The rapid growth of single-cell datasets poses major computational challenges for adaptive immune receptor repertoire (AIRR) analysis. Scirpy, a widely used Python toolkit and core scverse package, enables AIRR analysis at the single-cell level. However, its original implementation could not scale beyond one million cells due to runtime and memory constraints.

Here, we present a comprehensive re-engineering of Scirpy's performance-critical components to enable large-scale analysis. The major reimplementations focused on improving the scalability of pairwise sequence distance computation and cluster identification. We optimized CPU implementations for Hamming distance and for the newly introduced TCRdist metric using Numba for parallelized compiled execution, and developed GPU-accelerated versions using CuPy. We further redesigned the cluster identification algorithm by optimizing sparse matrix operations, and extended AIRR matching, clustering, and visualization functionalities.

The re-engineered Scirpy achieved substantial performance gains across multiple real-world AIRR datasets encompassing millions of cells. The optimized CPU Hamming implementation provided a speedup of ~15x, while cluster identification was accelerated by ~127x. Our implementations enabled end-to-end analysis of the Omniscope single-cell AIRR dataset with ~8 million cells (<=64 CPU cores, NVIDIA A30), with runtimes of ~3 min (GPU Hamming), ~33 min (CPU Hamming), ~10 min (GPU TCRdist), ~135 min (CPU TCRdist), and ~17 min (CPU cluster identification). The GPU implementations offered a speedup of ~11x for Hamming and ~13x for TCRdist compared to their optimized CPU counterparts.

These improvements overcome prior computational limitations, enable large-scale immune repertoire studies, and establish a foundation for systematically characterizing AIRR breadth and dynamics.

Translational research aims to convert scientific discovery into tangible health improvements. At TRON, we employ single-cell RNA-sequencing (scRNA-seq) analysis to understand molecular mechanisms that could be leveraged to develop novel immunotherapeutic treatments in cancer and other severe diseases. Natural killer (NK) cells play a crucial role in cancer immunosurveillance, yet their cytotoxic function is frequently impaired within the tumor microenvironment (TME) through mechanisms that remain incompletely understood. Here, we constructed a comprehensive pan-cancer single-cell atlas to systematically investigate tumor-NK cell crosstalk contributing to NK cell dysfunction. For this purpose, more than one million cells from 130 patients across 13 solid tumor were integrated and harmonized across tissues and platforms. Automated cell type annotation combined with reference mapping resolved the transcriptional heterogeneity of NK subsets within the TME. Compared to blood-derived NK cells, we identified tumor-enriched NK populations, highlighting TME-induced transcriptional NK states. Cell-cell communication analysis further delineated receptor-ligand pairs mediating tumor-NK interactions which are functionally tested in vitro to confirm their role in tumor-driven NK cell exhaustion and tissue adaptation. Together, our results demonstrate how large-scale single-cell atlases can be employed to identify cell-cell communication axes that are relevant and targetable in disease.

Astrocytes perform essential functions in the brain, including supporting neuronal survival, plasticity, and metabolic homeostasis. Yet most computational methods for inferring cell-cell communication from single-cell transcriptomics focus on ligand-receptor interactions, overlooking metabolite-mediated signaling between neural and glial cells.
BrainChat aims to expand the types of detectable intercellular communication by integrating metabolic exchange into interaction inference from single-cell RNA-seq data.
Using an in-house transcriptomic dataset from the rodent cortex, we evaluated two complementary computational approaches: scFEA, a graph neural network-based method for estimating metabolic fluxes through predefined biological modules, which we adapted to better capture brain-specific metabolic context; and NeuronChat, a flexible framework integrating genes encoding ligands, receptors, and complex molecular machineries to infer intercellular communication.
Both tools were benchmarked against the glutamate–glutamine cycle, a well-characterized pathway with known directionality between neurons and astrocytes. This benchmarking revealed important limitations of existing methods, notably their reliance on prior knowledge and limited scalability.
To address these gaps, we developed a machine learning-based approach to classify cell pairs by communication status, comparing automated feature selection with biologically-informed gene sets derived from GO pathways across approximately 17,000 genes. Both strategies achieve F1 scores above 0.8, while GO-based approaches additionally enable the identification of specific KEGG pathways most involved in neuron-astrocyte communication, providing direct biological interpretability.
BrainChat project ultimately seeks to leverage these approaches to detect pathway-level dysregulation brain pathologies such as neurodevelopmental disorders or epilepsy, with potential applicability across tissues and species.

Transcription termination and polyadenylation are crucial regulatory mechanisms that can be hijacked by viruses such as Herpes Simplex Virus 1 (HSV-1) to negatively impact host gene expression and consequently dampen host immune responses. In this study, we investigate how HSV-1 immediate early proteins ICP22 and ICP27 alter the usage of host polyadenylation sites (PAS). While ICP27 mediates the disruption of host transcription termination, ICP22 is known to interact with CDK9, a kinase that defines checkpoints for transcription initiation and termination.

Here, we employ a computational pipeline based on 3'-end sequencing data to precisely quantify PAS and identify distinct changes in polyadenylation patterns. An additional PAS preference metric allows quantifying the direction and strength of shifts in PAS usage across conditions.

Using this approach, we show that the expression of ICP27 alone leads to a significant shift in PAS usage, as the use of internal PAS is increased and terminal PAS usage is reduced. While ICP22 expression alone did not significantly alter host PAS usage, the combined expression of ICP22 and ICP27 showed a consistent but notably amplified pattern in PAS usage compared to ICP27 expression alone. This was despite the fact that ICP27 was less expressed and overall disruption of transcription termination was reduced.

In summary, this suggests that ICP22 and ICP27 synergistically affect PAS use during HSV-1 infection in a manner distinct from ICP27-induced disruption of transcription termination.

Dysregulation of gene expression is central to cancer development, and genes do not function in isolation but instead as part of an expansive gene regulatory network. Increasing complexity further, within the tumour is a highly diverse array of cell types. Understanding changes to cell type-specific gene regulatory networks (GRNs) during cancer development can reveal valuable insights with the potential to inform clinical decisions and treatments.
Cell type-specific GRNs can be reconstructed from single-cell RNA sequencing data by existing bioinformatic tools, however single cell data is often not available in the clinic for large sample sizes. Reconstructing GRNs from larger datasets increases statistical power which will also enable the analysis of GRNs in cancer subtypes beyond the most prevalent. Our research goal is to develop a tool to generate cell type-specific GRNs from bulk RNA-seq data.
Cell type-specific gene expression was predicted from bulk RNA-seq data by cell deconvolution with BayesPrism, followed by sample-specific GRN reconstruction with PANDA and LIONESS. To develop the method, a pseudobulk was created from single-cell RNA-seq data from a publicly available prostate cancer dataset (GSE141445) for the analysis. This allowed benchmarking of the cell-type specific GRNs to those reconstructed with cell types identified from the original single-cell RNA-seq data. Preliminary results show a strong correlation of the network metrics between predicted and real networks (Pearson's test, median r = 0.64). Future research will focus on applying the tool to additional cancer datasets to identify cell type-specific features of GRN dysregulation in the tumour environment.

Causal structure learning offers a promising approach to studying gene regulation in cells, aiming to provide deeper mechanistic insights than purely association-based methods. Theoretical groundwork indicating that interventions improve identifiability of causal structure motivates the use of causal structure learning methods in scenarios with interventional information.
This benchmark investigates the ability of existing causal structure learning algorithms to leverage the interventional information revealed by single-cell CRISPR screens to infer gene regulatory networks (GRNs). In this study both synthetic and experimental single-cell CRISPR perturbation data is leveraged, and a suite of causal structure learning algorithms is evaluated on metrics tailored to synthetic ground truth and real biological data respectively. On synthetic data, accurate recovery of GRNs is achieved under favourable conditions by score-based (GIES) and neural causal methods (DCDI, AVICI): strong interventions, large sample sizes, and low measurement noise. However, under more challenging conditions, prediction accuracy is shown to reduce, in particular when applying realistic technical noise. Similarly on real data, performance remains unreliable, limited by technical and biological noise, as well as algorithmic scalability. This demonstrates a current gap between theoretical potential and practical application of causal structure learning for GRNs.
The benchmark provides insight into algorithm strengths and limitations, showing simpler empirical baselines such as GRNBoost2 and Mean Difference remain competitive under realistic noise conditions, and offers groundwork for further methodological development. We also provide an accessible software package to leverage modern causal structure learning on custom datasets.

Disruption of gene regulation can lead to uncontrolled cell growth and ultimately cancer, with further alterations facilitating metastasis. To understand the driving processes of breast cancer metastasis we studied gene regulation in breast primary tumors and metastases.
We reconstructed sample-specific gene regulatory networks (PANDA+LIONESS) from bulk RNA sequencing data of 45 paired primary breast cancers and brain metastases. Paired analysis revealed significant dysregulation of interferon signaling in metastases, through repression (P < 0.01). This was validated in independent data (39 patient-matched primary tumors, 84 metastases). Rewiring of this pathway remained evident across different breast cancer subtypes, and was independent of immune cell infiltration.
To further understand in which cell type the interferon deregulation occurred, we analyzed single-cell RNA sequencing data from 20 patient-matched primary and lymph node metastatic tumors and constructed pseudobulk expression profiles to enable cell-type-resolved sample-specific network inference. This showed that, although both cancer and immune cells downregulate expression of interferon signaling genes, only cancer cells rewire the network around these genes.
Lastly, to assess whether dysregulation of interferon signaling may drive metastasis, we analyzed primary tumors from patients without metastasis at diagnosis (stage N0 (i-), TCGA), comparing patients who did not develop recurrence (n = 129), to those that progressed later (n = 15). We observed significant interferon signaling dysregulation in tumors that metastasized later (P < 0.01). Altogether, our results show that network rewiring in cancer cells reflects a tumor-intrinsic reprogramming of interferon signaling that may contribute to metastatic progression.

