transformer / foundation model

MAGI provides mechanistic insights into genetic variants by integrating genomic data and clinical annotations to improve variant interpretation and generate testable hypotheses.

The paper introduces a two-stage framework for gene expression prediction from cis-regulatory elements, showing improved performance across human cell types and tissues.

Vermeer generates synthetic microscopy images of protein localization using protein sequences and cell morphology, outperforming previous methods in quality and flexibility.

This paper introduces NFR, a prior-free algorithm that enhances microscopy resolution without requiring PSF estimation, achieving results beyond the Sparrow limit and improving super-resolution techniques.

DanioDecima improves gene expression prediction in zebrafish by transferring knowledge from human and mouse models, enhancing correlations and generating synthetic promoters with expected lineage-specific motifs.

The paper introduces ProteomeLM-HPI, a language model adapted to predict cross-species host-pathogen interactions using proteome context and fine-tuning.

The paper introduces CoMA-UNet, a method to synthesize tau PET from MRI using covariates, showing strong agreement with true PET and utility in clinical tasks.

This study identifies active transcriptome remodeling and epigenetic adaptations in Leishmania persisters, revealing a sophisticated survival strategy rather than passive dormancy.

This paper introduces TrASPr+BOS, a generative AI model for predicting and designing RNA to achieve tissue-specific splicing outcomes, showing improved performance over existing methods.

The paper introduces SE3CSP, an SE(3)-equivariant deep-learning framework for crystal structure prediction of nucleosides, achieving 57% accuracy and enabling the discovery of bioactive self-assembling materials.

The paper introduces STMDiT, a diffusion transformer that synthesizes H&E histopathology images from transcriptomic data, improving image quality and transcriptomic fidelity over existing methods.

ModCRE-NN predicts transcription factor DNA-binding motifs by integrating structural and evolutionary data, showing improved performance under low similarity and uncertainty.

LucaPhylo infers phylogenetic trees from unaligned protein sequences using hyperbolic geometry and protein language models, achieving accuracy comparable to alignment-based methods for divergent RNA viruses.

OneGenome-Rice (OGR) is a genomic foundation model for rice pre-trained on 422 genomes, outperforming existing models in 11 tasks and enabling downstream applications like trait analysis and gene expression prediction.

This study compares RAM-NJ and SAM-NJ phylogenies using protein foundation models to reconstruct species relationships without conventional substitution models or multiple sequence alignments.

This paper uses a DNA foundation model to prioritize promoter regulatory variants in colorectal cancer, identifying high-impact variants enriched in cancer-related pathways and GWAS loci.

Unimeth improves DNA methylation detection from nanopore reads across species and modification types by using a transformer framework with multi-phase training.

UcTCRp profiles MAIT- and iNKT-associated immune states using TCRβ sequences without requiring single-cell data or antigen-specific markers.

The paper introduces a grey-box Transformer framework for reconstructing sparse pharmacokinetic curves and initializing pharmacometric models, achieving high accuracy in drug concentration prediction.

The paper uses a genome language model to analyze plant Y chromosomes from Silene and Humulus, identifying shared genomic features and evolutionary patterns despite structural divergence.

The paper introduces CO_SCPLOWARBONC_SCPLOW, a family of efficient DNA language models that address challenges in genomic sequence modeling by adapting LLM techniques to DNA's unique properties.

scDynOmics is a transformer model for single-cell multiomics data that outperforms existing models and provides interpretable insights into cellular dynamics.

Genos-m is a foundation model for human-associated microbial genomes that outperforms existing models in gene-fitness prediction and community structure analysis.

The paper identifies late stomatal lineage targets of SPCH and FAMA, showing their roles in controlling guard cell division and expansion through temporal regulation.

PlasmidLM is a DNA language model that generates plasmid constructs from component specifications, achieving high fidelity and viability in generated sequences.

MeNet-AO corrects optical aberrations in intravital microscopy using a multi-encoder network, improving imaging in zebrafish and mouse tissues.

CASPE is a lightweight platform for site-directed protein engineering that improves the identification of beneficial mutations for thermostability and pH tolerance.

HiCP2GAN enhances low-resolution Hi-C data using a pretrained Vision Transformer as a discriminator, improving resolution over existing methods.

ProteomeLM predicts protein-protein interactions and gene essentiality across species using a proteome-scale language model, outperforming existing methods in speed and accuracy.

This paper introduces a feature-preserving conditional GAN for multi-site MRI harmonization, showing improved performance over existing methods.