Search PubMed⌕ Search

SEARCH · Search PubMed

Results for “Large language model”

Search indexed PubMed citations on genomics, clinical trials, systematic reviews and public health. Explore titles, authors and supplied subject terms, then open the PubMed record.

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

At least 73 records · Page 4Linked to original sources
PubMed2026

LAMBDA: a prophage detection benchmark for genomic language models.

Transformer-based genomic sequence models represent an emerging frontier in computational biology. Yet, their embeddings have not yet shown the same level of predictive power as natural and protein language models, highlighting a gap between current implementations and theoretical promise. Existing benchmarks for DNA language models primarily focus on classifying regulatory elements in eukaryotic genomes, leaving open the fundamental question of whether these models learn sequence-level features across whole genomes. We introduce LAMBDA, a benchmark designed to rigorously evaluate genome language model embeddings through phage-bacteria sequence discrimination across four categories of increasing complexity: probing tasks, fine-tuning assessments, diagnostic tests, and genome-wide prophage detection. Our comprehensive analysis of current genomic language models provides insight into the importance of training data selection relative to model size, the need for domain-specific training, and the capabilities and limitations of genomic language models for detecting prophage sequences. This benchmark represents a challenging genomic annotation task in the bacterial domain and addresses a key computational problem with direct relevance to microbiology and medicine.

Prophages↗
PubMed2025

The role of chromatin state in intron retention: A case study in leveraging large scale deep learning models.

Complex deep learning models trained on very large datasets have become key enabling tools for current research in natural language processing and computer vision. By providing pre-trained models that can be fine-tuned for specific applications, they enable researchers to create accurate models with minimal effort and computational resources. Large scale genomics deep learning models come in two flavors: the first are large language models of DNA sequences trained in a self-supervised fashion, similar to the corresponding natural language models; the second are supervised learning models that leverage large scale genomics datasets from ENCODE and other sources. We argue that these models are the equivalent of foundation models in natural language processing in their utility, as they encode within them chromatin state in its different aspects, providing useful representations that allow quick deployment of accurate models of gene regulation. We demonstrate this premise by leveraging the recently created Sei model to develop simple, interpretable models of intron retention, and demonstrate their advantage over models based on the DNA language model DNABERT-2. Our work also demonstrates the impact of chromatin state on the regulation of intron retention. Using representations learned by Sei, our model is able to discover the involvement of transcription factors and chromatin marks in regulating intron retention, providing better accuracy than a recently published custom model developed for this purpose.

Deep Learning↗
PubMed2026

GBFN: A gated bimodal fusion network leveraging foundation model embeddings for cancer drug sensitivity prediction.

Despite recent progress in deep learning for cancer drug sensitivity prediction, many existing models still rely on task-specific representation learning or relatively simple multimodal fusion, which may limit their ability to capture complex drug-cell interactions. To address this issue, we developed GBFN, a gated bimodal fusion network for continuous IC50 prediction that integrates pretrained drug and cell-line representations. Specifically, drug embeddings were obtained from SMI-TED, whereas cell-line embeddings were derived from transcriptomic profiles using BulkFormer. These two modalities were then combined through a dimension-wise gated fusion module and used to predict IC50 values in matched drug-cell line pairs. On the CCLE-based benchmark, GBFN outperformed representative neural baselines, including GraphDRP, TGSA, and TransEDRP, and achieved the best overall performance, with an R² of 0.8714 and an RMSE of 0.8938. Moreover, ablation analysis showed that the model using drug features and cell-line expression data with gated fusion performed better than the corresponding model using direct concatenation, indicating that the improvement was associated with the fusion strategy rather than with the input modalities alone. In addition, cell-line expression data were more informative than mutation data in the present setting, and adding mutation data to the model using drug features and expression data did not further improve performance. Across major cancer types, GBFN maintained generally high cell-line-level predictive performance, and perturbation-based attribution identified biologically relevant transcriptomic programs in selected drug-cell line settings. Together, these findings support GBFN as a compact and effective framework for continuous drug response prediction.

Humans↗
PubMed2025

Deep learning-based annotation of plant abiotic stress resistance genes for crops.

