Biology subjects

Boeva, V.

Publications and source records attributed to Boeva, V..

At least 19 recordsLinked to original sources

bioRxiv2026

Gene Embedding Hub: a standardized resource for human gene representations

Gene embeddings are vector representations of gene identity that provide machine-learning models with information from diverse biological sources. However, their use is hindered by a lack of standardization: inconsistent identifiers, formats, and preprocessing conventions. In this work, we present a unified resource of 67 human gene-embedding sets derived from protein, DNA and RNA sequences, biomedical text, bulk expression data, single-cell expression models, functional data, biological networks, knowledge graphs, regulatory genomics and multimodal models, covering up to 20,097 GENCODE-annotated protein-coding genes. The collection includes both published precomputed embeddings and representations generated or extracted from publicly available models and biological data. Each embedding set is organized in a folder containing the embedding matrix, the Ensembl gene identifiers, machine-readable metadata, and processing documentation. We validate the resource by measuring global and local similarity, cross-embedding consensus, and cross-embedding predictability. The collection is designed to allow community-contributed embeddings that meet predefined technical and documentation requirements, enables comparison of gene representations across modalities and provides standardized embeddings for training and downstream applications.

genomics↗

bioRxiv2026

GRASS-MIL: graph-based representation and discovery of phenotype-associated spatial structures with multiple instance learning

Tissue function is shaped by both cellular composition and the spatial organization of cells. Spatial single-cell technologies capture this organization across tissue sections, enabling tissue architecture to be linked to disease phenotypes, prognosis and treatment response. However, many current graph-based predictive models depend on assay-specific molecular features, limiting joint analysis of cohorts profiled with incompatible panels, and either aggregate local predictions uniformly or rely on post hoc importance scores that do not directly decompose the basis of a sample-level prediction. Here we introduce GRASS-MIL, a graph multiple-instance learning framework that learns the relative relevance and class-specific evidence of individual cellular neighborhoods, predicts phenotypes across platforms, and decomposes each sample-level prediction exactly into signed contribution from each local neighborhood. These contributions are terms of the prediction itself rather than a post hoc approximation and sum exactly to the predicted logit. In simulations, GRASS-MIL distinguished phenotypes with identical global cell-type composition but different spatial organization, whereas composition-based classifiers performed no better than chance. A single model trained jointly on CosMx spatial molecular imaging, Xenium and MERFISH datasets, harmonized only through a shared cell-type vocabulary, predicted treatment status in high-grade serous ovarian carcinoma with an area under the receiver operating characteristic curve of 0.74 and associated this prediction with two stromal architectures: an immune-rich niche admixed with malignant cells and a desmoplastic niche characterized by malignant-cell exclusion. Applied to spatial proteomics data from non-small-cell lung cancer, GRASS-MIL identified distinct niches associated with relapse (area under the receiver operating characteristic curve, 0.57), including myeloid-rich niches in relapsing tumors and adaptive immune niches in non-relapsing tumors. In a separate lung adenocarcinoma cohort, a similar lymphoid-myeloid polarity was observed when predicting 3-year mortality (area under the receiver operating characteristic curve, 0.62), with a B-cell and CD4+ T-cell niche associated with favorable outcome and a neutrophil and macrophage niche associated with adverse outcome. GRASS-MIL thus combines cross-platform phenotype prediction with an exact, signed decomposition of each prediction into the cellular neighborhoods on which it depends, providing a direct route from sample-level prediction to testable hypotheses linking tissue architecture with patient phenotypes.

