SEARCH · PubMed Health

Results for “machine learning feature selection”

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 55 records · Page 3Linked to original sources

PubMed2026

Radiogenomics predicts immune microenvironment heterogeneity and response to combination immunotherapy in hepatocellular carcinoma.

BACKGROUND: The combination of immune checkpoint inhibitors (ICIs) with anti-angiogenic agents is the preferred first-line therapy option for patients with advanced hepatocellular carcinoma (HCC), yet only a subset of patients responds, urging the quest for prediction biomarkers. We aimed to integrate genomics with radiology to propose an immune-derived radiogenomics biomarker of response to such combination immunotherapy and evaluate its added value in clinical context. METHODS: We integrated bulk RNA sequencing (RNA-seq) and proteomics data of 994 HCC patients with single-cell RNA-seq data of 11 samples across multiple datasets to identify an immune-related signature (IRS) that may influence sensitivity or resistance to such combined immunotherapy strategy, followed by verification of selected marker genes using immunohistochemistry and cytological experiments. We then trained/validated a cross-modality radiogenomics biomarker using machine learning based on TCIA database that was further tested in multi-scale independent cohorts covering 754 HCC patients. RESULTS: Integrative multi-omics analysis identifed a parsimonious 2-gene prognostic signature including KPNA2 and SMG5 that was significantly associated with immune heterogeneity and response to combination immunotherapy. Machine-learning pipeline exported the optimal 4-feature radiogenomics biomarker using support vector machine that significantly discriminated prognosis (hazard ratio 1.415–1.890; p < 0.05 for all) and modestly predicted response to ICI plus anti-angiogenic therapy (area under the curve 0.720–0.829) in independent retrospective series across major imaging modalities (computed tomography/magnetic resonance imaging). In a prospective neoadjuvant cohort, this biomarker also showed favorable performance for predicting pathological response and tumor recurrence, accompanied by biological validation through single-cell RNA-seq analysis of pre-treatment biopsies. CONCLUSIONS: Our study provides a cross-device-cross-modal radiogenomics biomarker that can improve patient selection for emerging ICI plus anti-angiogenic therapy with novel potential therapeutic targets in HCC.

Humans↗

PubMed2025

MyESL: A Software for Evolutionary Sparse Learning in Molecular Phylogenetics and Genomics.

Evolutionary sparse learning uses supervised machine learning to build evolutionary models where genomic sites loci are parameters. It uses the Least Absolute Shrinkage and Selection Operator with bi-level sparsity to connect a specific phylogenetic hypothesis with sequence variation across genomic loci. The MyESL software addresses the need for open-source tools to perform evolutionary sparse learning analyses, offering features to preprocess input phylogenomic alignments, post-process output models to generate molecular evolutionary metrics, and make Least Absolute Shrinkage and Selection Operator regression adaptable and efficient for phylogenetic trees and alignments. The core of MyESL, which constructs models with logistic regressions using bi-level sparsity, is written in C++. Its input data preprocessing and result post-processing tools are developed in Python. Compared to other tools, MyESL is more computationally efficient and provides evolution-friendly inputs and outputs. These features have already enabled the use of MyESL in two phylogenomic applications, one to identify outlier sequences and fragile clades in inferred phylogenies and another to build genetic models of convergent traits. In addition to the use in a Python environment, MyESL is available as a standalone executable compatible across multiple platforms, which can be directly integrated into scripts and third-party software. The source code, executable, and documentation for MyESL are openly accessible at https://github.com/kumarlabgit/MyESL.

Phylogeny↗

PubMed2025

Assessing individual genetic susceptibility to metabolic syndrome: interpretable machine learning method.

