Estimating RNA modifications from Nanopore direct RNA sequencing data is a critical task for the RNA research community. However, current computational methods often fail to deliver satisfactory results due to inaccurate segmentation of the raw signal. We have developed a new method, SegPore, which leverages a molecular jiggling translocation hypothesis to improve raw signal segmentation. SegPore is a pure white-box model with enhanced interpretability, significantly reducing structured noise in the raw signal. We demonstrate that SegPore outperforms state-of-the-art methods, such as Nanopolish and Tombo, in raw signal segmentation across three large benchmark datasets. Moreover, the improved signal segmentation achieved by SegPore enables better performances in m6A modification estimation.

Understanding tissue organization and cellular composition is key to deciphering organ function and disease mechanisms. Spatial transcriptomics enables genome-wide, spatially resolved RNA measurements, yet technologies like 10x Genomics Visium capture multiple cells per spot, requiring computational deconvolution to infer cell-type composition. We present spacedeconv, an R package that streamlines the use of multiple first- and second-generation deconvolution tools, supports the integration of multi-modal spatial data, and provides flexible visualization of cellular and molecular patterns. spacedeconv facilitates holistic, interpretable exploration of tissue architecture, revealing spatial niches, cellular interactions, and molecular determinants of tissue function. Using diverse 10x Visium datasets, we demonstrate how spacedeconv's integrated analyses and visualizations enable comprehensive, systems-level views of tissue complexity across organisms and organs, making spatial deconvolution more accessible and informative for the wider research community.

Availability and implementation: spacedeconv is available at https://github.com/omnideconv/spacedeconv

In silico cell-type deconvolution from bulk transcriptomics data is a powerful technique to gain insights into the cellular composition of complex tissues. While first-generation methods used precomputed expression signatures covering limited cell types and tissues, second-generation tools use single-cell RNA sequencing data to build custom signatures for deconvoluting arbitrary cell types, tissues, and organisms. This flexibility poses significant challenges in assessing their deconvolution performance.

Here, we comprehensively benchmark second-generation tools, disentangling different sources of variation and bias using a diverse panel of real and simulated data. Our results reveal substantial differences in accuracy, scalability, and robustness across methods, depending on factors such as cell-type similarity, reference composition, and dataset origin.

Our study highlights the strengths, limitations, and complementarity of state-of-the-art tools, shedding light on how different data characteristics and confounders impact deconvolution performance. We provide the scientific community with an ecosystem of tools and resources, omnideconv, simplifying the application, benchmarking, and optimization of deconvolution methods.

Timecourse single-cell RNA (scRNA) datasets have become indispensable biological tools for studying and uncovering the underlying regulatory
processes driving continuous cell-state transitions such as the epithelial-to-mesenchymal transition. However, these highly fluid, non-linear
trajectories remain difficult to disentangle. While deep neural network models can effectively map complex topologies, they are biological black
boxes, lacking the gene-level interpretability required by experimental practitioners. Here, we introduce scCont, a fully unsupervised contrastive
learning framework for extracting interpretable functional transition units from scRNA-seq data. Rather than relying on explicit temporal
labels or clustering priors, scCont uses an unsupervised k-nearest neighbors approach to project local transcriptomic neighborhoods into a
low-dimensional embedding. During post-training analysis, scCont employs Shapley network attribution to determine gene-latent relationships,
mapping latent dimensions directly into modular gene programs. We quantitatively and qualitatively validate this representation by demonstrating
that scCont captures biologically interpretable dynamics more effectively than baseline methods such as LDVAE. Evaluated across 13 diverse
EMT timecourse datasets, scCont exposes interpretable functional dynamics and effectively maps context-dependent functional groups, helping
to bridge the gap between deep learning representation and translational biological utility. scCont is an open source software available on GitHub
(https://github.com/nramani611/scCont).

Motivation: Differential gene expression (DGE) is a ubiquitous but ill-defined analysis. After transcription is initiated, most mRNAs are spliced immediately in a regulated manner yielding a defined mixture of different transcripts of the gene. Thus, a DGE analysis should also include an analysis of the changes in the transcript composition provided by differential alternative splicing (DAS) analysis.
Results: EmpiReS is an approach for model-free quantification of directed feature fold changes via Empirical error distributions estimated from Replicate Sequencing measurements. We comprehensively assess the performance of EmpiReS for DGE and the more complex doubly differential DAS. For both analyses, we could show that EmpiReS is highly reliable and has an excellent trade-off between sensitivity and precision. It outperforms state-of-the-art methods such as DEXSeq and is orders of magnitude faster which allows it to jointly analyse DGE and DAS which provides a more realistic view of the changes in the cell.
Availability and Implementation: the EmpiReS software is available at
https://github.com/zimmerlab/EmpiReS

Motivation : Beyond coding potential, the signals that make the lncRNA/mRNA boundary learnable remain poorly understood. This work uses interpretable multimodal learning to assess whether locus-derived features, such as transposable element composition as well as non-B DNA motifs and non-canonical secondary structures, can provide discriminative information, and whether hard-to-classify transcripts reflect model errors or intrinsic biological/annotation ambiguity.
Results : We present a structured interpretability experiment using β-LNC, a β-VAE with cross-modal attention over explicit feature-group tokens representing transposable elements and non-B/structure-related genomic contexts. We used this framework to link predictive performance with feature attribution, ablation analyses, and controlled synthetic benchmarks designed to test recovery of known discriminative signals. Interpretability analyses revealed that TE family composition provides a strong genomic-context signal for distinguishing lncRNAs from mRNAs, beyond what is captured by coding-potential metrics alone. Importantly, this signal is supported by two independent analyses: TE-related feature groups rank among the most discriminative by Fréchet distance and produce the largest loss of performance when ablated in controlled synthetic experiments. A cross-release benchmark on held-out GENCODE v47 and v49 transcripts against seven methods (namely: CPAT, CPC2, lncDC, RNAsamba, Orthrus 4-track, and lncRNA-BERT-based models) shows that the framework maintains strong predictive performance, with macro F1 around 0.97, while supporting the added value of multimodal genomic context relative to sequence-only approaches. Finally, transcripts misclassified by all methods reveal a symmetric boundary: some protein-coding transcripts carry lncRNA-like features, whereas some lncRNAs carry coding-like features, suggesting a biologically grounded classification ceiling shared across approaches.
Availability : Code: https://github.com/cbib/beta_vae_lnclassifier; Data: https://doi.org/10.5281/zenodo.18849718.
Contact: mikael.georges@ibgc.cnrs.fr — macha.nikolski@u-bordeaux.fr
Supplementary information: Supplementary Materials are available online.

