynamically updates the SEIR transmission parameter as a combined function of body temperature and oxygen saturation deviations from their clinical reference values. The proposed architecture is organized into four functional phases: measurement, communication, computational processing, and feedback. Five simulated clinical scenarios were evaluated, ranging from normal conditions (T = 36.5 {degrees}C, Sp
- Scalable deep-learning-based inference of time-varying transmission dynamics from outbreak phylogenies
Infectious disease dynamics can be inferred from pathogen genomic data using phylodynamic methods, but the applicability of many such approaches to large data sets is constrained by computational cost. Recent deep-learning approaches to phylodynamics have improved scalability, yet challenges remain when genetic divergence is limited during fast spreading outbreaks. To address this, we use pathogen-specific models to show that deep-learning models trained on outbreak-like phylogenies can accurately estimate the reproductive number (R) when both the birth-death model and the expected phylogenetic resolution are matched to the target pathogen, highlighting the importance of realistic training conditions. Focusing on three major respiratory pathogens of public health importance (SARS-CoV-2, seasonal human influenza virus, and respiratory syncytial virus (RSV)), we introduce PhyloRt, a scalable framework for estimating the time-varying reproductive number (Rt) from large outbreak phylogenie
- A Deep Learning Framework for Spatiotemporal Modeling of Visual Task fMRI
Characterizing the dynamic coordination of distributed brain regions during cognitive tasks remains challenging, as traditional fMRI analysis focuses on localized activations without revealing the underlying information flow that drives them. Here, we propose STREAM (Spatiotemporal Representation for Effective connectivity Analysis Model), a deep-learning framework that learns neural transition functions in task-fMRI to characterize effective connectivity and whole-brain information flow. Applied to visual category processing in 1074 participants, STREAM accurately reconstructs activation maps while further revealing that traditional activation regions are primarily driven by incoming signals. Moreover, the Default Mode Network acts as a high-level regulatory hub with extensive outgoing influence, challenging its passive characterization. Additionally, category-specific communication emerges from dynamic reconfiguration of signaling patterns among key hubs rather than static pathways.
- Machine learning cross-platform proteomic imputation enables protein quality scoring and replication of epidemiological associations
High-throughput affinity-based proteomics has advanced biomedical research, yet fundamental, persistent discordance between mainstream platforms (SomaScan and Olink) routinely undermines the replication of findings. This platform-driven non-replication complicates downstream biological validation and biomarker prioritization. Here, we develop a machine learning-based framework for cross-platform protein value imputation to resolve this translational bottleneck. Using paired proteomic data measured by both SomaScan and Olink from 5,325 participants of the Multi-Ethnic Study of Atherosclerosis, we developed models to impute cross-platform measurements and applied them to two independent and demographically distinct cohorts (Cardiovascular Health Study [N=3,171] and UK Biobank [UKB; N=41,405]) for external validation. Our bi-directional model 1) established an imputation performance-based protein fidelity index, validated against gold-standard measurements from Atherosclerosis Risk in Com
- A distributed neural architecture of sustained affect across external and internal experience
Sustained affect shapes well-being, yet its neural architecture across externally elicited and internally generated experience remains unclear. Using whole-brain functional connectivity during minutes-long naturalistic movie viewing, we derived positive and negative affective experience signatures and their underlying neural architecture. These signatures predicted valence-specific affective intensity and generalized to independent movie-viewing data and internally generated affect, discriminating sad memory and rumination from neutral distraction while tracking subjective experience. Importantly, their expression showed little relation to vigilance or cognitive demand. Characterization of these signatures revealed coherent community structure and a shared distributed backbone, alongside valence-preferential components, consistent with a partially separable architecture. Extending beyond experimentally evoked states, in four resting-state depression cohorts, these signatures distinguis
