Acquired treatment resistance in homologous recombination (HR)-deficient high-grade serous ovarian carcinoma (HGSOC) commonly involves restoration of HR, often through secondary somatic reversion mutations in HR genes such as BRCA1, BRCA2 and BRIP1. Recent work suggests that these resistance mechanisms are frequently subclonal, raising the question of how reversion-negative cells persist during treatment. To address this, we developed a single-cell sequencing approach that enables direct genotyping of individual cells and links genotype to transcriptomic phenotype. In a pilot experiment, we analysed two HGSOC cell lines with known reversion mutations by combining short-read whole-transcriptome single-cell RNA sequencing with long-read targeted single-cell cDNA sequencing on an Oxford Nanopore MinION using a 44-gene hybridisation capture panel. A custom bioinformatics pipeline incorporating Flexiplex resolved multiallelic events and annotated cell barcodes with driver, reversion and TP53 mutations. Genotypes were then transferred to the short-read dataset following outlier and doublet filtering, enabling differential expression analysis between cells with and without reversions. We have also applied this protocol to patient ascites and tumour samples collected before and after treatment. We successfully detected driver and reversion alleles at single-cell resolution in both cell lines. Rare wild-type heterozygous cells were also identified, suggesting back-reversion events or subclones that escaped loss of heterozygosity and would likely be missed by bulk sequencing. This approach enables detection of subclonal resistance mechanisms and characterisation of transcriptional programs associated with persistence of non-reverted cells in HR-deficient HGSOC.

Glioblastoma (GBM) remains the most lethal glioma, characterised by extensive epigenetic remodeling and dysregulation of super-enhancers (SE). SE-mediated cell fate determination is increasingly linked to transcribed superenhancer RNAs (seRNA); however, the specific functional roles of SE-transcribed lncRNAs, a distinct, stable subset of seRNAs, remain poorly understood in oncogenesis. To circumvent the scarcity of paired-normal brain tissue, we utilised population-scale Allele-Specific Expression analysis of public RNA-seq data (GBM and LGG) to identify candidate SE-lncRNAs with significant functional imbalance. We performed biophysical assessments of top recurrently-mutated SNVs in 19 candidates (14 GBM, 5 LGG) using structural prediction (RNAsnp, ViennaRNA), triplex-mediated DNA-binding modeling (Fasim-LongTarget, PATO), and subcellular localisation algorithms. Our pipeline recovered known drivers alongside novel unannotated loci. TMEM44-AS1 emerged as a uniquely significant driver in GBM. While triplex modeling showed no change in theoretical DNA-binding affinity, RNAsnp analysis revealed a GBM-specific SNV in TMEM44-AS1 (position 243) triggering a significant structural phase-transition (d=0.173, p=0.0387). We observed increased ensemble diversity and "local melting" characterised by a flattened centroid curve indicating high conformational frustration. Delta-accessibility profiling revealed a localised redistribution around the SNV, including a deep accessibility trough within the 245–261 region which potentially masks critical hnRNP-binding motifs. These results suggest that progression from LGG to GBM may be associated with mutational "structural switches" that occlude regulatory protein-binding sites, warranting further analyses and experimental validation of remaining high-priority candidates. Our findings provide a mechanistic outline for how non-coding mutations disrupt SE-lncRNA interactomes, offering a generalisable computational framework for understanding glioma malignancy.

Congenital heart disease (CHD) is the most common structural birth defect in humans, affecting around 1% of live births. While mutations in cardiac genes contribute to CHD, many disease-associated nucleotide variants map to noncoding genomic regions (including transcriptional enhancers). These regulatory elements play a crucial role in heart development by controlling gene expression, yet only a small fraction of heart enhancers have been experimentally validated.
As part of the HeartX research program, we aimed to further elucidate the regulatory mechanisms of CHD etiology by applying machine learning models to predict heart enhancers based on epigenetic and sequence features. As a training set, we used about 4,000 genomic elements tested for enhancer activity in vivo in transgenic mice (VISTA Enhancer Browser). Our model accurately distinguished heart enhancers from other elements, enabling genome-wide heart enhancer prediction. To refine these predictions, we integrated single-cell ATAC-seq chromatin accessibility data from embryonic mouse hearts, further augmented by newly generated data from HeartX cardiac organoids.
In addition to enhancer identification, we investigated the functional impact of noncoding variants associated with CHD. Using published deep learning methods, we compared wild-type and mutant regulatory sequences to identify potentially pathogenic changes. This approach allows for the prioritization of disease-associated regulatory variants and enhances our understanding of CHD etiology.
To facilitate further research, we have developed a publicly accessible web application that enables users to explore genome-wide heart enhancer predictions and evaluate the functional effects of noncoding variants.

Sequence-to-activity models are widely used in regulatory genomics to predict the activity of gene regulatory regions and the effects of non-coding variants. A general assumption is that, with a growing context window, these models model not only the region of interest but also the impact of the surrounding nucleotide context. Accordingly, a major trend in model development has been to increase input window sizes to 200 kb (Enformer), 500 kb (Borzoi), and even 1 Mb (AlphaGenome). However, growing evidence suggests that these models struggle to predict the effects of distal mutations and to capture long-range regulatory interactions, motivating systematic benchmarking.
Massively parallel reporter assays (MPRAs) provide a framework for this purpose, as MPRAs usually evaluate the activity of regulatory regions or regulatory mutation effects outside of the native genomic environment, enabling a direct test of model sensitivity to sequence context. Specifically, we evaluate whether models achieve better predictions if provided with the real sequence context (e.g., plasmid sequences) and whether they remain consistent across different contexts, including those deliberately mismatched including shuffled sequences, centromeric regions and unrelated regulatory elements).
We benchmark seq-to-activity models across multiple MPRA datasets (Kircher et al., 2019; Agarwal et al., 2025; Barbadilla-Martínez et al., 2026; Siraj et al., 2026). Surprisingly, providing the real context in general decreases performance for many models, excluding ChromBPNet, whose performance was context-agnostic, even in the centromeric context. Our findings suggest that the current models fail to reliably capture long-range interactions and likely underutilize or even misinterpret the local sequence context.

A recurring claim in the literature on lncRNAs is that a sizeable fraction of lncRNAs are highly specific to individual cell types (or a small number of cell types), whereas mRNAs tend to be more ubiquitously expressed. Evaluating this claim requires a quantitative measure. However, there is currently no consensus on how to define and measure cell type specificity of RNAs. We propose an adaptation of the Gini index, originally developed in economics to quantify income inequality among individuals in a population, and use it to compare the cell type specificity of lncRNAs and mRNAs in single-nuclei RNA seq data from mouse kidney. In contrast to previous work, where Gini indices are calculated from expression values of individual cells, we propose using values aggregated at the level of cell types or clusters. We derive Gini indices either from averaged expression levels or from the proportion of cells expressing a gene within each cell type, reflecting related but distinct notions of cell type specificity. We find a strong negative correlation between cell type specificity and mean expression level, with highly specific genes generally exhibiting low expression. After stratifying genes by expression level, differences in average Gini index between lncRNAs and mRNAs within each stratum are small, and in the lowest expression strata negligible. We address concerns about inflated Gini indices for lowly expressed genes and argue, based on permutation tests, that our adaptation remains informative across the full range of expression levels, including lowly expressed RNAs.

Male puberty involves extensive molecular reprogramming of reproductive tissues; however, the transcriptional and epigenetic changes during this process remain poorly characterised in cattle (Bos taurus). We present a multi-omics study of testis tissue from 113 prepubertal and 123 postpubertal Braunvieh bulls, integrating short-read RNA-seq for transcriptome profiling and PacBio HiFi whole-genome sequencing for variant detection and 5mC DNA methylation calling.
We identified 10,723 differentially expressed genes (DEGs), predominantly with increased expression postpubertally and enriched for reproductive processes. We characterised 7,535 differentially spliced genes (DSGs), of which 32.8% were also DEGs. This suggests that alternative isoform usage only partially explains the observed changes in gene expression. We detected 3,341 differentially methylated regions (DMRs), 91.4% of which were hypermethylated in prepubertal individuals. These regions were enriched for genes involved in reproductive processes and exhibited higher expression postpubertally, supporting the established inverse relationship between DNA methylation and gene expression.
Using over 18 million sequence variants, we detected 2- to 3-fold more eQTL and sQTL in postpubertal animals (12,322 eQTL; 6,725 sQTL) than prepubertal animals (5,516 eQTL; 2,390 sQTL), indicating enhanced genetic regulation of testicular gene expression after puberty onset. However, meQTL were balanced between age groups (449 vs 494 meQTL).
Integrating these layers, we identified 44 postpubertal-specific eQTL located within DMRs hypomethylated in postpubertal animals, including spermatogenesis-associated genes SYCP3 and ZSCAN2. This suggests that epigenetic remodelling facilitates the emergence of genetic regulatory effects during testicular maturation.
Our findings provide a multi-layered transcriptomic and epigenomic map of bovine puberty with implications for fertility.