Motivation: Transformer-based models are increasingly applied to large-scale single-cell transcriptomics, showing strong performance through self-supervised learning on millions of cells. However, most existing approaches treat genes as independent features, and largely ignore prior biological knowledge, which limits interpretability and robustness. In this paper, we explore whether explicitly incorporating gene regulatory information can improve both model performance and biological insight.
Results: We present scTransformer, the first Transformer-based approach that builds a priori knowledge of biological mechanisms into the model's attention patterns. By constraining information flow according to known regulatory structures, the model learns representations that are more biologically meaningful. We evaluate scTransformer on a disease-relevant single-nucleus RNA-seq dataset using supervised cell-type classification. Compared to standard Transformers, our approach improves classification accuracy, enhances separation of cell types in embedding space, and produces attention patterns consistent with known regulatory programs. Overall, our results demonstrate that embedding biological structure into Transformer models can enhance interpretability without
sacrificing performance, offering a principled step toward biologically grounded foundation models for single-cell omics.
Availability: scTransformer, is freely available with an open-source license at https://gitlab.com/sysbiobig/scTransformer
Contact: barbara.dicamillo@unipd.it

Motivation: Most existing workflows for quantifying bacterial gene expression from RNA-seq data rely on mapping reads to a (single) reference transcriptome, typically ignoring strain-level variation. When samples contain unknown or mixed strains, these workflows may introduce reference bias and fail to accurately capture strain-specific gene expression. Pan-transcriptomic approaches address this issue by using pan-transcriptomes as references, but existing solutions require multiple steps for pan-transcriptome construction, indexing, and expression quantification.

Results: We introduce PanXpress, a unified framework for bacterial pan-transcriptomics that performs pan-transcriptome construction and indexing directly from genomic FASTA and GFF annotation files, alignment-free mapping of reads to genes from FASTQ samples, and gene expression quantification. The index, a multi-way Cuckoo hash table storing gapped k-mers with associated genes, preserves diversity on the k-mer level.

Using simulated RNA-seq data from a mixture of Pseudomonas aeruginosa strains, PanXpress achieves mapping recall comparable to alignment-based methods such as Bowtie2 with higher precision and obtains accurate gene expression and log fold change estimates.

On real P. aeruginosa RNA-seq data, using PanXpress' pan-transcriptomic reference increases the proportion of mapped reads and discovered expressed genes. The index of PanXpress is smaller than that of other tools and it provides faster analysis with consistent results, compared to other tools (Salmon, Kallisto, Bowtie2). PanXpress is thus an accurate and efficient method for bacterial gene expression analysis in complex samples.

Plant regeneration is a highly orchestrated process characterised by dynamic cellular reprogramming and lineage plasticity. Although Arabidopsis thaliana demonstrates a remarkable capacity for regeneration following root tip injury, the specific cellular populations and temporal dynamics that govern this response remain poorly defined. This study presents an integrative deep generative framework to identify and characterise regenerative cell states from scRNA-seq data collected across five post-injury time points. Using the resulting representation, we employed a range of complementary strategies, including Latent Dirichlet Allocation (LDA)-based topic modeling, iterative semi-supervised classification with controlled false-discovery rates, entropy-based detection of transitional migratory states, and cluster-level Optimal Transport and Waddington-OT trajectory inference. The topic modeling identified 2,585 high-specificity regeneration-associated cells, which expanded to a total of 3,416 candidates under a constraint of 1\% false discovery. Entropy and transport analyses highlighted maximal lineage instability occurring between 4,9 and 14 hours post-injury, followed by subsequent phases characterised by expansion and stabilisation. The integration of orthogonal signals yielded a core of 906 high-confidence regenerative cells, with notable enrichment in the Cortex, Atrichoblast, and Columella LRC lineages. These findings substantiate a temporally structured, multi-phase model of regeneration that is derived directly from single-cell dynamics.

Motivation: Current position weight matrices and sequence logos may not be sufficient for accurately modeling transcription factor binding sites recognised by a mixture of homodimer and heterodimer complexes.
Results: To address this issue, we developed forkedTF, an R-library that allows the creation of Forked-Position Weight Matrices (FPWM) and Forked-Sequence Logos (F-Logos), which better capture the heterogeneity of TF binding affinities based on interactions and dimerisation with other TFs. Furthermore, we have enhanced the standard PWM format by incorporating additional information on co-factor binding and DNA methylation. Precomputed FPWM and F-Logos are made available in the MethMotif 2024 database, thereby providing ready-to-use resources for analysing TF binding dynamics. Finally, forkedTF is designed to support the TRANSFAC format, which is compatible with most third-party bioinformatics tools that utilise PWMs.
Availability: The forkedTF R-library is open source and can be accessed on GitHub at https://github.com/benoukraflab/forkedTF.