cancer biology↗

bioRxiv2026

Striping artifact removal in VisiumHD data through nuclear counts modeling

Motivation10x Genomics VisiumHD enables spatial transcriptomics at 2 m x 2 m resolution but exhibits slide-specific, non-periodic striping artifacts due to lane-width variability. These multiplicative row/column effects distort bin total counts and can bias downstream analyses. The state-of-the-art destriping approach is the normalization procedure used as a preprocessing step in bin2cell; it applies sequential high-quantile row- then column-wise normalization, which is asymmetric and can introduce edge effects/macro-stripes and distortions of large-scale total-count structure. ResultsWe propose a statistical destriping approach that leverages nuclei segmentation from the coregistered H&E image. Assuming transcript abundance is constant within each nucleus, we model bin counts with a negative binomial distribution whose mean is a product of a nucleus-specific concentration and row- and column-specific stripe-factors reflecting lane-width variation. We fit all parameters in a generalized linear modeling framework with cross-validated regularization on stripe-factors and iterative dispersion estimation, and use the fitted parameters to correct the observed counts into a destriped image. On synthetic data with known ground truth, our method improves stripe-factor estimation accuracy and reduces error in corrected counts relative to bin2cell and bin2cell-derived baselines. Across four public VisiumHD slides, it consistently lowers striping intensity while substantially better preserving biological signal present in the large-scale global count structure and avoiding the artifacts introduced by other methods. Availability and ImplementationAll source code and links to publicly available data used for this study are available at https://github.com/paolamalsot/destriping-GLM. Contactpaola.malsot@inf.ethz.ch, raetsch@inf.ethz.ch NoteThis manuscript extends the version submitted to Intelligent Systems for Molecular Biology (ISMB) 2026 by describing a new optimization algorithm that yields an approximately tenfold speedup. All plots and benchmarks in this manuscript use the updated implementation.

bioinformatics↗

bioRxiv2026

Chiron3D: an interpretable deep learning framework for understanding the DNA code of chromatin looping

MotivationThree-dimensional folding of the genome into structures such as chromatin loops is essential for gene regulation. Current experimental methods for mapping these structures, like Hi-C and HiChIP, are labor-intensive and require repeated assays to test hypothesized mutation effects. This motivates the need for predictive approaches that reveal the sequence determinants of chromatin loops. ResultsIn this work, we present a novel and interpretable computational pipeline for predicting CTCF-mediated chromatin loops. We propose Chiron3D, a DNA-only model trained in a cell-type specific manner to predict CTCF HiChIP contact maps. By leveraging pre-trained embeddings from a foundation model, our approach is competitive with baselines that take CTCF ChIP-seq as additional input, while enabling nucleotide-level attribution to the input DNA sequence. Using our framework, we provide likely mechanistic insights into the physical control of loop dynamics. Specifically, we find that the strength of the loop extrusion anchorage site is largely governed by the amount and binding affinity of CTCF sites at the boundaries. Furthermore, we reveal that loop stability is regulated by the amount of intra-loop CTCF binding sites, where fewer sites within the loop lead to a more stable domain. Using targeted, single-nucleotide edit simulations with Chiron3D, we show that both loop strength and stability can be precisely controlled. Together, these results provide novel mechanistic insights into the physical control of genome organization and highlight the potential of decoding the DNA sequence logic in silico. AvailabilityThe Chiron3D pipeline is made available at https://github.com/BoevaLab/Chiron3D. Supp. informationSupp. data are available at Journal Name online.

genomics↗

bioRxiv2025

DeepCAST-GWAS: Improving the Discovery of Genetic Associations Using Deep Learning-Based Regulatory SNP Prioritization

Genome-wide association studies (GWAS) have uncovered numerous variants linked to complex traits, yet power remains limited by the large multiple testing burden and the inclusion of many variants with minimal regulatory impact. We present Deep learning-based Chromatin Accessibility SNP Targeting for GWAS (DeepCAST-GWAS), a framework that integrates functional annotations derived from deep learning models to improve both the yield and the reliability of GWAS findings. DeepCAST-GWAS uses SNP Activity Difference (SAD) scores from in silico mutagenesis with the Enformer model to estimate the predicted effect of each variant on chromatin accessibility across tissues, allowing statistical testing to focus on variants with stronger regulatory evidence. Using conservative family-wise error rate (FWER) control, DeepCAST-FWER produces fewer associations than existing power-boosting approaches, but the associations it reports replicate in larger cohort GWAS at substantially higher rates. For applications where discovery count is more important, DeepCAST-sFDR increases the number of genome-wide significant findings above baseline GWAS by using the Enformer SAD scores for stratified False Discovery Rate (sFDR) control. DeepCAST-sFDR achieves performance comparable to the strongest competing method, while maintaining reliability on par with a standard GWAS. Subsampling analyses across a wide range of traits confirm these improvements in both sensitivity and replicability. DeepCAST-GWAS offers a principled way to incorporate sequence-based regulatory predictions into population-scale association testing, demonstrating that chromatin accessibility activity scores can improve the stability of GWAS discoveries. The framework is made available at https://github.com/BoevaLab/DeepCAST-GWAS.

