transformer / foundation model

MuAt2, a Transformer model, classifies tumour types and subtypes from somatic variants, improving accuracy and calibration across cancer types and cohorts.

This study identifies tissue- and age-specific transcriptional regulators of ginsenoside biosynthesis in Panax notoginseng using transcriptomics and metabolomics.

TweetyBERT is a self-supervised transformer model that parses birdsong without human labels, learning acoustic and temporal patterns in canary song.

This paper introduces OpenPheno, a multimodal model that predicts bioactivity using phenotypic profiles and natural language descriptions, achieving strong zero-shot performance on unseen assays.

CancerSTFormer analyzes spatial transcriptomes to study cancer niches and therapy responses at multiple scales.

This paper introduces VascFlexMap, a deep learning framework that reconstructs microvascular maps from low-frame-rate ultrasound data using a transformer-decoder network, achieving faster reconstruction with reduced data requirements.

The paper evaluates deep learning architectures for high-resolution chromatin accessibility prediction, finding that early feature extraction, not downstream architecture, primarily determines performance.

IDPFunNet predicts IDR functions using a hybrid deep learning model, outperforming existing methods in binding prediction and DFL identification across multiple benchmarks.

Influ-BERT improves IAV subtype identification by using domain-adaptive pretraining and achieves high F1-scores across underrepresented subtypes.

This paper uses GPT-4o to extract mechanistic insights from biomedical images related to COVID-19 and neurodegenerative diseases, revealing both established and novel biological connections.

The paper benchmarks automated detection and classification methods for gibbon calls using ARUs in Cambodia, finding BirdNET outperforms others with fewer samples.

This study designs peptides targeting the FtsQBL divisome complex in E. coli using structure-guided methods, showing they inhibit cell division.

This paper applies attention-based models to predict phenotype from genotype, showing improved performance in epistatic regimes and enabling transfer learning across environments.

This paper predicts synthetic lethal gene pairs using a foundation model-based framework with uncertainty calibration, improving performance without relying on PPI or GO data.

The paper develops a deep learning pipeline for analyzing 3D collagen architecture in cleared fibrotic tissues using optimized clearing, imaging, and a U-Net model.

This paper introduces ARCH3D, a foundation model for global genome architecture that captures spatial structure and interchromosomal interactions using a novel masked locus modeling task.

The paper introduces a diverse RNA structure database and a deep learning model, eFold, which improves RNA secondary structure prediction by emphasizing structural diversity over database size.

Across 14 sarcoma and mimic subtypes, six pathology foundation models beat ResNet50 but lost accuracy under multi-institutional staining; adding 121 externally stained slides narrowed the gap.

This study uses large language models to map how the brain encodes relational semantic information, showing distinct brain regions represent concepts and their relationships.

This paper introduces a transformer-based model for predicting and designing biosynthetic gene clusters, achieving high accuracy in domain prediction and validating novel gene clusters through experimental testing.

This work introduces a transformer-based model for lysine methylation site prediction, achieving state-of-the-art accuracy through multitask learning and validation with mass spectrometry.

EES-Transformer V2 classifies tissues and learns gene representations from extreme expression sets, achieving high accuracy and revealing tissue-specific gene interactions.

This study benchmarks foundation models for splice site and exon annotation, finding performance varies greatly depending on exon type and training data representation.

CrossAffinity predicts protein-protein binding affinity using a sequence-based cross-attention model, outperforming both structure-based and sequence-based methods on multiple test sets.

DynaBiomeX filters noise in sparse longitudinal microbiome data using a dual-strategy deep learning framework, achieving high sensitivity and zero false positives in detecting gut dysbiosis.

PEINT simulates realistic protein evolution by modeling sequence changes over time, generating functional proteins that retain enzymatic activity.

CLM-X is a multimodal single-cell foundation model that outperforms existing methods in cross-modal translation and perturbation prediction across scRNA-seq and scATAC-seq datasets.

This paper introduces RDBCycleGAN-CBAM, a deep learning model for low-dose CT image denoising, achieving higher PSNR and SSIM than existing methods.

This study identifies an auxin-BR antagonistic module that regulates plasmodesmal connectivity in response to viral movement proteins during early infection.

evoCancerGPT generates cancer progression patterns using transfer learning on single-cell RNA sequencing data, outperforming baselines in low-context scenarios.