PubMed Health⌕ Search

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 127 records · Page 7Linked to original sources
PubMed2025

Reconstructing the 3D genome organization of Neanderthals reveals that chromatin folding shaped phenotypic and sequence divergence.

Changes in gene regulation were a major driver of the divergence of archaic hominins (AHs)-Neanderthals and Denisovans-and modern humans (MHs). The three-dimensional (3D) folding of the genome is critical for regulating gene expression; however, its role in recent human evolution has not been explored because the degradation of ancient samples does not permit experimental determination of AH 3D genome folding. To fill this gap, we apply novel deep learning methods for inferring 3D genome organization from DNA sequence to Neanderthal, Denisovan, and diverse MH genomes. Using the resulting 3D contact maps across the genome, we identify 167 distinct regions with diverged 3D genome organization between AHs and MHs. We show that these 3D-diverged loci are enriched for genes related to the function and morphology of the eye, supra-orbital ridges, hair, lungs, immune response, and cognition. Despite these specific diverged loci, the 3D genome of AHs and MHs is more similar than expected based on sequence divergence, suggesting that the pressure to maintain 3D genome organization constrained hominin sequence evolution. We also find that 3D genome organization constrained the landscape of AH ancestry in MHs today: regions more tolerant of 3D variation are enriched for introgression in modern Eurasians. Finally, we identify loci where modern Eurasians have inherited novel 3D genome folding patterns from AH ancestors and validate folding differences in a high-frequency locus using Hi-C, revealing a putative molecular mechanism for phenotypes associated with archaic introgression. In summary, our application of deep learning to predict archaic 3D genome organization illustrates the potential of inferring molecular phenotypes from ancient DNA to reveal previously unobservable biological differences.

Journal Article↗
PubMed2027

Inferring Gene Regulatory Networks in Stem Cells: Methods and Applications.

Gene regulatory networks (GRNs) represent the complex interplay of transcription factors, regulatory elements, and target genes that orchestrate cellular identity and function, playing a crucial role in the differentiation and maintenance of stem cells. This chapter provides an overview of experimental and computational methodologies for inferring GRNs, with particular emphasis on single-cell approaches. We first review key experimental techniques for detecting transcription factor binding sites, chromatin accessibility, and DNA motifs, alongside essential databases that support GRN reconstruction. We then introduce computational inference methods that can be categorized into four principal frameworks: correlation-based approaches, regression and machine learning models, probabilistic and deep learning methods, and integrative or message-passing frameworks. To illustrate practical application, we present a case study applying the pySCENIC workflow to a peripheral blood mononuclear cell single-cell RNA sequencing dataset from mouse, demonstrating how regulon-based analysis can reveal cell-type-specific regulatory programs. This chapter aims to serve as a practical guide for researchers seeking to understand and implement GRN inference methodologies in stem cell biology and related fields.

Gene Regulatory Networks↗
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

MedImg: An Integrated Database for Public Medical Images.

The advancements in deep learning algorithms for medical image analysis have garnered significant attention in recent years. While several studies have shown promising results, with models achieving or even surpassing human performance, translating these advancements into clinical practice is still accompanied by various challenges. A primary obstacle lies in the availability of large-scale, well-characterized datasets for validating the generalization of approaches. To address this challenge, we curated a diverse collection of medical image datasets from multiple public sources, containing 105 datasets and a total of 1,995,671 images. These images span 14 modalities, including X-ray, computed tomography, magnetic resonance imaging, optical coherence tomography, ultrasound, and endoscopy, and originate from 13 organs, such as the lung, brain, eye, and heart. Subsequently, we constructed an online database, MedImg, which incorporates and systematically organizes these medical images to facilitate data accessibility. MedImg serves as an intuitive and open-access platform for facilitating research in deep learning-based medical image analysis, accessible at https://www.cuilab.cn/medimg/.

Humans↗
PubMed2025

A multi-modal transformer for cell type-agnostic regulatory predictions.

Sequence-based deep learning models have emerged as powerful tools for deciphering the cis-regulatory grammar of the human genome but cannot generalize to unobserved cellular contexts. Here, we present EpiBERT, a multi-modal transformer that learns generalizable representations of genomic sequence and cell type-specific chromatin accessibility through a masked accessibility-based pre-training objective. Following pre-training, EpiBERT can be fine-tuned for gene expression prediction, achieving accuracy comparable to the sequence-only Enformer model, while also being able to generalize to unobserved cell states. The learned representations are interpretable and useful for predicting chromatin accessibility quantitative trait loci (caQTLs), regulatory motifs, and enhancer-gene links. Our work represents a step toward improving the generalization of sequence-based deep neural networks in regulatory genomics.