BACKGROUND: Genome-wide association studies have provided profound insights into the genetic aetiology of metabolic syndrome (MetS). However, there is a lack of machine-learning (ML)-based predictive models to assess individual genetic susceptibility to MetS. This study utilized single-nucleotide polymorphisms (SNPs) as variables and employed ML-based genetic risk score (GRS) models to predict the occurrence of MetS, bringing it closer to clinical application. METHODS: Feature selection was performed using Least Absolute Shrinkage and Selection Operator. Six ML algorithms were employed to construct GRS models. A fivefold cross-validation was utilized to aid in the internal validation of models. The receiver operating characteristic (ROC) curve was used to select the better-performing GRS model. The SHapley Additive exPlanations (SHAP) was then applied to interpret the model. After extracting GRS, stratified analysis of BMI, age and gender was performed. Finally, these conventional risk factors and GRS were integrated through multivariate logistic regression to establish a combined model. RESULTS: A total of 17 SNPs were selected for analysis. Among the GRS models, the extreme gradient boosting (XGBoost) model demonstrated superior discriminative performance (AUC = 0.837). The XGBoost's optimal robustness was also validated through five-fold cross-validation (mean ROC-AUC = 0.706). The XGBoost-based SHAP algorithm not only elucidated the global effects of 17 SNPs across all samples, but also described the interaction between SNPs, providing a visual representation of how SNPs impact the prediction of MetS in an individual. There was a strong correlation between GRS and MetS risk, particularly observed among young individuals, males and overweight individuals. Furthermore, the model combining conventional risk factors and GRS exhibited excellent discriminative performance (AUC = 0.962) and outstanding robustness (mean ROC-AUC = 0.959). CONCLUSION: This study established a reliable XGBoost-based GRS model and a GRS prediction platform (https://metabolicsyndromeapps.shinyapps.io/geneticriskscore/) to assess individual genetic susceptibility to MetS. This model has high interpretability and can provide personalized reference for determining the necessity of primary prevention measures for MetS. Additionally, there may be interactions between traditional risk factors and GRS, and the integration of both in a comprehensive model is useful in the prediction of MetS occurrence.

Humans↗

PubMed2026

Benchmark of biomarker identification and prognostic modeling methods on diverse censored data.

The practices of identifying biomarkers and developing prognostic models using genomic data has become increasingly prevalent. Such data often features characteristics that make these practices difficult, namely high dimensionality, correlations between predictors, and sparsity. Many modern methods have been developed to address these problematic characteristics while performing feature selection and prognostic modeling, but a large-scale comparison of their performances in these tasks on diverse right-censored time to event data (aka survival time data) is much needed. We have compiled many existing methods, including some machine learning methods, several which have performed well in previous benchmarks, primarily for comparison in regards to variable selection capability, and secondarily for survival time prediction on many synthetic datasets with varying levels of sparsity, correlation between predictors, and signal strength of informative predictors. For illustration, we have also performed multiple analyses on a publicly available and widely used cancer cohort from The Cancer Genome Atlas using these methods. We evaluated the methods through extensive simulation studies in terms of the false discovery rate, F1-score, concordance index, Brier score, root mean square error, and computation time. Of the methods compared, CoxBoost and the Adaptive LASSO performed well in all metrics, and the LASSO and elastic net excelled when evaluating concordance index and F1-score. The Benjamini-Hoschberg and q-value procedures showed volatile performances in controlling the false discovery rate. Some methods' performances were greatly affected by differences in the data characteristics. With our extensive numerical study, we have identified the best performing methods for a plethora of data characteristics using informative metrics. This will help cancer researchers in choosing the best approach for their needs when working with genomic data.

Humans↗

PubMed2026

Spectral Transforms as a Tool to Optimize Digital Phenotyping in Biological Images.

Modern livestock breeding has mastered genotyping. Genome-wide association studies, genomic selection, and SNP arrays enable genetic merit prediction at lower cost. However, phenotyping remains the bottleneck, as manual measurement is slow, expensive, subjective, and unable to capture spatial or temporal trait organization. Digital phenotyping via artificial intelligence could resolve this, but deep learning requires thousands of labelled examples, impractical when phenotyping cost itself limits datasets to hundreds of individuals. This creates a paradox: AI could accelerate phenotyping but requires large numbers of samples to train the models. Here, we demonstrate that integrating computer vision with machine learning offers sample-efficient digital phenotyping using eggshell colour as a model system. Rather than learning features from scratch (deep learning), we engineer physically motivated features via Wavelet transforms that decompose images into multi-scale spatial components. Wavelet features captured 14.2 percentage points more variance (R2 = 0.976 vs. 0.834, p < 0.001) than standard colorimetry, with 50% better sample efficiency (achieving at n = 60 what colorimetry required n = 120). Variance decomposition revealed 77% of discriminative capacity derives from spatial patterns (bands, spots, gradients) invisible to scalar averages. Additionally, we identified "cryptic phenotypes" (3.3%) where spatial patterns contradicted average colour, cases where colorimeters failed but Wavelets succeeded. The underlying principle-that spatial decomposition can recover organizational information lost by scalar averaging-may be applicable to other traits with spatial or temporal structure, such as marbling, dermatitis, or pigmentation rhythms, although whether comparable performance gains would be observed remains to be tested empirically. Hence, for breeding programs implementing genomic selection, computer vision-based digital phenotyping captures complex trait variation without massive training datasets, addressing the bottleneck that increasingly limits genetic progress as genotyping becomes trivial.