- Linking live-cell behavior to transcriptional responses across perturbations using dynamic caging
Single-cell technologies, encompassing molecular, morphological, and functional assays, have emerged as cornerstones of modern biological research and discovery. However, current experimental methods often fail to explicitly link these 'omic' modalities, especially in live cells or longitudinally through time, impeding the study of multi-scale interactions and mechanisms of regulation. CellCage Enclosure (CCE) technology overcomes these limitations by dynamically compartmentalizing cells, allowing for scalable, live-cell, longitudinal exploration and simultaneous analysis of transcriptomic, proteomic, and morphological profiles. Using this novel technology, we generate previously inaccessible insights across various in vitro cellular systems under a diverse set of perturbations, including the discovery of morphological and proteomic features linked to immune suppressive gene set expression in human primary regulatory T cells (Tregs), as well as direct association of morphological and p
- STARMAP: A 3D-informed framework for mapping functional regions in proteins to regulatory and cellular phenotypes
Artificial Intelligence (AI) has transformed biology by revealing patterns in large-scale datasets and predicting regulatory relationships. Yet even the most advanced models often fail to identify biologically meaningful mechanisms from statistical associations. This limitation arises not from algorithmic capacity but from the lack of mechanistically grounded input features. Our structure-informed framework Structure-based Topological Analysis of Regulatory and Molecular Activity Patterns (STARMAP) embeds protein three-dimensional structure and population-scale functional genomics data into a unified representation for mechanistic inference. By mapping over 1.5 million naturally occurring variants across ~1,700 cancer cell lines onto protein structures, STARMAP was able to identify spatial clusters of variation associated with shifts in transcriptional regulatory networks and drug response phenotypes. This approach transforms natural genetic variation into a large-scale, structure-info
- A brain-inspired framework for memory prioritization in neural networks based on valence
Improving long-term memory in artificial neural networks remains an open challenge. To address this, we developed a novel brain-inspired framework for memory prioritization based on the principle of emotional valence. Our framework includes: (i) a valence-weighted cross-entropy loss that scales the learning signal by the valence magnitude, analogous to neuromodulation; (ii) an amygdala-inspired module that learns high-dimensional valence embeddings; and (iii) a hippocampus-inspired module that integrates valence embeddings into the attention mechanism to modulate information retrieval. We demonstrated the generalization of our framework across spatial, episodic, and language-based memory tasks, consistently improving memory prioritization and long-term retention of high-salience information. In addition to improving long-term memory, we also showed that our framework can help mitigate the lost-in-the-middle problem in language modeling. More generally, this research provides further ev
- TopoFuseNet: Hierarchical Graph Representation Learning with Multi-Scale Topological Features for Accurate Drug Synergy Prediction
Accurate prediction of drug synergy is paramount for developing effective combination therapies and advancing personalized medicine. Although methods based on graph neural networks (GNNs) have become a prevalent approach, they often treat molecules as flat graphs of connected atoms, thus overlooking their inherent hierarchical structure (i.e., atoms forming functional groups) and the critical topological information that governs molecular interactions. To address this limitation, we introduce TopoFuseNet, a novel hierarchical graph representation learning framework that integrates multi-scale topological features. The core innovations of TopoFuseNet include: 1) The first-ever application of "Group Centrality" from network science to cheminformatics, enabling the identification and quantification of functional groups crucial to drug activity; 2) A systematic, multi-path strategy to seamlessly integrate node-level (atom) and group-level (functional group) topological features into a Grap
- Towards Continuous Home Monitoring for Dementia: A Real-Time mmWave Radar Framework for Activity Classification and Tracking
