SEARCH · PubMed Health

Results for “Deep Learning”

Explore indexed PubMed citations for clinical trials, systematic reviews and public health research. Read source abstracts and follow each citation to its original PubMed record.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 109 records · Page 6Linked to original sources

PubMed2026

abCRISPR: deep learning-based design of abasic gRNA sequences for specific CRISPR-Cas9 genome editing.

SUMMARY: CRISPR-Cas9 has become a widely used tool for genome editing. However, its off-target cleavage caused by partial sequence matches with guide RNAs (gRNAs) remains a critical limitation. Recently, abasic gRNAs (ØXØ) have been developed to enhance target specificity, but their effects vary depending on the positional sequence context. Here, we present abCRISPR, a deep neural network (DNN) framework for the rational design of ØXØ sequences with minimized off-target activity. abCRISPR leverages informative few-shot training with paired datasets of abasic and unmodified gRNAs, using high-quality random mismatch target libraries, exhaustively sequenced for mismatched off-target substrates (n = 97583) in in vitro CRISPR-Cas9 cleavage experiments. Predicted off-target activities for both abasic and unmodified gRNAs showed strong correlation with experimental data (r ≥ 0.95, 10-fold cross-validation). Notably, these comprehensive training sets provide robust ground-truth negatives, enabling accurate and sensitive prediction of off-targets. For unmodified gRNAs, abCRISPR (AUC = 0.98) was validated to outperform existing deep learning-based methods (AUC = 0.45-0.68). When applied to the human genome, abCRISPR generated ØXØ sequences, covering 58 875 004 potent CRISPR-targetable sites with improved target specificity. Together, this work provides a comprehensive bioinformatics resource for safe and precise CRISPR-Cas9 genome editing. AVAILABILITY AND IMPLEMENTATION: The source code for abCRISPR and training data are available at https://doi.org/10.5281/zenodo.20398246. abCRISPR results for the human genome are available at http://clip.korea.ac.kr/abCRISPR/.

Deep Learning↗

PubMed2026

PEARL: integrative multi-omics classification and omics feature discovery via deep graph learning.

MOTIVATION: Integrating multi-omics data provides valuable insights into biological processes by capturing information across multiple molecular layers, enabling a comprehensive understanding of complex diseases and driving advancements in precision medicine. However, existing computational methods for multi-omics integration face significant challenges, such as low reliability and poor generalizability, due to the high dimensionality and low sample size nature of omics data. RESULTS: To address these challenges, we present PEARL (Pearson-Enhanced spectrAl gRaph convoLutional networks), a novel deep graph learning method for biomedical classification and functional important omics features identification. PEARL leverages a simple yet effective learning architecture to achieve superior and robust performance in high-dimensional, low-sample-size multi-omics settings. Our results demonstrate that PEARL significantly outperforms existing state-of-the-art methods on both synthetic and real biomedical datasets. Furthermore, applied to Alzheimer's disease (AD) brain multi-omics data, features prioritized by PEARL lead to functionally important genes that demonstrate significant enrichment in AD-related pathways. These findings highlight PEARL's practical utility in biomedical research and its potential to enhance biological interpretability in multi-omics studies. AVAILABILITY AND IMPLEMENTATION: The source code of our computational framework is available at https://github.com/zqq121017/PEARL.

Multiomics↗

PubMed2026

Deep learning-assisted, pathogenesis-informed lung histopathology scoring in preclinical mouse models of SARS-CoV-2 and influenza A infection.