Humans↗
PubMed2026

EvoSNR-Prom: Predicting promoters at single-nucleotide resolution with label-aware transfer learning of the pretrained EVO model.

The precise identification of promoters is crucial for understanding gene regulation. Deep learning methods have achieved considerable success in promoter prediction, yet most operate at the sequence level with coarse-grained labels. This means they label an entire DNA segment as either a "promoter" or "non-promoter," which results in a lack of the nucleotide-level resolution in prediction. In this study, we propose EvoSNR-Prom, a model designed for promoter prediction at single-nucleotide resolution. EvoSNR-Prom is built on the Evo foundation model and formulates promoter identification as a token-level sequence labeling problem, analogous to named entity recognition in natural language processing. To address the limited contextual information available in single-nucleotide tokenization, we introduce a lexicon-enhanced embedding strategy that incorporates biologically meaningful DNA lexicons, enriching contextual representations and improving the model's ability to capture complex sequence motifs. Furthermore, to enhance predictive performance on small size datasets, we integrate a label-aware transfer learning framework to leverage knowledge from well-annotated source species to a target organism. The results across various prokaryotic datasets show that EvoSNR-Prom achieves excellent performance. This work provides a valuable computational framework for the high-precision analysis of gene regulatory elements, contributing to the advancement of promoter prediction at single-nucleotide resolution.

Promoter Regions, Genetic↗
PubMed2026

Artificial intelligence for dental caries detection: An umbrella review.

Artificial intelligence (AI) has been proposed as a tool to improve dental caries detection across imaging modalities; however, its clinical value remains uncertain. This umbrella review aimed to synthesize and critically appraise systematic reviews evaluating AI for caries detection and diagnosis. An umbrella review was conducted following PRIOR guidance (PROSPERO CRD420261340728). Searches were performed in MEDLINE, Embase, Scopus, Web of Science, and Google Scholar up to 15 March 2026. Methodological quality was assessed using AMSTAR 2, and overlap of primary studies was quantified using the corrected covered area (CCA). Seventeen systematic reviews were included, of which five reported diagnostic test accuracy meta-analyses using bivariate or HSROC models. Across these meta-analyses, pooled sensitivity ranged from 0.76 to 0.94 and specificity from 0.85 to 0.91. Most systems were based on deep learning models applied to bitewing radiographs and intraoral photographs. However, substantial heterogeneity was observed in imaging modalities, lesion thresholds, analytical tasks, and evaluation metrics. In addition, a high degree of overlap across reviews and recurrent methodological limitations, including reliance on retrospective datasets, limited external validation, and inconsistent reporting, substantially weaken the reliability of the evidence. Although AI models demonstrate high diagnostic performance under experimental conditions, current evidence does not support their use as stand-alone diagnostic tools. Their clinical applicability remains limited, and implementation should be restricted to decision-support contexts until robust prospective validation demonstrates meaningful impact on clinical decision-making and patient outcomes.

Dental Caries↗
PubMed2025

OmicsTweezer: A distribution-independent cell deconvolution model for multi-omics Data.

Cell deconvolution estimates cell type proportions from bulk omics data, enabling insights into tissue microenvironments and disease. However, practical applications are often hindered by batch effects between bulk data and referenced single-cell data, a challenge that is frequently overlooked. To address this discrepancy, we developed OmicsTweezer, a distribution-independent cell deconvolution model. By integrating optimal transport with deep learning, OmicsTweezer aligns simulated and real data in a shared latent space, effectively mitigating data shifts and inter-omics distribution differences. OmicsTweezer is versatile, capable of deconvolving bulk RNA-seq, bulk proteomics, and spatial transcriptomics. Extensive evaluations on simulated and real-world datasets demonstrate its robustness and accuracy. Furthermore, applications in prostate and colon cancer showcase OmicsTweezer's ability to identify biologically meaningful cell types. As a unified deconvolution framework for multi-omics data, OmicsTweezer offers an efficient and powerful tool for studying disease microenvironments.

Humans↗
PubMed2026

UFold-X: an enhanced Dual & Dynamic U-Mamba model for long-range RNA secondary structure prediction.