Duchenne Muscular Dystrophy (DMD) is a genetically inherited, X-linked neuromuscular disease caused by mutations in dystrophin encoding DMD gene. It is characterized by progressive degeneration, fibrosis, and fatty replacement driven in part by fibro-adipogenic progenitors (FAPs), a cell population that supports regeneration but becomes pathogenic when dysregulated. Despite being essential stromal regulators of skeletal muscle homeostasis, their state transitions and spatial organization in neuromuscular disease remain poorly defined. Here, we developed a multi-step computational framework to study this gap in DMD by integrating single-nuclei RNA sequencing and 10x Visium spatial transcriptomics. We first analysed control and DMD snRNA-seq datasets separately, subsetted FAPs, and re-clustered them to define transcriptionally distinct states. Using the DMD single-nuclei atlas as a reference, we deconvoluted the Visium sections with cell2location and then applied non-negative matrix factorization to identify co-occurring cellular niches across the tissue. Neighborhood enrichment analysis at multiple spatial scales further quantified niche adjacency and segregation. This workflow resolved discrete FAP states, including inflammatory FAPs, MME⁺ FAPs and pre-adipogenic FAPs and mapped them to spatially distinct tissue domains. FAP-rich niches formed coherent spatial clusters and showed selective proximity to other cellular neighborhoods, whereas myofiber-dominant niches remained segregated across the section. Finally, a ligand-receptor analysis suggested that the FAP-rich niches act as active signalling hubs and drive neighboring cell recruitment in these microenvironments through immune and extracellular matrix signalling. Together, our workflow links FAP state diversification to spatial tissue organisation and communication, providing a computational framework for dissecting stromal remodeling in DMD.

Pooling patient samples for single-cell sequencing substantially reduces costs and batch effects, but requires reliable demultiplexing and other preprocessing steps for accurate biological interpretation. To address this, we developed Split-flow, a concordance-based Nextflow pipeline that integrates orthogonal demultiplexing strategies. We applied Split-flow to multiplexed single-nucleus multiome (snRNA-seq + snATAC-seq) data from patients with acute myeloid leukemia (AML) treated with the hypomethylating agent decitabine (DAC) and all-trans retinoic acid (ATRA) in the DECIDER trial. Incorporating ATAC-based demultiplexing further resolved RNA-based donor assignment discordance and increased per-donor cell recovery. Within the resulting demultiplexed cells, AML leukemic cells were identified by integrated RNA and ATAC somatic variant calling, and eight lineage-based cell states were resolved, including stem cell-like, progenitor-like, erythroid-like, and monocyte/macrophage-like populations whose proportions shifted markedly upon treatment. To characterize the complex AML blast response to treatment, we applied CellRank trajectory analysis, multi-omics factor analysis, and a newly developed approach to assess chromatin-transcription coherence as a proxy for cellular plasticity. This analysis revealed blast-state-specific depletion and emergence trajectories, coordinated transcriptomic and chromatin changes across treated patients, and treatment-associated shifts in regulatory coherence. Collectively, these results suggest that blast-state-specific therapy responses are governed by coordinated multi-layered epigenetic and transcriptional dynamics. They further establish Split-flow as a generalizable concordance-based preprocessing framework for multiplexed single-cell sequencing, with applicability beyond AML.

It is well known that alternative splicing is a crucial mechanism of post-transcriptional regulation, yielding mRNA variants with distinct biological functions. In prostate cancer (PCa), specific variants, such as the androgen receptor variant AR-V7, play a critical role in promoting disease progression and resistance to targeted therapies. However, large-scale transcriptomic analyses often face severe cross-study batch effects, limiting true transcript-level resolution. In the present study, we aim to systematically investigate hidden isoform-level regulations and map their specific correlations with oncogenic hallmark pathways.

To achieve this, we constructed a unified RNA sequencing atlas containing 1,365 clinical samples from 14 independent cohorts, uniformly re-analyzed to achieve an unprecedented level of data integration. Using the nf-core/rnaseq pipeline, we identified that excluding single-end sequencing libraries is crucial; subsequent transcript-level principal component analysis (PCA) showed that PolyA and Total RNA libraries were demonstrated to be well mixed without persistent batch effects.

Disease progression was modeled using pseudotime trajectory inference to strictly evaluate the relationship between isoform ratio/abundance and disease progression. We developed a novel Abundance-Weighted Splicing Index (AWSI) to evaluate the impact of isoform conversions on hallmark pathways. Furthermore, we successfully identified significant isoform switches strongly correlated with disease progression in key coding genes, for example ROCK2, as well as TRIM11, RNF43, and RELA. AlphaFold models and clinical data confirmed these splicing events cause major 3D structural changes and affect patient survival. For long non-coding RNAs (lncRNAs), mapping transcript structures to disease progression scores demonstrated that specific retained exons and k-mer motifs largely determine their pathological functions as PCa drivers or suppressors.

Taken together, our study highlights the necessity of rigorously excluding library-layout batch effects to achieve true transcript-level resolution. We demonstrate that isoform switches, primarily determined by specific exons and k-mer motifs, play critical roles in regulating PCa progression. All harmonized data and visualization resources are available at https://prostatecanceratlas.org.

The generation and maintenance of differentiated cell types in the intestine depends on genetic networks that remain poorly understood. The Drosophila midgut is a powerful model for studying these regulatory circuits. Its biology is conserved across species, and its genetic accessibility allows targeted perturbation of individual genes at scale.
Our research investigates how intestinal stem cells (ISCs) give rise to distinct cell types in the Drosophila midgut. We use single-cell transcriptomic data from in vivo CRISPR knockdown experiments targeting ISCs in healthy adult animals. Our goal is to explore how single gene perturbations affect cellular behaviour across the tissue. The dataset comprises scRNA-seq profiles of ~620 000 cells from 127 replicate experiments covering 49 unique genetic perturbations.
We perform an exploratory analysis to uncover molecular relationships between perturbation conditions. We group conditions based on similarity in cell type abundance shifts, particularly among differentiated enterocyte populations, identifying 6 phenotypically coherent clusters. For each cluster, we examine expression changes to detect shared molecular responses. We assess common pathways, transcription factors, and differentially expressed genes across perturbations and cell types to characterize convergent regulatory signatures in the intestinal tissue.
This analysis reveals candidate genetic network modules and generates testable hypotheses about how distinct perturbations converge on shared cellular outcomes.

Motivation: Many biological and clinical processes are dynamic, yet most datasets are cross-sectional, capturing populations at discrete states rather than tracking individuals over time. This makes it difficult to quantify how populations change across developmental, physiological, or disease-associated conditions. Existing trajectory and transport-based methods often rely on fixed feature spaces, assumptions tailored to transcriptomic time-course data, or approximately linear progression, limiting their ability to model heterogeneous and unbalanced transitions across diverse biomedical modalities. Flexible methods are needed that can infer directed population-level change from cross-sectional data while retaining biological interpretability.
Results: We present GOTFlow, a framework for learning directed population transitions from cross-sectional biomedical data using graph-constrained optimal transport in a learned latent space. GOTFlow integrates representation learning with unbalanced optimal transport to jointly estimate embeddings and transport couplings between biological states. This enables hypothesis-driven modelling of progression structures while accommodating non-linear geometry, branching relationships, and changes in population mass. From the inferred transport plans, GOTFlow derives interpretable summaries of dynamics, including drift vectors quantifying transitions, and feature-level transported changes that highlight molecular drivers of progression. In synthetic data, GOTFlow recovered known transitions with strong agreement between inferred and ground-truth drifts. Across three biological applications, endometrial remodelling, breast cancer risk progression, and prion disease, GOTFlow identified state-to-state transitions and biologically meaningful feature shifts reflecting impaired decidualisation, increasing cancer risk, and neurodegenerative progression. These results establish GOTFlow as a general and interpretable framework for analysing directed population dynamics from cross-sectional data.
Availability: Code available at: https://github.com/wgrgwrght/GOTFlow
Supplementary information: Available online.

Psoriasis is a chronic inflammatory skin disease involving coordinated remodeling across epithelial, stromal, and immune compartments. Although multiple single-cell RNA-sequencing studies of psoriatic skin have been reported, cross-study comparison remains limited by technical heterogeneity and inconsistent cell-type annotation. Here, we constructed a large-scale multi-study harmonized single-cell atlas of psoriatic skin, comprising approximately 270 samples from 126 donors across ten studies and spanning healthy, non-lesional, lesional, and treatment-associated conditions. We first established a core reference atlas from six studies representing untreated healthy, non-lesional, and lesional skin, and subsequently mapped four additional treatment cohorts spanning pre- and post-treatment samples onto this reference. To identify a robust integration strategy, we benchmarked six methods, including scVI, scANVI, scPoli, Harmony, Scanorama, and scGen, using metrics of biological conservation and batch correction. scGen showed the most favorable balance between biological conservation and batch mixing in our benchmark and was therefore selected for downstream analyses. To overcome discordant annotations across studies, we harmonized cell-type labels and performed hierarchical re-annotation at broad and fine-grained levels, establishing a unified label space for cross-study comparison. This atlas enabled systematic comparison of cell-type composition across disease states and revealed disease-associated shifts in immune and non-immune compartments, including lesion-associated enrichment of Th17/Tc17 subsets. We further quantified the fraction of inter-sample transcriptional variance explained by biological and clinical covariates in a cell-type-specific manner. Across major cell populations, anatomical location, tissue compartment, and treatment exposure explained substantial transcriptional heterogeneity, highlighting the importance of local tissue context and clinical state in multi-study analyses. Pseudobulk differential expression and pathway enrichment analyses identified psoriasis-associated inflammatory programs, including interferon and antiviral signaling, antigen presentation, chemokine signaling, and stromal remodeling signatures. Together, our atlas provides a reference-based framework for multi-study single-cell integration and a harmonized resource for exploring psoriasis-associated cellular and molecular remodeling.