Millimeter-wave radar can quietly monitor health and behavior at home, which is vital for supporting people living with dementia. Most studies, however, remain limited to short-term testing in controlled spaces. Realworld deployment requires robust activity classification as a prerequisite: vital-sign and behavioral sensing require fundamentally different processing pipelines, and absent periods need to be reliably distinguished from stationary states. Bridging the critical gap between controlled laboratory demonstrations and continuous home monitoring, this paper introduces a self-adapting radar framework that extracts meaningful behavioral segments from massive, unconstrained real-world data. The system performs continuous real-time activity classification (stationary, walking, and absent) and target localization, selectively directing downstream processing to the most informative segments. It addresses key real-world deployment challenges including adaptive thresholding across subje
- Evaluating the Sensitivity of Dry and Gel-Based Wearable EEG for Cognitive Load Estimation
Purpose: We present a large-scale (N=120) comparative study of gel-based and dry electroencephalography systems for cognitive load analysis in tasks involving information visualization stimuli. Although dry systems are increasingly adopted owing to their portability and fast setup, their sensitivity to cognitive-related measurements (as compared to gel-based systems) remains debated. This limits the understanding of whether dry systems provide sufficient sensitivity for cognitive load assessment under controlled task conditions. Methods: We analyzed a diverse set of signal quality metrics, such as signal-to-noise ratio and channel retention, combined with spectral features across frequency bands to evaluate the ability for each device to capture workload-related neural markers during information visualization tasks. Results: Although the gel-based device showed consistently better quality results than the dry one, the effect sizes suggest a small practical significance of the differenc
- Characterizing Resting-State Brain Dynamics with Frequency-Resolved EEG Microstates: Parallel Analyses of Psilocybin Microdosing and Acute Inhaled DMT
Electroencephalographic (EEG) microstates provide a compact framework for characterizing the temporal organization of large-scale brain activity, yet their sensitivity to altered brain states remains insufficiently explored. In this study, we applied broadband and frequency-resolved EEG microstate analysis to resting-state EEG data from two publicly available datasets acquired under markedly different altered-state conditions: psilocybin microdosing and acute inhaled N,N-dimethyltryptamine (DMT). The aim was to determine whether narrowband microstate analysis reveals structured alterations in resting-state brain dynamics beyond those captured by broadband analysis alone. Psilocybin microdosing was associated with relatively subtle effects, including reduced global field power and frequency-specific alterations in delta- and theta-band microstate parameters, while no significant broadband spatiotemporal changes were observed. In contrast, acute inhaled DMT was associated with broader mi
- A Differentiable dFBA Simulator for Scalable Bayesian Inference over Microbial Metabolic Models
Medium optimisation for bioprocess design remains challenging and costly: fermentation recipes typically contain ten or more components, the design space expands combinatorially as ingredients are added, and each batch experiment requires over 24 hours. High-throughput 96-well plate screening can reduce experimental cost, but extracting actionable predictions from growth curves requires a mechanistic model that links medium composition to cellular metabolism. In this paper, we present a differentiable simulator for dynamic flux balance analysis (dFBA) that enables scalable Bayesian inference over microbial metabolic models. A distinguishing feature is that inference is driven entirely by OD600 measurements, a simple optical proxy for biomass, without substrate or product assays; internal fluxes, substrate consumption, and secreted metabolite profiles are recovered as latent variables constrained by the metabolic network stoichiometry. We resolve the core differentiability barrier of cl
- CT Attenuation Map Derived Body Composition Is Associated with Cardiorespiratory Fitness in Multicenter External Validation
Aim: Exercise capacity is a powerful predictor of cardiovascular risk. In patients unable to exercise, body composition analysis can potentially be used to estimate cardiorespiratory fitness. We developed a body composition fitness score, then validated its utility in two external populations. Methods and Results: We included patients from four sites undergoing single photon emission computed tomography (SPECT) and twelve sites undergoing positron emission tomography (PET). We quantified body composition using deep learning. We evaluated associations between body composition and good exercise capacity (defined as completing [≥]7 minutes on a Bruce protocol) then developed a body composition fitness score. We then assessed the associations of fitness score with exercise capacity and all-cause mortality in external populations. In total, 36471 patients were included with median age 67 (interquartile range 58 - 74). Median skeletal muscle density was higher among patients with good exerci