Wavelet transform↗

PubMed2026

Machine Learning and Metabolomics to Characterize Warburg-Like Metabolic Subtypes in Human Retinal Endothelial Cells Exposed to Risk Factors Associated With Proliferative Diabetic Retinopathy.

PURPOSE: High glucose (HG), hypoxia (Hyp), and their combination are major risk factors for proliferative diabetic retinopathy (PDR). Although these conditions induce features of the Warburg-like metabolic reprogramming in human retinal endothelial cells (HRECs), it remains unclear whether they produce distinct metabolic and angiogenic subtypes. This study aimed to characterize the Warburg-like-associated metabolic heterogeneity induced by these PDR-related risk factors and evaluate the ability of supervised machine-learning models to distinguish these subtypes. METHODS: HRECs were cultured under normoglycemic, HG, Hyp (2% O2), and combined HG-Hyp conditions. Untargeted LC-MS/MS metabolomics quantified metabolites spanning carbohydrates, amino acids, nucleotides, and lipids. Principal component analysis (PCA) assessed overall metabolic variation, and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis identified metabolic pathways associated with angiogenesis. In vitro angiogenesis assays measured endothelial tube formation and branching. Nine supervised classifiers (decision tree, logistic regression, naïve Bayes, random forest, K-Nearest Neighbors, neural network, gradient boosting, AdaBoost, and Support Vector Machine) were trained on the highest-ranked metabolites selected by the Information Gain Ratio feature-ranking approach. Model performance was evaluated using 10-fold cross-validation, leave-one-out cross-validation (LOOCV), permutation testing, and a classifier stability analysis under biologically meaningful distributional shift using an independent chemically induced hypoxia model (CoCl2). RESULTS: PCA revealed partial separation of metabolic profiles across conditions, indicating different Warburg-like metabolic subtypes. The combined HG-Hyp condition exhibited enhanced angiogenic potential relative to either HG or Hyp alone. KEGG pathway enrichment analysis identified fatty acid biosynthesis and elongation among the most significantly enriched pathways in HRECs under combined HG-Hyp conditions, alongside amino sugar and nucleotide sugar metabolism, glycerophospholipid metabolism, the pentose phosphate pathway, and glycolysis/gluconeogenesis. Supervised machine-learning classifiers distinguished these metabolic subtypes, with AdaBoost and gradient Boosting showing the most balanced, reproducible performance across 10-fold cross-validation, LOOCV, and permutation testing, and remaining the most reliable classifiers under domain-shift testing (area under the curve = 0.88, P = 0.0061). CONCLUSIONS: In this exploratory analysis, HG, Hyp, and their combination drive metabolically and functionally distinct subtypes of Warburg-like metabolic reprogramming in HRECs, with HG-Hyp in combination producing a highly angiogenic phenotype. Boosting-based ensemble classifiers provide a promising framework for detecting these subtypes even under domain-shift conditions, warranting validation in larger independent datasets. TRANSLATIONAL RELEVANCE: Integrating metabolomics with machine-learning classification offers a strategy to identify Warburg-like metabolic subtypes in retinal endothelial cells, providing insights into angiogenic mechanisms and guiding the development of targeted diagnostics or therapeutics for PDR.

Humans↗

PubMed2026

Transfer Learning across Material Properties Using Center-Environment Features: From Energetics to Mechanical Properties in Multicomponent Mo Alloys.