Inference of cell-cell communication has become a standard approach in single-cell transcriptomics. However, as datasets grow in scale and complexity, researchers face the increasing challenge of distinguishing true biological signals from technical noise. Multi-condition experimental designs add a further layer of difficulty, as standard tools often struggle to isolate condition-specific shifts while accounting for sample-to-sample variability. While many models have been proposed, some lack biological interpretability, functioning as opaque “black boxes,” while others fall short in statistical rigor, consequently, there remains a need for transparent, biologically meaningful, and statistically robust frameworks.



To address these challenges, we built upon CellPhoneDB, a well-known tool for inferring interactions based on curated ligand-receptor pairs and cell cluster's mean expression. We extend this framework by integrating it with the statistical power of pseudo-bulk differential expression analysis (DEA) by DESeq2. By aggregating expression at the sample level, our approach ensures that candidate ligands and receptors are consistently differentially expressed across biological replicates. We refined the standard CellPhoneDB interaction scores by weighting them with LOG2FC results from the DEA. This allows for a conservative ranking of interaction dynamics.

We applied this workflow to a multi-sample, multi-condition dataset, where our approach effectively pruned the interaction space by discarding over 80% of the candidate signaling shifts that lacked support by DEA. By ensuring that the prioritized interactions were driven by robust gene-level shifts, this pipeline provides a reproducible and interpretable framework to prioritize the most responsive interactions for targeted follow-up studies.

Decades of molecular biology research have shed light on the alterations that gene regulatory networks undergo, influencing cancer progression and outcomes. While sample-specific gene regulatory networks (ssGRNs) allow us to map the gene regulatory landscape for individual patients, their ability to predict clinical outcomes has not been fully explored. In this project, we aim to predict Acute Myeloid Leukemia (AML) patient treatment response and outcomes by developing robust statistical models that incorporate information derived from ssGRNs. Here, we demonstrate the generalization capabilities of ssGRNs across 120 different prediction tasks, using various network-building strategies suitable for this purpose in a 5-fold cross-validation setting. We show that in some cases, despite network metrics not showing an increase in predictive signal overall, crucial regulatory drivers are highlighted by the network features and not by the expression modality. For example, TP53, a transcription factor known to be associated with venetoclax response, is highlighted as a stable predictor in the network data but not in the expression data. This suggests that network-based predictive modelling could complement gene expression baselines through added interpretability of the regulatory mechanisms behind drug response.

African Swine Fever (ASF) is a high mortality viral disease in pigs, currently leading to high losses in pig populations and a crisis in pork farming worldwide. The pathogen, ASFV, is a large double-enveloped, double-stranded DNA virus with a strong tropism for macrophages, causing immune dysregulation and cytokine storm.
To elucidate the underlying mechanisms on how the virus infects macrophages and subsequent dynamics that lead to host death, we sequenced immune cell populations at a single-cell resolution (scRNA), with a focus on macrophages, monocytes, and dendritic cells from the spleen of nine infected pigs, over three time points - from day 0, until pig death at day 5.
To observe the progression of infection over time, we focussed on the on the trajectory and profiles of macrophage subsets. However, viral gene expression at day 5 amounted to upwards of 92% of total reads per cell, making their origin indistinguishable and leading to the clustering of all infected cells into one group. In order to identify the origin of infected macrophages in this cluster, we used linear regression approaches (elastic net) to define their unique transcriptomic signature before infection and allowing us to characterize specific infected sub-populations and their likely origin. Tissue-resident macrophages were shown to be strongly depleted over time, and monocyte-derived macrophages showed a cycle of recruitment, differentiation into macrophages, infection and depletion. In addition, we show the progression of the innate immune response in infected and bystander cells and their cell specific signatures, elucidating infection response mechanisms.

Cell-cell interactions are essential for maintaining tissue homeostasis, and their dysregulation is closely associated with the development of a wide range of diseases, including cancer and autoimmune disorders. Therefore, a comprehensive understanding of the underlying molecular mechanisms is critical for the development of novel therapeutic strategies.
Ligand proteins, which play a central role in cell–cell communication, bind to cell surface receptors and activate intracellular signaling pathways. While interactions between neighboring cells have been extensively validated experimentally, the spatial range over which secreted ligands exert their effects remains insufficiently understood.
In this study, we first used Visium spatial transcriptomics data to model the spatial diffusion of ligands. Assuming a Gaussian distribution, we constructed a model to describe the relationship between the spatial spread of ligand signals and target gene expression. By setting target gene expression levels as the response variable and estimating the diffusion distance parameter, we estimated the effective range of each ligand.
Furthermore, by utilizing high-resolution Stereo-seq data, we enabled analysis at the single-cell level. This advancement has made it possible to elucidate in greater detail the spatial dynamics of intercellular signaling and the influence of the microenvironment on signaling pathways.
Our analytical framework successfully detected known interactions between plasmacytoid dendritic cells (pDCs) and their neighboring cells. Furthermore, we present examples of this method applied to both normal and tumor tissues.

Functional or Pathway Enrichment Analysis (FEA or PEA) is a cornerstone of modern genomics, essential for providing biological meaning from high-dimensional omics datasets. Biological insights on regulation of pathways like cellular metabolism, signaling and immune responses can be obtained by FEA. Despite its ubiquity, with over 170,000 related publications from 2022 to 2026, the reliability of FEA remains compromised by widespread methodological flaws. This indicates that the reproducibility crisis is prevalent in bioinformatics and genomics. While FEA has become a routine part of workflows via myriad software packages and easy-to-use websites, mistakes can easily creep in due to poor tool design and unawareness among users of pitfalls. Some critical issues related to statistically flawed and lower reproducibility of enrichment studies include the lack of context-specific background correction in gene list analysis, lack of p-value correction when carrying out parallel testing, and a general lack of methodological details. Critical examination of hundreds of published articles describing the use of FEA, including articles from respected high-impact journals like Nature Communications and Nucleic Acid Research, had high error rates for essential parameters. In our article, we outline the top ten mistakes that undermine the effectiveness of FEA, which we have commonly observed in published research articles. We also share practical solutions that can be implemented to mitigate non-reproducible methodologies. We aim to tackle the reproducibility crisis in the life sciences sector and subsequently improve the statistical rigour and biological validity of enrichment analysis results in life sciences research.

Gene expression is regulated by multiple interacting mechanisms, including DNA methylation, genetic variation, and alternative splicing, whose effects are often distributed across many weak and correlated elements. These regulatory influences are highly context-dependent, yet most studies rely on pairwise association testing to identify cis-regulatory elements, limiting the characterization of gene-specific regulatory architecture across cancer states.

We introduce GeneCL (Gene-anchored Contrastive Learning), a representation learning framework that models genes and their associated regulatory elements as nodes in a heterogeneous graph constructed from statistically supported associations. GeneCL learns context-specific embeddings using a contrastive objective, where proximity reflects shared regulatory context rather than marginal associations. Applied to primary tumors from The Cancer Genome Atlas (TCGA) and metastatic tumors from the Hartwig Medical Foundation, using matched RNA-seq and WES/WGS datasets, GeneCL captures gene–anchored relationships and enables systematic characterization of gene-level regulatory architecture across cancer contexts.

GeneCL provides a scalable and extensible framework for learning context-specific regulatory architecture beyond cis-regulatory for integrating and comparing multiple molecular modalities.

Spatial transcriptomics (ST) enables the linking of gene expression profiles with spatial coordinates. In particular, imaging-based methods such as CosMx or MERSCOPE offer subcellular resolution, thereby enabling the investigation of a wide range of biological questions. However, these methods are inherently limited to targeted marker panels, leaving crucial markers for downstream analyses such as cell type annotation unmeasured.
Various approaches have been developed to close this information gap, broadly falling into two categories: computational prediction of missing gene expression on the one hand, and reference-based mapping of single-cell RNA sequencing (scRNA-seq) data onto ST datasets on the other. However, both strategies have shown only limited success, as the inherent technical differences between scRNA-seq and ST assays restrict reliable information transfer. So-called foundation models, pre-trained on large amounts of biological data, offer a promising alternative. Unlike the previously mentioned approaches, foundation models learn robust, generalizable transcriptomic representations. In the field of single-cell research, such models have already demonstrated their effectiveness. Recently, models trained on combined SC- and ST-data have emerged (Nicheformer, scGPT-spatial, sCT), whose broad transcriptomic representations effectively compensate for the restricted marker panels of imaging-based ST assays.
We demonstrate our approach on a complex dataset spanning multiple conditions and various tissue samples, where previously established methods have shown limited success. By combining a foundation model with targeted fine-tuning via Low-Rank Adaptation on a small subset of the original data, we extract enriched embeddings that serve as base for an informed clustering, yielding more structured clusters and an improved downstream analysis.

Diabetes is a chronic metabolic disorder affecting approximately 589 million adults worldwide; it is characterised by persistently elevated blood glucose levels due to impaired insulin production, utilisation, or sensitivity. Type 2 diabetes mellitus (T2DM) is the most prevalent form, and is defined by beta cell dysfunction and insulin resistance. Single-cell transcriptomics offers valuable insight into T2DM disease mechanisms at cell-type resolution; however, analysing these datasets typically requires considerable bioinformatics expertise. To lower this barrier, we developed T2DIAbetes, an intuitive, web-based application designed to accelerate research by making complex single-cell data accessible to a broader scientific community.