bioinformatics↗

bioRxiv2025

Sparse Autoencoders Reveal Interpretable Features in Single-Cell Foundation Models

Single-cell foundation models (scFMs) hold promise for applications in cell type annotation, data integration, and prediction of the effects of cell perturbations, but their internal mechanisms remain poorly understood. We investigate the structure of these models by training sparse autoencoders (SAEs) on the hidden representations of three widely used scFMs: scGPT, scFoundation, and Geneformer.The learned features reveal diverse and complex biological and technical signals, which emerge even in pre-trained models. We also observe that the encoding of this information differs between scFMs with distinct training protocols and architectures. Finally, we demonstrate that SAE-derived features are functionally related to model behavior and can be intervened upon. Suppressing batch-associated features reduces unwanted technical variation and improves data integration while preserving the core biological signal. Activating drug-encoding features steers control cells toward drug-perturbed states in a concentration-dependent manner. These findings provide a path toward more interpretable and controllable single-cell foundation models.

bioinformatics↗

bioRxiv2025

Exploring Augmentation-Driven Invariances for Graph Self-supervised Learning in Spatial Omics

Spatial omics technologies provide rich insights into biological processes by jointly capturing molecular profiles and the spatial organization of cells. The resulting high-dimensional data can be naturally represented as graphs, where Graph Neural Networks (GNNs) offer an effective framework to model interactions in the tissue. Self-supervised pretraining methods such as Bootstrapped Graph Latents (BGRL) and GRACE leverage graph augmentations to build invariances without costly labels. Yet, the design of augmentation strategies remains underexplored, particularly in the context of spatial omics. In this work, we systematically investigate how different graph augmentations affect embedding quality and downstream performance in spatial omics. We evaluate a suite of existing and novel augmentations, including transformations tailored to biological variation, across two representative tasks: unsupervised domain identification in healthy tissue and supervised phenotype prediction in cancer tissue. Our results show that carefully chosen augmentations substantially improve performance, whereas poorly aligned or overly complex augmentations may fail to help or even degrade performance. These findings highlight the central role of augmentation design in enforcing meaningful invariances for graph contrastive pretraining in spatial omics.

bioinformatics↗

bioRxiv2025

H&Enium, Applying Foundation Models to Computational Pathology and Spatial Transcriptomics to Learn an Aligned Latent Space

Bridging the gap from transcriptomic to imaging data at single-cell resolution is essential for understanding tumor biology and improving cancer diagnostics. Spatial transcriptomics enables mapping gene expression onto H&E images of segmented single cells, but remains limited by cost and throughput. We introduce H&Enium, a contrastive alignment framework that projects image and gene expression embeddings from foundation models into an aligned latent space using projection heads and a novel soft alignment target. This alignment enriches image-derived embeddings with transcriptomic context improving downstream tasks such as cell type classification and gene expression prediction. Additional evaluations on independent pathology datasets demonstrate superior generalization of our aligned representations over unaligned baselines. Our method offers a scalable path to enhance the utility of standard H&E imaging in both research and clinical settings.

pathology↗

bioRxiv2025

Exosomal cargo genes as biomarkers and potential mediators of exosome-driven cell communication in Pleural Mesothelioma