Transfer learning (TL) provides a viable approach to mitigate data scarcity in materials informatics. While conventional TL focuses on predicting identical properties across different systems, this work demonstrates a cross-property extension of TL from energy to mechanical properties via end-to-end model weight pre-training and fine-tuning: knowledge learned from predicting substitution energies is transferred to predict distinctly different mechanical properties, substantially improving computational efficiency given the typically higher cost of acquiring target-domain data. To accelerate computational alloy design, machine learning models using center-environment (CE) features were first developed to predict substitution energies of alloying elements in molybdenum (Mo)-based alloys. The Random Forest models achieved the optimal performance and transferability-R2 = 0.97, 〈MAE〉 = 0.11 eV, and 〈RMSE〉 = 0.16 eV-against the density functional theory (DFT) benchmark. The model dependency of feature selection and importance analysis was discussed. The transferability of the energy models was validated on unknown systems with new elements. Subsequently, the energy models were fine-tuned using limited mechanical property data to construct energy-to-property (E2P) TL models capable of predicting elastic properties, including bulk modulus, Young's modulus, shear modulus, and elastic constants, achieving an improved accuracy over the non-transferred ML by ∼10-30%, with its transferability verified by additional DFT calculations. This cross-property E2P transfer learning framework opens a new avenue for accelerating computational materials discovery and may be extended to other multiproperty predictions governed by similar physical principles.

center-environment feature↗

PubMed2025

A comparative study highlights superiority of LSTM in crop genomic prediction.

We systematically evaluated three key determinants affecting prediction accuracy and the algorithm performance differences based on fifteen state-of-the-art GP methods, and found LSTM suitable for capturing additive and epistatic effects. Genomic prediction (GP) has been developed as an important method supporting crop breeding. By utilizing the phenotype values result from GP, breeders could make decisions in the seedling stage that consequently benefit for cost saving. In recent years, machine learning emerged as an efficient technology to solve modeling problems in many fields, including crop breeding. However, numerous modeling approaches have hindered the application of GP since breeders struggle to choose. Therefore, a comprehensively methodological research with guiding significance is extremely necessary. In the present study, we systematically evaluated three key determinants affecting prediction accuracy and the algorithm performance differences based on fifteen state-of-the-art GP methods. As for genomic feature processing, we found feature selection (SNP filtering approach) performed better than feature extraction (PCA method). Specifically, the feature relationship dependent methods (GBLUP, RNN, and LSTM) as well as DNN architecture showed superior performance with feature selection. Marker density analysis showed positive correlation with prediction accuracy in a limited threshold. Comparison on effect of population size demonstrated a positive correlation between trait genetic complexity and the optimal population size required. By testing fifteen modeling methods, we found LSTM network displayed superior performance, achieving the highest average STScore (0.967) across six datasets. Further research using all cell states or the latest cell states of LSTM inputs demonstrated its architecture particularly adept with capturing additive and epistatic QTL effects among SNPs. In conclusion, our findings provide basic principles for implementing GP in breeding project to maximize prediction accuracy while maintaining cost-effectiveness.

Plant Breeding↗

PubMed2026

Artificial intelligence for anticancer drug discovery from natural products of macroalgae and sponges: A systematic review.

Marine natural products (MNPs) from macroalgae and marine sponges have inspired clinically important anticancer agents, including the cytarabine pharmacophore and the eribulin scaffold, while cyanobacterial dolastatin chemistry supplies the auristatin payloads of several marine-inspired antibody-drug conjugates (ADCs) such as brentuximab vedotin. Artificial intelligence (AI) methods, encompassing both classical machine learning (ML) with hand-engineered features and modern deep learning (DL) with many-layered neural networks, are increasingly supporting key decisions in natural-product anticancer drug discovery, including bioactivity prediction, target identification, absorption, distribution, metabolism, excretion and toxicity (ADMET) filtering, generative analogue design, and the selection of preclinical candidates. DL architectures relevant to this field include graph neural networks, transformer-based molecular generators, diffusion models for protein-ligand docking, and convolutional networks for mass spectrometry, while classical ML contributes interpretable fingerprint-based bioactivity models and molecular networking for dereplication. This review follows a systematic literature review methodology to organize the landscape of AI methods now applied to MNP anticancer discovery, distinguishing ML and DL approaches where relevant, situating them within the chemical context of macroalgal and sponge-derived oncology leads, and critically examining published case studies, including validation level (computational, in vitro, in vivo, clinical). The principal bottleneck for medical translation has shifted partly from algorithmic capability toward data infrastructure and experimental validation. Sparse, heterogeneous, and taxonomically biased bioactivity records limit what current models can learn and reduce the reliability of AI-prioritized candidates entering the preclinical pipeline. A roadmap is proposed that prioritizes open MNP-specific benchmarks, symbiont-aware modeling, and active learning loops with synthesizability and ADMET constraints. These AI workflows may accelerate the prioritization of marine-derived anticancer leads and support earlier, more evidence-based translational decisions in oncology drug development.