The declining costs of DNA sequencing have expanded genomic data, crucial for understanding plant abiotic stress responses and crop improvement. However, accurate gene annotation remains challenging. To address this limitation, we propose the PASRGA, a deep learning approach that leverages transfer learning and contrastive learning to annotate genes related to drought, salt, cold, and UV resistance. PASRGA achieves high F1-scores, area under the receiver operating characteristic (AUROC), area under the precision-recall curve (AUPRC), and Matthews correlation coefficient (MCC) in annotating stress resistance genes, significantly outperforming the general protein annotation model CLEAN, the plant phosphatase gene annotation model PF-NET, the top-ranked model in the CAFA5 challenge NetGO 4.0, and four traditional machine learning methods. Its effectiveness was further validated with a salt stress treatment experiment in Eutrema salsugineum. To facilitate crop breeding practices, we utilized PASRGA to annotate the genomes of 17 major crops. To improve accessibility and utility, we incorporated both manually curated and PASRGA-predicted gene data, together with the PASRGA tool, into the PlantASRG database (https://bioinfor.nefu.edu.cn/PlantASRG/). This comprehensive resource aims to support crop breeding initiatives and ensure food security.

Crops, Agricultural↗
PubMed2025

mamp-ml: A deep learning approach to epitope immunogenicity in plants.

Eukaryotes detect biomolecules through surface-localized receptors, key signaling components. A subset of receptors survey for pathogens, induce immunity, and restrict pathogen growth. Comparative genomics of both hosts and pathogens has unveiled vast sequence variation in receptors and potential ligands, creating an experimental bottleneck. We have developed mamp-ml, a machine learning framework for predicting plant receptor-ligand interactions. We leveraged existing functional data from over two decades of foundational research, together with the large protein language model ESM-2, to build a pipeline and model that predicts immunogenic outcomes using a combination of receptor-ligand features. Our model achieves 73% prediction accuracy on a held-out test set, even when an experimental structure is lacking. Our approach enables high-throughput screening of LRR receptor-ligand combinations and provides a computational framework for engineering plant immune systems.

Journal Article↗
PubMed2026

Agentic AI for Spatial Omics.

This highlight summarises recent advances in agentic artificial intelligence (AI) systems for spatial omics analysis. These systems are compared along two central tensions: autonomy versus accountability, and adaptability versus reproducibility. We argue that progress will depend not on maximising automation, but on defining where autonomy is appropriate.

Artificial Intelligence↗
PubMed2026

Hard to Halt: Automation Bias in Agent-Driven Sequencing Prior Authorization Workflows.

PURPOSE: Prior authorization (PA) for exome or genome sequencing is a time-consuming process that impedes timely rare disease diagnosis. Large language model-based browser agents offer potential for automating these workflows, but their clinical reliability remain uncharacterized. METHODS: We developed a sandbox compromising a simulated ES/GS PA submission payer portal and a synthetic EHR containing 836 patient records spanning compliant profiles and deficient profiles with different types of issues. Gemini 3 Pro, Gemini 3 Flash, and Claude Opus 4.5 were evaluated on task completion rate, form completion accuracy, and appropriate withholding for deficient profiles. RESULTS: Larger models achieved much higher task completion rates (Gemini 3 Pro 95.45%, Claude Opus 4.5 93.67%) compared to Gemini 3 Flash (56.05%), but nearly universally failed to withhold submission for deficient profiles whereas Gemini 3 Flash ironically demonstrated superior withholding performance (17.33%). In a non-agentic setting, Gemini 3 Pro correctly identified 91% of the issues in deficient profiles, indicating that withholding failure is attributable to the browser interaction rather than the model's reasoning limitations. CONCLUSION: Current LLM-based browser agents exhibit a systematic bias towards form submission that poses risks in PA workflows. A modular, multi-agent architecture with human supervision is necessary for a safe clinical deployment.

Journal Article↗
PubMed2026

Predicting genome-wide functional constraints with GPN-Star.