INTRODUCTION: SARS-CoV-2 and influenza A virus (IAV) cause viral pneumonia, yet their lung lesions evolve with distinct spatial organization and resolution-phase architecture. In preclinical murine studies, H&E histopathology is a primary endpoint, but burden-focused semiquantitative scoring can miss pathogen- and phase-specific differences in lesion topology, compartmental involvement, inflammatory organization, and repair. We aimed to define virus- and phase-specific morphologic signatures and translate them into a practical, pathogenesis-informed scoring guide, supported by whole-slide convolutional neural network (CNN) analysis with class activation mapping (CAM). METHODS: Mice were infected under standardized conditions and evaluated during the early, peak-injury, and late phases of infection, corresponding to 2~3, 5~8, and 14 days post-infection (dpi), respectively. Lungs were assessed by H&E with semiquantitative scoring and by immunostaining to map viral antigen distribution and epithelial tropism. Whole-slide CNN models were trained for virus- and phase-specific classification, and CAM localized discriminative regions. RESULTS: Dose titration established reproducible lethal and sublethal infection conditions for both viruses. Viral antigen kinetics diverged, with SARS-CoV-2 peaking early and declining toward clearance by the resolution phase, whereas IAV peaked later and declined by the resolution phase, paralleling distinct injury-repair trajectories. CNN/CAM analysis distinguished virus- and phase-specific histologic patterns across the early, peak-injury, and resolution phases of infection and highlighted spatial signatures consistent with expert review. At the peak-injury phase, SARS-CoV-2 lungs showed broad alveolar/interstitial involvement, whereas IAV exhibited bronchocentric inflammatory organization. During the resolution phase, IAV showed prominent epithelial regeneration with remodeling-forward architecture, while SARS-CoV-2 more often retained localized residual inflammatory foci. Across both infections, tissue inflammatory composition shifted over time, with higher neutrophil representation during the peak-injury phase and a relative increase in lymphocytic representation during the resolution phase. Integrating lesion topology/distribution, edema, epithelial injury-regeneration, remodeling features, and lymphocyte predominance, we proposed a pathogen-resolved, phase-informed histopathology scoring guide with recommended evaluation windows for each model. CONCLUSION: Together, these findings define virus- and phase-specific morphologic programs that inform respiratory virus pathogenesis in mice and can be translated into practical scoring criteria for preclinical respiratory virus studies.

Animals↗

PubMed2026

Predicting 5-Year Mortality in Non-Small-Cell Lung Cancer Using the Korean Central Cancer Registry: Model Development and Validation Study.

BACKGROUND: Non-small-cell lung cancer (NSCLC) is one of the most common cancers and a leading cause of cancer-related mortality, making prognostic prediction clinically essential. Machine learning models are increasingly used to assess prognosis; however, developing systems that combine high discrimination with clear, clinically interpretable reasoning remains challenging. OBJECTIVE: This study aimed to develop deep learning models that predict 5-year mortality in NSCLC using data from the Korea Central Cancer Registry and quantify feature importance through permutation testing. METHODS: We identified 3144 patients diagnosed between 2014 and 2017 who had complete clinical data, pulmonary function test results, histological information, genomic data, and staging details. After preprocessing, the cohort was divided into stratified training, validation, and test sets in a 70%-15%-15% ratio. Five models were tuned using Hyperband across 10 predefined feature groups. The primary evaluation metric was the area under the receiver operating characteristic curve (AUC); additional metrics included accuracy, F1-score, precision, and recall. Groupwise permutation importance was calculated for each model, and the concordance of importance rankings was assessed using the Friedman test. RESULTS: All 5 models yielded comparable discrimination values on the test set (AUC=0.875-0.879). Model A was selected as the primary model and achieved an AUC of 0.879, an accuracy of 0.806, an F1-score of 0.824, and a Brier score of 0.142. Permuting the stage resulted in the largest decrease in AUC (0.217), followed by the pulmonary function test (0.016). Gene mutation had a modest overall impact but became more influential within the adenocarcinoma subset. The Friedman test showed no statistically significant differences in importance rankings across the models (P=.93). CONCLUSIONS: A grouped-input deep learning framework achieved discrimination comparable to a conventional Cox proportional hazards model using the same routine clinical variables for 5-year mortality prediction in NSCLC. Group-level permutation importance provided stable and reproducible insights into the clinical factors influencing risk, which may guide future model refinement and clinical decision-making.

Humans↗

PubMed2026

Deep learning-based multimodal pathogenomics integration for precision cancer prognosis.