Sharing clinical transcriptomic data across institutions is essential for advancing artificial intelligence in precision oncology, yet stringent privacy regulations severely constrain access to patient-level datasets. Synthetic data generation (SDG) has emerged as a promising strategy to alleviate these constraints, but existing benchmarks have focused almost exclusively on statistical fidelity, leaving the biological utility of synthetic transcriptomic data largely uncharacterised.
Here, we present SynOmicBench, the first systematic benchmark for high-dimensional clinical transcriptomic cancer data. Our framework combines standardized preprocessing with multidimensional evaluation, prioritizing biological validation alongside statistical fidelity and attack-based privacy assessment. Biological utility is assessed through six downstream bioinformatic tasks, including differential gene expression, gene set enrichment, immune cell deconvolution, survival analysis, and predictive modelling. Privacy risk is quantified via attack-based metrics aligned with European Data Protection Board guidelines rather than distance-based proxies.
We evaluate five SDG methods spanning deep generative models and statistical approaches: CTGAN, TVAE, Gaussian Copula, Synthpop, and Avatar, across three independent immunotherapy cohorts: clear cell renal cell carcinoma, melanoma, and non-small cell lung cancer. Across these cohorts, no single method excelled across all dimensions, reflecting inherent trade-offs between fidelity, utility, and privacy. Gaussian Copula achieved the most balanced performance, followed by Avatar. Synthetic data consistently reproduced biomedical signal directionality but with attenuated effect sizes, supporting hypothesis generation while cautioning against overinterpretation. Inter-replicate variability highlights the necessity of multi-seed synthesis.
SynOmicBench (https://trinhthechuong.github.io/SynOmicBench/) provides a reproducible decision-support tool for method selection and promotes biologically informed, privacy-aware adoption of synthetic data in precision oncology.

Despite recent therapeutic advances, a substantial proportion of melanoma patients fail to achieve durable responses to anti-Programmed cell Death protein 1 (PD-1) therapy. To investigate molecular determinants of this variable response, transcriptomic analysis is combined with structural modelling approaches. The single-cell RNA dataset comprised 48 samples and 16291 cells of melanoma patients treated with Pembrolizumab. Differential analysis is applied and the significant differences are investigated between two patient groups: responders and non-respondents. Poly (ADP-ribose) polymerase family member 14 (PARP14) was significantly overexpressed in non-responders, highlighting its potential involvement in anti-PD1 resistance mechanisms. It is presented as a promising target of resistance but lacks structural data on its binding modes with inhibitors.
To refine new biomarker identification, variance-based gene filtering is applied followed by machine learning approaches, including regression models (linear regression, Ridge regression), penalised regression (Least Absolute Shrinkage and Selection Operator or Lasso, ElasticNet) and classification models (eXtreme Gradient Boosting or XGBoost and Random Forest) to a bulk dataset of 122 patients treated by anti-PD1. Lasso and Wilcoxon-test enabled the identification of 14 genes associated with the treatment response. These signatures are further assessed in an independent bulk and a single-cell datasets to support their robustness and reproducibility.
Finally, to explore the structural relevance of anti-PARP14 and anti-PD1, transcriptomic analyses are combined with comparative molecular docking approaches using deep learning model or gradient optimization method. We proved inhibitors' high specific interactions with their targets. These bispecific ligands hold promise for enhancing targeted therapy efficacy and decreasing resistance in melanoma.

Single-cell RNA sequencing (scRNA-seq) has revolutionized the field of genetics, uncovering nuanced transcriptional profiles that can be lost with whole-tissue RNA sequencing. Despite the advantages, challenges remain in the analysis, interpretation, and reproducibility of scRNA-seq data. One such challenge is cell-type annotation. Traditionally, cluster identities are annotated based on the expression of only a few genes, which is a time-consuming and biased approach. While numerous computational annotation tools have been developed, benchmark studies have shown that they provide inconsistent annotations, limiting their reliability. We developed celltypeEnrich, a web application that determines a consensus cell-type annotation for a given list of input marker genes. The tool uses the hypergeometric test to calculate enrichment of cell-type-specific genes within an input gene list against any combination of 26 reference datasets, including two consortium datasets: Tabula Sapiens and the Human Protein Atlas. Enrichment results across reference datasets are then compared from broad to specific cell-type levels, and recurring matches are used to assign a consensus cell-type annotation. We benchmarked celltypeEnrich using scRNA-seq datasets from three tissues and two species and found that it correctly annotated 63-70% of the clusters using the marker gene lists. In comparison, the three other annotation tools tested were limited by either low accuracy, a lack of models for the tissues evaluated, or the need for parameter optimization. Furthermore, celltypeEnrich's accuracy is maintained when down-sampling input gene lists to 25% of their original size. In summary, celltypeEnrich provides a robust, consistent, and accurate approach for scRNA-seq cluster annotation.

Ex vivo cell culture systems are widely used in biological research, but their ability to accurately reflect in vivo biology remains unclear. We developed a computational framework to quantify how experimental model context shapes gene expression. Across various datasets, we identified consistent baseline shifts indicating systematic biases introduced by ex vivo culture. Next, by analyzing interaction effects between perturbations and the experimental model, we found that baseline differences propagate into perturbation responses, leading to substantial divergence in perturbation effects between ex vivo and in vivo contexts. Notably, some perturbations showed discordant or even opposing perturbation effects, posing challenges for prediction across experimental models.
Using AI-based prediction tools, we next attempted to infer in vivo perturbation effects from baseline and perturbed ex vivo data, together with baseline in vivo profiles, but performance remained limited, particularly for perturbations with opposing trends. Linear modelling showed that models trained across perturbations performed poorly in prediction task, suggesting that it is challenging to predict perturbation-specific responses.
Taken together, we show that baseline transcriptional differences encode systematic, generalizable shifts that can inform ex vivo model adaptations and cross-prediction (computational) modelling tools, but accurate prediction of perturbation effects requires incorporating perturbation-specific structure rather than relying on global patterns. Overall, these findings evidence the need for experimental model- and perturbation- aware modelling strategies to improve the translation of ex vivo perturbation data to in vivo biology.

Prostate cancer (PC) is the most common cancer in men, and despite initially responding to androgen receptor (AR) targeted therapy, an incurable castration-resistant PC (CRPC) frequently develops. Cancer cells exhibit high transcriptional activity to sustain the expression of pro-proliferative and anti-apoptotic genes that typically have short half-lives. It is therefore unexpected that the major transcription elongation factor, cyclin-dependent kinase 12 (CDK12), is inactivated in aggressive CRPC and confers even a growth advantage to these cells.
We have employed multi-omics strategy to study how the decrease in CDK12 activity affects AR function by assessing nascent transcription, alternative splicing, intronic poly-adenylation, and overall transcriptional program. Our results show that inhibiting CDK12 along with AR stimulation remodels the RNA polymerase II interactome by recruiting specific factors to the polymerase to rescue PC cells from the transcription elongation defect caused by decrease in CDK12 activity. In addition, functional loss of CDK12 activity together with AR hyper-activation leads to upregulation of specific pathways that support cancer cell survival.
In brief, our data show that pharmacological inhibition of CDK12 activity alters the AR-driven cellular program and remodels the core transcription machinery in an adaptive manner. Understanding this remodelling should enable the design of rational therapies that selectively eliminate the CDK12-mutant cancer cells.

Deep learning models have shown strong performance in modelling transcriptional expression patterns but often lack interpretability. In settings where perturbations follow an ordinal structure, such as dosage levels, promoting ordinality in the latent space may improve both robustness and biological insight.
We analyse single-nuclei RNA-seq data from 131,613 mouse liver cells exposed to nine concentration levels of TCDD using a variational autoencoder augmented with ordinal losses. We compare three ordinal loss functions (binomial cross-entropy, ordinal encoding, and cumulative logistic link) against a standard cross-entropy baseline, with random forest classifiers and regressors as additional baselines. Ablation studies evaluate the effect of classification loss weighting and VAE architecture. We further incorporate a pathway-constrained decoder, using Bayesian differential factor analysis to assess pathway activity in the latent space.
While cross-entropy achieves the highest predictive performance across ordinal-unaware and ordinal-aware metrics (including balanced accuracy, MSE, Kendall's tau, and AUOC), ordinal losses substantially alter the latent space structure. Binomial cross-entropy and Cumulative Logistic Link induce a continuous gradient aligned with increasing dosage, whereas Cross-entropy produces clustered representations and Ordinal Encoding partially preserves ordering. These structured embeddings correspond to biologically meaningful pathway activity patterns, including cholesterol biosynthesis. The pathway-constrained decoder improves performance across all models without changing their relative rankings.
These results show that encoding ordinality reveals biologically relevant structure in the latent space, highlighting a trade-off between predictive performance and interpretable representation learning. Future work will explore ordinal latent representations for generating expression profiles of unseen dosage perturbations.