Triple-negative breast cancer (TNBC) frequently develops resistance to chemotherapeutic agents, yet the transcriptional programs underlying treatment adaptation remain incompletely understood. Here we have used an integrative computational analysis combining transcriptomic profiling, pathway enrichment, and machine-learning–based feature selection to identify regulatory programs associated with chemotherapy response in APC-deficient TNBC cells. RNA-seq data were generated from parental MDA-MB-157 cells and two APC knockdown derivatives exposed to control, cisplatin, and paclitaxel conditions. After variance-based filtering of genes, discriminant transcriptomic features were identified using a consensus machine-learning strategy integrating Random Forest permutation importance and Partial Least Squares–Discriminant Analysis variable-importance scores. Intersection of the top-ranked features from both models produced a robust 43-gene discriminant signature separating genotype–treatment states. Functional enrichment analysis of transcriptomic profiles revealed coordinated activation of DNA-damage response pathways, mitochondrial oxidative phosphorylation, inflammatory signaling, and chromatin-regulatory processes in APC-depleted cells. The resulting gene signature included regulators of cell-cycle control and DNA repair (CCNB3, ORC1, E2F2, UNG), cytokine signaling (CXCL2, IL11), and metabolic support, reflecting transcriptional programs associated with chemotherapeutic adaptation. Together, these results demonstrate how integrating pathway-level analysis with machine-learning feature selection can uncover coherent regulatory programs underlying treatment-dependent cellular states. This computational framework may provide a generalizable strategy for identifying transcriptional signatures associated with adaptive responses to therapy in cancer. The code and supplementary data are available at https://github.com/MurliNair/APC-response-programs-TNBC-supplemental-data

Motivation: Enhancer-derived RNAs (eRNAs) and their fusion with protein-coding genes represent a crucial yet understudied layer of transcriptional regulation. eRNAs are typically expressed at low levels, which makes fusion events difficult to detect with conventional fusion detection tools. In addition, these tools are not designed to capture fusion transcripts arising from spatial proximity between distal regulatory elements and gene loci. Reads spanning such regions are also frequently filtered as mapping artifacts. As a result, computational approaches for systematically identifying spatially mediated enhancer-exon fusion transcripts remain lacking.

Methods: We developed ChiMER, a graph-based framework for detecting ChiMeric Enhancer RNAs from short-read RNA-seq data. ChiMER constructs splice graphs with chromatin contact information to introduce enhancer-exon edges and uses graph alignment to search for potential transcriptional paths. A ranking-based scoring module then prioritizes high-confidence events. Evaluations on simulated and real RNA-seq datasets show that ChiMER achieves higher sensitivity than conventional linear fusion detection methods while maintaining low false-positive rates.

Results: Applied to cancer cell line RNA-seq datasets, ChiMER identified multiple enhancer-exon chimeric transcripts, several associated with super-enhancer regions. Multi-omics analysis further shows that fusion transcripts occur in transcriptionally active regulatory environments and frequently coincide with strong R-loop signals, suggesting a potential role of RNA-DNA hybrid structures in facilitating long-range transcriptional joining events.

Single-cell RNA sequencing (scRNA-seq) datasets generated across laboratories and experimental conditions often exhibit batch effects that obscure biological variation. Numerous computational methods for batch integration have been developed, making rigorous benchmarking critical. Evaluation metrics are central to assessing method performance; however, existing metrics capture only partial aspects of integration quality and often rely on implicit assumptions about cell distributions in the embedding space. Consequently, benchmarking studies frequently report discordant rankings of batch integration methods across metrics, complicating interpretation and method selection. Here, we systematically evaluate widely used metrics under controlled scenarios that isolate common integration challenges, including imbalanced batch composition, partial cell-type overlap, and varying cluster geometries. By stress-testing metrics under these scenarios, we identify the conditions under which each metric succeeds or fails. Based on these observations, we introduce sBEE (single-cell Batch Effect Evaluator), a unified metric that jointly evaluates cross-batch distance relationships and local neighborhood batch composition. Across diverse scenarios, sBEE provides stable assessments of mixing quality and remains robust to failure modes that affect existing metrics. Together, our work provides a systematic evaluation of batch integration metrics and introduces a unified metric for a more reliable assessment of integration quality.

Spatial transcriptomics has enabled unprecedented insights into tissue organization by linking gene expression to histological context; however, the high cost, technical complexity, and limited coverage of sequencing-based technologies restrict their widespread application. Inferring biologically meaningful molecular states directly from routine histology remains a major challenge. Here, we introduce PATH (PAthway acTivation from Histology), a deep learning framework for predicting biologicalpathway activation from hematoxylin and eosin (H&E) stained images. PATH
is trained on paired spatial transcriptomics and histology data to learn pathway-level representations rather than gene-level expression. PATH
leverages a pre-trained Vision Transformer with Low-Rank Adaptation (LoRA) for efficient fine-tuning, combined with an adversarial learning
strategy to remove patient-, slide-, and dataset-specific confounding signals from the learned embeddings. By operating at the pathway level,PATH reduces noise and biological ambiguity inherent to gene-level prediction while improving generalization across datasets. Across multiplespatial transcriptomics datasets, PATH substantially outperforms gene-expression and pathway-based baselines, while exhibiting markedly reduced batch effects. We show that PATH captures biologically relevant processes, including immune signaling, cell cycle regulation, and oncogenic
pathways, and generalizes to high-resolution Visium HD slides containing tens of thousands of spatial locations. Overall, PATH demonstrates
that pathway-centric modeling enables robust, interpretable, and scalable inference of molecular states from histology, opening new avenues for
leveraging routine pathology images to study tissue biology and disease.

Motivation: Tumor-educated platelets (TEPs) represent a promising non-invasive source of cancer identification. However, existing transcriptomic approaches often rely on hundreds or thousands of genes, limiting interpretability and clinical translatability. We propose a network-based strategy to identify compact and biologically meaningful gene signatures derived from topological rewiring between healthy and cancer conditions.
Results: We constructed Spearman correlation networks from TEP RNA-seq data and quantified condition-specific topological changes using differential degree and betweenness centrality. Genes exhibiting consistent local and global rewiring were selected as candidates. The method was applied to gliomas, non-small cell lung cancer (NSCLC), and breast cancer (BrCa). The resulting genes were highly compact (29, 26, and 20 genes, respectively). PERMANOVA analyses confirmed that selected genes captured significant structural differences between groups, independent of dispersion effects. Across multiple classifiers and several independent external datasets, the proposed genes achieved superior or competitive predictive performance relative to larger DEG-based models, while substantially reducing dimensionality. Functional enrichment highlighted coherent cancer-related programs, including WNT/beta-catenin signalling in gliomas and translational machinery in NSCLC. Downstream analyses in gliomas further suggested a putative lncRNA/miRNA--FKBP5 regulatory axis linked to immune evasion mechanisms. Overall, differential centrality-based rewiring enables compact, interpretable, and generalizable TEP-derived biomarker panels across cancers.

