transformer / foundation model

HViLM is a foundation model for viral genomics that predicts pathogenicity, transmissibility, and host tropism across multiple viral families with high accuracy.

GenBio-PathFM is a foundation model for histopathology that achieves state-of-the-art performance with less training data by using a diverse data curation pipeline and a dual-stage learning strategy.

Pro2RNA generates species-adapted mRNA coding sequences from protein sequences using a multimodal language model, outperforming existing methods in benchmark tests.

The paper introduces scCChain, a transformer-based framework for mapping spatial cell-cell communication programs in spatial transcriptomics data, identifying hotspots at single-cell and spot resolution.

This study maps immune cell dynamics during human sequential lymph node metastasis using multi-omics approaches and identifies immune alterations linked to patient survival.

This paper introduces a framework for high-plex spectral imaging using DNA-barcoded labeling and programmable signal amplification to improve panel design and unmixing accuracy.

This paper introduces BertMS, a transformer-based spectral similarity framework that improves compound identification in metabolomics by better capturing structural relationships in MS/MS data.

SpeciefAI generates species-specific mRNA and protein sequences for antibodies and nanobodies using a transformer model trained on six species.

MetaReact predicts drug metabolites and enzyme selectivity using a reaction-aware Transformer, outperforming existing methods in accuracy and generalizability.

The paper introduces CytoType and ESM-CE, species-specific small models for cell type classification that achieve performance comparable to large foundation models with fewer parameters.

SpatialFusion identifies spatial niches through pathway activation patterns, revealing pre-malignant and malignant microenvironments in cancer tissues.

This paper develops a vision transformer framework to classify CMT2A disease phenotypes in iPSC motor neurons using kymographs, showing improved accuracy over traditional metrics.

This paper investigates whether a signed error derived from an XOR-based neuronal mechanism can serve as a learning signal in a binary autoencoder, showing improved reconstruction accuracy compared to traditional methods.

VarDCL predicts pathogenic missense variants by integrating multimodal protein language model embeddings and self-distilled contrastive learning, achieving high performance on clinical variants.

The paper introduces a transformer-based language model to identify computable phenotype definitions in biomedical literature and integrates it into a platform for literature curation.

The paper introduces pertTF, a transformer-based model for predicting genetic perturbation responses, showing improved performance in expression prediction and cell identity shifts across cellular contexts.

This paper introduces ARIES, a method using protein language model embeddings to construct multiple sequence alignments, achieving higher accuracy than existing methods, especially in low-identity regimes.

mnDINO is a vision transformer model for accurate micronuclei segmentation in DNA-stained images, achieving high accuracy and generalization across cell lines and microscopes.

This paper develops an automated system using deep learning to estimate morphometric measurements and predict weight of juvenile Chinook Salmon from 2D images with high accuracy.

FishMamba-1 is a linear-complexity foundation model designed for polyploid Cyprinid genomes, achieving high precision in gene annotation without RNA-seq evidence.

CDS-BART is a foundation model for mRNA sequence analysis that can handle sequences up to 4kb, pre-trained on nine taxonomic groups and available under MIT License.

HAETAE predicts tissue-specific histone modifications with high accuracy using a 5-base framework and long-read sequencing data.

This paper introduces a method to infer experimental culture conditions from spheroid images using structured learning, achieving high accuracy in protocol reconstruction.

PAVR enables high-resolution volumetric imaging of cells using a physics-aware framework trained on simulated data, demonstrating applications in tracking subcellular structures and analyzing cardiac cell dynamics.

ODIEU pairs human odor texts for 10,124 molecules with olfaction-tuned Sentence-BERT evaluators; CIRANO, a frozen GPT-2 with a molecular prefix, generates descriptions approaching human agreement.

This paper introduces a Transformer-based framework to predict full molecular dynamics trajectories from limited initial data, showing it captures long-range temporal structures and improves sampling compared to traditional methods.

PerturbGen predicts how genetic perturbations affect cellular trajectories across time in human datasets, showing effects on immune responses, hematopoiesis, and skin development.

This paper introduces InertialGenome, a transformer-based framework for resolution-agnostic chromosome 3D structure reconstruction, showing improved performance across multiple resolutions and experimental validation.

ROCIT classifies tumor or non-tumor origin of reads from bulk long-read sequencing with high accuracy, improving somatic variant calling in tumors.

This paper improves camera trap classification models by incorporating self-attention to leverage information from multiple animal detections, enhancing species classification accuracy in both mono- and multi-species scenarios.