A riboswitch is a functional RNA region, typically found in the untranslated regions (UTRs) of mRNA, that regulates gene expression by undergoing structural changes in response to ligand binding or other stimuli. While the functional mechanisms of riboswitches have been extensively elucidated in bacteria, their identification and functional characterization in eukaryotes—which possess complex gene regulatory mechanisms—remain significant challenges in Biology. The comprehensive identification of riboswitches is crucial for deepening our understanding of gene regulatory networks. To address this, SwitchFinder was developed as a method to predict riboswitches from RNA sequences, utilizing thermodynamic calculations based on energy models of RNA secondary structure (Khoroshkin et al., 2024). However, this method is limited by its computational requirements; specifically, it is inapplicable to approximately 60% of the sequences in the Rfam riboswitch database, and challenges remain regarding its prediction accuracy. Therefore, the aim of this study was to construct a deep learning model capable of accurately predicting riboswitches across all available data, free from such limitations. Specifically, we designed a model based on a convolutional neural network that uses not only thermodynamic features of RNA secondary structure but also RNA sequence information and predicted secondary structure information as inputs. As a result, our method achieved an improvement in prediction accuracy (AUC) compared to SwitchFinder. Furthermore, we applied the model to human transcriptomes to identify candidate riboswitches and conducted experimental validation to assess their riboswitch activity.

Adequate diagnostics and prognostication of prostate cancer remains challenging due to the multifaceted heterogeneity of the disease. Prognostic signatures derived from bulk transcriptomics assess the risk of adverse disease progression. Moreover, the impact of the tumor microenvironment in prostate cancer is only being uncovered. It is unresolved whether certain cell types contribute more to the prognostic signatures than others, and whether this signal reflects distinct spatial patterns in the biopsy.
Here, we report our work-in-progress towards jointly analyzing prostate cancer spatial transcriptomics (ST) and histopathology imaging data. We collected ST data from 48 prostate cancer patients using the 10x Genomics Xenium platform. We compared Xenium's cell segmentation to transcriptome-based methods Proseg and Segger. We find that Proseg consistently segments larger cells, while Segger shows more conservative cell size estimates. In dense tissue areas, Segger identifies more overlapping cells. We identified expected cell types in prostate biopsy tissue including endothelial, epithelial, fibroblast, smooth muscle, and immune cells.
We then investigated the spatial origin and patterns underlying bulk transcriptomic signatures and found that the cellular origins of the signals differ by the cell type. Four out of five prognostic signatures were epithelial-driven; only one signature captured expression from stromal and immune cells.
Next, we will investigate the spatial patterns of the signatures, correlate our findings to the clinical data, and validate our findings in the iCAN Digital Precision Cancer Medicine Flagship project. Our results will contribute to improved clinical diagnostic and prognostic tools for prostate cancer.

Ribosome profiling (Ribo-seq) enables high-resolution measurement of translation by quantifying ribosome occupancy along coding sequences. A key challenge in comparative computational biology is to assess whether ribosome occupancy profiles differ at specific positions across species or conditions in a statistically rigorous manner. Such analyses are complicated by sequence divergence, differences in gene length, and the count-based, autocorrelated structure of Ribo-seq data.

Here, we present a computational framework for comparative Ribo-seq analysis that integrates sequence alignment with flexible statistical modeling. Gene-specific ribosome occupancy profiles are constructed and aligned using multiple sequence alignment (MSA), enabling position-resolved comparisons across species. We model translation dynamics using generalized additive mixed models (GAMMs) with explicit autocorrelation structures, allowing estimation of smooth occupancy profiles, quantification of uncertainty, and detection of local deviations in ribosome density. The framework enables systematic comparison of ribosome occupancy profiles across species, capturing both global trends and localized differences, including deviations from gene-specific mean profiles.

Overall, the approach provides a flexible and reproducible strategy for cross-species analysis of ribosome profiling data and facilitates the study of translational regulation and its evolutionary dynamics.

Bone marrow transplantation is the only curative treatment for transfusion-dependent thalassemia, yet the cellular dynamics of haematopoietic reconstitution after the procedure remain poorly characterised at single-cell resolution. A key analytical challenge is that post-transplant sequencing libraries always contains cells from multiple individuals - the patient and a donor - which must be separated computationally before any biological interpretation is possible.
We present an integrated bioinformatics framework addressing this in two steps. First, we applied SCSM-HLA, a probabilistic model that extends the Vireo demultiplexing framework, assigning single cells to their donor of origin by jointly modelling SNP profiles, HLA allele expression - each person carries a unique set of immune genes - and sex-linked read counts. Incorporating experimental pool design as prior information enables accurate assignment even in complex multi-donor samples.
Second, we developed a comprehensive scRNA-seq analysis pipeline for the bone marrow dataset comprising thalassemia patients and stem cell donors. Following quality control, doublet removal, and Harmony-based batch correction, unsupervised clustering identified 20 haematopoietic cell populations spanning erythroid, myeloid, lymphoid, and progenitor compartments. Differential expression analysis revealed hallmarks of ineffective erythropoiesis in thalassemia patients, including upregulation of fetal haemoglobin genes (HBG1/HBG2), a stress response in erythroid precursors, and dysregulation of immediate-early transcription factors in haematopoietic progenitors compared to healthy donors.
Together, these tools provide a reproducible and biologically interpretable framework for single-cell analysis of complex bone marrow transplant samples, with direct applicability to other multiplexed clinical cohorts.

Alternative splicing (AS) is a key post-transcriptional mechanism that contributes to the diversification of transcripts and proteins by generating multiple mRNA isoforms from a single gene. AS has been observed to impact a significant proportion of multi-exonic genes, operating within a context-dependent manner and contributing to a variety of outcomes, including alterations in protein function, RNA localization, or degradation. Major AS event types include exon skipping, mutually exclusive exons, alternative splice site usage, and intron retention, all of which play important roles in development and disease. RNA sequencing has facilitated large-scale transcriptome analysis, yet AS isoforms are often underexplored due to inherent analytical challenges. Computational tools are therefore essential for transcriptome-wide AS detection, with event-based approaches generally outperforming isoform-based methods. However, reliable detection remains difficult with short-read sequencing, leading to high false-positive rates. Existing benchmarking studies are limited by inconsistent evaluation criteria, tool selection, and dataset variability.
Here, we present a systematic evaluation of AS detection tools using both simulated and real RNA-seq data. From 66 identified tools, 14 were selected through a rigorous multi-step filtering process. Performance was assessed primarily using precision and recall, prioritizing precision to minimize false positives. Additional metrics included inter-tool agreement and the impact of filtering strategies. Results were integrated into a comprehensive framework that also considers usability factors such as installation and documentation quality. This benchmark provides the community with an evidence-based guide for selecting optimal AS detection workflows and highlights opportunities to improve the accuracy and robustness of AS analysis in RNA-seq studies.

Translation efficiency (TE) captures how effectively mRNA is translated into protein and represents a key regulatory layer of gene expression beyond mRNA abundance. Identifying genes with differential translation efficiency (DTE) across cell types and tissue regions is essential for understanding spatial variation in protein synthesis. Emerging spatial imaging technologies, such as STARmap and RIBOmap, enable measurement of transcriptional and translational signals at single-cell resolution within intact tissues, providing new opportunities to study translational regulation using spatial multi-omics data. However, most existing statistical methods for DTE gene detection are developed for bulk sequencing data. Current methods for DTE analysis in single-cell data remain limited and primarily rely on ratio-based metrics that are highly sensitive to data sparsity and modality-specific noise. In particular, RIBOmap signals are inherently sparser than STARmap signals, causing TE estimates to be undefined or unstable for the majority of gene-cell pairs, leading to biased or incomplete detection of differentially translated genes.
To address this, we propose a statistical framework for DTE gene detection that jointly models spatial transcriptomic and translatomic data within a hierarchical probabilistic model. Our approach explicitly accounts for excess zeros and measurement noise in both modalities, and uses the transcriptomic signal to inform the interpretation of translatomic zeros. This work contributes a statistical approach to DTE gene detection in spatial single-cell data, with potential applicability across spatial multi-omics platforms for studying spatially resolved translational regulation.

Technological advances led to the generation of large, often under-exploited collections of transcriptional profiles from single-cells, enabling the detection of transcriptional heterogeneity in seemingly homogeneous cell populations. Sparse evidence suggests that this is associated with development, ageing, and diseases, such as cancer. We thus hypothesize that transcriptional heterogeneity is regulated, subject to natural selection and exploited by cancer cells to increase their evolvability and plasticity, enhancing their chances of developing resistance to therapies. To address these hypotheses, we comprehensively mapped transcriptional heterogeneity across tissues, age groups, and diseases. We developed a computational pipeline for the re-analysis and robust estimation of transcriptional variability from over 1000000 published single-cell profiles. Our analyses revealed 6000 hyper-variably transcribed genes (HVG) in immune cells of the spleen, but almost none in pancreatic cells, emphasizing potential biological implications of this phenomenon. We further identified a switch from stable to variable transcription at 541 genes in aging hepatocytes. We are currently training random forest and regression models to identify the upstream regulators of such patterns, combining transcription factor binding, chromatin state, and alternative splicing data as predictors. Initial results are promising (AUC 0.67); we therefore aim to refine these models and extend them to detect cell type-, age-, and disease-specific drivers of heterogeneity. Preliminary models on HVG across different cell types also yielded promising results (AUC 0.72), highlighting general regulators of transcriptional variability. These results may advance our understanding of transcription regulation mechanisms, their correlation with age, as well as the evolvability and plasticity of cancer cells.