Biological Products↗

PubMed2025

Efficient Detection and Characterization of Targets of Natural Selection Using Transfer Learning.

Natural selection leaves detectable patterns of altered spatial diversity within genomes, and identifying affected regions is crucial for understanding species evolution. Recently, machine learning approaches applied to raw population genomic data have been developed to uncover these adaptive signatures. Convolutional neural networks (CNNs) are particularly effective for this task, as they handle large data arrays while maintaining element correlations. However, shallow CNNs may miss complex patterns due to their limited capacity, while deep CNNs can capture these patterns but require extensive data and computational power. Transfer learning addresses these challenges by utilizing a deep CNN pretrained on a large dataset as a feature extraction tool for downstream classification and evolutionary parameter prediction. This approach reduces extensive training data generation requirements and computational needs while maintaining high performance. In this study, we developed TrIdent, a tool that uses transfer learning to enhance detection of adaptive genomic regions from image representations of multilocus variation. We evaluated TrIdent across various genetic, demographic, and adaptive settings, in addition to unphased data and other confounding factors. TrIdent demonstrated improved detection of adaptive regions compared to recent methods using similar data representations. We further explored model interpretability through class activation maps and adapted TrIdent to infer selection parameters for identified adaptive candidates. Using whole-genome haplotype data from European and African populations, TrIdent effectively recapitulated known sweep candidates and identified novel cancer, and other disease-associated genes as potential sweeps.

Selection, Genetic↗

PubMed2025

Identification of potential biomarkers and mechanisms for keloid disorder based on comprehensive bioinformatics analysis and machine learning algorithms.

BACKGROUND: Keloid disorder (KD) encompasses a spectrum of fibroproliferative dermal conditions, the pathogenesis remains complex and incompletely understood. This study sought to identify biomarkers and potential therapeutic targets for KD through an integrative bioinformatics approach and machine learning analysis of RNA sequencing data. METHODS: RNA sequencing was performed on skin tissue samples from 13 patients with KD and 14 healthy controls. Using weighted gene co-expression network analysis and differential expression analysis revealed differentially expressed key module genes, and the CytoHubba plugin identified candidate genes. Subsequently analyzed using least absolute shrinkage and selection operator (LASSO) and support vector machine recursive feature elimination (SVM-RFE) methods to pinpoint feature genes associated with KD. Following this, biomarkers were determined through expression level validation, enrichment analysis, and immune infiltration analysis. RESULTS: A total of 420 differentially expressed key module genes were identified, and the top 10 genes with DMNC values were selected as candidate genes. Five feature genes were selected through LASSO and SVM-RFE, with NID2, MFAP2, COL8A1, and P4HA3 showing significant expression differences between KD and control samples, along with consistent expression patterns across datasets, identified as potential biomarkers. These four biomarkers were proved to possess high diagnostic potential, and they were found to exhibit significant positive correlations with one another. Functional enrichment analysis indicated that the primary KEGG pathways associated with these biomarkers included "steroid hormone biosynthesis" and "cytokine-cytokine receptor interaction." Moreover, immune infiltration analysis revealed that the four biomarkers were negatively correlated with type 17 T helper cells and positively correlated with 15 immune cell types, including activated B cells and central memory CD4 T cells. CONCLUSION: In conclusion, NID2, MFAP2, COL8A1, and P4HA3 were identified as key biomarkers for KD, offering new avenues for more targeted and effective diagnostic and therapeutic strategies for managing this condition.