BACKGROUND: Recent studies have revealed valuable prognostic insights in haematoxylin and eosin (H&E)-stained histological sections and transcriptomic profiles, suggesting potential applications in machine learning. However, existing methods lack sufficient intra- and inter-modal interactions, and face challenges in clinical validation due to incomplete multimodal data. METHODS: We proposed PathoGems (PathoGenomics-based integrative survival prediction), a weakly-supervised, interpretable multimodal learning framework that integrates histology and genomic profiles for precise cancer prognosis prediction. To evaluate the robustness of PathoGems, we initially curated a dataset of 1965 cases across four cohorts from The Cancer Genome Atlas (TCGA), including breast, colorectal, glioblastoma, and esophageal cancers. For external validation, PathoGems was further evaluated on four independent cohorts, consisting of 76 breast cancer and 41 esophageal squamous cell carcinoma cases from Zhejiang Cancer Hospital, as well as 102 colorectal cancer and 58 glioblastoma cases from the Clinical Proteomic Tumor Analysis Consortium (CPTAC). RESULTS: PathoGems effectively stratified patients into favorable and unfavorable risk groups, revealing significant differences in histological patterns, genomic features, and overall survival (log-rank test, p < 0.05). Moreover, the model’s predictions are further supported by visualization and transcriptomic analysis, enhancing interpretability and reliability. CONCLUSIONS: By fusing histological and clinicogenomic multimodal models, PathoGems will provide a solid foundation for developing an innovative tool that aids clinicians in making informed decisions and selection personalized treatment strategies for cancer patients.

Humans↗

PubMed2026

Non-destructive prediction of lead content in oilseed rape leaves by fluorescence hyperspectral technology based on neural network.

Based on fluorescence hyperspectral imaging (FHSI), this study targeted rapid, non-destructive quantification of lead (Pb) content in oilseed rape leaves treated with varying silicon (Si) concentrations, acquiring fluorescence spectra over the 484.43-1001.61 nm wavelength range. To optimize spectral data quality, preprocessing methods (Savitzky-Golay smoothing, first derivative, detrending) were comprehensively compared. Characteristic wavelengths were then selected via interval variable iterative shrinkage, which effectively compressed data dimensionality and reduced computational load. A hybrid SE-CL1DA model, fusing a 1D convolutional neural network, a long short-term memory network and SE attention mechanism was constructed, with Bayesian optimization tuning hyperparameters to boost stability. The BO-SE-CL1DA outperformed both traditional machine learning and insufficiently optimized deep learning model (Rp2=0.9609, RMSE = 0.0377 mg/kg, RPD = 5.1736), thus enabling accurate Pb estimation, supporting Si-regulated heavy metal stress management and facilitating agricultural contamination monitoring.

Plant Leaves↗

PubMed2026

Data-driven approaches in green microbiology: strategies for plant growth-promoting bacteria.

Plant growth-promoting bacteria (PGPB) are gaining attention as scalable biological solutions to enhance crop productivity and resilience. However, accurately identifying and characterizing PGPB remains challenging, particularly under variable environmental conditions where microbial functions are context-dependent and shaped by complex plant-microbe interactions. Advances in high-throughput sequencing have shifted the field from culture-dependent approaches to genome-informed strategies, enabling large-scale taxonomic and functional profiling. Although trait-based databases support the prediction of plant-beneficial genes, they capture only a fraction of the underlying biological complexity and often require labor-intensive analyses. Machine learning (ML) and deep learning (DL) have emerged as powerful tools to integrate genomic, physiological, and ecological data, enabling the prioritization of candidate strains with plant growth-promoting potential. To evaluate advances in the field, we conducted a systematic review of studies integrating ML and DL with PGPB characterization, assessing algorithm selection, performance, and target plant systems. Across 248 observations, only 6.0% of studies directly addressed PGPB screening, whereas the majority (77.4%) focused on plant disease detection, revealing a substantial gap in the application of AI to beneficial microorganisms for plant growth. Convolutional neural networks (CNNs) were the most frequently applied algorithms, largely driven by image-based phenotyping tasks. Overall, the field is constrained by limited datasets, high computational demands, and challenges in modeling multispecies and host-associated interactions. We highlight the need for integrative and interpretable ML and DL frameworks that bridge genomic data and functional validation. Such approaches represent a promising path toward scalable, data-driven discovery and deployment of bioinoculants in sustainable agriculture.

Agriculture↗

PubMed2026

Exposome influences: a multi-omics perspective on the combined toxic effects of pharmaceuticals and personal care products in Alzheimer's disease.