Chronic low-grade inflammation is a hallmark of aging, promoted by factors like accumulation of senescent cells and myeloid skewing of the immune compartment. Macrophages are central regulators of tissue homeostasis, yet in aging they fuel pro-inflammatory feedback loops promoting cancer and autoimmune diseases. Understanding the cellular processes that maintain macrophage homeostasis may reveal strategies to prevent or reverse the transition to pathogenic states. Aging can increase transcriptional heterogeneity, and the rise of gene expression noise is known to have regulatory impact on cell differentiation and cancer cell survival. We therefore hypothesize that loss of precise transcriptional control can drive spurious expression of pro-inflammatory genes, contributing to the breakdown of macrophage homeostasis. To test this, we propose a computational framework that identifies regulatory mechanisms of transcriptional noise. To this end, we stratified genes from scRNA-seq of homeostatic macrophages based on their noise and expression levels, isolating biological from technical sources of noise via a custom pipeline. As a proof of concept, the stratified gene groups revealed distinct noise patterns, for example, between immune response and housekeeping genes. To uncover the regulatory features that distinguish the noise-stratified gene groups, we trained supervised machine learning models on epigenetic and RNA-level data. In our model, TATA box and chromatin accessibility were identified as important predictors which were previously shown to modulate transcriptional variability. Incorporating additional features such as protein and kinetic-based estimates will further deepen our understanding of the regulatory determinants of transcriptional noise.

Deciphering the regulatory grammar of promoters and enhancers is essential for understanding transcription initiation mechanisms and predicting the functional impact of regulatory sequence variants. While genomic deep learning models, from Enformer to AlphaGenome, are successful at predicting promoter activity across most tissues, analysis of underrepresented tissues remains challenging. This applies to human skeletal muscles, characterized by profound transcriptional diversity but undersampled in existing omics atlases.
We present Arinette, a novel deep learning model for predicting muscle-specific chromatin accessibility and transcription initiation at single-nucleotide resolution, utilizing an extensive set of CAGE-Seq from FANTOMUS (fantomus.autosome.org) and DNase-Seq from ENCODE.
Compared to existing models, Arinette features reduced parameter and time complexity, while maintaining competitive accuracy. The model reached a Pearson r of 0.72 for the aggregate promoter activity and 0.38 for single-nucleotide profiles. AlphaGenome's embeddings further boost model performance to 0.86 and 0.57, respectively. Arinette demonstrates strong performance in variant effect prediction, validated on the PromoterAI MPRA eQTL (AUROC 0.85), CAGI7 lentiMPRA data (r = 0.430 if tailored for HepG2), and muscle allele-specific variants (AUROC 0.76 for high-effect SNPs). Interpretability analysis confirms the model's ability to recognize individual transcription factor motifs. All in all, Arinette offers a powerful new tool for modeling human regulatory regions.

In omics analysis, traditional regression struggles with high dimensionality and noise. We introduce BI-SCUDO Regression, an explainable rank-based algorithm for predicting continuous clinical variables. It utilizes a genetic algorithm to extract patient-specific signatures compiled into a biomarker, whose statistical significance is validated by permutation testing (n = 10000), generating distance-weighted predictions.
Performance was compared against LASSO, Random Forest Regression (RFR), and XGBoost, by considering increasingly noisy synthetic datasets. In addition, TCGA transcriptomics (n=3464) was used to test the methods in a real-world scenario, trying to predict overall survival across 32 cancer projects stratified by cancer stage (whole dataset, stages I-IV, II-IV, and III-IV). Predictive power was evaluated using 10-fold cross-validation R² and the Ratio of Performance to Deviation (RPD). Feature selection was assessed via Fisher's exact test for cancer-gene enrichment.
In synthetic datasets, BI-SCUDO demonstrated superior robustness (p≤0.0003), maintaining a median R² of 0.78 at maximum noise, whereas LASSO collapsed (R²≈0) and RFR and XGBoost degraded to 0.70 and 0.69. On TCGA data, BI-SCUDO achieved significant biomarkers in 95% of the datasets (p < 0.05) and was the only method yielding positive median R² values (0.23–0.25) across all stratifications. XGBoost was the only method identifying an equal or greater number of significantly cancer-gene-enriched biomarkers compared to BI-SCUDO. However, given that XGBoost biomarkers are on average twice as long as those from BI-SCUDO and XGBoost's poor predictive performance, this result supports BI-SCUDO as the method that best balances biological relevance with predictive stability.

Spatial transcriptomics is currently revolutionizing discovery across biology and medicine, particularly in neuroscience and oncology. It is now possible to determine which genes are (co-) expressed in specific regions or cell types, where they are located in tissue, and how neighboring cells relate. However, extracting this information is computationally demanding, whether using chip-based platforms for genome-wide coverage or microscopy-based systems for high spatial resolution. A common compromise is to preselect genes of interest, known as spatially variable genes (SVGs).
We present Image Correlation of Expression ProfileS (ICEPS), a new method for SVG detection based on spatial autocorrelation length. ICEPS compares favorably with existing approaches: it is more sensitive to compact, spatially extended expression patterns than methods focused only on spot-to-spot variance, and more robust to random single-spot expression spikes, likely technical artifacts, than the widely used Moran's I statistic. Although some more sophisticated methods, such as nnSVG, may outperform ICEPS in certain settings, ICEPS is substantially faster because it relies on the classic FFT algorithm. We recently extended ICEPS with pairwise cross-correlation (BICEPS) and additional correlation-based features that distinguish different spatial pattern classes.
Auto- and cross-correlation techniques are widely used and well-understood tools in classical image analysis. With its favorable characteristics compared to several current tools, ICEPS and BICEPS should find their niche in the spatial transcriptomics toolbox.
[1] Weber, L.M., et al. Nat Commun 14, 4059 (2023).

Recent advances in single-cell technologies have greatly improved our understanding of biological systems. However, accurate cell type identification remains challenging due to heterogeneous transcriptomic profiles and batch effects.
To address these challenges, we propose REFLECT (REpresentation of Features via Latent Embedding and Clustering Training), a two-stage deep learning framework for robust and generalizable cell type classification in single-cell RNA-seq data. REFLECT incorporates our in-house feature selection method, PreLect, which identifies stable and biologically meaningful genes in highly sparse data. These genes serve as anchors to enhance cross-dataset consistency and reduce variability in highly variable genes.
In the pretraining stage, REFLECT applies biologically motivated multi-view augmentation to simulate gene expression variability and sequencing dropout, and leverages a VICReg-based objective to learn invariant yet diverse representations, preventing collapse and promoting a well-dispersed latent space. In the fine-tuning stage, REFLECT combines supervised classification with a triplet-center loss to enforce structured latent organization.
We evaluated REFLECT on three PBMC datasets and the Human Lung Cell Atlas, achieving performance comparable to scANVI and superior to CellAssign and SingleR in F1-score. Notably, REFLECT reduces training and inference time compared to scANVI and CellAssign, while producing structured latent representations with improved class separability and stability.
Overall, REFLECT provides an automated and robust framework for cell type annotation, reducing the need for manual curation and facilitating downstream single-cell analysis.

Why mammals cannot regenerate limbs like amphibians do presents a long‑standing puzzle in biology. To uncover the underlying differences, we compared amputation responses of embryonic mouse (Mus musculus) and Xenopus laevistadpole limbs. Lowering environmental oxygen or stabilizing the oxygen‑sensitive hypoxia‑inducible factor 1A (HIF1A) induced rapid wound healing in mouse limbs, accompanied by altered cellular mechanics, metabolism, and a histone landscape that primed regenerative cell states. Comparative transcriptomic and epigenomic analyses revealed a conserved, amphibian‑like features enriched for wound‑healing, histone modifications, and metabolic functions, activated under low oxygen conditions in mouse. Conversely, Xenopus tadpole limbs retained these features even under high oxygen levels; their reduced oxygen‑sensing capacity was associated with decreased expression of HIF1A‑regulating genes, as identified through cross‑species transcriptomic profiles. Our results thus propose species‑specific oxygen‑sensing capacity as a fundamental, targetable mechanism that can unlock latent regenerative programs in mammals. By integrating multi‑species scRNA‑seq, time‑resolved scMultiomics, and epigenomic profiling, we provide a bioinformatically informed, genome‑scale framework linking oxygen sensing to the activation of regenerative cell states in across vertebrates.

Rice (Oryza sativa L.) is a key model for understanding cereal floret development.The transition from vegetative to reproductive growth involves sequential conversion of the shoot apical meristem into inflorescence (IM), branch (BM), spikelet (SM), and ultimately floret meristems (FM). While several transcriptional regulators governing these transitions are known, the dynamic interplay between transcriptional and post-transcriptional regulation remains poorly understood.
We defined three developmental stages-Early (0.2–0.5 cm), Middle (0.5–1 cm), and Late (1–2 cm) based on panicle length and performed RNA-seq and nucleoproteome (LC-MS) profiling. These datasets were integrated with published small RNA and degradome data to achieve a comprehensive multi-layered analysis. Integrative profiling identified 6,368 nuclear proteins, of which 4,631 exhibited significant temporal dynamics. Transcription factor families displayed distinct regulatory trends: MADS-box proteins increased during floral organ differentiation, whereas SPLs and RFL showed progressive decline, indicating stage-specific regulatory reprogramming. Notably, modest mRNA–protein correlation highlighted extensive post-transcriptional control.
Meta-analysis of miRNA and alternative splice isoform abundances identified 104 inversely correlated mRNA–miRNA pairs, supported by degradome evidence. Isoform-level analysis revealed widespread alternative splicing and stage-specific isoform switching independent of total transcript abundance, often correlating with protein-level changes, exemplified by PCF1.
To move beyond descriptive analysis, we constructed a multi-layered gene regulatory network integrating transcriptional, post-transcriptional, and isoform-level regulation. Furthermore, we propose an AI-driven framework integrating user-generated transcriptomic data with public datasets to infer context-specific regulatory networks.
Together, this study establishes a systems-level understanding of rice floret development and introduces a scalable, AI-assisted paradigm for decoding complex gene regulatory architectures.