Humans↗

PubMed2026

From Variability to Consensus: Rescoring Harmonizes Peptide Identification across Diverse Search Engines and Data Sets.

Peptide-spectrum match (PSM) rescoring has become standard in proteomics workflows, improving peptide identification accuracy across diverse search engines. Despite the availability of multiple rescoring strategies, systematic comparisons spanning several search engines, data sets, and database configurations remain limited. Here, we benchmarked seven publicly available search engines, evaluating standard target-decoy-based false discovery rate (FDR) estimation alongside Percolator, MS2Rescore, and Oktoberfest across four data sets acquired on different mass spectrometry platforms in data-dependent mode and searched against protein databases of varying size and composition. Rescoring substantially increased identification consensus and reduced variability between search engines, with prediction-based approaches yielding the largest gains. While database size had limited impact for human data sets, it significantly affected identification rates on a metaproteomic data set. Entrapment-based evaluation indicated generally adequate FDR control across methods, although prediction-based rescoring exhibited a higher tendency toward FDR underestimation in specific configurations. Overall, advanced rescoring strategies harmonize peptide identification outcomes across search engines, thereby enhancing robustness and comparability in proteomics analyses. However, careful feature selection and appropriate database choice remain essential to ensure reliable FDR control and optimal performance across diverse experimental settings.

Search Engine↗

PubMed2026

Integrative dual-track transcriptomics reveals stage-specific coordination, regulatory divergence, and HSP90AA1-associated remodeling in human folliculogenesis.

Human folliculogenesis depends on coordinated yet non-identical developmental remodeling in the oocyte and its surrounding granulosa cells. When these two compartments remain synchronized and when they diverge into lineage-specific regulatory states, however, remains incompletely resolved. Here we performed an integrative dual-track re-analysis of the human RNA-seq dataset GSE107746, modeling oocytes and granulosa cells as distinct but developmentally linked compartments across follicular progression. Analysis of 148 sequencing libraries showed that compartment identity was the dominant source of transcriptomic variation, supporting compartment-aware downstream interpretation. Within this framework, oocytes followed a relatively continuous developmental trajectory, with substantial transcriptional remodeling already evident across adjacent stages, whereas granulosa cells showed weaker early-stage contrasts but markedly stronger late-stage reorganization, particularly around the antral and preovulatory transitions. Functional enrichment indicated that oocyte maturation was associated with RNA-processing and broader genome-regulatory remodeling, whereas granulosa maturation was dominated by progressive mitochondrial and bioenergetic activation. Co-expression analysis showed that both compartments contained strong late-stage programmes together with inverse early-state modules, indicating a shared systems-level architecture of maturation, although the hub-gene composition and biological content of these programmes were largely compartment-specific. Machine-learning validation reinforced this asymmetry: oocyte stage classification was best recovered from a compact eigengene-based representation, whereas granulosa stage discrimination was better resolved by a broader differential-expression-derived feature set. At the gene level, HSP90AA1 emerged as a stage-associated marker with compartment-specific behavior, showing progressive attenuation across oocyte development, assignment to the selected oocyte blue module, and sharper transitional dynamics in granulosa cells. Together, these findings support a model in which human folliculogenesis proceeds through coordinated but non-equivalent transcriptomic remodeling, with shared developmental logic at the systems level but distinct molecular execution in germline and somatic compartments.

Co-expression networks↗

PubMed2026

Cancer of unknown primary: the evolution of tissue of origin identification in the artificial intelligence era.