Genomic language models have emerged as a powerful approach for learning genome-wide functional constraints directly from DNA sequences1. However, standard genomic language models adapted from natural language processing often require large model sizes and computational resources, yet still fall short of classical evolutionary models in predictive tasks2-4. Here we introduce a genomic pretrained network with species tree and alignment representations (GPN-Star), which is a biologically grounded genomic language model featuring a phylogeny-aware architecture that leverages whole-genome alignments and species trees to model evolutionary relationships explicitly. Trained on alignments spanning vertebrate, mammal and primate evolutionary timescales, GPN-Star achieves state-of-the-art performance across a wide range of variant effect prediction tasks in both coding and non-coding regions of the human genome. Analyses across timescales show task-dependent advantages of modelling more recent versus deeper evolution. To demonstrate its potential to advance human genetics, we show that GPN-Star substantially outperforms previous methods in prioritizing pathogenic and fine-mapped genome-wide association study variants, yields strong enrichments of complex trait heritability and improves power in rare variant association testing5. Extending beyond humans, we train GPN-Star for five model organisms-Mus musculus, Gallus gallus, Drosophila melanogaster, Caenorhabditis elegans and Arabidopsis thaliana-demonstrating the robustness and generalizability of the framework. Taken together, these results position GPN-Star as a scalable, powerful and flexible tool for genome interpretation, well suited to leverage the growing abundance of comparative genomics data.

Journal Article↗
PubMed2026

GUANinE v1.1 reveals complementarity of supervised and genomic language models.

There has been much debate about the benefits of supervised versus unsupervised learning on genomes. Determining which is better in what contexts requires developing comprehensive benchmarks spanning functional and evolutionary tasks. Importantly, such benchmarks need large sample sizes to enable well-powered ranking of models. Having developed and applied such a benchmark here (GUANinE v1.1), we conclusively demonstrate each paradigm offers key advantages and outperforms on certain tasks. In accordance with training, supervised sequence-to-function models exhibit strong performance when annotating functional states characterized by chromatin accessibility or histone marks, while self-supervised language models outperform on evolutionary conservation. Our hundreds of new evaluations in this v1.1 expansion provide evidence for a tradeoff between input context size and model parameter count for a fixed compute budget, which we depict with new metrics such as kiloparameters/base pair. We also construct two new large-scale variant interpretation tasks in v1.1: cadd-snv measuring deleteriousness, and clinvar-snv measuring clinical pathogenicity. We find that conservation scores, and by extension, genomic language models, predict deleteriousness well, but successfully translating deleteriousness predictions to pathogenicity remains challenging. GUANinE v1.1 newly evaluates dozens of pretrained genomic models, and we conclude that moderate-context hybrid or post-trained language models may define the next era of machine learning in genomics.

Genomics↗
PubMed2026

Genome- and peak-informed two-stage framework for scATAC-seq cell type identification.

MOTIVATION: Accurate cell type annotation is essential in scATAC-seq analysis, as it underpins the characterization of cellular heterogeneity, the identification of regulatory elements, and downstream biological discovery. However, current annotation methods still face major challenges. First, although some approaches attempt to integrate genomic sequence information, they typically rely on shallow sequence representations and thus fail to capture the long-range dependencies and regulatory signals encoded in DNA. Second, substantial batch effects introduced by different platforms, sequencing batches, or tissue sources remain insufficiently addressed. Existing models often lack robust distribution alignment and domain generalization capabilities, leading to confounding non-biological variation and reduced annotation accuracy across datasets. RESULTS: To overcome these limitations, we propose seqAlignATAC, a two-stage intra-modality annotation framework that integrates sequence-derived embeddings with domain adaptation. In the first stage, we employ a large-scale pretrained nucleotide language model to extract low-dimensional, biologically informative representations from the genomic sequences of chromatin-accessible peaks. In the second stage, these embeddings are fed into a supervised neural network equipped with an adaptive alignment module to mitigate batch effects and harmonize feature distributions between labeled reference and unlabeled target datasets. Extensive experiments across multiple settings demonstrate that seqAlignATAC achieves competitive accuracy and robustness, effectively leveraging genome-level information while alleviating batch-induced distributional discrepancies. AVAILABILITY AND IMPLEMENTATION: The source code of seqAlignATAC is available at: https://github.com/BioCS-Lab/seqAlignATAC.

Humans↗
PubMed2006

When language breaks into pieces. A conflict between communication through isolated signals and language.