Pleural mesothelioma (PM) is a rare yet aggressive and heterogeneous cancer type with very poor survival rates. Due to its long latency period and nonspecific symptoms, the disease is usually detected at advanced stages, limiting available treatment options and leading to poor survival rates. So far, the disease can only be confirmed through invasive thoracoscopic biopsy. Moreover, the proposed circulating protein biomarkers lack sensitivity and specificity, highlighting the urgent need for novel approaches. In our previous work, we characterized the transcriptomic profile of extracellular vesicles secreted by PM cells and demonstrated their feasibility as a source of circulating biomarkers. To further investigate the role of circulating RNA in tumor progression and its potential use in non-invasive testing for PM, here we present a detailed characterization of RNA cargo carried by PM exosomes - a distinct subset of extracellular vesicles known to serve as key mediators of intercellular communication. By utilizing primary cell cultures established from tissue samples of 11 PM and 6 non-PM patients, followed by exosome isolation, and total RNA sequencing of RNA from isolated exosomes, matching cells, and tissues, we provide new evidence on exosomal RNA secretion, regulation of PM-exosome cargo genes, and their potential functions in recipient cells. We show that PM-exosomal cargo is enriched in genes associated with proliferation, whose key transcriptional factors are SETDB1, FOXM1 and GATA2. We identified Pcbp2, Srsf1 and Srsf9, RNA-binding proteins, which may be involved in selective cargo sorting into PM-exosomes. Our analysis identified six secreted genes, including GAS5, AL031666.3, RSLD1D1, AC103740.2, ADAM10, and AC020892.1, whose exosomal expression was associated with patient survival, as promising biomarkers for patient diagnosis, stratification, and prognosis assessment. Finally, we identify potential target genes of candidate LncRNAs biomarkers and processes in which they are involved.

cancer biology↗

bioRxiv2025

High-Resolution Time-Lapse Imaging of Droplet-Cell Dynamics via Optimal Transport and Contrastive Learning

Single-cell analysis is essential for uncovering heterogeneous biological functions that arise from intricate cellular interaction. Microfluidic droplet arrays enable precise dynamic data collection through cell encapsulation in picoliter volumes. The time-lapse imaging of these arrays can reveal functional kinetics and cellular fates, but accurate tracking of cell identities across time frames remains challenging when droplets move significantly. Specifically, existing machine learning methods often depend on labeled data or require neighboring cells as reference; without them, these methods struggle to track identical objects across long distances with complex movements. To address these limitations, we developed a pipeline combining visual object detection, feature extraction via contrastive learning, and optimal transport-based object matching, which minimizes reliance on labeled training data. Our approach was validated across various experimental conditions and was able to track thousands of water-in-oil microfluidic droplets over large distances and long (> 30 min) time-separated frames. We achieved high precision in previously untraceable scenarios, tracking small, medium and large movements (corresponding to ~126, ~800 and ~10,000 {micro}m, respectively) with a success rate of correctly tracked droplets of > 90% for average movements within 212 object diameters, and > 60% for average movements of > 100 object diameters. This workflow lays the foundation for high-resolution, dynamic analysis of droplets and cells in both spatial and temporal dimensions without relying on visual labeling, allowing high-accuracy tracking in samples, where the uniqueness of the sample makes repeating experiments infeasible.

bioinformatics↗

bioRxiv2025

CDState: an unsupervised approach to predict malignant cell heterogeneity in tumor bulk RNA-sequencing data

Intratumor transcriptional heterogeneity (ITTH), defined by the coexistence of diverse cell states within one tumor, complicates cancer treatment by contributing to variable therapeutic responses. Although single-cell RNA sequencing can resolve this complexity, its cost and technical demands limit its large-scale use. Bulk RNA-seq data provide a scalable alternative, but most deconvolution methods depend on predefined references, restricting their ability to detect novel malignant states. Unsupervised approaches avoid these constraints but are not tailored to capture heterogeneity within the malignant compartment. To address these limitations, we introduce CDState, an unsupervised method for inferring malignant cell subpopulations from bulk RNA-seq data. CDState utilizes non-negative matrix factorization improved with sum-to-one constraint and a cosine similarity-based optimization to deconvolve bulk gene expression into distinct cell state profiles. We demonstrate robustness of CDState on bulkified single-cell RNA-seq datasets from five cancer types, showing that it outperforms existing unsupervised deconvolution methods in the estimation of both cell state proportions and gene expression profiles. Applied to 33 cancer types from The Cancer Genome Atlas, CDState reveals recurrent gene programs, including epithelial-mesenchymal transition, MYC targets, and oxidative phosphorylation, as major contributors to malignant cell ITTH. We further link malignant states to patient clinical features, identifying states associated with poor prognosis. We propose an intratumor heterogeneity index and show its association with patient survival, clinical characteristics, and therapeutic response. Finally, we identify mutations and copy number alterations in genes such as TP53, KRAS, PIK3CA, SOX2, and SATB1 as potential genetic drivers of malignant cell ITTH across cancer types.

bioinformatics↗

bioRxiv2025

Evaluation of deep learning approaches for high-resolution chromatin accessibility prediction from genomic sequence