MYC deregulation in PDAC has been shown to drive tumorigenesis and immune evasion by limiting tumor antigen expression, reducing T cell infiltration, and maintaining an immunosuppressive tumor microenvironment (TME). However, the mechanisms by which MYC orchestrates immune evasion remain incompletely understood.
To address this, we combined bulk, single-cell, and spatial transcriptomics with metabolomics to characterize tumor-immune interactions in an orthotopic KPC model of pancreatic cancer. Analysis across these modalities revealed that MYC-mediated immune evasion does not merely depend on suppression of antigen presentation. Instead, MYC depletion restores amino acid availability in the TME. This is associated with increased expression of amino acid-responsive gene programs across multiple immune cell populations.
Mechanistically, chromatin binding and transcriptomic analysis identifies MYC/MIZ1 driven repression of lysosomal genes as a key mediator of this phenotype. MYC increases tumor reliance on extracellular amino acids by limiting lysosomal protein turnover. Consistently, MYC variants that fail to repress lysosomal genes maintain proliferation in vitro, but do not support immune evasion or sustained tumor growth in vivo.
Finally, by selectively and acutely inhibiting transport of soluble amino acids into tumors, we restore amino acid availability in the TME, enhance immune activity, and induce rapid, CD4+ T cell-mediated tumor clearance, resulting in long-term survival.
Together, these results demonstrate how integrative analysis of multi-modal datasets can uncover functional couplings between tumor-intrinsic transcriptional programs and immune cell metabolic states. This framework links MYC/MIZ1-dependent regulation to ecosystem-level metabolic constraints and suggests that disrupting these regulatory interactions may reverse immune suppression in PDAC.

Despite longstanding focus on single nucleotide variation, longer and more complex variant types are expected to underlie a disproportionate amount of mammalian trait variation. However, identifying functional, larger variants has proved challenging. We used the Survey of Regulatory Effects (SuRE) approach, a massively parallel reporter assay that tests the ability of individual genomic DNA fragments to initiate transcription in an otherwise promoterless plasmid, to quantify the transcriptional activity of >1.5 billion fragments from across both cattle subspecies. We implemented a novel genome graph-based, haplotype-aware SuRE analysis pipeline to screen >15 million genetic variants in primary bovine cells, and generated a catalogue of >150,000 regulatory variants, including a number potentially affecting production-relevant traits. Because SuRE tests many overlapping genomic fragments, our analysis could fine-map likely causal variants at high resolution, and disentangle clustered but independent regulatory effects. Our graph approach considerably improved identification of genome fragments carrying longer genetic variants, enabling us to include in our analysis 1.9 million insertion/deletion mutations, longer substitutions and structural variants, and demonstrate that larger variants are up to four times as likely to have a regulatory impact. We also identified regulatory effects from multi-allelic variants, highlighting that these represent functional diversity often missed by standard fine-mapping approaches. Transfecting the cattle libraries into matched human cells revealed general conservation of regulatory impact between species, but identified several variants with species-specific effects. Our work has produced a high-resolution atlas of cattle regulatory variants and demonstrates a valuable alternative approach for identifying and prioritising more complex causal variants.

iNKT cells represent a powerful T-cell subset orchestrating innate and adaptive immune responses, making them attractive targets for immunotherapies; however, their transcriptomic characterization following expansion can be influenced by bioinformatic tool choice. To establish a definitive characterization of iNKT cells expanded in IL-2 vs. IL-15, we implemented a high-resolution pipeline to validate identities across independent computational frameworks. Single-cell RNA sequencing data were analyzed using a parallel-processing architecture within Seurat . To eliminate algorithmic bias, the dataset was processed using two normalization techniques—Log-Normalization and SCTransform —and two independent integration workflows: Seurat and Harmony. Cellular annotations were derived through a tripartite consensus strategy incorporating automated classification via SingleR, manual marker profiling, and reference-based mapping to Mavers et al. (2025) data. Despite divergent mathematical assumptions, both strategies yielded a converged identification of ten clusters across four lineages: Th1-, Th17-, NK-like, and Th2-like iNKT cells. Quantitative analysis consistently demonstrated that IL-15 expansion modulates the cellular product, inducing marked expansion of the Th17-like population (48.5% in IL-15 vs. 35.7% in IL-2) and enriching a high-potency NK-like cytotoxic cluster (8.3%). Sub-clustering of the proliferating fraction (Cluster 4) confirmed that IL-15 drives a superior effector-memory trajectory. Pseudobulk analysis confirmed that IL-15-expanded cells exhibit a robust cytotoxic profile (GZMB, PRF1) while downregulating exhaustion indicators like BTG1. This multi-layered validation ensures maximum confidence in the functional quality of these therapeutic cell products.
Funding: Charles University Prague, SVV – 2025 No 260773 and Ministry of Health of Czechia in cooperation with the Czech Health Research Council under project No NW24-03-00079.

Cancer is a complex disease with diverse causes and clinical manifestations, making the identification of effective treatments a persistent challenge. Comprehensive genomic profiling assays are currently the gold standard for guiding targeted therapies; however, many patients fail to respond to recommended drugs despite detailed characterization of their mutational profiles. This limitation highlights the need for more informative predictive approaches.
Experimental and computational drug screening methods can support Molecular Tumor Boards in prioritizing treatment strategies, yet most existing predictors rely primarily on omics and gene expression data and show limited performance. Notably, alternative splicing-recognized as a key hallmark of cancer and a driver of drug resistance-remains largely underexplored in this context.
In this ongoing project, we aim to address this gap by developing a drug response and resistance prediction framework that integrates alternative splicing analysis with interpretable machine learning. Our work leverages large-scale public resources, including Cancer Cell Line Encyclopedia and The Cancer Genome Atlas, to systematically incorporate splicing-derived features alongside conventional molecular data. Current efforts focus on data integration, preprocessing pipelines, and baseline model development, while subsequent steps will evaluate the added predictive value of alternative splicing.
By incorporating this overlooked layer of regulation into predictive models, this study aims to improve the prioritization of therapeutic options and provide more informative, interpretable tools to support clinical decision-making.

Background:
Alzheimer's disease (AD) shows pronounced sex differences, but underlying mechanisms remain unclear. Estrogen receptor beta (ERβ; Esr2) is neuroprotective, yet its role in sex-specific amyloid pathology remains poorly defined. Using the AppNL G F mouse model, we examined how Esr2 loss affects hippocampal processes in males and females.
Methods:
Six-month-old male and female AppNL G F and AppNL G F/Esr2KO mice underwent behavioral testing (Y-maze, elevated plus maze, and fear conditioning). Pathology was assessed via immunostaining, cytokine assays, and RT-qPCR. Single-cell RNA sequencing of hippocampal tissue enabled cell-type annotation, ligand-receptor inference, and sex-stratified differential expression and pathway analyses.
Results:
Esr2 loss impaired associative memory in males only. Male AppNL G F/Esr2KO mice showed increased hippocampal Aβ plaques and higher insoluble Aβ42/Aβ40 ratios, while females were largely unaffected. Single-cell analysis revealed disrupted endothelial communication, particularly in males, with a shift toward APP- and PSAP-related signaling. Astrocytes and vascular cells showed both shared and sex-specific responses. Transthyretin (Ttr) was the most strongly sex-modulated gene, reduced specifically in male KOs and associated with increased amyloid burden.
Conclusions:
ERβ protects the male hippocampus by maintaining microglial, astrocytic, and vascular homeostasis, potentially via TTR-mediated Aβ clearance. Females appear resilient, possibly through ERα compensation. ERβ-TTR signaling emerges as a sex-specific neuroprotective axis in AD.

RNA-binding proteins (RBPs) are central regulators of post-transcriptional gene expression. While transcriptome-wide RBP binding studies have focused on untranslated regions and introns, recent CLIP-seq data indicate that coding sequence (CDS) binding is widespread and, for a subset of proteins, exceeds UTR binding, pointing to an underexplored layer of regulation. To characterize this protein class, we developed a computational workflow to identify CDS-binding RBPs from public human CLIP-seq data and reveal their shared and distinct regulatory features.
By defining CDS-binding RBPs as those with more than 40% of binding sites in coding regions, we find that their transcript-wide profiles show recurrent enrichment near the CDS start, consistent with roles in translation initiation and early elongation. Additionally, comparative motif discovery across CDS and non-CDS binding sites reveals distinctive motifs, while clustering across all motifs identifies motif families associated with distinct target-gene functions - including translation, transport, and RNA metabolism. For many RBPs, motifs identified in their binding sites map to distinct functional clusters, evidence of multiple regulatory modes per protein. Across CDS-binding RBPs, a purine-rich motif family is shared, suggesting a common sequence feature of CDS binding. Finally, by placing motif-associated sites in coding coordinates and examining their codon context, we identify cases in which binding recurs at the same codon, suggesting that some CDS-binding events may interface with codon-dependent regulation.
Together, these results establish CDS-binding RBPs as a heterogeneous class of post-transcriptional regulators and reveal sequence, positional, and codon-linked features that may shape their regulatory activity within coding regions.