Here, we study a communication model where signals associate to stimuli. The model assumes that signals follow Zipf's law and the exponent of the law depends on a balance between maximizing the information transfer and saving the cost of signal use. We study the effect of tuning that balance on the structure of signal-stimulus associations. The model starts from two recent results. First, the exponent grows as the weight of information transfer increases. Second, a rudimentary form of language is obtained when the network of signal-stimulus associations is almost connected. Here, we show the existence of a sudden destruction of language once a critical balance is crossed. The model shows that maximizing the information transfer through isolated signals and language are in conflict. The model proposes a strong reason for not finding large exponents in complex communication systems: language is in danger. Besides, the findings suggest that human words may need to be ambiguous to keep language alive. Interestingly, the model predicts that large exponents should be associated to decreased synaptic density. It is not surprising that the largest exponents correspond to schizophrenic patients since, according to the spirit of Feinberg's hypothesis, i.e. decreased synaptic density may lead to schizophrenia. Our findings suggest that the exponent of Zipf's law is intimately related to language and that it could be used to detect anomalous structure and organization of the brain.

Animals↗
PubMed2025

Predicting functional constraints across evolutionary timescales with phylogeny-informed genomic language models.

Genomic language models (gLMs) have emerged as a powerful approach for learning genome-wide functional constraints directly from DNA sequences. However, standard gLMs adapted from natural language processing often require extremely large model sizes and computational resources, yet still fall short of classical evolutionary models in predictive tasks. Here, we introduce GPN-Star (Genomic Pretrained Network with Species Tree and Alignment Representation), a biologically grounded gLM featuring a phylogeny-aware architecture that leverages whole-genome alignments and species trees to model evolutionary relationships explicitly. Trained on alignments spanning vertebrate, mammalian, and primate evolutionary timescales, GPN-Star achieves state-of-the-art performance across a wide range of variant effect prediction tasks in both coding and non-coding regions of the human genome. Analyses across timescales reveal task-dependent advantages of modeling more recent versus deeper evolution. To demonstrate its potential to advance human genetics, we show that GPN-Star substantially outperforms prior methods in prioritizing pathogenic and fine-mapped GWAS variants; yields unprecedented enrichments of complex trait heritability; and improves power in rare variant association testing. Extending beyond humans, we train GPN-Star for five model organisms - Mus musculus, Gallus gallus, Drosophila melanogaster, Caenorhabditis elegans, and Arabidopsis thaliana - demonstrating the robustness and generalizability of the framework. Taken together, these results position GPN-Star as a scalable, powerful, and flexible new tool for genome interpretation, well suited to leverage the growing abundance of comparative genomics data.

Journal Article↗
PubMed2026

HyLnc: a hybrid deep learning and feature-based approach for long non-coding RNA prediction.

Long non-coding RNAs (lncRNAs) play important roles in gene regulation, development and disease, yet accurate identification of lncRNAs from transcriptomic data remains a major computational challenge. Existing methods often rely either on handcrafted sequence features or deep learning approaches, each with their inherent limitations in capturing the full complexity of RNA sequences. In this study, we proposed HyLnc, a computational framework that integrates transformer-based contextual embeddings with biologically meaningful sequence features for improved lncRNA prediction. A custom BERT-based model was first pre-trained on a large corpus of metazoan RNA sequences using a masked language modelling strategy to learn contextual nucleotide dependencies. The model was subsequently fine-tuned on curated datasets of lncRNAs and protein-coding transcripts and 256-dimensional deep sequence embeddings were extracted. Parallelly, 348 handcrafted features, including ORF characteristics, untranslated region (UTR) properties, nucleotide composition and Fickett scores, were computed. A multi-stage feature selection strategy was applied to identify the most informative features, resulting in optimized hybrid feature sets. Multiple machine learning classifiers were evaluated, with the RF model achieving the best performance. The proposed framework attained an accuracy of 91.30%, F1-score of 91.23% and MCC of 82.60 on an independent validation dataset, outperforming several existing lncRNA prediction tools. Thus, HyLnc demonstrates that integrating deep contextual representations with biologically interpretable features enhances lncRNA prediction. This approach provides a robust and scalable solution for large-scale transcriptome annotation and can be extended to other sequence-based prediction.

