Valentina Boeva

The paper introduces a standardized resource of human gene embeddings derived from diverse biological data sources, enabling comparison and use across modalities.

GRASS-MIL predicts phenotypes from spatial single-cell data and decomposes predictions into specific cellular neighborhoods, showing improved performance over composition-based methods.

The paper introduces a statistical destriping method for VisiumHD spatial transcriptomics data using nuclear counts modeling, improving stripe-factor estimation and preserving biological signal.

UniversalEPI predicts cell type-specific and differential chromatin interactions using DNA sequence and ATAC-seq data, showing robust performance across cell types and conditions.

Chiron3D predicts CTCF-mediated chromatin loops using a DNA-only model, revealing sequence determinants of loop strength and stability through interpretable deep learning.

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

DeepCAST-GWAS improves GWAS by using deep learning to prioritize regulatory SNPs, increasing the reliability and replicability of genetic associations.

This study uses sparse autoencoders to analyze hidden representations in single-cell foundation models, revealing interpretable biological and technical features and demonstrating their functional relevance.

This study evaluates how different graph augmentations affect embedding quality and downstream performance in spatial omics, showing that well-designed augmentations improve results while poor ones can harm performance.

H&Enium aligns spatial transcriptomics and H&E imaging data using a contrastive framework to improve cell type classification and gene expression prediction.

The paper analyzes multiomic data from primary and metastatic esophageal adenocarcinoma tumors to identify cell states and their spatial interactions, linking them to clinical outcomes.

The paper characterizes exosomal RNA cargo in pleural mesothelioma, identifies potential biomarkers, and explores their role in cell communication and prognosis.

CDState infers malignant cell subpopulations from bulk RNA-seq data, revealing gene programs and clinical associations in 33 cancer types.

This paper develops a method to track microfluidic droplets over time using optimal transport and contrastive learning, achieving high accuracy in tracking without labeled data.

This study benchmarks eight self-supervised learning methods on single-cell data across three tasks and eight datasets, finding SimCLR and VICReg to be the top performers.

CancerFoundation is a single-cell RNA-seq foundation model that outperforms existing models in batch integration and drug response prediction, even when trained on a smaller dataset.

The paper investigates the impact of mitochondrial content filtering in cancer single-cell studies, finding that high pctMT malignant cells are viable and functionally significant.

The paper introduces ANS, a new gene signature scoring method for scRNA-seq, improving cell annotation by addressing limitations in existing methods and demonstrating strong classification performance.

SurvBoard benchmarks multi-omics cancer survival models, showing statistical models outperform deep learning in calibration and that pan-cancer training improves performance.

scROSHI identifies cell types without training data by using pre-existing gene sets and hierarchical relationships, outperforming methods in low-data and high-diversity scenarios.

CanSig benchmarks methods for identifying transcriptional signatures in cancer using scRNA-seq data, finding BBKNN as a leading method with clinically relevant outcomes.

The paper sparsifies high-dimensional multi-omics survival models using knowledge distillation with 1-penalized linear models, achieving performance comparable to complex teachers while using fewer features.

This paper introduces a hierarchical autoencoder for multi-omics cancer survival analysis, showing improved performance through soft modality selection.

CHROMATIX reconstructs many-body chromatin interactions from Hi-C data and identifies functional features in transcriptionally active loci.

FUN-LDA predicts tissue-specific functional effects of noncoding variants by integrating epigenetic data, showing higher accuracy than existing methods in eQTL and GWAS analyses.