A common technical artifact in droplet-based single-cell RNA sequencing (scRNA-seq) is the capture of two cells in the same droplet, termed doublets. Doublets may present as hybrid cell types, and their removal is a critical step during quality control. Another form of contamination is ambient RNA, which originates from other cells in suspension. Like doublets, the presence of ambient RNA results in unexpected gene expression patterns. Despite similarities between ambient contamination and doublets, no formal investigation on how doublet detection is affected by ambient RNA correction has been conducted.
Here, we evaluate the impact of ambient RNA correction on the performance of doublet detection using a probe-based, multiplexed dataset. We apply three of the most widely-used methods for removing ambient RNA (decontX, soupX, cellbender), in conjunction with six doublet detection methods (scDblFinder, scds, bccds, cxds scrublet, DoubletDetection). We find that, with the exception of cellbender, more true doublets are detected when doublet detection is performed on data without ambient RNA correction. Particularly for decontX, ambient RNA correction leads to an increased proportion of doublets retained in the dataset. We also find that decontX removes proportionally more ambient RNA from doublets that are retained after decontX processing, suggesting that ambient RNA correction may partially mask doublet signal, leading to reduced doublet detection performance.
Our results show that doublet detection is impacted by the choice of the ambient RNA correction tool and we suggest that in many settings doublet detection should be performed prior to ambient RNA correction.

Public repositories such as the Gene Expression Omnibus (GEO) contain many transcriptomics experiments that could be reused for downstream analyses such as validation and meta-analysis. However, reuse depends on identifying compatible experiments, which requires a clear representation of their experimental context. In practice, this context is often incompletely and inconsistently recorded across heterogeneous free-text metadata fields and linked publications. Although recent language model- and ontology-based methods have improved the extraction and normalisation of key experimental variables, it remains unclear whether their outputs are sufficiently complete and specific to assess compatibility for the intended reuse. We investigate this problem using human microarray studies from the GEO GPL570 platform, including a subset with manually curated, ontology-mapped annotations from the Gemma transcriptomics database. Using this curated subset, we evaluate extraction methods in terms of coverage, relevance, and experimental-group assignment: whether they recover the annotated experimental information, avoid biological details that are not relevant for reuse, and assign extracted variables to the experimental groups they describe. We then apply the resulting pipeline across the broader collection of GPL570 studies to examine where relevant information is reported and how often the documentation provides enough information to assess experiment compatibility. This work provides a foundation for developing and evaluating metadata-driven methods to identify compatible transcriptomics experiments and support scalable reuse of public transcriptomics data.

Single-cell and spatial transcriptomics have dramatically changed how we profile RNA from heterogeneous biological samples. Combining these modalities with long-read RNA-Seq promises to enable the discovery and quantification of individual RNA isoforms at the single-cell and spatial level. However, highly multiplexed experiments and high-resolution spatial modalities, such as Visium HD, generate a limited number of reads for each cell or spot. This constitutes a major challenge for transcript discovery and quantification with existing approaches, which often lack the scalability for large cohorts and the statistical power for samples with low single-cell or spots sequencing depth.

Here, we present Bambu-Pipe, a Nextflow pipeline that performs end-to-end transcript discovery and quantification from multi-sample single-cell and spatial long-read RNA-Seq data. Bambu-Pipe enables the parallel processing of large sample cohorts and introduces an improved Expectation-Maximization (EM) algorithm that utilizes information from both individual cells and cell clusters. By leveraging cluster-level information—such as cell types or regional spots in Visium HD—the pipeline overcomes data sparsity to provide more accurate quantification estimates at the cluster level than methods relying solely on single-cell depth. Furthermore, Bambu-Pipe is specifically optimized to handle the computational demands and low signal-to-noise ratios inherent in high-resolution spatial datasets containing millions of spots. Together, Bambu-Pipe provides an easy-to-use, efficient, and accurate framework for analyzing isoform expression across multiple datasets and replicates from long-read RNA-Seq.

Gene regulatory networks (GRNs) provide conceptual frameworks for understanding how cells sustain their identity and respond to disease. Coexpression analysis is a foundational approach in GRN inference, providing statistical associations that reflect the coordinated activity of transcription factors (TFs) and their putative targets. However, the reproducibility of TF-gene coexpression patterns across datasets and their conservation across species remain poorly characterized. To address this, we designed a multi-level evidence framework that prioritizes coexpression patterns through the systematic integration of functional genomics data. Using a harmonized collection of 19 million single nuclei from 4,537 human brain samples spanning 23 studies, we first prioritized robust and reproducible cell-type-specific coexpression patterns. To refine these candidates, we integrated TF-binding evidence from Unibind, functional perturbation data from Gemma database, and cell-type chromatin stat data from CATlas. Finally, we performed comparative analyses with mouse orthologous networks to quantify evolutionary conservation, prioritizing regulatory relationships that are most likely to translate across species. Applying this framework to Alzheimer's disease, we prioritized coexpression rewiring within microglial regulatory states, where specific TFs shift their target associations in the disease context. These results show that while core coexpression patterns remain stable, disease states show loss of coexpression patterns and the rise of new, disease-driven relationships. This work provides the first systematic atlas of reproducible human brain TF-gene patterns, creating a resource for prioritizing the drivers of brain cell identity and their changes in disease.

Wilms tumor (WT) is the most common renal malignancy in children, with overall survival exceeding 90%. However, approximately 20% of patients relapse, and post-relapse survival drops to 50%. Standard treatment involves chemotherapy and surgery. Chemotherapy causes substantial short- and long-term adverse effects. Histologically, WT comprises blastemal, epithelial, and stromal components, and anaplasia can occur in each. Tumors with blastemal predominance and diffuse anaplasia are classified as high-risk and are responsible for 60% of relapses. About 40% of relapses occur in patients not classified as high-risk histologically. Therefore, there is an urgent need to improve outcome prediction and to identify more specific and less toxic therapeutic targets.
Our goal is to identify markers predictive of relapse and clinical outcome by integrating computational pathology and spatial transcriptomics. In the first aim, we apply a histopathology foundation model trained on patient-derived H&E images to recognize distinct cellular phenotypes within WT. Identified phenotypes are subsequently linked to transcriptional profiles through integration with spatial transcriptomics data.
In the second aim, we characterize the spatial organization of the WT tumor microenvironment using the CosMx Spatial Molecular Imaging platform, profiling approximately 6,000 genes at subcellular resolution. Starting with a pilot cohort of 14 patients, we will expand to compare spatial and transcriptional differences within the blastemal compartment between relapsed and non-relapsed patients, as well as patients with and without metastasis.
This integrative approach aims to uncover novel, spatially resolved molecular drivers of WT relapse, with the potential to refine risk stratification and identify candidate therapeutic targets.

Single-cell RNA sequencing (scRNA-seq) has revolutionized the study of complex biological systems by uncovering cellular heterogeneity and intercellular communication at unprecedented molecular resolution, far surpassing the capabilities of Microarray and bulk RNA-seq technologies. The recent emergence of scRNA-seq atlases has further accelerated these advances.
Here, we present a comprehensive scRNA-seq atlas of 27 inflammatory skin diseases (ISDs) and corresponding healthy skin, developed as part of the Human Cell Atlas consortium and primarily utilizing 10X Genomics data. ISDs affect up to 25% of the global population, yet the shared pathogenic mechanisms underlying these diseases remain poorly understood and only few studies have comparatively investigated multiple ISDs. Atlas integrations provide computational challenges, nevertheless they are invaluable big-data resources, driving the discovery of new cellular and molecular mechanisms. Our atlas integrates in-house and public scRNA-seq datasets from 50 studies, encompassing 441 independently sequenced samples and over two million cells. These extensive numbers enabled the identification of even rare cell types. Thereby, we were not only able to replicate established findings related to cell abundance and gene expression patterns, but could distinguish conserved inflammatory signatures from disease specific alterations. Through downstream analyses, including differential abundance testing, cell-cell communication analysis, and the discovery of shared and disease-specific gene programs, we provide novel insights into ISD pathology and validated novel findings using spatial transcriptomics. This atlas offers a comparative resource not only between healthy and diseased states but also across diverse inflammatory diseases revealing conserved and disease specific mechanisms.

Metabolic RNA labeling has enabled the study of transcriptional regulation and RNA dynamics at scale, leveraging nucleotide analog incorporation into nascent transcripts to quantify newly synthesized and pre-existing RNA fractions. Current analysis frameworks account for the stochastic nature of incorporation but assume constant incorporation rates over labeling intervals. However, a growing body of studies suggests that incorporation rates are not constant but increase monotonically over the labeling interval.

Our improved GRAND3 framework does not only address this issue by modeling nucleotide incorporation kinetics as a function of labeling time. It exploits the increasing frequency of 4sU induced nucleotide substitutions to achieve temporal super-resolution differentiating transcriptional events that occur early or late during the labeling time in single cells.

