transformer / foundation model

SpaAxFormer improves spatial transcriptomics by integrating gene-level and cell-level representations, achieving better imputation and spatial-domain identification in some datasets.

LingoCell interprets gene perturbations using natural language and tissue niches, revealing synergistic disease responses and shared fibroblast programs in cancers.

This study shows ApoE isoforms influence antioxidant pathways differently across aging, with ApoE4 showing stronger transcriptional responses but less effective protein production, linking to neurodegenerative risk.

The paper identifies how medial frontal cortex integrates semantic content and sensory evidence for decision-making using human neurosurgical data and a vision-language model.

This paper introduces MirrorDNA, a DNA language model with a strand-equivariant tokenizer that improves prediction consistency across DNA strands.

This paper introduces Brieno, a self-supervised model for raw nanopore signals that improves read classification and methylation detection without reference genomes.

This paper introduces PDLMs, protein-DNA language models that jointly represent protein and DNA sequences to predict and engineer DNA recognition specificity.

The paper compares transcriptomic profiles across kidney organoid, assembloid, and transplantation models to evaluate their maturity and suitability for research.

CellReg3D integrates DNA language models and 3D genome structure to model cell-type-specific gene regulation, achieving state-of-the-art performance in predicting epigenetic marks and gene expression.

MicrobeRT is a framework of fine-tuned language models for metagenomic analysis, achieving strong performance in taxonomic classification, antibiotic resistance detection, and pathogenicity assessment.

The paper shows that neural populations adjust their optimal temporal resolution during stimulus processing, becoming faster and more nonlinear when decoding information is most accurate.

The paper decodes protein N-glycosylation sites from protein language model embeddings using a zero-shot framework, revealing biological insights and disease associations.

VIGAT uses a vision-language model to quantify gene activity from single-cell ATAC-seq by analyzing accessibility patterns and genomic annotations across marker genes.

This paper identifies two failure modes in extended reasoning models within bioinformatics tasks and introduces DTDR to improve accuracy and reduce errors.

MaxCAT-Net improves skin lesion classification and segmentation with hybrid architecture and calibration, showing performance on benchmark datasets but with preliminary results.

This paper introduces a deep learning method that integrates genomic and trait data to improve species delimitation in complex speciation scenarios.

ESMfluc predicts local backbone variability in proteins using attention-based evidence decomposition, showing how sequence positions contribute to predictions of single or multi-PB states.

ctDNA-BERT scores individual bisulfite reads for cell-free DNA methylation analysis, achieving high accuracy in cancer staging without cohort-specific retraining.

This paper introduces Borzoi-SE, a deep learning framework that identifies super-enhancers and detects pathogenic variants directly from DNA sequence across multiple human cell lineages.

This paper shows that biological concepts can be represented as directions in foundation model embeddings, enabling interpretable and reusable analysis of diverse biological properties.

The paper introduces Cerberus, an improved regulatory sequence model using bidirectional state space blocks, achieving higher accuracy and efficiency than existing models like Borzoi.

This study evaluates genomic foundation models for viral sequence classification, finding that NTv3-650M achieves high accuracy comparable to BLAST on standard-length sequences.

CheckAMG identifies auxiliary viral genes by combining annotation-based prediction with a genome-language model, improving accuracy and uncovering novel genes with no sequence similarity to references.

Minerva extracts RNA base-pairing, protein-contact and repeat maps from a genome language model without alignments, flagging unannotated structured loci and a phage ncRNA array reverse-transcribed into short cDNAs.

Continued pretraining of an RNA language model on picornaviral 5′ UTRs alone predicts IRES base pairs with median precision 0.80 versus 0.47 for Infernal, and reveals a Type II subclass.

DeepSCENIC uses transfer learning from sequence-to-function models to infer causal gene regulatory networks from single-cell multiome data.

This study shows that Brf1-mediated RNA polymerase III activity restricts murine gammaherpesvirus spread by enhancing host antiviral responses during low MOI infections.

REN-former identifies gene regulators of kidney disease transitions by combining single-cell modeling and human genetics.

LucaCell is a sequence-centric foundation model that improves cross-species single-cell analysis by using mRNA sequence embeddings instead of fixed gene identifiers.

This paper introduces an environment-aware DNA language model to prioritize genomic regions associated with stress response in maize using stress-specific prompts and transformer-based modeling.