According to WHO data, approximately 57 million people worldwide were affected by dementia in 2021, with prevalence projected to rise. Alzheimer's disease (AD), responsible for 60%-80% of dementia cases, continues to be a leading cause of mortality, with current treatments offering limited efficacy and disease-modifying therapies lacking widespread adoption or conclusive safety evidence, shifting the focus toward prevention and risk modification. Risk factors for AD include both non-modifiable elements, such as age, genetics, and gender, and modifiable factors, like environmental pollution, health status, and diet. While age remains the primary non-modifiable risk factor, early-onset dementia represents only up to 9% of cases. Addressing modifiable factors is essential, as it could prevent or delay almost half of dementia cases, with interventions-such as increased physical activity, smoking cessation, alcohol limitation, and overall health management-being significantly associated with a reduced risk. In this context, the exposome approach offers a comprehensive, integrative framework in which both modifiable and non-modifiable risk factors interact to influence individual susceptibility. Within the neural exposome, chronic low-dose exposure to xenobiotics-such as industrial chemicals, pesticides, metals, pharmaceuticals and personal care products (PPCPs), and air pollutants-may induce neurodegeneration via mechanisms including oxidative stress, neuroinflammation, proteinopathies, and epigenetic modifications, although establishing causality remains challenging. Integration of genomics, transcriptomics, proteomics, metabolomics, and lipidomics, combined with artificial intelligence (AI) techniques such as machine learning (ML) and deep learning (DL), provides promising avenues for biomarker discovery, enhanced preventive strategies, early non-invasive diagnosis, and therapeutic target identification by integrating multi-layered biological data with exposure profiles. This review highlights emerging AD risk factors-including PPCPs-underscoring complex, multifactorial nature of AD and exposome, and the requirement for an interdisciplinary research approach, while also addressing several critical research gaps and methodological limitations.

Alzheimer’s disease↗

PubMed2023

Active learning of enhancer and silencer regulatory grammar in photoreceptors.

Cis-regulatory elements (CREs) direct gene expression in health and disease, and models that can accurately predict their activities from DNA sequences are crucial for biomedicine. Deep learning represents one emerging strategy to model the regulatory grammar that relates CRE sequence to function. However, these models require training data on a scale that exceeds the number of CREs in the genome. We address this problem using active machine learning to iteratively train models on multiple rounds of synthetic DNA sequences assayed in live mammalian retinas. During each round of training the model actively selects sequence perturbations to assay, thereby efficiently generating informative training data. We iteratively trained a model that predicts the activities of sequences containing binding motifs for the photoreceptor transcription factor Cone-rod homeobox (CRX) using an order of magnitude less training data than current approaches. The model's internal confidence estimates of its predictions are reliable guides for designing sequences with high activity. The model correctly identified critical sequence differences between active and inactive sequences with nearly identical transcription factor binding sites, and revealed order and spacing preferences for combinations of motifs. Our results establish active learning as an effective method to train accurate deep learning models of cis-regulatory function after exhausting naturally occurring training examples in the genome.

Journal Article↗

PubMed2026

BaGGLS: a Bayesian shrinkage framework for interpretable modeling of interactions in high-dimensional biological data.

MOTIVATION: Biological data is often high dimensional, noisy, and governed by complex interactions among sparse signals. This poses major challenges for interpretability and reliable feature selection. Tasks such as identifying motif interactions in genomics exemplify these difficulties, as only a small subset of biologically relevant features (e.g. motifs) are typically active, and their effects are often non-linear and context-dependent. While statistical approaches often result in more interpretable models, deep learning models have proven effective in modeling complex interactions and prediction accuracy, yet their black-box nature limits interpretability. RESULTS: We introduce BaGGLS, a flexible and interpretable probabilistic binary regression model designed for high-dimensional biological inference involving feature interactions. BaGGLS incorporates a Bayesian group global-local shrinkage prior, aligned with the group structure introduced by interaction terms. This prior encourages sparsity while retaining interpretability, helping to isolate meaningful signals and suppress noise. To enable scalable inference, we employ a partially factorized variational approximation that captures posterior skewness and supports efficient learning even in large feature spaces. In extensive simulations, we compare BaGGLS to frequentist probit regressions (unconstrained and with L1-penalty) as well as a probit model with Markov Chain Monte Carlo (MCMC) sampling under a horseshoe prior. We can show that BaGGLS outperforms the other methods with regard to interaction detection and is many times faster than MCMC sampling under the horseshoe prior. We also demonstrate the usefulness of BaGGLS in the context of interaction discovery from motif scanner outputs (e.g. Find Individual Motif Occurrences (FIMO)) and noisy attribution scores from deep learning models. This shows that BaGGLS is a promising approach for uncovering biologically relevant interaction patterns, with potential applicability across a range of high-dimensional tasks in computational biology. AVAILABILITY: Code is available at gitlab.com/dacs-hpi/baggls.