We used GRAND3 to analyze scSLAM-seq data of human fibroblasts infected with human cytomegalovirus (HCMV) sampled in 2-hour-intervals over the first 8 hours of infection. The model estimates precisely captured the highly coordinated viral transcriptional program during early infection and revealed previously unrecognized temporal coordination in the host defense response. Single-cell trajectories further revealed that temporal heterogeneity in this response is associated with divergent infection outcomes.

Enhancing the temporal resolution of metabolic labeling data computationally without the need to modify existing protocols, establishes GRAND3 as a broadly applicable tool to study RNA dynamics and transcriptional regulation.

Inferring gene interaction networks from transcriptomic data is a central challenge in systems biology. Most existing approaches emphasize population level relationships, thereby obscuring biologically meaningful inter‑individual heterogeneity. While single‑sample network inference methods aim to address this limitation, many rely on the contrasts between reference and perturbed aggregate networks, potentially reducing the signals from the fundamental gene-gene relationships shared across samples. Here, we introduce ssNITE (Single‑Sample Network Inference Through Trees in Ensemble), a novel framework inspired by GENIE3, a top-performing population level gene regulatory network inference method in the DREAM challenges, for constructing sample‑specific gene interaction networks. ssNITE leverages Random Forest regression models trained on bulk transcriptomic data to extract individualized gene dependencies by tracing sample‑specific decision paths in the trees of the forest. Applied to breast cancer transcriptomic data from The Cancer Genome Atlas, ssNITE produces networks that better aligns with known clinical information than its peers. Particularly, features derived from ssNITE networks exhibited stronger associations with clinical phenotypes, improved predictive performance in molecular subtype classification and tumour purity estimation, and enhanced prognostic value for overall survival. By directly coupling predictive modelling with individualized network reconstruction, ssNITE provides a scalable and interpretable approach for sample‑specific network inference, advancing methodological foundations for individualised network medicine.

Background
Facial palsy is the most common cranial neuropathy in children, frequently caused by neuro-borreliosis or Bell's palsy. Differentiation requires lumbar puncture to assess cerebrospinal fluid (CSF) cell count and Borrelia burgdorferi antibody index. We aimed to investigate CSF and plasma proteomics to explore the potential for early, accurate, non-invasive diagnostics.

Methods
Samples were collected from 129 children with facial palsy at the first hospital visit. The causal disease was later diagnosed as by Bell's palsy, neuroborreliosis, unverified aetiology, or Ramsay Hunt syndrome. Mass spectrometry was applied to measure the proteomes of each sample. Differential abundance was assessed between (i) facial palsy of unverified aeti-ology and definite neuroborreliosis, (ii) Bell's palsy and neuroborreliosis, and (iii) neuroborrelio-sis with and without facial palsy, using multivariate linear model corrected for sex and age. Functional interpretation was carried out through enrichment analysis and tree-based cluster-ing of enriched terms. A plasma-based diagnostic classifier was developed using machine learning with feature selection and cross-validation across 14 algorithms.

Results
Proteomic profiles indicated that children with facial palsy of unverified aetiology likely repre-sent neuroborreliosis, supporting empirical treatment. Neuroborreliosis showed significantly elevated CSF proteins related to immune pathways, whereas those with Bell's palsy had ele-vated CSF proteins associated with extracellular matrix organization. Within neuroborreliosis, patients without facial palsy showed stronger enrichment of immune and transport pathways. A diagnostic classifier based on 14 features achieved AUROC of 0.77 and MCC of 0.56, with CD248 as the strongest contributor, supporting a diagnostic step-down from lumbar puncture to plasma-based diagnostic screening.

Copy number-driven cancers are characterised by extensive chromosomal instability, creating challenges for molecular quantitative trait loci (molQTL) analysis. Sample-specific breakpoints can separate a gene from nearby regulatory variants, while allele-specific copy number changes alter the contribution of the reference and alternative alleles. Modelling variant effects in these tumours therefore requires accounting for structural context and allelic imbalance.

To address this, we developed SegmentQTL, which integrates sample-specific segmentation with allele-resolved dosage information. For each gene-variant pair, only samples in which the variant and gene remain on the same segment are considered. SegmentQTL models total local dosage and allelic imbalance, distinguishing effects driven by overall copy number from those arising through preferential retention or amplification of one allele.

SegmentQTL includes a finemapping mode using a modified Elastic Net and stability selection to accommodate segment-filter-induced missingness. Because different variants remain linked to the gene in different samples, standard multi-variable models would require either imputation or restricting analysis to samples shared across variants. SegmentQTL instead jointly analyses cis variants without imputation or sample loss, prioritising reproducible signals across bootstraps.

Applied to high-grade serous carcinoma, SegmentQTL highlighted several survival-associated genes. Among these, high EIF2AK1 expression was associated with poorer overall survival (adj. p = 0.01). SegmentQTL identified three independent variants regulating EIF2AK1, increasing adjusted R² by 0.07 beyond gene dosage alone (adj. R² = 0.54). Experimental knockdown increased sensitivity to paclitaxel and carboplatin, linking genetically driven EIF2AK1 upregulation to chemotherapy resistance. SegmentQTL thus offers a path from association to mechanism.

Cardiotoxicity is a severe complication of anthracycline therapy, yet mechanisms driving myocardial vulnerability remain partially unknown. Patients with underlying cardiac disease show higher incidence and severity of cardiotoxic events, underscoring the need for human models that capture such risk modifiers. Here, we generate 3D engineered cardiac tissues using melt electrowriting (MEW) for the systematic assessment of Doxorubicin injury. By combining human induced pluripotent stem cell-derived cardiomyocytes and cardiac fibroblasts within fibrin-reinforced MEW scaffolds, we show that clinically relevant Doxorubicin exposure reproduces key hallmarks of toxicity, including diminished viability, impaired contractility and sarcomere disarrangement. To model a common comorbidity, we induced a fibrotic microenvironment using stimulation with TGFβ, generating tissues that exhibit structural remodeling and functional decline characteristic of diseased myocardium. A transcriptomic analysis indicated upregulation of extracellular matrix-related processes and pro-fibrotic regulators, with a supression of cardiomyocyte contractile and vascular-associated programs, closely mirroring remodeling processes from ischemic cardiomyopathy in vivo. When exposed to doxorubicin, fibrotic tissues display significantly greater toxicity than non-fibrotic constructs, reflecting the heightened susceptibility observed in patients with pre-existing fibrosis. The integration of transcriptomic profiling provided mechanistic insight into these differential responses, revealing enhanced p53 signaling, repression of contractile identity and early matrix destabilization in response to Doxorubicin, while sharing a canonical anthracycline response characterized by cell-cycle arrest, stress and mitochondrial injury.

In summary, our all-human in vitro model recapitulates the enhanced cardiotoxicity profile of cardiac patients, positioning itself as a promising translational platform for mechanistic elucidation and management of cancer therapy-related cardiovascular risk.

Attaining sufficient maturation and biomimetism to resemble adult myocardium remains a key challenge in cardiac tissue engineering. Here, we consider a multi-omic approach, including scRNA-seq and Spatial Transcriptomics (Visium technology), to improve our understanding of 3D engineered heart tissues (EHTs) generated using Melt Electrowriting and hiPSC-derived cardiomyocytes (hiPSC-CMs). We obtained datasets from hiPSC-CMs and cardiac fibroblasts (hiPSC-CFs) at day 0, and their combination in a 9:1 proportion in 3D EHTs cultured for 30 days. We optimized the downstream computational workflow, primarily using Seurat, and identified differential gene regulatory networks (GRN) via SCENIC and the machine learning algorithm SimiC. Subsequently, we collected and curated embryonic human cardiac datasets to benchmark our EHTs, define gene signatures for various cardiac phenotypes, and identify corresponding developmental stages.

We uncovered cellular heterogeneity within the initial hiPSC-CMs, including an off-target fibroblast population that persists after 30 days in EHTs, while most of the initially seeded hiPSC-CFs disappear. Cardiomyocytes at day 30 resemble in vivo embryonic ventricular and pacemaker phenotypes, showing features of maturation relative to day 0. However, GRN analysis revealed that these EHTs are closer to the earliest in vivo reference stages in terms of maturity. Spatial transcriptomics identified a uniform distribution of cardiomyocyte phenotypes while displaying an enrichment of fibroblasts near scaffold fibers.

This study highlights the power of integrating single-cell, spatial transcriptomics and computational approaches to resolve tissue heterogeneity, cell-type distribution and to identify the key targets required to advance in tissue design, maturation and resemble adult myocardium with higher fidelity.

Background:
Next Generation Sequencing (NGS) technologies, such as microarray and RNA sequencing, are key technologies in biomedical research. Pathway analysis is widely used to interpret such data, yet it remains unclear whether systematic differences exist between ‘omics technologies and the pathways they identify. This study investigates potential biases by analyzing over 1,500 datasets from GREIN and Expression Atlas using ReactomeGSA.

Methods:
Public ‘omics datasets were analyzed with ReactomeGSA using two pathway algorithms, PADOG and Camera. PADOG down-weights genes shared across pathways to improve specificity, while Camera accounts for gene–gene correlations, enhancing enrichment accuracy. The analysis focused on how technology type and algorithm choice influence pathway detection.