Analysis of single-cell transcriptome data has provided valuable insights into biological mechanisms at molecular resolution. However, learning
phenotype-informative patient-level representations from single-cell data while supporting interpretable, cell-type-resolved gene prioritization
remains challenging. Key obstacles include (i) capturing heterogeneity in cellular states and their interaction-structured molecular context over
large gene networks, (ii) learning patient-level predictors under scarce supervision despite abundant cells, and (iii) navigating a combinatorial
search space over genes and cell types for interpretable prioritization. In this work, we introduce MIL2Het, a deep learning framework released as
a comprehensive Python package for phenotype prediction and post-hoc interpretation from scRNA-seq with patient labels. Across scRNA-seq
cohorts spanning breast cancer, COVID-19, and asthma, MIL2Het showed strong and consistent classification performance. The same trained
framework further yielded biologically coherent, cell-type-resolved gene prioritization that varied with disease context and clinical setting.
Together, these results suggest that MIL2Het provides a practical framework for patient-level prediction and context-aware biomarker hypothesis
generation from scRNA-seq.

Motivation: Wide adoption of Massively Parallel Reporter Assays, MPRAs, provides a rich ground for the application of deep learning to model gene regulatory regions and predict their activity from a DNA sequence. Yet, the development of new and improvement of existing models is hindered by the lack of a conveniently standardized dataset encompassing the data from multiple species and alternative MPRA designs or a repertoire of uniformly tested baseline models.
Results: Here we present MPRA-MNIST, a curated benchmarking dataset collection, encompassing results of 14 MPRA studies from human, yeast, fly, and bacteria. The underlying data were uniformly processed and converted to a standardized format with predefined and data-leakage–free ""training-validation-test"" splits, enabling rigorous and reproducible comparison of modelling approaches.
Each dataset is equipped with a vignette illustrating the basic model training from scratch to establish the baseline performance. Technically, MPRA-MNIST is a lightweight Python package providing dataset loaders, sequence transformations, and common deep learning utilities, while its modular structure allows for including extra user-defined datasets and models. We illustrate the practical usage of the framework with comparative benchmarking of several deep learning architectures across all datasets, which also provides standardized evaluation baselines. All in all, MPRA-MNIST establishes a convenient toolbox and educational resource for further development of reproducible MPRA-based deep learning models in regulatory genomics.

Bulk RNA sequencing methods for molecular subtyping of Head and Neck Squamous Cell Carcinoma (HNSCC) frequently fail to resolve intra-tumoral heterogeneity, often resulting in ambiguous subtype assignments across samples. Our previous bulk RNA analysis revealed significant heterogeneity within preclinical xenograft HNSCC models, demonstrating that clear subtype definitions are elusive in most cases. Studies involving spatial transcriptomics (ST) and single-cell profiling frequently show variability, further highlighting limitations of current subtyping methods.
In this study, consecutive tissue sections were systematically prepared. The sequence included slides for MALDI proteomics combined with H&E, followed by ST paired with H&E, and concluded with hypoxia marker staining (pimonidazole). This design facilitates future integration of proteomic and radiomic analyses using histological images (H&E) to predict molecular clusters. Currently, 20 samples have passed quality control and are under pathologist-guided annotation for downstream spatial analysis where a pipeline has already been designed.
We aim to develop a molecular profiling strategy explicitly focused on tumor microenvironment (TME) markers, particularly hypoxia-related signatures. This method combines histopathological and TME annotations with ST data to characterize tumor regions more precisely. Our approach seeks to establish robust TME-specific molecular profiles with translational potential, enhancing prognostic predictions and guiding targeted therapy in clinical HNSCC.

Understanding the regulatory logic behind cell fate decisions remains a major challenge, due to high dimensionality and noise of single-cell data and dynamic, context-dependent nature of transcriptional regulation. Addressing this challenge is essential for explaining how cells transition between states during development and disease. Transcription factors (TFs) orchestrate cell state transitions through gene regulatory networks (GRNs) that define stable cell identities and differentiation trajectories. However, current GRN inference methods, including multi-omic approaches, often struggle to predict cell states and transitions under perturbations.

Here, we present a novel framework for GRN inference from single-cell transcriptomics data. We first identify marker TFs per cell type via differential expression, defining the nodes of a Boolean network. Starting from a random network, we iteratively refine its regulatory rules through stochastic local search, selecting modifications that improve a composite score. This score evaluates whether the network's Boolean attractors recapitulate observed cell type expression profiles while also favouring biologically plausible network properties.

Our method correctly identifies the stable cell types and predicts cell-state outcomes under genetic perturbations in synthetic and real perturbation data, outperforming the established methods GENIE3 and GRNBoost2.

Our results indicate that incorporating and enforcing cell lineage information improves GRN inference. By evaluating the network as a system rather than independent edges, our framework better identifies regulatory circuits driving cell-state transitions under perturbations. By focusing on attractor structure rather than expression trends, our approach captures core regulatory logic underlying cell transitions improving our ability to model and predict regulatory control of cell fate decisions.

Background: Split Open Reading frames (Split-ORFs) exist on transcripts containing at least two open reading frames, each of which encodes a part of the same full-length protein. These multiple open reading frames arise from alternatively spliced transcript isoforms. The phenomenon of Split-ORFs has been observed for the SR protein family of splicing factors, where the Split-ORF proteins play important autoregulatory roles.
Aims: The aim of this study was to investigate the translation and expression of Split-ORFs in cardiovascular cell types.
Methods: We built a pipeline that predicts potential Split-ORFs for a user supplied set of transcripts and determines DNA sequences and peptides unique to the potential Split-ORFs. These unique sequences are absent from protein coding transcripts. The translation of the predicted Split-ORFs can be investigated by finding significant Ribo-seq coverage in their unique regions or their unique peptides in proteomics data.
Results: Split-ORF transcripts, their unique regions and peptides were predicted for the cell-type specific transcriptomes of cardiomyocytes and endothelial cells derived from PacBio long read sequencing data. A meaningful share of the unique DNA regions had significant ribosome coverage in Ribo-seq data of the same cell types. Additionally, proteomics data corroborated a substantial fraction of the unique Split-ORF peptides.
Outlook: These results suggest that the occurrence of Split-ORFs is more widespread than previously assumed and that they might play a role in the cardiovascular system. This paves the road for further functional investigations of the validated Split-ORF candidates, mechanisms of their biogenesis and their involvement in cardiovascular disease.