- Automated Brain and CSF Volume Assessment in Infant Hydrocephalus Using Deep Learning
Accurate brain and cerebrospinal fluid (CSF) volume assessment is essential for pediatric hydrocephalus management. Current clinical practice relies on linear measurements that fail to capture complex three-dimensional ventricular morphology, while quantitative volumetric assessment remains limited by laborious processing and lack of clinically optimized automated tools. This study developed a rapid, automated AI-based intracranial segmentation model suitable for clinical workflows. We retrospectively analyzed 167 T2-weighted MRI scans from infants with hydrocephalus, randomly split into training (60%), validation (20%), and hold-out test (20%) sets. All scans were manually segmented into CSF, brain parenchyma, and background. Our model integrates DenseNet and U-Net architectures with feature smoothness regularization to enhance generalizability. Performance was evaluated using Dice scores and absolute relative volume error (ARVE) compared with state-of-the-art methods. The AI model ac
- Simpler is not always better: Phylodynamic misspecification and deep-learning corrections
Phylodynamics bridges the gap between epidemiology and pathogen genetic data by estimating epidemiological parameters from time-scaled pathogen phylogenies. Multi-type birth-death (MTBD) models are phylodynamic analogies of compartmental models in classical epidemiology. They serve to infer the average number of secondary infections R and the infection duration d. Moreover, more complex MTBD models add extra parameters, such as the average length of the incubation period or the proportion of superspreaders in the infected population. However, these additional parameters come at an important computational cost: Apart from the simplest, BD, model, MTBD models do not have a closed-form solution and require numerical methods for their likelihood computation. This leads to increased computational times and potential numerical errors. Therefore, the BD model remains the favorite researchers' choice for real dataset analyses, and is often applied even in cases where more complex epidemiologic
- DentaCoPilot: An LLM-Augmented Next-Procedure Recommender for General Dentistry, Designed for Dentist Augmentation
Background. Commercial dental artificial intelligence in 2026 is overwhelmingly diagnostic: caries, calculus, periapical, and bone-level detection on radiographs. The clinically harder question that follows every diagnosis-given a patient's chart and most recent procedure, what should the dentist do next-remains unsolved at general-dentistry scale. The closest published system, MultiTP (Chen et al., 2024), is a CNN-RNN restricted to partial-edentulism cases and provides neither calibrated uncertainty, structured rationale, nor an evaluation that treats the model as decision support instead of an autonomous classifier. Methods. We introduce DentaCoPilot, a recommender that, given a structured chart, returns (i) a calibrated top-K probability distribution over Current Dental Terminology (CDT) codes for the next procedure,(ii) a verbalised confidence label, (iii) an explicit abstain flag when context is insufficient, and (iv) a chart-grounded rationale. We compare four classical baselines
- Transforming Semi-structured Variant Assessments into Computable Clinical Assertions: A Pilot Study for AI-Assisted Curation
Genomic medicine relies on expert evaluation of genomic variants, but this process is dramatically slowed by a lack of readily-accessible genomic knowledge. Although genomic knowledge resources such as ClinVar and CIViC support structured data sharing and provide interfaces for adding structure, much of the variant interpretation data generated upstream of these resources is not readily interoperable with these resources, limiting the ability of clinical labs to share data and creating knowledge silos. Here we evaluate a strategy for breaking down these knowledge silos in a pilot study to transform semi-structured variant classification knowledge into computable clinical assertions leveraging the Global Alliance for Genomics and Health (GA4GH) Genomic Knowledge Standards specifications. We programmatically mapped previously captured somatic cancer clinical significance classifications from spreadsheets to the GA4GH Variant Annotation specification. For diagnostic classification data, t
- Reproducible Biochemical Clusters Embedded Within a Continuous Neurochemical Landscape of Autism Spectrum Disorder Revealed by NeuroCLAD