Results:
Preliminary findings indicate variation in the number of detected pathways across technologies and algorithms. PADOG and Camera differed in both the number and type of significant pathways identified, with PADOG generally detecting more pathways, suggesting higher sensitivity. Differences between technologies were also observed, pointing to potential platform-specific biases. Further analysis is ongoing to assess the extent and implications of these differences.

Conclusions:
Both the choice of technology platform and analysis algorithm influence pathway analysis results, with each method having its own strengths and limitations. PADOG appears to offer higher sensitivity, while Camera may provide more conservative estimates, highlighting more robust pathways. These initial findings suggest that cross-validation using multiple platforms and algorithms is essential to ensure reliable biological interpretations.

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

Transcriptomic analysis has allowed researchers to investigate how an individual's gene expression is modulated by Mycobacterium tuberculosis infection and antimycobacterial drugs. However, most studies using short-read RNA sequencing analyse whole gene expression, only one of several gene regulatory layers captured by this technology. Whole gene expression analysis cannot ascertain how alternative transcript usage and 3' UTR lengths potentially impact a gene's post-transcriptional function and localisation. To investigate these new data modalities, we performed paired bulk RNA sequencing on peripheral blood mononuclear cells (PBMCs) collected from 24 Tanzanian patients immediately before and five months after treatment for pulmonary TB. We then bioinformatically compared bulk expression, splicing patterns and 3' UTR shifts across treatment using DESeq2, apatools and rMATs.

Our bulk transcriptomic analysis confirmed previously published gene signatures, notably the downregulation of immunoglobulin-associated genes after treatment. We also identified a number of non-differentially expressed genes with variations in 3' UTR length and splicing patterns across treatment, revealing regulatory signals complementary to expression-based analyses. More specifically, we observed a highly constrained splicing pattern that was relaxed with treatment, as well as a global shortening of 3' UTR lengths post-treatment. This suggests a reduced post-transcriptional regulatory pattern upon recovery. These findings highlight the value of non-expression bulk analyses in expanding our understanding of gene expression shifts. This multipronged approach could deepen our understanding of tuberculosis biology.

Computational analysis of high-resolution CLIP-seq data has enabled precise
mapping of RBP binding sites and cis-regulatory elements underlying post-
transcriptional RNA regulation.

Recently, the field has shifted with the emergence of RNA foundation models trained on vast unlabeled RNA sequences that enable holistic modeling of RNA function and in silico hypothesis generation.
Leveraging such models represents the next frontier, allowing prediction of
multiple RNA regulatory processes, as well as in silico design of RNA-based
therapeutics.
Building on the premise that RNA function is encoded in its interaction partners,
we developed Parnet, a multi-task model that densely represents the RNA
interactome. Extending our earlier single-task model RBPNet, Parnet is trained
end-to-end on raw CLIP-seq profiles from hundreds of RBPs to predict
genome-wide binding profiles directly from sequence. In contrast to
unsupervised RNA language models, Parnet learns embeddings that capture
the combinatorial RBP codeunderlying post-transcriptional regulation.
As a result, Parnet generalizes across downstream tasks including splicing,
intron retention, and mRNA translation and degradation often with minimal or
no fine-tuning. In a therapeutic context, we show the potential of Parnet to
computationally prioritize high-impact ASO hotspots in concrete applications
and to design mRNA UTRs that achieve desired expression, stability, and
regulatory profiles.

ASAP is a web platform for reproducible single-cell and bulk transcriptomics analysis. Recent work improves scalability, the user experience, and alignment with community metadata standards, with emphasis on how users see, explore, and annotate their data. The portal is built on Ruby on Rails 8.0.

ASAP is an SIB resource and a core participant in scFAIR, promoting FAIR practices for single-cell data and metadata. Together with partners including Bgee, CELLxGENE, and the EBI Single Cell Expression Atlas we align on conventions; cell lines use Cellosaurus nomenclature, and an online validator against the scFAIR schema is being integrated. A generic pipeline framework runs single-cell and bulk RNA-seq and is designed to extend toward spatial transcriptomics, metabolomics, and proteomics, supporting exploration from quantitative matrices to shared, visually grounded discovery.

In March 2026 we released a new interface centered on interactive embedding maps in reduced dimensional space, with coloring by sample metadata or gene expression and smooth navigation. This new release also includes a wider list of file formats for data input, a revised and more intuitive analysis workflow, and enhanced capabilities for collaborative annotations in the context of analysis results. ASAP enables teams annotating a given dataset to converge on labels such as cell types and to keep a joint record of decisions as the work evolves. Projects can be made public so analysis and annotation are visible for community feedback. Once public, community members can add further annotations, enriching the resource beyond the original team's work.

mRNA therapy offers promising capabilities in biomedicine, although its efficacy and safety depend on off-target effects, including non-specific expression of the protein load.
5' and 3' UTRs affect mRNA translation efficiency and stability through interactions with miRNAs and RNA-binding proteins, potentially enabling rational design of sequences with cell-type-specific activity.
Here we employed a rich dataset obtained with a massively parallel reporter assay, MPRA, which profiled 5' and 3' UTR activity in multiple cell lines (MDA-MB-231, HepG2, Jurkat, NALM6, SW480) to train predictive and generative deep learning models and develop PARADE (Prediction And RAtional DEsign of UTRs). Running TF-MoDISco with PARADE along with classic motif discovery with MPRA data revealed hundreds of known and dozens of novel RNA motifs.
Next, we trained and compared generative deep learning models, including cGAN, Fast SeqProp, Flow Matching, and cold diffusion; the latter was chosen for validation alongside a classic genetic algorithm. The newly designed UTRs showed significant correlation between the desired and achieved activity and a major improvement in cell-type-specificity. The most successful designs were validated in vitro in traditional luciferase reporter assays and in vivo in mice injected with nanoparticles containing Fluc-encoding mRNAs bearing PARADE‑engineered UTRs.

Gene regulatory networks convey the relationship between regulators such as transcription factors and the expression of individual genes. They are useful for understanding changes in cell phenotype, particularly for complex diseases such as cancer.

Although methods to generate gene regulatory networks from bulk gene expression data are common, it is more challenging to deal with single cell and spatial transcriptomics data due to their sparsity. Here we present an approach to generate spatially resolved gene regulatory networks from spatial transcriptomics data called STOAT (Spatial TranscriptOmics to Assess Transcriptional regulation), building on the methods previously developed in the lab (PANDA, LIONESS) and also a recently published tool that helps with the generation of prior networks (SPONGE). STOAT adapts the LIONESS approach of generating sample-specific networks to the spatial transcriptomic reality of very sparse gene expression through the use of gene expression averaging.

We show supporting evidence for the validity of this method and also its application to a publically available datasets and samples from our collaborators, ranging from healthy samples to triple-negative breast cancer. In particular, we highlight the differences in the classification of individual spots using gene regulatory networks and gene expression data. We also identify a subset of spot clusters which are identified using both approaches, and discuss why this isn't the case for all of them and how our approach could be further improved.

In recent years, spatial proteomics and transcriptomics technologies have rapidly advanced, enabling the quantification of tissue architecture with the overarching goal of informing clinical decision-making. A widely used spatial feature for quantifying tissue organization is the pairwise neighbor preference (NEP) of cell types, commonly referred to as co-occurrence or colocalization. Various methods to infer NEP have proved their utility in spatial omics studies, but despite their broad usage, no clear guidelines exist for selecting one method over the other. Here we show the diversity of NEP method results and present a comprehensive guide through existing NEP analysis methods. We evaluate combinations of their underlying analysis steps and introduce COZI, a novel combination of analysis steps not previously described. We assess existing NEP methods and COZI on two aspects: (1) their ability to distinguish different tissue architectures and (2) their ability to recover the directionality of NEPs using two tissue simulations and two biological datasets. We highlight method-specific abilities and limitations in creating biologically interpretable insights and find that COZI uniquely enables both sensitivity and directionality to perform NEP analysis. I will present COZI and its application in cardiology and oncology.

Spatial transcriptomics has enabled gene expression profiling in a spatial context, but accurately mapping cell types in situ remains challenging. Although numerous deconvolution methods have been developed, comprehensive benchmarking across diverse tissue architectures and technologies is still lacking. Here, we present spDDB (https://github.com/Zafar-Lab/spDDB), benchmarking study for 21 state-of-the-art spatial deconvolution methods, including 7 newly developed methods, evaluated across an exhaustive repository of 37 datasets. Existing studies often rely on oversimplified simulation strategies, necessitated by the absence of ground-truth cell-type proportions. To address this limitation, we developed SynthST, which leverages a deep graph attention network to generate synthetic cell-type proportions matrices that preserve spatial dependencies learned from real spatial transcriptomic data. These simulated proportions are then used to construct paired spatial gene-expression datasets by sampling expression profiles from matched single-cell data and modeling UMI count distributions using Gaussian process framework.
To enable robust evaluation, we further compiled a curated dataset repository spanning brain, cancer, and organ tissues across multiple technologies, for which realistic simulated datasets were generated using SynthST. In addition, spDDB introduces a comprehensive evaluation framework comprising five spatial bivariate metrics (including a novel bivariate Geary's C), three rare cell-type metrics, and five cell-shape characterization metrics, enabling more holistic assessment of deconvolution methods. Further, we introduce a benchmarking pipeline evaluating 18 spatial domain detection methods across 36 datasets (21 real and 15 simulated) spanning six technologies. Finally, we provide practical guidelines and recommendations to assist researchers in selecting optimal methods across diverse experimental settings.