The common octopus, despite its very ancient evolutionary divergence from vertebrates, has a very large and complex brain with a cell number comparable to mammals. This leads to a wide variety of sophisticated behaviors such as recognizing itself, using tools or learning through observation. To better understand how this brain complexity is emerging during the development of this coleoid cephalopod species, whole embryo single-nuclei RNA was sequenced and analyzed at three different stages of organogenesis. Based on marker genes specific to cell types, the nuclei were clustered into multiple groups: retinal, glial, neuronal progenitors, differentiated neurons, mesodermal and epithelial. In order to establish the ontogeny of certain cell types, trajectory analysis was performed to track developmental progression along different embryonic stages. This analysis revealed the cell fates at the end of organogenesis and suggested roles of the different sub-types through identification of marker genes which vary along the cell type trajectories. This works brings a deeper insight in the development of coleoid cephalopod cell type diversity throughout organogenesis, and will be an important resource for future studies in this field.

A comparative analysis is presented, based on sequences from the Eukaryotic promoter database EPD and covering four metazoan phyla: Cnidarians, Nematodes, Arthropods and Chordates. Eukaryotic promoters contain sequence motifs that are over-represented at specific distances from the transcription start site (TSS). Two types are distinguished: (i) core promoter elements (CPEs) occurring at fixed locations, and upstream elements (UPEs), spread over a larger upstream region. CPEs interact with basal transcription factors (TFs), while UPEs interact with regulatory TFs found at enhancers as well. This study focuses on UPEs. The CCAAT- and GC-boxes were the first UPEs identified in humans and were later found to be recognized by NFY and SP1 proteins. Additional vertebrate UPEs include ETS, NRF-1, and bZIP binding motifs. When Drosophila promoters were systematically analyzed, it came as a surprise that none of the vertebrate UPEs were found. Instead, two new motifs popped up, later assigned to proteins M1BP and BEAF-32. This suggested that UPEs, unlike CPEs, are not conserved across distant metazoan phyla. The results presented here suggest otherwise. Surprisingly, honeybee shows strong over-representation of vertebrate UPEs, including NFY and NFR-1, but none or only marginal over-representation of Drosophila motifs. Perhaps even more surprisingly, the starlet sea anemone, the most primordial species analyzed, featured all vertebrate UPEs except SP1. These findings point to the existence of an ancestral metazoan UPE inventory, which has been conserved in many branches of the animal tree but lost and replaced in others, including flies and nematodes.

The combination of multi-omics data and network-based representations has been increasingly used to model complex biological relationships. We aim to apply this method to understand the signaling events upon Ca2+ influx in mast cells.
In mouse peritoneal mast cells (PMCs), we demonstrated that stimulation with agonists, including adenosine (ADO), compound 48/80 (C48/80), or antigen, triggers Ca²⁺ influx via ORAI1 and ORAI2 channel proteins. We also found that influx Ca2+ governed by ORAI channels drives immediate release of preformed mediators (e.g., histamine) in PMCs. However, its role in newly synthesized inflammatory mediators and metabolic shifts remains unclear.
To unravel these Ca2+-entry-dependent mechanisms, we employed a multi-omic strategy comparing wild-type and Ca2+-entry-deficient (Orai1/2-double-knockout) PMCs under various stimuli. We integrate ATAC-Seq, bulk RNA-Seq, and miRNA-Seq to profile changes in chromatin accessibility and transcription, while utilizing proteomics, secretomics, and metabolomics to define the functional consequences of this reprogramming. These datasets are then analyzed using advanced computational frameworks and interpreted through knowledge-guided Graph Neural Networks to resolve the underlying regulatory circuits.
To date, we have defined an ADO-specific transcriptional program distinct from C48/80 and antigen responses in PMCs using bulk RNA-Seq (published). Using ADO and its analogues for stimulation, we identified over 500 Ca²⁺-entry-dependent genes. By combining ATAC- and miRNA-Seq data, we constructed a Ca²⁺-entry-dependent TF-miRNA-gene regulatory network.
This analysis will further integrate proteomics, secretomics, and metabolomics data to define the specific Ca2+-entry-dependent mechanisms, serving as the basis for an unprecedented level of understanding of these disease-relevant signaling pathways in mast cells.

The mammalian immune system relies on precise mRNA stability control to balance pathogen defence and prevent damaging hyperinflammation. AU-rich elements (AREs) in 3' UTRs are the key regulatory elements of mRNA decay in the immune system. However, ARE positions and, importantly, their functional variants in the human genome remain largely unknown. This limited data availability precludes the diagnostic and therapeutic exploitation of AREs in precision medicine. To fill this knowledge gap, we aim to annotate AREs functionally and their genetic variants using deep learning models in combination with experimental systems.

AREs drive degradation of mRNAs via ARE-mediated mRNA decay (AMD). AMD employs RNA-binding proteins (RBPs) that bind AREs and directly or indirectly promote RNA degradation. Dysfunctional AMD leads to immune disorders and failure of immune homeostasis in mice. We integrate our resources on AMD in mice, including transcriptome-wide mRNA stability data and binding sites of AMD-active RBPs, such as Zfp36, in immune cells, with human CLIP-Seq datasets and RNA structure modelling to train deep learning models. As a result, we present a comprehensive deep learning framework that will eventually allow us to predict how genetic variation alters AMD in immune genes.

Our prototype achieves an AUPRC of 0.9571 and an AUROC of 0.9987 on held-out test data. Clinical relevance spans rheumatoid arthritis, lupus, IBD, and hepatitis C clearance conditions, where disrupted AMD is directly implicated.
This framework establishes new standards for interpreting non-coding regulatory variants and delivers actionable tools for personalised therapies in inflammatory and autoimmune disease.

Traditional statistical approaches, such as co-expression network analysis and differential expression analysis, may misrepresent gene importance by overemphasizing statistical significance while overlooking functional relevance. As a result, biologically important genes can be underestimated, whereas statistically significant but less relevant genes may be overestimated.