T2DIAbetes consolidates multiple publicly available single-cell RNA sequencing (scRNA-seq) datasets, each pre-processed to generate UMAP embeddings for interactive visualisation. Users can examine how T2DM affects distinct cell populations across tissues, with datasets further integrated into tissue-specific atlases to support targeted investigation. To explore disease mechanisms, the application enables comparison of gene expression levels and cell-type proportions between control and T2DM conditions. Key features include visualising the expression of user-selected genes, examining cell-type distribution shifts between conditions, and identifying differentially expressed genes across cell clusters.

Future development will extend the platform to support user-uploaded datasets, leveraging the integrated atlases to refine cell-type annotations, enable cross-sample comparisons, and map cell-to-cell signalling networks. Together, these capabilities will deepen our understanding of T2DM pathophysiology at single-cell resolution.

Single-cell RNA sequencing has transformed our understanding of tumour microenvironment (TME), yet cell type abundance alone cannot capture spatial organisation governing tumour-immune interactions. Tumour-proximal zone, the interface for tumour-stroma and tumour-immune interactions, is the site of immune exclusion and infiltration. Its spatial architecture determines anti-tumour immune response, disease progression, and therapy outcomes. Characterising cell type abundances at this interface, and how proximal zone composition shifts across disease states requires spatial transcriptomics at single-cell resolution, paired with statistical frameworks that distinguish boundary-proximal architecture from background TME.
Existing methods designed for whole tissue slides present challenges when applied to tissue microarray (TMA) cores. TME structures such as tertiary lymphoid structures with spatial patterns independent of tumour proximity, thus confounding boundary-homing biology in TMA cores. To overcome this, we developed MARGIN, a computational framework for spatial enrichment analysis at the tumour-stroma interface. MARGIN computes per-cell distances to detected tumour patches, and quantifies proximal zone enrichment using permutation specific for each TMA core to derive the null distribution by shuffling cell type labels while keeping spatial coordinates fixed. This implicitly captures tissue-specific features including cell density before statistical testing.
We applied MARGIN to 232 cores from 142 melanoma patients to show that the melanoma-stroma boundary consists of two major tumour-associated macrophage populations (TAMs): IFN+ and SPP1+ TAMs. The latter is consistently enriched intratumourally and at the tumour-stroma interface, with transcriptional programs linked to matrix remodelling, angiogenesis, and T-cell exclusion. Therefore, MARGIN provides a principled, scalable approach to spatial analysis at the tumour-stroma boundary.

Transcriptional noise, the stochastic variation in gene expression among cells of the same type, arises from the inherent randomness of transcription, chromatin dynamics, and molecular interactions. This variability encodes information about the state and stability of gene regulatory networks and has been associated with aging and loss of cellular identity. However, how transcriptional noise is altered across Alzheimer's disease progression and whether regional- and cell-type-specific alterations occur remains largely unexplored.
Here we address these questions by assessing transcriptional noise across multiple integrated single-nucleus RNA-seq datasets encompassing several brain regions, Braak-Tau stages, and independent patient cohorts. We apply robust statistical frameworks to obtain reliable estimates while accounting for the structure of the data.
Our results reveal that transcriptional noise changes with disease progression in a region- and cell-type-specific manner, with distinct patterns observed across glial and neuronal populations. Key associations replicate across independent datasets and brain regions, supporting their biological relevance.
By integrating multiple datasets and analytical approaches, this work provides a characterization of how transcriptional stability is reshaped in AD, offering a complementary view to traditional differential expression that highlights cell-type-specific vulnerability signatures in neurodegeneration.

Background: Target validation in inflammatory cardiomyopathy requires more than bulk expression evidence: a credible target should be localized to pathology-relevant regions, track with disease stage, and respond appropriately to therapeutic intervention. Imaging-based spatial transcriptomics offers a framework to resolve all three axes simultaneously. We developed an integrated Xenium-based pipeline to contextualize candidate target genes across the inflammatory-to-fibrotic progression of experimental autoimmune myocarditis (EAM) and demonstrate its utility as a target-validation platform.
Methods: We applied 10x Genomics Xenium in situ sequencing (488-gene panel) to 20 murine hearts spanning Healthy, EAM, Placebo, and Treated arms across three timepoints (D14, D21, D42). Cells were segmented with Baysor, and lesions were defined transcriptionally from inflammatory and fibrotic marker activity and correlated with paired histology. The pipeline integrates pseudobulk differential expression (edgeR, dual-tier thresholds), k-nearest-neighbour cell–cell colocalization, niche composition, and compositional forest plots to profile each candidate target across disease stage, tissue region, cellular source, and treatment response.
Results: The pipeline successfully distinguished candidate targets by their temporal dynamics (acute versus chronic), lesion-specificity, cellular localization (immune vs stromal vs cardiomyocyte), and treatment reversibility. Contrasting target profiles emerged – some normalized under anti-inflammatory treatment while others persisted – clarifying which targets map onto composition-reducing versus tissue-remodelling therapeutic hypotheses.
Conclusions: Xenium spatial transcriptomics, coupled with this integrated analytical framework, provides a powerful target validation tool that contextualizes candidate genes across the inflammatory-fibrotic continuum of cardiac disease.

Tumour hypoxia promotes increased proliferation, migration, and invasion of cancer cells. Expression of the hypoxia-induced enzyme carbonic anhydrase IX is associated with an unfavourable prognosis in breast cancer patients, highlighting the need for novel chemotherapeutic strategies to enhance the sensitivity of hypoxic tumour cells to chemoradiotherapy. Notably, combinations of triterpenes with carbonic anhydrase inhibitors (CAIs) have been shown to induce high cytotoxicity, apoptosis, and radiosensitization in breast cancer cells.

Here, we perform RNA-seq experiments to investigate the transcriptional response of the triple-negative breast cancer cell line MDA-MB-231 to a novel betulinic-acid-based sulfamate (CAI3) in a quantitative, fold-change-specific manner. We find that CAI3 simultaneously orchestrates a broad spectrum of biological pathways, including mitosis, DNA replication and repair, and chromatin regulation. The coordination principles underlying this complex orchestration remain poorly understood, but we observe that (i) genes associated with the same pathway respond to CAI3 with remarkably similar fold changes and that (ii) these fold changes differ systematically between pathways.

These findings suggest the existence of a transcriptional logic that governs the coordinated, pathway-specific regulation of gene expression in a fold-change-dependent manner. To gain initial insight into this coordinated response, we developed a novel pathway-enrichment approach that identifies pathways shared by groups of genes exhibiting similar response magnitudes, e.g., pathways associated with weakly, moderately, or strongly differentially expressed genes. Applying this method to CAI3-treated MDA-MB-231 cells, we identify numerous pathways enriched in a fold-change-specific manner, several of which remain undetected by conventional pathway enrichment approaches.

The evolution from low grade to metastatic tumour is a major determinant of cancer mortality. Cancer evolution involves a complex interplay between intrinsic genetics and transcriptional alterations and the external microenvironmental factors. To define mechanisms underpinning metastatic development, we focused on metastatic castration-resistant prostate cancer (mCRPC) and employed single-cell multi-omics and whole-genome sequencing to deeply profile 34 metastatic lesions from 9 patients by rapid autopsy. We found that intra-tumour heterogeneity is an indicator of key evolutionary processes, characterised by recurrent tumour populations acting as critical functional components of the tumour ecosystem, irrespective of clonal and microenvironmental backgrounds. Unexpectedly, microenvironments only played a limited role while clonal evolution primarily promoted transcriptional noise. Intra-patient functional convergence of tumour ecosystems was observed across metastases, showing system-level selection pressures that drive the heterogeneity landscape of mCRPC. Our findings reveal functional evolutionary convergence of metastatic disease into units of intra-tumour heterogeneity, identifying critical determinants for therapeutic targeting.

Adrenergic (ADRN) and mesenchymal (MES) are two major transcriptional states of Neuroblastoma cells. Transitions between these states play a critical role in drug resistance and tumor progression. Notably, tumor cells can switch between ADRN and MES identities without genetic alterations, indicating that these transformations are driven by epigenetic and transcriptional reprogramming. However, existing network-based and perturbation analyses often overlook the dynamic and heterogeneous nature of this process.

Here, we integrated public snRNA-seq and snATAC-seq datasets from six neuroblastoma cell lines (~50,000 cells, 135,000 peaks). Our analyses revealed inter-tumor heterogeneity and a continuous trajectory from ADRN to MES states, marked by decreasing ADRN and increasing MES signatures. Network inference and in silico perturbation analyses highlighted candidate transcription factors (TFs) that stabilize either the ADRN or MES states. Ongoing work aims to refine this regulatory network, distinguish tumor-specific regulation from normal sympathoadrenal development, and explore strategies to reprogram malignant cells toward stable, therapy-sensitive states.

Single-cell and spatial transcriptomics technologies have transformed the study of biological systems and disease mechanisms. One challenging step in analyzing such high-dimensional data is to accurately identify cell types within a new dataset. Most approaches rely on comparing cells to collections of annotated references; however, collecting and integrating these heterogeneous reference datasets is time-consuming.

Here, we present a structured database designed to centralize and manage diverse reference datasets, originating from bulk, single-cell, and spatial transcriptomics technologies. These datasets can come from public repositories or controlled-access sources.

Building on this foundation, we have developed a suite of advanced tools that:
(i) Harmonize free-text annotation of reference datasets using ontologies and controlled vocabularies
(ii) Enhance data accessibility and usability, notably by providing intuitive search capabilities and visualization
(iii) Suggest to the users the most accurate reference datasets for annotating new experiments

These developments improve the process of reference-based cell type annotation of scRNA-seq data by improving dataset discoverability and interoperability. For the research community this addresses a key bottleneck in many projects using single-cell data.