Abstract Background Autism Spectrum Disorder (ASD) is marked by pronounced biological heterogeneity, yet most neurochemical studies have relied on single-analyte comparisons that cannot capture coordinated variation across neurotransmitter systems. Whether ASD blood neurotransmitter profiles reflect discrete subtypes, a continuous landscape, or something in between remains unresolved. Methods We applied NeuroCLAD, a structured multivariate analytical framework, to peripheral blood neurotransmitter profiles from 261 children with ASD (mean age 6.98 [SD 3.13 years]; 78.5% male). The pipeline incorporated z-score normalisation, natural cubic spline residualisation for age and sex, principal component analysis, k-means clustering, consensus stability assessment, Gaussian mixture modelling, Cohen's d enrichment analysis, and clinical symptom mapping. Cross-compartment consistency was explored using urine neurotransmitter profiles from the same cohort. Results Twelve reproducible biochemical
- SHANK3-anchored reverse phenotyping identifies a rare-variant-enriched cognitive-motor subgroup of autism
Rare deleterious variants in SHANK3 are established causes of autism spectrum disorder (ASD), but the extent to which they define a phenotypically and genetically coherent ASD subgroup remains unclear. Using the SPARK cohort, we identified 132 SHANK3 variant carriers; 108 had phenotype data and were compared with 47,555 non-carrier ASD cases. SHANK3 damaging variant carriers showed lower cognitive ability, poorer motor coordination, and delayed developmental milestones. Protein-truncating variant and deletion carriers showed similarly severe phenotypic profiles, whereas duplication carriers did not differ from non-carriers. A combined threshold of intelligence quotient (IQ) < 70 and impaired motor coordination (DCDQ total score) < 35 defined a discriminative cognitive-motor phenotype among cases meeting this cognitive-motor phenotype. Beyond SHANK3, SLC6A1 was the only additional gene reaching false discovery rate significance, while pathway analyses implicated synaptic and chromatin-r
- GlazyBench: A Benchmark for Ceramic Glaze Property Prediction and Image Generation
Developing ceramic glazes is a costly, time-consuming process of trial and error due to complex chemistry, placing a significant burden on independent artists. While recent advances in multimodal AI offer a modern solution, the field lacks the large-scale datasets required to train these models. We ...
- Recursive Agent Optimization
We introduce Recursive Agent Optimization (RAO), a reinforcement learning approach for training recursive agents: agents that can spawn and delegate sub-tasks to new instantiations of themselves recursively. Recursive agents implement an inference-time scaling algorithm that naturally allows agents ...
- Can RL Teach Long-Horizon Reasoning to LLMs? Expressiveness Is Key
Reinforcement learning (RL) has been applied to improve large language model (LLM) reasoning, yet the systematic study of how training scales with task difficulty has been hampered by the lack of controlled, scalable environments. We introduce ScaleLogic, a synthetic logical reasoning framework that...
- MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems
Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-tri...
- SkillOS: Learning Skill Curation for Self-Evolving Agents
LLM-based agents are increasingly deployed to handle streaming tasks, yet they often remain one-off problem solvers that fail to learn from past interactions. Reusable skills distilled from experience provide a natural substrate for self-evolution, where high-quality skill curation serves as the key...
- The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity
Despite the prevalence of the attention sink phenomenon in Large Language Models (LLMs), where initial tokens disproportionately monopolize attention scores, its structural origins remain elusive. This work provides a \textit{mechanistic explanation} for this phenomenon. First, we trace its root to ...
- StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction
Large language models (LLMs) are increasingly used as interactive agents, but optimizing them for long-horizon decision making remains difficult because current methods are largely purely reactive, which weakens both exploration and credit assignment over extended trajectories. In this work, we pres...
- Concept-Based Abductive and Contrastive Explanations for Behaviors of Vision Models
*Concept-based explanations* offer a promising approach for explaining the predictions of deep neural networks in terms of high-level, human-understandable concepts. However, existing methods either do not establish a causal connection between the concepts and model predictions or are limited in exp...