Bayes Theorem↗

PubMed2026

GraphyloVar: predicting the impact of non-coding variants using a multi-species sequence model.

MOTIVATION: Understanding the functional impact of genetic variants is a key problem for precision medicine. Tools like CADD, PhyloP, and PhastCons are useful, but they often look at each position in the genome in isolation. This means they can miss important information from the evolutionary history that connects different species. In this paper, we extend our previous model, Graphylo, to predict the effects of variants. Our new model, GraphyloVar, is built to directly utilize the phylogenetic tree that relates the species. RESULTS: GraphyloVar is a deep learning model that considers both DNA sequence and evolutionary patterns from many species. It uses two main components: Graph Convolutional Networks (GCNs) to process the phylogenetic tree, and Transformer encoders to extract features from the DNA sequences. Pre-trained to predict population-level allele frequencies on the TOPMed whole-genome sequencing cohort, GraphyloVar achieves an AUROC of 0.6246 zero-shot on ∼149M held-out variants, and an ensemble with CADD reaches 0.6442 (+0.020, P<10-15). Fine-tuned GraphyloVar achieves the highest AUROC across all 13 MPRA benchmark datasets. By integrating deep learning with explicit phylogenetic input, GraphyloVar offers a powerful and complementary approach to variant effect prediction that utilizes the full evolutionary history from many species to better identify and prioritize important non-coding variants. AVAILABILITY AND IMPLEMENTATION: Code and datasets are available at https://github.com/DongjoonLim/GraphyloVar under DOI: 10.5281/zenodo.20616818.

Phylogeny↗

PubMed2025

Leveraging protein language models for cross-variant CRISPR/Cas9 sgRNA activity prediction.

MOTIVATION: Accurate prediction of single-guide RNA (sgRNA) activity is crucial for optimizing the CRISPR/Cas9 gene-editing system, as it directly influences the efficiency and accuracy of genome modifications. However, existing prediction methods mainly rely on large-scale experimental data of a single Cas9 variant to construct Cas9 protein (variants)-specific sgRNA activity prediction models, which limits their generalization ability and prediction performance across different Cas9 protein (variants), as well as their scalability to the continuously discovered new variants. RESULTS: In this study, we proposed PLM-CRISPR, a novel deep learning-based model that leverages protein language models to capture Cas9 protein (variants) representations for cross-variant sgRNA activity prediction. PLM-CRISPR uses tailored feature extraction modules for both sgRNA and protein sequences, incorporating a cross-variant training strategy and a dynamic feature fusion mechanism to effectively model their interactions. Extensive experiments demonstrate that PLM-CRISPR outperforms existing methods across datasets spanning seven Cas9 protein (variants) in three real-world scenarios, demonstrating its superior performance in handling data-scarce situations, including cases with few or no samples for novel variants. Comparative analyses with traditional machine learning and deep learning models further confirm the effectiveness of PLM-CRISPR. Additionally, motif analysis reveals that PLM-CRISPR accurately identifies high-activity sgRNA sequence patterns across diverse Cas9 protein (variants). Overall, PLM-CRISPR provides a robust, scalable, and generalizable solution for sgRNA activity prediction across diverse Cas9 protein (variants). AVAILABILITY AND IMPLEMENTATION: The source code can be obtained from https://github.com/CSUBioGroup/PLM-CRISPR.

CRISPR-Cas Systems↗

PubMed2025

HINN: Hierarchical Input Neural Network identifies multi-omics biomarker for cognitive decline.