RNA post-transcriptional modifications strongly influence the context-dependent processes dictating RNA localization, translation and organelle interactions. Yet, large-scale experimental mapping of RNA modifications is infeasible, in particular when considering different cellular contexts. Accurate and scalable computational methods for inferring potentially modified sites are thus essential.
RNA language models (RNA-LMs), trained on millions of unlabelled sequences by predicting hidden nucleotides have shown promise for modeling RNA regulation. However, the biological information they capture at nucleotide resolution remains largely unexplored. Here, we systematically evaluate RNA-LM representations and show that these models learn universal RNA features, while also encoding fundamental biological properties. We find that RNA-LMs operate in two distinct regimes: low-confidence predictions driven by simple sequence properties, and high-confidence predictions informed partly by structural context. Chemically modified sites affect model confidence, for example, m⁶A sites have a much lower probability of being predicted as Adenosine compared with randomly sampled positions. We introduce a visualization tool to interrogate internal representations and show that RNA-LMs distinguish m⁶A sites, recovering the canonical DRACH motif and additional non-canonical motifs. Building on this, we develop a RNA modification predictor achieving high accuracy and enabling a genome-wide map of predicted modifications, including previously unannotated sites and combinatorial motif patterns.
These results establish RNA-LMs as powerful tools for decoding regulatory logic. Future work will explore the potential of RNA-LMs to explore the combinatorial code through which chemical modifications and RBP binding collectively specify RNA metabolism, including localization and translation.

A-to-I RNA editing represents an important modification enriched within primate-specific Alu elements, particularly in 3′ untranslated regions (3′ UTRs), where pairing of inverted alu elements enables the formation of double-stranded RNA structures, creating favourable substrates for ADAR enzymes. In this study, we investigated the patterns and co-occurrence of RNA editing events within such structured Alu elements.
In human RNA-seq data with depleted SRP9/14 or DHX9 and control samples, we identified high-confidence A-to-I editing sites. By combining genome annotation resources with repeat element mapping, we focused specifically on full-length Alu elements located in 3′ UTRs that form sense–antisense pairs. To allow comparative analysis, sequences were grouped by Alu families and aligned into a unified positional reference space.
A custom analytical approach was implemented to reconstruct read-level alignments relative to Alu consensus sequences, allowing precise tracking of editing events across individual reads. This enabled quantification of editing intensity and variability at single-molecule resolution. Furthermore, we applied statistical modelling to investigate dependencies between editing sites, revealing non-random co-occurrence patterns suggesting coordinated editing.
Our findings provide insights into the complexity of RNA editing within repetitive elements and highlight the importance of analysing RNA editing at both site-specific and read-level resolution to better understand its regulatory potential in non-coding regions. This project was supported by research grant APVV-24-0054

Gene expression is a complex process involving multiple regulated steps. These molecular controls are especially important for animal development, during which genes interact in temporally and spatially precise patterns. Although mRNA levels are used as a proxy for protein expression, global correlations between mRNA and protein levels are often modest. Here, we analysed and compared gene expression patterns from time-course quantitative mass spectrometry and RNA-seq data during 24h of Drosophila embryogenesis. The correlations between mRNA and protein levels are moderate, consistent with prior observations. We observe a delay in expression between mRNAs and their corresponding proteins, in line with previous work. Ancient genes experience higher delay than modern genes, suggesting they have evolved a higher degree of translational control. Unbiased clustering of transcript and protein expression recapitulated the temporal profile of embryogenesis and revealed different delay modes at different development stages. We observe functional enrichment in gene clusters with different delays. For example, genes involved in muscle structure development and cuticle development exhibit little expression delay between mRNA and protein, whereas proteins involved in pattern specification process and ribosome biogenesis are expressed noticeably later in embryogenesis than their corresponding mRNAs. Maternal and zygotic genes also show distinct temporal regulation in transcriptomics and proteomic profiles. These results highlight the importance of integrative approaches to understand complex biology, and allow us to explore the mechanisms that define the timing of gene regulation during development.

Paired single-cell multiome profiling of chromatin accessibility and gene expression is rapidly becoming routine, with PubMed query counts suggesting a sharp increase in multiome publications from ~74 (2022) to ~202 (2024), comparable to unimodal single-cell accessibility studies (~93 in 2022; ~194 in 2024). This growth highlights an urgent need for methods that analyze multiome data jointly rather than treating modalities independently. Accessibility alone cannot distinguish merely open elements from those that actively regulate transcription, nor capture temporal delays between chromatin opening and RNA changes

We present SULALA, a background-corrected, game-theoretic framework for enhancer–gene linking from paired scATAC+scRNA data. SULALA fits a bootstrapped non-linear support vector regression (SVR) model to predict gene expression from local accessibility and uses SHAP-based feature attribution to quantify region-level contributions. To preserve single-cell heterogeneity beyond cell-type averages, SULALA infers links at the level of pseudobulks along pseudotime, enabling within-cluster regulatory variation. It further incorporates trajectory-aware modeling to accommodate temporal lag between accessibility and expression and controls spurious associations via background-calibrated link scoring.

Across brain and epiblast multiome datasets, SULALA yields higher-precision, high-confidence enhancer–gene links than correlation-based methods and distance-only baselines (precision lift: 1.53 and 1.42, respectively). Analyses reveal stage-specific increases in enhancer activity at key developmental transitions, including glutamatergic neuron differentiation and critical epiblast stages (day 7, amniotic early/late, definitive endoderm, neuroectoderm). By producing a peak–gene link strength matrix across pseudotime and enabling motif-level aggregation, SULALA supports dynamic, fine-grained gene regulatory network and transcription factor activity inference beyond canonical cell types.

Among the various stromal cell types within the tumor microenvironment, cancer-associated fibroblasts (CAFs) are the predominant component, exhibiting diverse oncogenic functions. Despite their potential as therapeutic targets, no effective strategy exists to neutralize CAF activity due to the lack of specific markers and their heterogeneous nature. Current anti-tumor approaches indiscriminately deplete CAFs, overlooking the existence of both pro- and antitumor subpopulations, which makes these therapies toxic to the organism. A more promising strategy involves reprogramming CAFs into a uniform anti-tumor state, necessitating the identification of CAF subsets with exclusively tumor-suppressive properties, which is the objective of this study.
In this study, we examined distinctive features of the CAF transcriptome from 11 different types of tumors using bioinformatics analysis of a large RNA sequencing dataset. The creation of our own CAF marker discovery algorithm allowed us to identify a set of genes specific to CAFs across 11 tumor types. In ovarian adenocarcinomas, we identified a CAF population whose higher abundance correlates with a favorable prognosis. This population specifically expressed genes activated in response to IFNγ, which are distinctively upregulated in ovarian tumors compared to normal ovarian tissue. Additionally, we demonstrated that the abundance of IFNγ-responsive CAFs is conserved between primary tumors and metastases within the same patients. Spatial transcriptomics analysis revealed that these fibroblasts are localized within IFNγ aggregates, enriched with epithelial cells, macrophages, and endothelial cells exhibiting enhanced IFNγ response gene expression, as well as CD8+ T cells.
The study is supported by the Russian Science Foundation (No. 25-15-00520).

Translation of messenger RNA (mRNA) is a central regulatory layer determining which genomic information is expressed as functional protein. In this process, cis-regulatory sequence elements (CREs) interact with a wide range of regulatory factors to control RNA processing and protein output depending on contexts, such as cellular state, cell type and species. The 5' untranslated region (5UTR) of mRNA, with a median length of around 200 nucleotides in vertebrates, has been identified as a major determinant of ribosome recruitment, accounting for much of the cis-regulatory information determining translation initiation efficiency. Recent studies applied massively parallel reporter assays (MPRAs) of short in-vitro transcribed reporter constructs to quantify the effect of 5UTR sequence variation on translation initiation for a large number of arbitrary sequences and design synthetic sequences for tailored protein output using deep neural networks (DNNs). However, limitations in 5UTR sequence composition, length and experimental scale have limited efforts to systematically interpret learned representations of regulatory phenomena.
We developed a stochastic multi-species 5UTR MPRA platform, which connects (A) ~112.000 tiles of 200 nucleotides from all human and zebrafish 5UTRs and (B) massive randomised libraries of >10^7 million expected high-confidence reporters to a polysome profiling readout of translation initiation in human cell lines. By interpreting supervised sequence-to-translation models trained on a combination of native and randomised MPRA reporters, we aim to uncover cis-regulatory interactions, reveal mechanisms behind disease-causing 5UTR mutations and engineer 5UTRs for tailored translation initiation.

MicroRNAs (miRNAs) regulate gene expression at the post-transcriptional level, yet interpreting their function at the pathway level remains challenging. Existing enrichment analysis tools predominantly adopt network node-centric approaches that focus on gene expression profiles, neglecting the regulatory information encoded in miRNA-gene interactions (MGIs) that constitute network edges. This omission introduces analytical bias and limits biological interpretability, underscoring the need for network edge-based enrichment analysis methods that explicitly incorporate MGIs. Therefore, we present miREA, a network-based tool for miRNA enrichment analysis that leverages MGIs to characterize miRNA function at the pathway level. miREA contains five edge-based enrichment methods that integrate paired miRNA-gene expression and interactome profiles with pathway networks to perform MGI overrepresentation, MGI scoring-based, network topology-aware, and network propagation analyses. Benchmarking across multiple cancer types shows that the edge-based methods outperform node-based methods in improving sensitivity to identify relevant pathways and biological interpretability while maintaining controlled false positive rates. We further demonstrate the utility of miREA in elucidating miRNA-gene-pathway regulatory mechanisms in bladder cancer. miREA is a versatile enrichment analysis tool that provides pathway-level interpretation of human miRNA function and facilitates mechanistic hypothesis generation for experimental validation.