Fine-grained prediction of chromatin accessibility from DNA sequence is a foundational step in modeling gene expression changes resulting from sequence variants. Yet, few methods operate at the resolution necessary to capture subtle effects of single-nucleotide changes. Furthermore, it remains unclear which architectural components--such as residual connections, normalization strategies, or attention mechanisms--drive performance in these high-resolution predictions. To address these knowledge gaps, we systematically evaluate classic architectural choices and introduce ConvNeXt V2 blocks, originally developed for computer vision, as high-resolution feature extractors in deep learning models for genomic data. Integrated into diverse architectures--CNNs, LSTMs, dilated CNNs, and transformers--ConvNeXt V2 blocks consistently improve performance, leading to similar prediction accuracy across these different model types. This reveals that early feature extraction, rather than downstream architecture, is the primary determinant of prediction accuracy. A comprehensive evaluation of these models on ATAC-seq signal prediction at 4 bp resolution in a cell type-specific manner identifies the ConvNeXtbased dilated CNN as the most robust performer, better preserving the signals shape. Our codebase and benchmarks provide practical tools for high-resolution chromatin modeling.

genomics↗

bioRxiv2024

UniversalEPI: Harnessing Attention Mechanisms to Decode Chromatin Interactions in Rare and Unexplored Cell Types

Enhancer-promoter interactions (EPIs) play a central role in gene regulation, but experimental techniques such as Hi-C for mapping these interactions remain costly and labor-intensive. Computational methods have been developed to predict EPIs in silico from DNA sequence and chromatin information; however, there are major challenges with the generalizability and accuracy of predictions by existing methods across cell types and conditions unseen during model training. We developed and validated UniversalEPI, an attention-based deep ensemble model that predicts EPIs up to 2 Mb apart using only DNA sequence and chromatin accessibility (ATAC-seq) data. Unlike models that reconstruct full Hi-C contact maps, UniversalEPI focuses on biologically relevant, sparse chromatin interactions between accessible regulatory elements. It generalizes across both bulk and single-cell ATAC-seq-derived pseudo-bulk datasets, delivering state-of-the-art performance while using fewer input modalities than existing approaches. By modeling predictive uncertainty, UniversalEPI enables statistically robust differential analysis of chromatin interactions across conditions. We demonstrate its utility by tracking dynamic EPIs during human macrophage activation and identifying regulatory differences between cancer cell states in esophageal adenocarcinoma. By providing precalculated Hi-C predictions for 157 ENCODE datasets, UniversalEPI expands the scope and applicability of in silico 3D genome modeling for studying gene regulation in development and disease.

genomics↗

bioRxiv2024

Benchmarking Self-Supervised Learning for Single-Cell Data

Self-supervised learning (SSL) has emerged as a powerful approach for learning biologically meaningful representations of single-cell data. To establish best practices in this domain, we present a comprehensive benchmark evaluating eight SSL methods across three downstream tasks and eight datasets, with various data augmentation strategies. Our results demonstrate that SimCLR and VICReg consistently outperform other methods across different tasks. Furthermore, we identify random masking as the most effective augmentation technique. This benchmark provides valuable insights into the application of SSL to single-cell data analysis, bridging the gap between SSL and single-cell biology.

bioinformatics↗

bioRxiv2024

CancerFoundation: A single-cell RNA sequencing foundation model to decipher drug resistance in cancer

We present CancerFoundation, a novel single-cell RNA-seq foundation model (scFM) trained exclusively on malignant cells. Despite being trained on only one million total cells, a fraction of the data used by existing models, CancerFoundation outperforms other scFMs in key tasks such as zero-shot batch integration and drug response prediction. During training, we employ tissue and technologyaware oversampling and domain-invariant training to enhance performance on underrepresented cancer types and sequencing technologies. We propose survival prediction as a new downstream task to evaluate the generalizability of single-cell foundation models to bulk RNA data and their applicability to patient stratification. CancerFoundation demonstrates superior batch integration performance and shows significant improvements in predicting drug responses for both unseen cell lines and drugs. These results highlight the potential of focused, smaller foundation models in advancing drug discovery and our understanding of cancer biology. Our code is available here1.