RNA, Long Noncoding↗
PubMed2026

Protein Language Model Decoys for Target Decoy Competition in Proteomics: Quality Assessment and Benchmarks.

Large-scale proteomics relies heavily on target-decoy competition for false discovery rate estimation in peptide identification, and the performance of this strategy depends strongly on the design of the decoy database. Classical generators such as reversal and shuffling remain widely used. Here, we introduce the first protein language model-based (PLM) decoy generation for peptide identification and benchmark it against classical strategies. We evaluate these approaches using three complementary quality-control layers: sequence-based separability, search-engine-agnostic spectral-space diagnostics, and end-to-end mass spectrometry benchmarks, including pipelines with rescoring. Across these analyses, PLM-based decoys are harder for sequence-only neural networks to distinguish than most classical generators, suggesting fewer obvious sequence-level artifacts. However, this signal is only weakly informative for search performance. Spectral diagnostics further show that short peptides occupy a particularly crowded target-decoy space and are therefore especially prone to local collisions across all generators. In full search pipelines, reverse decoys remain a strong baseline, and current PLM-based generators do not yet provide a clear overall advantage. We therefore view PLM-based decoys not as universal replacements for reverse decoys but as tunable tools for benchmarking, diagnostics, stress testing, and future adaptive decoy optimization, with increasing value as search models become more expressive.

Proteomics↗
PubMed2025

Predicting the First Onset of Suicidal Thoughts and Behaviors in Adolescents Using Multimodal Risk Factors: A 4-Year Longitudinal Study.

OBJECTIVE: Suicide is one of the leading causes of death among youth worldwide, yet existing studies that aimed to predict the first onset of suicidal thoughts and behaviors (STB) included a limited number of data modalities and/or focused on adult populations. This study aimed to prospectively predict first-onset STB across 4-year follow-ups in adolescents using an existing STB history classification model that was previously applied to baseline data and a new machine learning model with 195 biopsychosocial features. METHOD: Participants were 7,503 unrelated adolescents (54.5% female, ages 9-11 years at baseline) from the multisite, longitudinal Adolescent Brain Cognitive Development (ABCD) Study. An existing baseline STB history classification model was applied to predict longitudinal first-onset STB in adolescents compared with healthy controls and clinical controls (individuals with a mental health disorder but no STB). A new elastic net logistic regression model with 195 features was trained on data from 14 sites (n = 5,220), and the resulting top 15 features were validated at 7 independent sites (n = 2,283). RESULTS: The previously developed model to classify STB lifetime history also prospectively predicted first-onset STB in adolescents with an area under the curve (AUC) [95% CI] of 0.73 [0.70, 0.75], p < .001, compared with healthy controls and AUC [95% CI] of 0.63 [0.60, 0.66], p < .001, compared with clinical controls. The newly trained model with top 15 features performed similarly with AUC [95% CI] of 0.73 [0.71, 0.76], p < .001, and AUC [95% CI] of 0.64 [0.60, 0.66], p < .001, for the same comparison groups. The most consistent predictors across models included female sex, sleep disturbances, and maladaptive home and school environments. CONCLUSION: The models predicted first-onset STB in adolescents with moderate accuracy. This study also confirmed the roles of well-established psychological risk factors for STB and identified several novel neurocognitive and brain imaging risk factors. Future studies should validate these models in large-scale diverse samples before clinical translation. PLAIN LANGUAGE SUMMARY: This study followed over 7,500 adolescents for 4 years and tested 2 machine learning models using psychological, social, and brain data to identify those at risk of experiencing suicidal thoughts or behaviors. Both models predicted first-time suicidal thoughts or behaviors with moderate accuracy. Key risk factors that were identified included being female, experiencing sleep problems, and negative home and school environments. DIVERSITY & INCLUSION STATEMENT: We worked to ensure sex and gender balance in the recruitment of human participants. We worked to ensure race, ethnic, and/or other types of diversity in the recruitment of human participants. We worked to ensure that the study questionnaires were prepared in an inclusive way. Diverse cell lines and/or genomic datasets were not available. One or more of the authors of this paper self-identifies as a member of one or more historically underrepresented racial and/or ethnic groups in science. One or more of the authors of this paper self-identifies as a member of one or more historically underrepresented sexual and/or gender groups in science. We actively worked to promote sex and gender balance in our author group. One or more of the authors of this paper received support from a program designed to increase minority representation in science. We actively worked to promote inclusion of historically underrepresented racial and/or ethnic groups in science in our author group. While citing references scientifically relevant for this work, we also actively worked to promote sex and gender balance in our reference list. While citing references scientifically relevant for this work, we also actively worked to promote inclusion of historically underrepresented racial and/or ethnic groups in science in our reference list. The author list of this paper includes contributors from the location and/or community where the research was conducted who participated in the data collection, design, analysis, and/or interpretation of the work.