Understanding complex diseases requires models that can integrate diverse layers of biological data while yielding insights that are biologically interpretable. Although multi-omics integration with machine learning (ML) has advanced disease prediction and biomarker discovery, most existing approaches overlook the hierarchical and regulatory relationships that connect these molecular layers. Here, we present the Hierarchical Input Neural Network (HINN), a deep learning framework that incorporates known cross-omics relationships directly into its architecture, capturing the flow of information from genomics to epigenomics, transcriptomics, and downstream biological processes. By embedding these relationships, HINN improves both predictive performance and biological interpretability. We applied HINN to blood-derived multi-omics data from individuals with Alzheimer's disease or mild cognitive impairment to predict cognitive scores from standardized assessments. HINN outperformed both baseline and state-of-the-art models and pinpointed multi-omics biomarkers-including SNPs and promoter-region CpG sites in ATP6V1C1 and RCHY1 -that were significantly correlated with plasma p-Tau181 levels. These features map to biologically relevant processes with potential implications for cognitive decline. Our findings demonstrate how combining deep learning with biological knowledge can uncover interpretable, blood-based biomarkers for cognitive decline due to complex diseases such as Alzheimer's. All code and data are openly available at https://github.com/bozdaglab/HINN.

Alzheimer’s disease↗

PubMed2026

Machine learning-driven spleen imaging and genomics uncover a splenic connection to coronary artery disease.

Despite advances in managing traditional risk factors, coronary artery disease (CAD) remains the leading cause of mortality. Circulating hematopoietic cells influence risk for CAD separately from traditional risk factors, but the role of a key regulating organ, the spleen, is unknown. The understudied spleen is a representation of the hematopoietic system optimally suited for unbiased radiologic investigations toward mechanistic insights. Here, we leveraged deep learning to extract 107 splenic radiomic features from abdominal magnetic resonance imaging (MRI) scans of 42,059 UK Biobank participants and of 2745 Mass General Brigham Biobank (MGBB) participants. Of these, 10 features from UK Biobank were associated with CAD. Genome-wide association analysis of CAD-associated features identified 219 loci, including 9p21. Variants at 9p21, the strongest yet mechanistically elusive CAD locus, were associated with splenic features such as run-length nonuniformity, reflecting heterogeneity of continuous texture regions. Research MRI findings were consistent internally, but external clinical validation highlighted challenges in translating analyses of abdominal MRI scans to routine clinical practice because of variability in imaging protocols and greater clinical heterogeneity among patients. Our study, combining deep learning with genomics, presents a framework to uncover potential splenic involvement in CAD and emphasizes translational gaps between research and clinical radiomics.

Humans↗

PubMed2025

Active learning of enhancers and silencers in the developing neural retina.

Deep learning is a promising strategy for modeling cis-regulatory elements. However, models trained on genomic sequences often fail to explain why the same transcription factor can activate or repress transcription in different contexts. To address this limitation, we developed an active learning approach to train models that distinguish between enhancers and silencers composed of binding sites for the photoreceptor transcription factor cone-rod homeobox (CRX). After training the model on nearly all bound CRX sites from the genome, we coupled synthetic biology with uncertainty sampling to generate additional rounds of informative training data. This allowed us to iteratively train models on data from multiple rounds of massively parallel reporter assays. The ability of the resulting models to discriminate between CRX sites with identical sequence but opposite functions establishes active learning as an effective strategy to train models of regulatory DNA. A record of this paper's transparent peer review process is included in the supplemental information.

Retina↗

PubMed2026

Flexible use of conserved motifs constrains genome access in cell type evolution.

Cell types can be organized into related families, but the regulatory mechanisms that define and maintain these families across deep evolutionary time remain unknown. Here, combining single-nucleus multi-omic sequencing with deep learning to analyse the accessible genomes of two groups of vastly divergent animals including flatworms and vertebrates, we find that hundreds of accessibility-dictating sequence motifs partition into distinct yet conserved sets, or 'vocabularies', each associated with a specific cell type family. However, combinatorial relationships among these motifs preferred by individual cell types are largely species specific. Deep-learning models trained on one species accurately predict family-level chromatin accessibility in distantly related species, albeit frequently rely on different motifs from shared vocabularies to reach convergent predictions. By contrast, models trained on individual cell types within a family lose cross-species predictive power, indicating that the regulatory syntax governing cell type-level identity evolves rapidly. We propose a 'collective maintenance' model in which motif vocabularies defining cell type families are evolutionarily stable, while recombination of these motifs generates cell type-specific regulatory programmes. This suggests that family identity is maintained collectively by large, conserved pools of regulatory factors, analogous to the logic of developmental homology, where character identity persists through network-level conservation despite extensive rewiring.