Immunotherapy has expanded cancer treatment, but its benefits remain restricted to certain cancers. In colorectal cancer (CRC), most cases are microsatellite-stable (MSS), poorly immunogenic, and refractory to immune checkpoint inhibitors (ICI). Here, we analyze patients with heavily-pretreated MSS metastatic CRC enrolled in a phase 1b trial (C-800-01; NCT03860272) of botensilimab (BOT; Fc-enhanced anti–CTLA-4) ± balstilimab (BAL; anti–PD-1), which demonstrated responses in a subset of otherwise ICI-resistant tumors.

We performed bulk transcriptomic profiling on 65 metastatic tumor biopsies across diverse anatomical sites, including lesions from liver, lung, and abdomen. To address the high transcriptional heterogeneity of the data, we applied self-organizing map (SOM) machine learning. SOM reduces dimensionality to a two-dimensional landscape of tumor microenvironment (TME) and tumor states, capturing dynamic transcriptional programs while preserving gene-level information for downstream functional interrogation.

Four major transcriptional programs concordant with pan-cancer classifications were identified: liver-like, proliferative (PRO), inflammatory (INF), and mesenchymal (MES). Each was associated with distinct cellular compositions, TME states, and clinical outcomes. The liver-like program was characterized by metabolic reprogramming and immunosuppressive signatures. The PRO, INF, and MES programs corresponded to immune-depleted, immune-enriched, and fibrotic states, respectively. INF and MES programs exhibited improved overall survival, although neither represented conventional hot tumors with high mutational burden.

Analyses also showed that metastatic tumors distributed along an immunogenicity axis marked by interferon-γ expression. Evaluation of matched pre- and post-treatment biopsies (available for 16 patients) revealed that BOT±BAL therapy promoted shifts toward immune-enriched states, including enhanced antigen presentation, T-cell recruitment, and cytotoxic activity.

Understanding the molecular basis of complex diseases such as Alzheimer's disease (AD) demands more than cataloguing differentially expressed genes; it requires situating those genes within a biologically coherent, disease-relevant framework. Although conventional gene set tools map genes onto broad ontologies, they are not designed to capture the heterogeneous, multifactorial nature of AD, limiting cross-study synthesis and disease-contextual interpretation.

To address this gap, we developed Genedex, a dedicated context engine. Genedex is built around a facet architecture that annotates gene lists across complementary biological dimensions: functional facets capturing disease-relevant processes such as neuroinflammation; cellular facets resolving signatures to specific brain cell types and states; tissue facets spanning brain regions and blood; and genetic facets linking variants to disease risk or protection. These facets are systematically mapped within the AD molecular continuum, which places an input gene set into a staged, multidimensional disease context. The current release indexes 309 gene lists from 23 manually curated publications against 92 facets.

Applied to infection-stratified bulk RNA-seq data from AD brains, Genedex defined upregulated genes in infected AD cases within astrocytic signatures and identified a previously unknown overlap with astrocytic resilience-associated downregulated genes, including AQP6 and VAV3. Downregulated genes converged on resilience-associated oligodendrocyte and neuronal programs. This pattern is consistent with systemic infection shifting the AD transcriptome away from resilience-associated states toward astrocytic activation.

Genedex supports systematic, disease-contextualized interrogation of AD molecular signatures and candidate gene prioritization for therapeutic investigation.

High-dimensional gene expression datasets present unique challenges in cancer biomarker discovery and classification. Here, we propose a novel deep learning framework incorporating principal component analysis (PCA) for dimensionality reduction and a custom optimizer, PCAGroupAdam, for effective gradient scaling. The framework was tested on multiple gene expression datasets, achieving superior classification performance compared to traditional optimizers (Adam, RMSprop, SGD). Key findings include the identification of biologically relevant genes such as AGR2, TSPAN8, and GAPDH, which were linked to cancer progression using SHAP analysis and validated through functional annotation (GO/KEGG) and STRING protein-protein interaction analysis and an unknown functioned lncRNA found to be correlated with breast cancer. Our approach demonstrates strong performance like high accuracy, f1 scores and significantly reduced loss values, interpretable results, and scalability to various high-dimensional omics datasets.

Understanding the role of transcript isoforms is crucial for dissecting disease mechanisms. TAR DNA binding protein-43 (TDP-43) is a key regulator of RNA splicing, and its dysfunction in neurons is a hallmark of some neurodegenerative diseases such as amyotrophic lateral sclerosis (ALS) and frontotemporal degeneration (FTD). Specifically, TDP-43 maintains proper splicing by preventing the aberrant inclusion of cryptic exons into mRNA, thereby preserving normal transcript isoforms. Although TDP-43-dependent cryptic exons have been implicated in disease pathogenesis, an approach to investigate how cryptic exons disrupt transcript isoforms has yet to be established. To address this, we developed IsoRefiner, a novel method for identifying full-length transcript structures using long-read RNA-seq. Our results show that IsoRefiner outperforms existing long-read analysis tools. Leveraging this method, we conducted long-read RNA-seq, guided by prior short-read RNA-seq, to comprehensively resolve the full-length structures of aberrant transcripts caused by TDP-43 depletion in human induced pluripotent stem cell (iPSC)-derived motor neurons. This led to the discovery of a novel TDP-43-dependent cryptic exon in the MNAT1 gene, along with its full-length transcript structure. Furthermore, we confirmed the presence of the MNAT1 cryptic exon in tissues derived from patients with ALS and FTD. Our findings deepen understanding of TDP-43 proteinopathy, and our approach provides a powerful framework for investigating splicing mechanisms across diverse cellular and disease contexts.

The high-dimensional and heterogeneous nature of transcriptomics data from RNA sequencing (RNA-Seq) experiments presents a persistent challenge for standard downstream pipelines, such as differential expression and enrichment analysis. In preclinical research, these challenges are often compounded by small cohort sizes dictated by financial and practical constraints. In light of recent studies on the low replicability of preclinical cancer research, it is essential to understand how the combination of population heterogeneity and underpowered cohort sizes affects the replicability of RNA-Seq research. Using 18,000 subsampled RNA-Seq experiments based on gene expression data from 18 high-quality real-world datasets, we systematically measured the impact of sample size on replicability, precision, and recall. We find that results from underpowered experiments are unlikely to replicate well. However, low replicability does not inherently imply low precision, as the analyzed datasets exhibit a wide range of possible outcomes. In fact, 10 out of 18 datasets yield high median precision despite low recall and replicability when the cohort size is five or more. To assist researchers constrained by small cohort sizes in estimating the expected performance regime of their datasets, we introduce a simple bootstrapping procedure that accurately predicts observed replicability and precision metrics. We conclude with actionable recommendations to improve the robustness of underpowered RNA-Seq studies.

Uncovering the regulatory interactions underlying various cellular processes remains a central challenge in systems biology. Our group has developed a set of methods to facilitate the reconstruction and analysis of gene regulatory networks. We present the SwissRegulon collection of tools and services, designed for the automated inference of regulatory interactions from high-throughput data, including gene expression (RNA-seq) and chromatin accessibility (ATAC-seq, ChIP-seq, DNase-seq).

The core of our inference framework is the Motif Activity Response Analysis (MARA) model. This approach uses genome-wide predictions of regulatory sites to infer the activity of specific regulators that best explain observed changes in gene expression or chromatin states across experimental conditions. Our automated web services ISMARA, CREMA, and CRUNCH provide results in a comprehensive, interactive format, characterizing key system regulators, their target genes, and associated biological pathways.

In addition to these web services, we provide access to genome-wide regulatory sites predictions and a curated set of regulatory motifs through SwissRegulonDB. Furthermore, we offer specialized tools for binding site prediction, such as Motevo and PhyloGibbs, which incorporate phylogenetic information to increase accuracy. These resources provide a set of tools for researchers to move from raw sequencing data to insights into gene regulation.

N6-methyladenosine (m6A) is the most prevalent modification of eukaryotic mRNA, yet many questions of its precise role in post-transcriptional regulation remain. In D. melanogaster, m6A has been implicated in gene regulation, but its relationship to alternative splicing and transposon stability is poorly characterized. While there are several sequencing methods that can be used to detect m6a, there is currently a lack of systematic comparison of said technologies, each of which differs in sensitivity and resolution.

Here, we used matched drosophila samples to compare three complementary sequencing approaches for m6A site detection: eTAM-seq (short-read, enzymatic marking), GLORI-seq (short-read, chemically induced conversion), and Oxford Nanopore Technologies (ONT) long-read sequencing with m6A-aware basecalling. We evaluate method-specific filtering strategies and characterize the strengths and limitations of each platform in terms of site-calling accuracy, reproducibility, and transcriptome coverage. Using these methods, we investigate whether m6A sites act cooperatively or competitively, and examine their downstream effects on alternative splicing and isoform usage.

Our results provide a guide for selecting methods to detect m6A and establishes a foundation for understanding how RNA methylation shapes the fly transcriptome.