Cancer of Unknown Primary (CUP) presents substantial diagnostic and therapeutic challenges owing to its heterogeneous nature and the absence of an identifiable primary tumor site. This review provides a structured search of the pathogenesis, epidemiological characteristics, and limitations of traditional diagnostic and therapeutic approaches for CUP, with an emphasis on the evolution of Tissue of Origin (TOO) identification techniques. Recent advances in precision medicine have accelerated the development of machine learning-based TOO identification tools, representing a paradigm shift in CUP diagnostics. Deep learning (DL) algorithms that integrate multi-omics data (such as genomics and transcriptomics) with clinical features have markedly enhanced the accuracy of tracing tumor origin, and artificial intelligence (AI) driven TOO models are increasingly being incorporated into clinical practice, offering new insights for pathological diagnosis, treatment selection, and prognostic evaluation. Nevertheless, several challenges remain, including issues of data standardization, model generalizability, and interpretability. Ethical considerations related to data privacy, algorithmic fairness, and clinical implementation also warrant careful attention. Future research should focus on establishing standardized multi-center databases, developing more interpretable AI models, and fostering multidisciplinary collaborative strategies for CUP management. Through continued refinement of technical solutions and regulatory guidelines, TOO identification is anticipated to progress from research to routine clinical application, ultimately supporting precise and personalized care for patients with CUP.

Artificial intelligence↗

PubMed2025

Efficient detection and characterization of targets of natural selection using transfer learning.

Natural selection leaves detectable patterns of altered spatial diversity within genomes, and identifying affected regions is crucial for understanding species evolution. Recently, machine learning approaches applied to raw population genomic data have been developed to uncover these adaptive signatures. Convolutional neural networks (CNNs) are particularly effective for this task, as they handle large data arrays while maintaining element correlations. However, shallow CNNs may miss complex patterns due to their limited capacity, while deep CNNs can capture these patterns but require extensive data and computational power. Transfer learning addresses these challenges by utilizing a deep CNN pre-trained on a large dataset as a feature extraction tool for downstream classification and evolutionary parameter prediction. This approach reduces extensive training data generation requirements and computational needs while maintaining high performance. In this study, we developed TrIdent, a tool that uses transfer learning to enhance detection of adaptive genomic regions from image representations of multilocus variation. We evaluated TrIdent across various genetic, demographic, and adaptive settings, in addition to unphased data and other confounding factors. TrIdent demonstrated improved detection of adaptive regions compared to recent methods using similar data representations. We further explored model interpretability through class activation maps and adapted TrIdent to infer selection parameters for identified adaptive candidates. Using whole-genome haplotype data from European and African populations, TrIdent effectively recapitulated known sweep candidates and identified novel cancer, and other disease-associated genes as potential sweeps.

Journal Article↗

PubMed2024

PredIL13: Stacking a variety of machine and deep learning methods with ESM-2 language model for identifying IL13-inducing peptides.

Interleukin (IL)-13 has emerged as one of the recently identified cytokine. Since IL-13 causes the severity of COVID-19 and alters crucial biological processes, it is urgent to explore novel molecules or peptides capable of including IL-13. Computational prediction has received attention as a complementary method to in-vivo and in-vitro experimental identification of IL-13 inducing peptides, because experimental identification is time-consuming, laborious, and expensive. A few computational tools have been presented, including the IL13Pred and iIL13Pred. To increase prediction capability, we have developed PredIL13, a cutting-edge ensemble learning method with the latest ESM-2 protein language model. This method stacked the probability scores outputted by 168 single-feature machine/deep learning models, and then trained a logistic regression-based meta-classifier with the stacked probability score vectors. The key technology was to implement ESM-2 and to select the optimal single-feature models according to their absolute weight coefficient for logistic regression (AWCLR), an indicator of the importance of each single-feature model. Especially, the sequential deletion of single-feature models based on the iterative AWCLR ranking (SDIWC) method constructed the meta-classifier consisting of the top 16 single-feature models, named PredIL13, while considering the model's accuracy. The PredIL13 greatly outperformed the-state-of-the-art predictors, thus is an invaluable tool for accelerating the detection of IL13-inducing peptide within the human genome.

Humans↗

PubMed2025

Alzheimer's subtypes A supervised, unsupervised, multimodal, multilayered embedded recursive (SUMMER) AI study.