Journal Article↗

PubMed2025

Improving RNA Secondary Structure Prediction Through Expanded Training Data.

In recent years, deep learning has revolutionized protein structure prediction, achieving remarkable speed and accuracy. RNA structure prediction, however, has lagged behind. Although several methods have shown some success in predicting RNA secondary and tertiary structures, none have reached the accuracy observed with contemporary protein models. The lack of success of these RNA structure prediction models has been proposed to be due to limited high-quality structural information that can be used as training data. To probe this proposed limitation, we developed a large and diverse dataset comprising paired RNA sequences and their corresponding secondary structures. We assess the utility of this enhanced dataset by retraining on a deep learning model, SincFold. We find that SincFold exhibited improved generalization to some previously unseen RNA families, enhancing its capability to predict accurate de novo RNA secondary structures. The RNASSTR dataset provides a substantial advance for RNA structure modeling, laying a strong foundation for the development of future RNA secondary structure prediction algorithms.

Journal Article↗

PubMed2026

Artificial intelligence (AI) uses in stereotactic radiosurgery (SRS): diagnosis with brain metastasis (BM) - A systematic review.

BACKGROUND: Brain metastases (BM) are the most common intracranial tumors in adults, and stereotactic radiosurgery (SRS) has become a mainstay of management. However, several diagnostic challenges persist in the SRS pathway, particularly the differentiation of radiation necrosis (RN) from true tumor progression, which conventional MRI and even advanced imaging techniques often cannot reliably resolve. Recent advances in artificial intelligence (AI) offer the potential to address these diagnostic limitations. This systematic review synthesizes current literature on AI applications for MRI-based diagnostic decision support in BM patients undergoing SRS, with a focus on radiomics and deep learning tools for distinguishing RN from progression, classifying molecular and histologic subtypes, and predicting treatment response. METHODS: A systematic review was performed in accordance with PRISMA guidelines. PubMed, Web of Science, and Scopus were searched using a targeted query combining terms related to AI, brain metastasis, diagnosis or imaging, and SRS. After screening 483 records and applying strict inclusion and exclusion criteria, 18 studies published between 2015 and 2025 were included. Data were extracted on study design, cohort characteristics, imaging modality, AI methodology, validation strategy, and reported diagnostic performance. RESULTS: Among the 18 included studies, AI models demonstrated strong performance across diagnostic tasks in the BM-SRS pathway. The differentiation of RN from true tumor progression was the most extensively studied application, addressed by 14 of 18 studies, with reported AUCs ranging from 0.71 to 0.94. Support vector machines, random-forest ensembles, convolutional neural networks, and transformer-based multimodal architectures were widely used. The literature evolved from single-sequence radiomic classifiers in 2018 to multimodal deep learning frameworks fusing imaging with clinical and genomic data in 2025. Contrast-enhanced T1-weighted MRI was the dominant imaging input, and texture-based radiomic features (GLCM, GLSZM, GLDM, and wavelet-derived features) were the most consistently predictive. The highest-performing models reached AUCs of 0.85-0.91 through multimodal integration of imaging with clinical and genomic features, and consistently outperformed expert neuroradiologist read on matched cases. Remaining studies addressed longitudinal segmentation-based detection of local failure and adverse radiation effects, BRAF mutation status in melanoma BM, early Gamma Knife treatment response, and primary tumor histology classification, with more variable performance. CONCLUSION: AI models, particularly those integrating MRI-derived radiomic features with clinical and genomic data, show high accuracy in supporting diagnostic decisions for BM patients treated with SRS. The post-SRS differentiation of radiation necrosis from true tumor progression has reached the greatest level of maturity and is closest to clinical translation, with potential to reduce unnecessary biopsies, personalize surveillance intervals, and rationalize treatment-pathway decisions. Other diagnostic applications, including molecular subtyping and primary tumor histology classification, remain exploratory and require further multicenter validation. Integration of AI tools into multidisciplinary tumor-board workflows, combined with prospective validation and standardized reporting, will be essential to realize the full clinical benefits of AI in SRS for brain metastases.

Humans↗