The molecular complexity of human disease has long outpaced our ability to resolve it. Bulk transcriptomics obscures cellular heterogeneity and can mask rare but clinically relevant cell states governing tissue function. At GSK, we have built an integrated single-cell long-read sequencing capability designed to translate molecular resolution directly into actionable drug discovery insights.
We describe an omics capability combining single-cell RNA sequencing and long-read sequencing across our drug discovery programmes. Long-read sequencing moves us beyond gene-level resolution to isoform-level characterisation, enabling detection of disease-relevant splice variants, fusion transcripts, and allele-specific expression patterns invisible to short-read approaches. This is particularly consequential in tissues such as the liver and brain, where isoform diversity is a primary driver of cellular identity and disease biology.
We are applying this capability to construct high-resolution cellular atlases of diseased and healthy tissue from clinical biopsies and in-house disease models. For target identification and prioritisation, this enables confident target nomination, patient stratification modelling, and isoform-level tractability assessment with direct implications for modality selection and safety profiling.
Underpinning this is a scalable computational infrastructure connected to GSK's translational and clinical data assets, enabling rapid interrogation against human genetic evidence and existing compound portfolios. We are actively deploying this capability across our high priority disease areas to generate disease-specific atlases that anchor targets in high-resolution human biology to compress the timeline from biological question to development-ready insight.

In this study, we evaluated 90 bioinformatics pipelines using RNA-Seq datasets from Rats, Zebrafish and Mice. The analysis was conducted in the context of weak signals, including exposure to metallic particles (tungsten), low-dose radiation or medical treatment. RNA-Seq data analysis involves several critical steps, from quality control to differential expression analysis, each offering multiple algorithmic options. Selecting the optimal pipeline is particularly challenging in complex scenarios with weak signals.
We applied a dual strategy based on two complementary approaches to rank and evaluate the performance of these pipelines. The first approach, a widely used method, is based on the correlation between RNA-Seq and qRT-PCR expression data to ensure the direct validation of RNA-Seq results. The second approach leverages machine learning classifiers to rank the pipelines based on their ability to distinguish between exposure groups. This dual strategy was designed to identify the most reliable pipelines capable of providing accurate biological insights, with the top-performing pipeline highlighting key biological processes linked to a weak signal.
Our results highlight the crucial role of pipeline selection in RNA-Seq studies, as it influences both analysis efficiency and biological insights. While our findings are particularly relevant for studies with weak RNA-Seq signals, the ranking methods we employed can be applied in other fields to identify the most appropriate pipeline for generating biologically meaningful data. We provide practical recommendations for bioinformaticians to select robust pipelines, ensuring reliable and insightful outcomes across various research contexts, including environmental exposures.

Frequently, tumour cells exert aberrant somatic copy number variants (CNV), which are thought to play an important role in tumourigenesis and carry significant prognostic value. Historically detected by array-based SNP chip technologies, CNVs are now increasingly identified through whole-genome (WGS) or whole-exome (WES) sequencing. These global chromosomal rearrangements – traditionally characterised by expert karyotypes – may be partially substituted by NGS pangenomic data analysis complemented with RNA-seq data for translocation events.
To test this, CNVs of the T cell lymphoma cell line DERL-7 were analysed by comparing WGS, WES, SNP array, and RNA-seq data via CNVpytor, Control-FREEC, CNVkit, AraCNA, ASCAT complicated by lacking data of matched-normal controls. Read depth and mutations were computed and results were evaluated against the known karyotype of DERL-7. The CNV callers could partially verify known chromosomal gains and duplications at chr7 (i(7q)) and chr8 (+8), whereas SNP array analysis detected one alterations (i(7q)) only. In addition, a deletion at chr10 and an amplification at chr19 could be tracked – neither of which were described in the reference karyotype. The translocation involving chr1 with chr16 (t(1;16)(q12.2;p13.3)) could not be inferred via FusionCatcher, possibly due to insufficient sequencing depth.
Overall, the NGS-based approaches demonstrated higher resolution and sensitivity for CNVs compared to conventional karyotyping, enabling precise localisation of breakpoints and a more comprehensive representation of the DERL-7 pangenome. These findings suggest that NGS data analysis is a powerful tool for overcoming the limitations of manually derived karyotypes and for capturing the full complexity of tumour genomes.

Spatial transcriptomics enables exploration of gene expression profiles linked to spatial context. Current technologies such as 10x Genomics Visium assume that each spot captures ~10 cells, which is often inaccurate in complex and heterogeneous tissues. In practice, the true cellular density per spot remains unknown, and total transcript counts are frequently used as a proxy. This approximation introduces biases in downstream analyses, including deconvolution, differential expression, while masking biologically relevant variability associated with cell density.

Here, we introduce Nucount, an AI-based framework for predicting cell numbers per spot from histological images. Nucount leverages an ensemble learning strategy, integrating multiple segmentation and detection models to generate robust consensus estimates. We benchmarked Nucount using combinations of StarDist, Cellpose, and YOLO on the PanNuke dataset, which includes thousands of annotated nuclei across diverse tissue types.

Our results demonstrate the value of accurate cell count estimation by incorporating Nucount predictions into deconvolution pipelines. By constraining methods such as CytoSPACE to respect the number of cells per spot, we achieve more precise and physically grounded cell-type assignments, supported by validation analyses. Finally, we explore the interplay between cell density and transcriptional activity in pathological tissues (cancer and auto-immunity). Using Bivariate Local Lee's spatial statistics, we identify regions of concordance and discordance between cell density and total RNA abundance. These regions exhibit distinct cellular compositions, revealing a novel morpho-transcriptomic landscape that links tissue structure to gene expression variability. Overall, Nucount provides a robust and flexible framework enabling more accurate and biologically meaningful interpretations of tissue heterogeneity.

The development of computer science opens exciting avenues for personalized medicine and disease prevention. An example of a disease that could be better understood with the support of computational methods is chronic lymphocytic leukemia (CLL) – the most commonly diagnosed blood cancer among European adults.
CLL is characterized by the clonal expansion of B cells, which disrupts normal immune function. A cell clone is a group of genetically identical cells that descend from a common ancestral cell. Monoclonal B-cell lymphocytosis (MBL), an asymptomatic precursor to CLL, shows similarity to CLL in mechanisms of autonomous B-cell receptor signaling. Studying the connection between CLL and MBL has the potential to pave the way for early intervention strategies and personalized therapeutic approaches.
This research uses single-cell RNA sequencing (scRNA-seq) to compare genetic profiles of expanded and non-expanded B-cell clones within the same individuals, in order to minimise inter-individual variability. The aim is to investigate whether MBL drives premalignant expansion or malignant progression of CLL. By analyzing scRNA-seq data, we developed and validated a novel somatic variant detection pipeline, utilizing germline and expanded clone count tables and statistical testing to identify true variants. Our approach, incorporating a binomial likelihood model and Genotype Quality (GQ) scores to analyse single nucleotide polymorphisms (SNPs), identified losses of heterozygosity as predominant genetic alterations.
While current findings are limited by relatively shallow sequencing depth, they provide insights into our understanding of CLL pathogenesis. Further work will employ long-read sequencing with targeted deeply sequenced regions to validate these observations.

Tumor drug response is shaped by cellular heterogeneity, making single-cell resolution increasingly important for precision oncology. Although large-scale cell-line screens provide abundant drug-response labels at bulk resolution, transferring predictive models to patient-derived single-cell data remains a major methodological challenge.

Recent methods address this problem with deep learning–based domain adaptation, aiming to transfer predictive signal from bulk cell-line profiles to single-cell patient data without requiring target-domain labels. However, it remains unclear whether these methods provide a measurable advantage over simpler, well-controlled baselines, as existing studies often differ in preprocessing, training, evaluation protocols, and baseline comparisons.

Here, we present a systematic benchmark of four representative domain adaptation methods against two gradient-boosting baselines across 19 single-cell datasets and 10 drugs. Across this evaluation, deep domain-adaptation methods do not consistently outperform simpler baselines. We further demonstrate that a significant portion of the performance gains can be attributed to target-informed hyperparameter tuning and sparse label supervision, rather than to the adaptation mechanisms themselves.

These results suggest that current bulk-to-single-cell drug-response models do not yet reliably bridge the conceptual and distributional gap between cell-line screens and patient-derived single-cell data. Our study highlights the need for transparent benchmarking and leakage-aware evaluation to support meaningful progress in translational pharmacogenomics.

Based on a manuscript currently under peer review and available on bioRxiv: https://doi.org/10.64898/2026.02.24.707713

During tissue repair, alveolar cells transiently dedifferentiate, increasing plasticity before proliferating and redifferentiating to restore epithelial integrity. This process is tightly regulated by interactions with the surrounding microenvironment, including stromal and immune cells. In cancer, these mechanisms may be hijacked, as chronic inflammation can stabilize plastic cellular states and promote tumor progression. In lung adenocarcinoma (LUAD), cancer stem-like cells (CSC) are associated with poor prognosis, yet their relationship with the tumor microenvironment (TME) remains poorly characterized. We hypothesized that distinct TME features, including cellular composition and spatial organization, are associated with CSC phenotypes.
To address this, tumor microarrays from 50 patients were analyzed using Xenium spatial transcriptomics (10x Genomics). Among >350 genes, 9 CSC-associated markers were evaluated. Cells were segmented and annotated, and cellular neighborhoods (CN) were defined. Transcriptomic neighborhoods (TN) were additionally inferred using GraphSAGE modeling. Associations between CN and TN compositions and CSC marker expression were assessed.
CN and TN analyses yielded consistent results. Stem cell markers were differentially expressed in cancer cells and organized into co-expression modules with distinct spatial distributions, each associated with specific niches. Notably, CD44 expression was enriched in cancer cells located within myeloid- and fibroblast-rich neighborhoods. Ongoing analyses aim to identify interacting partners and signaling pathways underlying these associations, while additional markers will be evaluated to assess their relationship with broader phenotypic alterations.
These results suggest that distinct spatial niches within the TME are associated with specific CSC-related transcriptional programs. This investigation may enable the identification of therapeutic vulnerabilities within these microenvironmental dependencies.