RNA secondary structure is essential for understanding the functions of non-coding RNAs, ribosomal RNAs, and viral genomes. However, accurate prediction of long RNA structures remains challenging due to complex long-range interactions and the limited availability of long-RNA training data. We present UFold-X, a dual-branch deep learning framework that combines a convolutional encoder for local structure modeling with a Mamba-based Visual State Space Module for capturing long-range dependencies. A dynamic gating mechanism adaptively integrates the two branches according to sequence length. UFold-X was evaluated on multiple benchmark datasets containing RNAs up to 5000 nucleotides. To rigorously assess generalization, we introduced a cross-clan benchmark for long RNAs. Under this stringent setting, UFold-X achieved performance comparable to state-of-the-art classical approaches while achieving the best performance among deep learning-based methods. Additional cross-family and within-family evaluations further demonstrated robust transferability and competitive predictive performance. UFold-X also maintained excellent computational efficiency, requiring only 0.08 s per sequence on average. To assess biological consistency, we developed a SHAPE-based reactivity prediction variant (UFold-X-R) and an integrated metric, the Hybrid Reactivity-Pairing Score (HRPS). UFold-X-R showed strong agreement with experimental icSHAPE data and achieved the highest HRPS among all evaluated methods. A user-friendly web server is available at https://ufold-x.ai4bread.com.

Nucleic Acid Conformation↗
PubMed2025

CoxKAN: Kolmogorov-Arnold networks for interpretable, high-performance survival analysis.

MOTIVATION: Survival analysis is a branch of statistics that is crucial in medicine for modeling the time to critical events such as death or relapse, in order to improve treatment strategies and patient outcomes. Selecting survival models often involves a trade-off between performance and interpretability; deep learning models offer high performance but lack the transparency of more traditional approaches. This poses a significant issue in medicine, where practitioners are reluctant to use black-box models for critical patient decisions. RESULTS: We introduce CoxKAN, a Cox proportional hazards Kolmogorov-Arnold Network for interpretable, high-performance survival analysis. Kolmogorov-Arnold Networks (KANs) were recently proposed as an interpretable and accurate alternative to multi-layer perceptrons. We evaluated CoxKAN on four synthetic and nine real datasets, including five cohorts with clinical data and four with genomics biomarkers. In synthetic experiments, CoxKAN accurately recovered interpretable hazard function formulae and excelled in automatic feature selection. Evaluations on real datasets showed that CoxKAN consistently outperformed the traditional Cox proportional hazards model (by up to 4% in C-index) and matched or surpassed the performance of deep learning-based models. Importantly, CoxKAN revealed complex interactions between predictor variables and uncovered symbolic formulae, which are key capabilities that other survival analysis methods lack, to provide clear insights into the impact of key biomarkers on patient risk. AVAILABILITY AND IMPLEMENTATION: CoxKAN is available at GitHub and Zenodo.

Humans↗
PubMed2026

A comparative evaluation of multiple enlarged perivascular space segmentation tools.

BACKGROUND: Enlarged perivascular spaces (ePVS) are a marker of cerebral small vessel disease, potentially reflecting reduced waste clearance. Because manual quantification is unfeasible in large datasets, we developed and evaluated an automated tool. METHODS: Detection Of Regions of Enlarged perivascular Spaces (DORES), a 3D nnU-Net-based deep learning algorithm was developed for ePVS segmentation using T1-weighted and fluid-attenuated inversion recovery magnetic resonance imaging (MRI). DORES was developed in two stages: an initial model trained on 35 manually segmented scans and a final model on 1460 pseudo-labeled sessions from the Vanderbilt Memory and Aging Project (VMAP). A subset of VMAP participants with 3 T brain MRI underwent whole-brain manual ePVS tracing (n = 35, 73 ± 9 years, 51% male) and visual rating (n = 388, 71 ± 8 years, 54% male) by a neuroradiologist. DORES was evaluated and compared against three other segmentation tools using Dice and F1 scores, absolute volume and element differences, correlation, and agreement. External validation used an Alzheimer's Disease Neuroimaging Initiative 3 subset with manual tracings (ADNI3, n = 18, 73 ± 9 years, 67% female). RESULTS: DORES achieved Dice scores of 0.61 ± 0.16 (white matter) and 0.72 ± 0.08 (basal ganglia) in VMAP, with strong correlations and agreement for ePVS count and volume. Performances modestly declined in ADNI3 across algorithms. Scanner-stratified analyses showed stronger correlations for Philips versus Siemens images in the basal ganglia, indicating scanner-dependent differences in measurement consistency. CONCLUSIONS: DORES provides a multimodal nnU-Net-based pipeline for ePVS segmentation in older adults. The model demonstrates robust within-cohort performance and reasonable external validity, though scanner-related effects limit application across sites.