To address this limitation, we propose a novel interpretable AI framework that integrates the Nucleotide Transformer—a biological foundation model that encodes nucleotide sequences—into probabilistic circuits, enabling theoretically grounded probabilistic interpretability. Using this framework, we trained a classification model to distinguish tumor cells from normal cells by jointly leveraging statistical signals from scRNA-seq data and biological representations derived from promoter sequences.


Importantly, the model provides gene-level probabilistic contributions for each prediction, enabling interpretation of how individual genes influence classification outcomes. Our analysis reveals that 1,524 out of 9,540 genes exhibit contradictory behavior: some genes are statistically more frequent in normal cells but contribute strongly to tumor classification, while others are more frequent in tumor cells yet contribute more to normal classification. This bidirectional discrepancy suggests that incorporating biological representations enables the model to overcome purely statistical biases and capture functionally relevant signals.
Notably, these genes include several well-established cancer-related genes, such as ITGA5, SIGLEC9, NOTUM, and TP73. These findings indicate that our approach can uncover biologically meaningful signals that are overlooked by conventional statistical methods.


Overall, this work goes beyond traditional gene expression–based analyses and provides a new framework for integrating biological knowledge with statistical data, offering deeper insights into gene function in cancer research.

More than 600 distinct skeletal muscles constitute up to 40% of the total mass of the human body. They differ in anatomical position, origin, and function, but the diversity of muscle molecular phenotypes remains poorly explored.
Here, we present FANTOMUS, a large-scale molecular atlas of human skeletal muscles. To construct FANTOMUS, we performed cap analysis of gene expression (CAGE-Seq) using 222 autopsy samples of 75 skeletal muscles from 4 individuals, complemented by 22 matched proteomes obtained with mass spectrometry. As a result, we identified 37001 transcribed regulatory elements (TREs), including promoters of 18329 genes, and estimated the abundance of 1804 protein groups encompassing 1895 proteins.
Differential expression analysis revealed that more than 80% of genes and proteins are non-uniformly expressed across different muscles, with the most distinct molecular profiles found in extraocular muscles, tongue, and diaphragm. In particular, we observed a significant differential expression of FKRP, whose mutations cause dystroglycanopathy, between the leg muscles having different susceptibility to this disease.
Next, we performed motif activity response analysis of FANTOMUS CAGE-Seq data and detected motifs of hundreds of transcription factors with tissue-specific activity. Finally, by analyzing the allelic imbalance of CAGE-Seq reads, we discovered 6653 allele-specific single-nucleotide variants marking TREs with allele-specific activity. These SNPs often coincided with eQTLs and muscle-related GWAS SNPs, including muscle volume.
All in all, we created a detailed resource of transcriptomic and proteomic molecular profiles of diverse human skeletal muscles, facilitating further studies of gene regulation and heritable pathologies of these tissues: https://fantomus.autosome.org.

Pooled single-cell CRISPR screens have become an important resource in computational biology for linking the effect of genetic perturbations to molecular phenotypes and thereby advancing our understanding of gene regulation. Since many perturbations are profiled simultaneously in such pooled screens, a key challenge is the assignment of individual cells to the perturbation they received based on the guide RNAs (gRNAs) detected in each cell. Despite its central role, this step lacks systematic benchmarks and accessible tools for comparing assignment strategies. To address this gap, we developed crispat, a Python package that helps researchers evaluate and select appropriate gRNA-cell assignment methods. Applying crispat across multiple published datasets revealed substantial differences across methods in both the number and quality of assigned cells highlighting that assignment choice can substantially influence downstream analyses. Moreover, no single method consistently performed best across data sets or metrics. Building on these insights, we introduce crispero, a probabilistic factor model, that avoids committing to a fixed assignment altogether. Our model jointly integrates gRNA counts with molecular readouts to infer perturbation effects while estimating the probability that each cell received a functional perturbation, thereby capturing assignment uncertainty and identifying non-responding cells. crispero further separates general sources of variation - such as cell state, cell cycle, and batch - from perturbation-specific effects that capture gene modules co-regulated by targeted genes. Finally, interaction terms enable detection of perturbation responses restricted to specific cellular subpopulations and help dissect heterogeneous responses in complex tissues.

Intratumoral heterogeneity (ITH) remains one of the principal obstacles to effective cancer treatment, yet its relationship with drug response across cancer types is still poorly characterized. We introduce the Therapeutic Cancer Cell Atlas (TCCA), a pan-cancer single-cell resource that consolidates approximately 1.8 million transcriptomes from 537 patients and 183 cancer cell lines encompassing 34 distinct tumor types. By integrating single-cell transcriptomics with copy-number alteration inference and computational drug-response modelling, we provide a systematic characterization of therapeutic heterogeneity at subclonal resolution across a broad cancer landscape. Through this approach, we define ten recurrent therapeutic clusters that capture both shared and lineage-specific drug vulnerabilities across tumor types. Interestingly, therapeutic heterogeneity shows limited coupling to genomic or transcriptomic diversity, arising instead from discrete functional transcriptional programs and tumor microenvironment (TME) configurations. Cross-referencing with transcriptional metaprograms and TME archetypes reveals how stress-response pathways, proliferative states, lineage identity, and immune context collectively modulate drug sensitivity independently of tissue origin. We further establish the clinical relevance of TCCA by associating therapeutic clusters with patient survival outcomes and validating predicted drug vulnerabilities through pharmacogenomic datasets, including actionable findings in aggressive cancer subtypes. Collectively, TCCA constitutes a multidimensional atlas linking subclonal states, microenvironmental context, and therapeutic response, providing a scalable foundation for therapeutic prioritization, drug repurposing, and rational combination strategies in precision oncology.