Motivation: Tumor-educated platelets (TEPs) are emerging as a promising liquid biopsy source for cancer detection and monitoring. However, it remains unclear whether their transcriptomic alterations reflect tumor-specific molecular mechanisms or broader systemic responses. In particular, their relationship with the molecular network underlying glioblastoma (GBM) has not been systematically investigated.
Objective: The aim of this study is to assess whether transcriptomic alterations in TEPs from GBM patients are functionally and topologically associated with the GBM disease module in the human interactome.
Methods: RNA-seq data from TEPs (GSE68086) were analyzed to identify differentially expressed genes (DEGs) between GBM patients and healthy controls (adjusted p-value < 0.01, |log2FC| ≥ 0.5), yielding 1401 DEGs. After gene mapping and filtering, 1298 non-overlapping DEGs were projected onto a human protein-protein interaction network from BioGRID. A curated set of 48 GBM seed genes from IntOGen defined the disease module. Network proximity was computed as the mean shortest-path and compared against 500 random gene sets. Random walk with restart (0.7) was applied to prioritize DEGs.
Results: DEGs showed significant proximity to GBM genes (mean distance 1.41 vs 1.66 expected; Z = -18.69; p < 0.002). Functional enrichment highlighted telomere maintenance, DNA repair, transcriptional regulation, cell cycle, and stress response pathways.
Conclusion: TEP transcriptomic alterations in GBM are non-randomly organized and closely linked to the GBM disease module.

Understanding how regulatory elements (CREs) control gene expression in disease remains a central challenge in epigenomics. While perturbation experiments can reveal causal relationships, systematically exploring combinatorial perturbations of CREs is experimentally infeasible due to the rapidly growing search space, which reaches billions of possible perturbation sets even for small combinations.

We propose a model-derived in silico perturbation framework to identify a small set of regulatory regions whose perturbation shifts a disease gene expression state towards a healthy profile. Our approach builds on gene-specific models that predict gene expression from epigenetic activity. As a representative implementation, we use random forest models trained on IHEC data, with candidate ENCODE CREs as input features. These models induce a bipartite graph linking regulatory regions to genes. Assuming approximate additivity of single-perturbation effects, we formulate the selection of k regions as a combinatorial optimization problem that minimizes distance (RMSE) between predicted disease and observed healthy expression profiles. This formulation enables efficient exploration of the search space using integer linear programming (ILP).

We assess the validity of the additivity assumption by comparing predicted combinatorial perturbation effects to a model-derived ground truth, and further validate prioritized regions using independent CRISPRi perturbation data.

Our framework enables systematic prioritization of regulatory regions for combinatorial perturbation and provides a scalable strategy for guiding targeted experimental interventions in disease.

CSF2 is consistently elevated in psoriatic lesions and supported by preclinical studies, yet anti-CSF2 therapy failed to produce meaningful clinical benefit in a phase II trial. To investigate this translational discrepancy, we developed an integrative multi-modal transcriptomic framework combining bulk discovery and validation cohorts (n=110 and n=90), single-cell, spatial, and longitudinal treatment data from psoriatic lesions.
CSF2 axis activity, defined by coordinated ligand–receptor expression scoring, marked a myeloid-enriched inflammatory lesion state characterised by antigen presentation and cytokine–chemokine signalling. Bayesian network structure learning positioned the CSF2 axis downstream of NF-κB-associated signalling with strong support (bnlearn strength=0.99), and this relationship was confirmed by structural equation modelling across independent cohorts (β=0.52 and 0.46, both p<0.001). At single-cell resolution, functional CSF2 receptor co-expression was concentrated in dendritic cell and monocyte populations and was essentially absent from keratinocytes, contrasting with IL-17 and TNF receptor accessibility in keratinocyte effectors and IL-23 receptor accessibility in pathogenic T-cell populations.
Spatial transcriptomics showed CSF2 co-receptor-high regions were spatially segregated from barrier-dominant epidermal zones and enriched within immune-associated tissue niches (Wilcoxon p=1.9×10⁻⁵, 16/18 sections). Longitudinal anti-TNF data showed CSF2 axis scores declined with inflammatory resolution but did not clearly predict baseline treatment response.
Together, these findings support a model in which the limited therapeutic relevance of CSF2 in psoriasis may reflect its downstream inflammatory positioning and receiver–effector mismatch. More broadly, this integrative transcriptomic framework provides a generalisable strategy for evaluating cytokine target tractability beyond preclinical disease associations.

Differential co-expression networks provide a powerful framework to highlight changes in gene-gene interactions between biological conditions, going beyond the analysis of individual gene activity. However, their interpretability is often limited to binary (considering only whether changes are significant) or, signed representations (positive/negative edge to consider also if the change is in favour of one specific condition), which may introduce inconsistencies in edge interpretation.
To overcome these limitations, we developed a Co-Expression Transition Network Model (CTNM) extending signed network theory by modelling transitions in correlation structure that can occur when a pathological condition is compared to a control one. Edges are classified into four types: appearance (0+) and disappearance (+0) of positive correlation, and appearance (0-) and disappearance (-0) of negative correlation).
The framework was applied to publicly available transcriptomic data from glioblastoma patients and healthy controls. The analysis of the obtained CTNM showed that large cliques contain only 0+ or +0 edges, never both, suggesting cooperative but mutually exclusive behaviour, while 0- and -0 transitions act antagonistically, not appearing in these dense structures.
Based on these observations, we developed a clustering strategy based on modularity optimization, generalizing its definition to account for the four transition types. Each cluster is assigned a dominant cooperative transition (0+ or +0) and optimization maximizes intra-cluster coherence while favouring inter-cluster antagonism.
The resulting structure revealed 3 functional clusters, two deactivating (immune system and neurological pathways) and one activating (extracellular organization, cell death pathways) with emerging connections toward the others.

Rationale
Metastatic melanoma exhibits marked transcriptional plasticity under microenvironmental cues and therapeutic pressure. An epithelial-to-mesenchymal transition–like program can drive melanoma cells from a proliferative state toward a mesenchymal-like phenotype associated with quiescence, dormancy, and treatment resistance. However, the biological and clinical relevance of melanoma dormancy remains poorly defined.
We sought to identify a transcriptomic biomarker that predicts melanoma propensity toward proliferative versus dormant phenotypes and to assess its prognostic value. This endeavor was carried out via Clinical Bioinformatics, optimized Biobanking, and automated statistical modeling.
Methods:
Primary cultures from metastatic melanoma biopsies were classified as Proliferants, enriched for dominant NRAS/BRAF-mutated clones with high proliferative activity or Dormants, characterized by prevalent NRAS/BRAF wild-type clones, low proliferation, loss of detectable driver mutations and enrichment of pathways related to extracellular matrix remodeling, migration, de-differentiation, and dormancy. RNA sequencing identified a stable seven-gene signature, including four genes upregulated in dormant cells. A continuous dormancy score was generated and validated in independent melanoma cohorts. TROP2 protein expression was also validated by immunohistochemistry.
Results:
The dormancy scores stratified melanomas along a dormancy–proliferation continuum and were associated with shorter progression-free survival, reduced overall survival, and lower response to immune checkpoint inhibitors. Dormant-like tumors displayed mesenchymal features despite low proliferative activity, consistent with a slow-cycling invasive phenotype. TROP2 emerged as a key biomarker, overexpressed at both RNA and protein levels. Single-cell analyses identified TROP2-positive cells as a rare, quiescent, therapy-resistant malignant subpopulation. The dormancy score and TROP2 are promising biomarkers that warrant prospective clinical validation.

The mammalian placenta supports foetal development by mediating the exchange of gases and nutrients at the foetal–maternal interface. DNA methylation is a repressive epigenetic modification, established by DNA methyltransferases DNMT3A and DNMT3B after embryo implantation in mammalian development. The placenta is uniquely characterised by global DNA hypomethylation compared to embryonic tissues, and loss of DNA methylation leads to severe malformations in placental formation and embryonic lethality. However, how DNA methylation impacts transcriptional networks and cell fate commitment in the trophoblast lineage remains poorly understood. Here, we apply single-nuclei multiome sequencing (RNA-seq and ATAC-seq) to E9.5 mouse placentae across four genotypes (wildtype, Dnmt3a knockout, Dnmt3b knockout, and Dnmt3a/3b double knockout) to characterise the impact of DNA methylation loss in trophoblast. Our analyses reveals that the differentiation trajectory of the Syncytiotrophoblast Layer II (SynT-II) lineage, which is essential for maternal–foetal exchange, is specifically disrupted in the Dnmt3a/3b double knockout, indicating impaired progression toward terminal differentiation. SCENIC+ analysis identified regulons linked to WNT signalling and known transcription factors, such as GCM1, within this lineage, and several CpG-containing motifs that could potentially be DNA methylation-sensitive. Given that WNT signalling has previously been implicated in trophoblast stem cell to SynT-II transitions, these observations raise the possibility that loss of DNA methylation may contribute to disrupted responsiveness to this pathway during differentiation toward SynT-II. Together, this work reveals how DNA methylation shapes trophoblast cell fate by modulating key regulatory pathways and provides new insights into the epigenetic mechanisms underlying healthy placental development and pregnancy.