Since Alzheimer's disease (AD) is a heterogeneous disease, different subtypes may have distinct biological, genetic, and clinical characteristics, requiring tailored interventions. While several proposed subtypes of AD exist, there is still no clear consensus on a definitive classification. By leveraging complementary AI approaches, including supervised and unsupervised learning, within a recursive pipeline (SUMMER) that integrates multimodal datasets encompassing MRI measurements, phenotypes, and genetic data, our goal was to generate robust scientific evidence for identifying AD subtypes. Data was downloaded from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database and included neuroimaging data (MRI), genetics (SNPs), clinical diagnosis, and demographics. 1133 European American participants' images, aged 55-95, were included in this study. The analysis was multi-fold, where the first step involved applying an unsupervised application to a subset of the MRI sample (AD + cognitively normal (CN) aged matched groups, 100 men aged 68-85 years, and 76 women aged 68-85 years). The MRI brain gray matter was segmented into 44 regions of interest (ROIs) according to a standard atlas, and 618 features were extracted, including ROI voxel intensity measurements such as minimum, maximum, and histogram variables. Results identified a cluster of subtype AD men and a cluster of subtype AD women that were distinct from the rest of their respective samples. In the next step, the integrity of the identified subtype AD clusters was investigated using the XGBoost supervised machine learning application with genetic features (SNPs, N=36,724) and labels: the identified subtype AD cluster vs. the rest of the sample, stratified by sex. A significant AD subtype men model (accuracy=0.85, F1=0.72, AUC=0.83) and a significant women AD subtype model (accuracy=0.81, F1=0.81, AUC=0.81) were built, confirming the homogeneity of the isolated AD subtype clusters. Discriminative biomarkers were extracted from the significant models, including selected ROIs and SNPs. Finally, the subtype models were tested on an unseen subset of ADNI data. The genetic-based models identified clusters of AD subtype participants consisting of 34% of the men AD group and 47% of the women AD group. Phenotypic analysis indicates that lower body weight was associated with the women's AD subtype. Complex diseases like AD demand a sophisticated, multimodal approach for precise diagnosis. Effectively identifying disease subtypes enhances the potential for personalized treatment, ultimately improving patient outcomes.

Journal Article↗

PubMed2026

NMR metabolomics and glycomics for cancer detection in patients with non-specific symptoms: a prospective observational cohort study.

BACKGROUND: Early cancer diagnosis in patients with non-specific symptoms is limited by the lack of discriminatory tests. Within the Oxfordshire Suspected CANcer (SCAN) pathway, exploratory biomarker work showed that serum 1H NMR-based metabolomics can identify cancer with high accuracy. SCAN2 evaluated whether integrating metabolomics with glycomics provides complementary molecular information and improves discrimination in a clinically complex, real-world population. METHODS: Serum from 369 SCAN patients (59 cancers) was analysed using AXINON® System-derived NMR metabolomics and HPLC-MS glycomics. Machine-learning models were trained to predict cancer status, with performance assessed by receiver operating characteristic (ROC) analysis of pooled cross-validated predictions. To place cancer risk in a broader clinical context, a second classifier modelling alternative non-cancer diagnosis was incorporated, and mean predicted probabilities from both models were jointly projected into a two-dimensional space, maintaining strict separation of training and test data. FINDINGS: In the full cohort, integration of glycomics with metabolomics achieved an AUC of 0.814 (95% CI 0.808-0.820). In a refined sub-cohort excluding major comorbidities and selected cancer types (32 cancers, 277 non-cancers), performance improved to an AUC of 0.884 (95% CI 0.879-0.890). Discriminatory features included cancer-associated biantennary fucosylated glycans alongside amino acid metabolites (glutamate, histidine) and lipoprotein-related measures. A classifier distinguishing metastatic from non-metastatic disease (n = 29 vs. 30) achieved an AUC of 0.80. Joint probability analysis in the full cohort preserved cancer-associated signatures across comorbidity burden, with projection-based classification achieving an accuracy of 89.2% (95% CI 85.7-92.6). INTERPRETATION: These findings validate the SCAN1 metabolomic signature in a more clinically complex cohort and indicate that integrating glycomics with metabolomics provides complementary biological information for cancer discrimination. Joint probability analysis provides an interpretable framework for cancer risk stratification within multimorbid diagnostic pathways, supporting the clinical potential of scalable multi-omics blood testing. FUNDING: EPSRC, EU Horizon 2020, Wellcome/MLSTF, Novo Nordisk Foundation.

Humans↗