Humans↗
PubMed2026

Unveiling tumor heterogeneity by single cell RNA-sequencing: From basic considerations to clinical applications.

Tumor heterogeneity-encompassing diverse cellular phenotypes, genomic alterations, and microenvironmental contexts-is a principal barrier to effective cancer therapy. Single-cell RNA sequencing (scRNA-seq) has transformed our ability to resolve this complexity by capturing transcriptomes at single-cell resolution. Here, we review the technical foundations required for high-quality scRNA-seq studies. We then trace the evolution of scRNA-seq platforms from manual micromanipulation to high-throughput systems, and describe the computational pipelines that enable reliable data interpretation. The application of scRNA-seq is exemplarily shown in the context of lung cancer, where single-cell profiling has revealed (i) the clonal and sub-clonal architecture of tumors, (ii) extensive remodeling of the immune microenvironment, iii) key mechanisms underlying resistance to targeted agents and immune-checkpoint blockade, and (iv) the dynamics of neo-antigen-specific T-cell responses. Integrating machine-learning techniques-such as deep-learning classifiers and graph-based models-with single-cell transcriptomic data has markedly sped up biomarker discovery, produced more accurate risk-stratification scores, and enabled the generation of patient-specific therapeutic predictions. We surveyed the major trial registry ClinicalTrials.gov and identified ∼380 ongoing or completed studies that explicitly incorporate scRNA-seq as a correlative or pharmacodynamic endpoint. Overall, the analysis shows that scRNA-seq becomes an increasingly important component of modern trials, providing high-resolution cellular and molecular readouts that complement conventional imaging and bulk-omics endpoints. While key challenges remain, ranging from costs, scalability and need for rigorous validation before routine clinical deployment, ongoing technological advances continue to expand the potential of scRNA-seq as a cornerstone of precision medicine.

Humans↗
PubMed2026

Prediction of bacterial protein-compound interactions with only positive samples.

MOTIVATION: Prediction of Compound-Protein Interactions (CPI) in bacteria is crucial to advance various pharmaceutical and chemical engineering fields, including biocatalysis, drug discovery, and industrial processing. However, current CPI models cannot be applied for bacterial CPI prediction due to the lack of curated negative interaction samples. RESULTS: We propose a novel Positive-Unlabeled (PU) learning framework, named BIN-PU, to address this limitation. BIN-PU generates pseudo positive and negative labels from known positive interaction data, enabling effective training of deep learning models for CPI prediction. We also propose a weighted positive loss function that weights to truly positive samples. We have validated BIN-PU coupled with multiple CPI backbone models, comparing the performance with the existing PU models using bacterial cytochrome P450 (CYP) data. Extensive experiments demonstrate the superiority of BIN-PU over the benchmark models in predicting CPIs with only truly positive samples. Furthermore, we have validated BIN-PU on additional bacterial proteins obtained from literature review, human CYP datasets, and uncurated data for its reproducibility. We have also validated the CPI prediction for the uncurated CYP data with biological and biophysical experiments. BIN-PU represents a significant advancement in CPI prediction for bacterial proteins, opening new possibilities for improving predictive models in related biological interaction tasks. AVAILABILITY AND IMPLEMENTATION: The source code and data are available at https://github.com/datax-lab/CYP.

Bacterial Proteins↗
PubMed2026

Representation learning for multi-modal spatially resolved transcriptomics data.

MOTIVATION: Spatial transcriptomics enables in-depth molecular characterization of samples on a morphology and RNA level while preserving spatial location. Integrating the resulting multi-modal data is an unsolved problem, and developing new solutions in precision medicine depends on improved methodologies. RESULTS: We introduce AESTETIK, a convolutional deep learning model that jointly integrates spatial, transcriptomics, and morphology information to learn accurate spot representations. AESTETIK yielded substantially improved cluster assignments on widely adopted technology platforms (e.g. 10x Genomics™, NanoString™) across multiple datasets. We achieved performance enhancement on structured tissues (e.g. brain) with a 21% increase in median ARI over previous state-of-the-art methods. Notably, AESTETIK also demonstrated superior performance on cancer tissues with heterogeneous cell populations, showing a 2-fold increase in breast cancer, 79% in melanoma, and 21% in liver cancer. We expect that these advances will enable a multi-modal understanding of key biological processes. AVAILABILITY AND IMPLEMENTATION: AESTETIK is implemented in Python 3 and is available as open source software at http://www.github.com/ratschlab/aestetik. The Snakemake pipeline for reproducing the results is available at http://www.github.com/ratschlab/st-rep.