Heterogeneous tissue samples from infected organs consist of diverse cell types whose proportions and transcriptional profiles change dynamically during pathogen invasion. RNA sequencing (RNA-seq) of mixed-cell populations captures average gene expression across all cells, obscuring cell type-specific responses, while single cell RNA-seq (scRNA-seq) provides higher resolution but suffers from technical biases, high costs, and limited applicability to large cohorts. In silico cell type deconvolution bridges this gap by estimating the proportions and expression profiles of cell types within RNA-seq data, using reference expression signatures. Although numerous deconvolution methods exist, their application to infectious disease research is limited, partly due to missing matched single-cell references under infection conditions and sensitivity to gene expression shifts between conditions.
We present Quasar, a novel cell deconvolution tool tailored for infectious disease transcriptomics. Quasar adapts the reference to match the condition of the bulk samples before estimating proportions. Bootstrapping assesses uncertainty. We tested Quasar on single cell datasets from peripheral blood mononuclear cells (PBMCs) of individuals with COVID‑19 and dengue infections, using pseudo‑bulks generated from held‑out test cells to benchmark performance against established methods.
In simulation studies based on single-cell data from infected individuals, Quasar achieves accurate cell-type proportion estimates across multiple cell types and performs as well as or better than existing methods for certain cell types. Future work will assess Quasar’s ability to estimate cell-type-specific expression profiles and evaluate its performance in capturing condition-specific expression changes in bulk samples when the reference and target conditions differ.

A major challenge in current sequence-to-activity modelling is the reliable interpretation of non-coding variants for clinical prioritisation. As a result, generating system-specific, high-throughput datasets—such as those derived from in vitro differentiation—has become a common strategy to gain clinically relevant insights. However, even with tailored datasets, deep learning models in regulatory genomics face interpretability challenges: it is not yet fully understood how architectural and methodological choices influence learned representations, potentially limiting reliable variant effect predictions.

We addressed these challenges in the context of congenital myopathies, elucidating the early regulatory events of skeletal muscle development and identifying key elements driving progenitor differentiation. We collected single-cell Multiome and bulk histone modification data across four critical time points in an in vitro human skeletal muscle differentiation system. We used these data to train a deep convolutional sequence-to-activity model that predicts cell-type- and timepoint-resolved DNA accessibility profiles at base-pair resolution. We used model explainability techniques to uncover sequence patterns that drive the model's predictions, as well as how multitask learning aids the model's explainability. We leveraged in silico mutagenesis to prioritise non-coding variants that disrupt transcription factor binding sites and utilised complementary 3D-interaction data to further link identified enhancers to target genes. This study provides the first comprehensive map of cis-regulatory drivers in human skeletal myogenesis by combining multi-omic data and explainable deep learning to unravel how time-dependent DNA regulatory activity drives cell fate decisions.

Dynamic differentiation processes determine the physiological and pathological behavior of all multicellular organisms. While single-cell RNA sequencing (scRNA-seq) enables the investigation of such processes, computational trajectory inference models are required to infer dynamics from snapshot gene expression measurements in the form of cell ordering and branching structure. However, such methods often neglect the underlying gene regulatory interactions. Differential equations can incorporate such interactions to model differentiation mechanistically and enable the simulation of gene expression dynamics. Since deterministic ordinary differential equation approaches cannot capture diverging differentiation from a shared initial state driven by the inherent stochasticity of gene expression, we present bowside, a stochastic differential equation (SDE) model based on gene regulatory networks (GRNs). Our model computes the rate of change of a gene's expression from the expression of its transcription factors using a weighted adjacency matrix. The GRN adjacency matrix structure is derived from a transcription factor database. Model parameters, including adjacency matrix weights, bias terms, and degradation rates, are optimized to minimize a distributional loss suitable for bifurcating differentiation that quantifies the difference between simulated sequences and cell distributions sampled from scRNA-seq data along (pseudo)time. We show that bowside infers the mutual inhibition of the fate-determining transcription factors GATA1 and SPI1 from a hematopoiesis scRNA-seq dataset and reconstructs diverging expression dynamics of the core GRN from a shared progenitor state to distinct cell fates. This work represents an important improvement over deterministic ordinary differential equation approaches to infer inherently stochastic differentiation dynamics from scRNA-seq data.

Single-cell multi-omics data offer a powerful opportunity to resolve how gene regulatory networks (GRNs) are rewired during cell-fate transitions. However, existing approaches typically infer GRNs as piecewise static networks, do not explicitly model temporal dependencies, and often rely on linear assumptions, limiting both accuracy and temporal coherence. Here we present DRAFT (Dynamic Gene Regulatory Network Inference with Graph-Recurrent Attention from single-cell multi-omics Trajectories), a deep learning framework for dynamic GRN inference from time-resolved single-cell multi-omics data. DRAFT is built on a prior transcription factor (TF)–regulatory element (RE)–target gene (TG) graph derived from motif and genomic distance information, and jointly models gene expression, chromatin accessibility and temporal dependencies through the integration of graph attention networks (GATs) and gated recurrent units (GRUs). DRAFT thereby captures nonlinear regulatory interactions and learns time-varying latent representations of TFs, REs and TGs, enabling reconstruction of smoothly evolving GRNs along cellular trajectories.
In simulations, DRAFT outperforms existing methods in both ground-truth network recovery and temporal smoothness. In mouse cellular reprogramming and Drosophila testis development, DRAFT recovers regulatory interactions corroborated by independent experimental evidence, including ChIP-seq, TF perturbation and Hi-C data, and identifies known stage-specific regulators and regulatory programs. In Drosophila, DRAFT further nominates Mondo as a candidate early-stage germline regulator, whose stage-specific expression pattern is supported by smFISH. Further, DRAFT enables temporal alignment across co-developing lineages, suggesting dynamic communication between germline and somatic lineages.

Biological sex-specific differences in complex diseases are difficult to study because transcriptomic differences between sexes conflate pre-existing baseline differences in gene regulation and genuine disease-specific responses that differ by sex. Standard between-sex differential expression cannot separate these components, making it impossible to determine whether observed differences are driven by baseline sex differences, disease-specific responses or their interaction. To assess genuine sex differences in disease biology, we developed a framework modelling the sex-by-disease interaction term, which decomposes gene-level expression into baseline sex effects, within-sex disease effects and interaction effects.