bioinformatics↗

bioRxiv2024

Filtering cells with high mitochondrial content removes viable metabolically altered malignant cell populations in cancer single-cell studies

BackgroundSingle-cell transcriptomics has transformed our understanding of cellular diversity in biological systems. However, systematic noise, often introduced by low-quality cells, can obscure biological signals if not properly accounted for. Thus, one of the common quality control steps involves filtering out cells with a high percentage of mitochondrial RNA counts (pctMT), as high pctMT typically indicates cell death. Yet, commonly used filtering thresholds, primarily derived from studies on healthy tissues, may be overly stringent for malignant cells, which often naturally exhibit higher baseline mitochondrial gene expression. We analyzed public single-cell RNA-seq and spatial data to investigate if malignant cells with high pctMT are viable and functionally significant subpopulations. ResultsWe analyzed nine single-cell RNA-seq datasets from uveal melanoma, breast, lung, kidney, head and neck, prostate, and pancreatic cancers, including 439,507 cells from 151 patients. Malignant cells exhibited significantly higher pctMT than nonmalignant cells without a significant increase in dissociation-induced stress signature scores. Malignant cells with high pctMT showed metabolic dysregulation, including increased xenobiotic metabolism, which is implicated in cancer therapeutic response. Our analysis of pctMT in cancer cell lines uncovered associations with resistance and sensitivity to certain classes of drugs. Additionally, we observed a link between pctMT and malignant cell transcriptional heterogeneity as well as patient clinical features. ConclusionsThis study provides a detailed exploration of the functional characteristics of malignant cells with elevated pctMT, challenging current quality control practices in single-cell RNA-seq analyses of tumors. Our findings have the potential to improve data interpretation and refine the biological conclusions of future cancer studies.

bioinformatics↗

bioRxiv2024

Cell states and neighborhoods in distinct clinical stages of primary and metastatic esophageal adenocarcinoma

Esophageal adenocarcinoma (EAC) is a highly lethal cancer of the upper gastrointestinal tract with rising incidence in western populations. To decipher EAC disease progression and therapeutic response, we performed multiomic analyses of a cohort of primary and metastatic EAC tumors, incorporating single-nuclei transcriptomic and chromatin accessibility sequencing, along with spatial profiling. We identified tumor microenvironmental features previously described to associate with therapy response. We identified five malignant cell programs, including undifferentiated, intermediate, differentiated, epithelial-to-mesenchymal transition, and cycling programs, which were associated with differential epigenetic plasticity and clinical outcomes, and for which we inferred candidate transcription factor regulons. Furthermore, we revealed diverse spatial localizations of malignant cells expressing their associated transcriptional programs and predicted their significant interactions with microenvironmental cell types. We validated our findings in three external single-cell RNA-seq and three bulk RNA-seq studies. Altogether, our findings advance the understanding of EAC heterogeneity, disease progression, and therapeutic response.

cancer biology↗

bioRxiv2023

ANS: Adjusted Neighborhood Scoring to improve assessment of gene signatures in single-cell RNA-seq data

In the field of single-cell RNA sequencing (scRNA-seq), gene signature scoring is integral for pinpointing and characterizing distinct cell populations. However, challenges arise in ensuring the robustness and comparability of scores across various gene signatures and across different batches and conditions. Here, we evaluated the stability of established methods such as Scanpy, UCell, and JASMINE in the context of scoring cells of different types and states. On eight cancer and healthy scRNA-seq datasets, we reported that none of the existing methods provide fair gene signature scores that can be used in unsupervised cell state annotation based on the highest signature values. Addressing this challenge, we introduced a new scoring method, the Adjusted Neighbourhood Scoring (ANS), that builds on the traditional Scanpy method and improves the handling of the control gene sets. We further exemplified the usability of ANS scoring in differentiating between cancer-associated fibroblasts and malignant cells undergoing epithelial-mesenchymal transition (EMT) in four cancer types, and evidenced excellent classification performance (AUCPR train: 0.95-0.99, AUCPR test: 0.91-0.99). In summary, our research introduces ANS as a robust and deterministic scoring approach that enables the comparison of diverse gene signatures and score-based annotation of cell types and states. The results of our study contribute to the development of more accurate and reliable methods for analyzing scRNA-seq data.

bioinformatics↗