Spatial Transcriptomics↗
PubMed2021

Deep generative neural network for accurate drug response imputation.

Drug response differs substantially in cancer patients due to inter- and intra-tumor heterogeneity. Particularly, transcriptome context, especially tumor microenvironment, has been shown playing a significant role in shaping the actual treatment outcome. In this study, we develop a deep variational autoencoder (VAE) model to compress thousands of genes into latent vectors in a low-dimensional space. We then demonstrate that these encoded vectors could accurately impute drug response, outperform standard signature-gene based approaches, and appropriately control the overfitting problem. We apply rigorous quality assessment and validation, including assessing the impact of cell line lineage, cross-validation, cross-panel evaluation, and application in independent clinical data sets, to warrant the accuracy of the imputed drug response in both cell lines and cancer samples. Specifically, the expression-regulated component (EReX) of the observed drug response achieves high correlation across panels. Using the well-trained models, we impute drug response of The Cancer Genome Atlas data and investigate the features and signatures associated with the imputed drug response, including cell line origins, somatic mutations and tumor mutation burdens, tumor microenvironment, and confounding factors. In summary, our deep learning method and the results are useful for the study of signatures and markers of drug response.

Antineoplastic Agents↗
PubMed2025

Refining sequence-to-activity models by increasing model resolution.

Decoding the cis-regulatory syntax that controls gene expression is essential for improving our understanding of cell differentiation and disease. To identify regulatory motifs and their regulatory syntax, deep learning based sequence-to-activity (S2A) models learn transcription factor binding motifs and their combinations from DNA sequence by modeling measured chromatin accessibility. Previously, we developed AI-TAC, a S2A model that predicts chromatin accessibility across various immune cell types in multi-task fashion, effectively decoding the regulatory syntax underlying immune cell differentiation. While ATAC-seq is commonly used to measure regional accessibility, it also provides high-resolution profiles, the distribution of Tn5 insertion sites, that offer additional insights into the precise location and strength of TF binding sites. Here we demonstrate that modeling ATAC-seq profiles alongside accessibility consistently improves predictions of differential chromatin accessibility across cell types. Moreover, we also find that multi-task learning across related immune cell types consistently outperforms single-task models. To understand what additional information bpAITAC learns from ATAC-seq profiles, we systematically compare sequence attributions from models trained with and without ATAC-seq profiles. We identify novel motifs with strong effect sizes that emerge only when profile data is included. Our findings suggest that modeling ATAC-seq at base-pair resolution enables the model to learn a more nuanced and sensitive representation of the cis-regulatory syntax driving immune cell-specific chromatin landscapes.

ATAC-seq↗
PubMed2025

Understanding the sources of performance in deep drug response models reveals insights and improvements.

MOTIVATION: Anti-cancer drug response prediction (DRP) using cancer cell lines (CLs) is crucial in stratified medicine and drug discovery. Recently, new deep learning models for DRP have improved performance over their predecessors. However, different models use different input data types and architectures making it hard to find the source of these improvements. Here we consider published DRP models that report state-of-the-art performance predicting continuous response values. These models take chemical structures of drugs and omics profiles of CLs as input. RESULTS: By experimenting with these models and comparing with our simple baselines, we show that no performance comes from drug features, instead, performance is due to the transcriptomics CL profiles. Furthermore, we show that, depending on the testing type, much of the current reported performance is a property of the training target values. We address these limitations by creating BinaryET and BinaryCB that predict binary drug response values, guided by the hypothesis that this reduces the noise in the drug efficacy data. Thus, better aligning them with biochemistry that can be learnt from the input data. BinaryCB leverages a chemical foundation model, while BinaryET is trained from scratch using a transformer-type architecture. We show that these models learn useful chemical drug features, which is the first time this has been demonstrated for multiple testing types to our knowledge. We further show binarizing the drug response values causes the models to learn useful chemical drug features. We also show that BinaryET improves performance over BinaryCB, and the published models that report state-of-the-art performance. AVAILABILITY AND IMPLEMENTATION: Code is available from https://github.com/Nik-BB/Understanding_DRP_models.

Humans↗