Adolescent↗
PubMed1999

Words in the brain's language.

If the cortex is an associative memory, strongly connected cell assemblies will form when neurons in different cortical areas are frequently active at the same time. The cortical distributions of these assemblies must be a consequence of where in the cortex correlated neuronal activity occurred during learning. An assembly can be considered a functional unit exhibiting activity states such as full activation ("ignition") after appropriate sensory stimulation (possibly related to perception) and continuous reverberation of excitation within the assembly (a putative memory process). This has implications for cortical topographies and activity dynamics of cell assemblies forming during language acquisition, in particular for those representing words. Cortical topographies of assemblies should be related to aspects of the meaning of the words they represent, and physiological signs of cell assembly ignition should be followed by possible indicators of reverberation. The following postulates are discussed in detail: (1) assemblies representing phonological word forms are strongly lateralized and distributed over perisylvian cortices; (2) assemblies representing highly abstract words such as grammatical function words are also strongly lateralized and restricted to these perisylvian regions; (3) assemblies representing concrete content words include additional neurons in both hemispheres; (4) assemblies representing words referring to visual stimuli include neurons in visual cortices; and (5) assemblies representing words referring to actions include neurons in motor cortices. Two main sources of evidence are used to evaluate these proposals: (a) imaging studies focusing on localizing word processing in the brain, based on stimulus-triggered event-related potentials (ERPs), positron emission tomography (PET), and functional magnetic resonance imaging (fMRI), and (b) studies of the temporal dynamics of fast activity changes in the brain, as revealed by high-frequency responses recorded in the electroencephalogram (EEG) and magnetoencephalogram (MEG). These data provide evidence for processing differences between words and matched meaningless pseudowords, and between word classes, such as concrete content and abstract function words, and words evoking visual or motor associations. There is evidence for early word class-specific spreading of neuronal activity and for equally specific high-frequency responses occurring later. These results support a neurobiological model of language in the Hebbian tradition. Competing large-scale neuronal theories of language are discussed in light of the data summarized. Neurobiological perspectives on the problem of serial order of words in syntactic strings are considered in closing.

Association Learning↗
PubMed1998

Molecular modeling information transfer with VRML: from small molecules to large systems in bioscience.

The suitability of the Virtual Reality Modeling Language (VRML) for the communication of scientists via the internet is demonstrated with recent results from computer assisted cancer research: I. Substrate channels in cytochrome P450 enzymes. II. Binding properties of the wild type and mutated p53 tumor suppressor protein. Complex 3D molecular models were used to visualize new insights in the active site access of cytochrome P450 enzymes and in the p53 protein-DNA binding achieved by the use of computational methods. These 3D models of biomolecular systems were transferred into VRML scenarios. Additional implemented features allow users to receive related information interactively. With these examples it is shown that VRML provides an efficient method for scientific information exchange by the use of complex 3D molecular models.

Binding Sites↗
PubMed2001

Constructing influence views from data to support dynamic decision making in medicine.

A dynamic decision model can facilitate the complicated decision-making process in medicine, in which both time and uncertainty are explicitly considered. In this paper, we address the problem of automatic construction of a dynamic decision model from a large medical database. Within the DynaMoL (a dynamic decision modeling language) framework, a model can be represented in influence view. Thus, our proposed approach first learns the structures of the influence view based on the minimal description length (MDL) principle, and then obtains the conditional probabilities of the model by Bayesian method. The experiment results demonstrate that our system can efficiently construct the influence views from data with high fidelity.

Algorithms↗