We apply this framework to whole blood RNA-seq data from a publicly available cohort of ankylosing spondylitis (AS) patients and healthy controls. AS is a chronic inflammatory arthritis of the axial skeleton with well-documented sex differences in presentation and outcome, yet existing studies have relied on direct between-sex comparisons, leaving the question of how disease biology itself differs by sex largely unaddressed.

Fitting our biological sex-aware interaction model, we distinguish transcriptomic sex differences that are present at baseline from those that arise specifically in disease, representing genuine interaction effects that would be obscured in a pooled analysis. Pathway enrichment analysis within the framework further identified disease-relevant processes offering potential mechanistic explanations for open questions in AS pathogenesis that have not previously been explored.

In summary, we present the first application of a biological sex-aware interaction model to AS transcriptomic data, demonstrating that it captures sex-specific disease signals and mechanistic insights that standard differential expression approaches would miss.

Single-cell RNA-sequencing (scRNA-seq) is a rapidly evolving technology that provides deep insights into gene expression and the regulation of cellular processes. Most commercially available sequencing kits rely on oligo-dT primers to reverse transcribe RNA molecules into cDNA for subsequent sequencing. These primers bind to poly-A sequences, which requires a poly-A tail in the target RNA. Additionally, the resulting reads are primarily located at the 3' end of the original RNA, which does not always capture sufficient transcriptomic information. Other approaches, such as the Parse Evercode chemistry, utilize random hexamer primers in addition to oligo-dT primers to enable binding across the entire RNA molecule. Parse also employs a split-pool approach during library preparation, which requires specialized read processing methods. While Parse provides a pipeline for the alignment and processing of this data, this Trailmaker pipeline is cloud-based and cannot be run on a local computing environment.
To address this, we developed a pipeline for processing of split-pool scRNA-seq data utilizing STARsolo. Our pipeline accounts for different primer types and automatically detects cellular barcodes from the reads. It is easily adaptable and its output is fully compatible with common scRNA-seq analysis frameworks like Scanpy and Seurat.
Using publicly available split-pool scRNA-seq data, we showed that our pipeline yields results highly concordant with Trailmaker at cellular, cell type, and gene levels. Furthermore, we investigated differences and biases between oligo-dT and random hexamer primers with regard to their binding patterns within a gene as well as between genes.

The evolution of robots has promoted both technological development and biological inquiry, but little to no effort has been placed on exploring gene interactions within evolving artificial genetic encodings. Studying gene interactions is fundamental to understanding how biological systems function, and the same applies to artificial evolvable systems. As autonomous systems become more complex, reliance on black-box encodings—where it is unclear how the genome produces phenotypes—poses limitations for efficacy, explainability, and safety. This study investigates epistatic gene interactions—a non-additive effect of two gene knockouts on a trait–in the evolution of robots encoded with artificial Gene Regulatory Networks. We evolve robot populations in a physics-based simulation and apply techniques from experimental biology to identify and quantify the evolution of epistasis in these robotic systems. The main contribution of this work is to demonstrate how epistasis can be quantified and analyzed to reveal insights into artificial genotype–phenotype mappings. The results show that selection pressure influences epistasis, though further studies are needed to determine whether epistasis itself is being favored or if it emerges as a by-product of selection acting on morphology and control.

Sexual dimorphism has been observed in a variety of biological processes. Sex-biased genes, which are defined as genes that are expressed at different levels in males and females, are central to these processes. Although cross-species comparative analyses of sex-biased genes have been conducted using bulk RNA sequencing (RNA-seq), the results have exhibited inconsistency across studies. Furthermore, previous research has focused on one-to-one orthologs. Consequently, the physiological roles and evolutionary processes of sex-biased genes remain poorly understood. We hypothesized that discrepancies were affected by variations in cell type composition. The objective of this study is to elucidate the mechanisms by which sex-biased genes acquire sex-biased expression patterns and the functions they fulfill.
A comprehensive analysis of publicly available single-cell RNA-seq (scRNA-seq) data from human kidneys, livers, and lungs, as well as those of mouse lemurs, mice, and rats, was conducted. We identified sex-biased genes at the cell type level. We performed an ortholog analysis that included complex relationships using DomClust and DomRefine. Our findings revealed that, while sex-biased expression was frequently not conserved within ortholog groups, the cell types expressing sex-biased genes were conserved. These results suggest that sexual dimorphism is evolutionarily conserved at the cellular level, even as the specific genes involved acquire species-specific sex-biased expression.
Currently, we are focusing on characterizing conserved regulatory landscapes, including transcription factor binding motifs, to elucidate the upstream mechanisms of these sex-biased patterns.

Foundation models (FMs) are redefining single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (SRT) analysis by learning transferable representations of cellular states and tissue organisation. While some adopt large language model (LLM) architectures, others rely on graph-based or hybrid multi-modal designs, yet all aim to generalise across biological contexts and analytical tasks. These models now drive diverse applications, from spatial domain discovery to cross-modality integration and therapeutic response prediction. Despite rapid progress, there remains no unified framework for systematically evaluating their capabilities across the breadth of single-cell and spatial analyses.
Here, we present a comprehensive benchmark and, to the best of our knowledge, the first systematic evaluation of foundation models jointly across single-cell and spatial transcriptomics, comparing six state-of-the-art architectures: Nicheformer, CellPLM, scGPT-spatial, GenePT, scELMo, and Novae. Performance is evaluated across multiple scRNA-seq and SRT datasets encompassing diverse diseases, species, and platforms, and assessed on key tasks including zero-shot and continually pretrained cell type clustering, cell type annotation, differential gene expression analysis, and perturbation prediction.
Our results demonstrate that preprocessing strategies, tokenisation schemes, and biologically informed priors profoundly influence model performance. Importantly, we highlight the urgent need for methods that address domain shifts arising from platform heterogeneity and biological variability. We also uncover persistent limitations in generalisation and interpretability, underscoring the challenge of building models that are both robust and biologically meaningful. This study provides actionable guidance for FM selection and establishes a standardised, extensible benchmarking framework to accelerate the next generation of single-cell and spatial foundation models.

Single-cell RNA sequencing (scRNA-seq) enables detailed characterization of gene expression across diverse cell types and conditions. While most analyses focus on differential expression and cell-type abundance, changes in gene–gene interaction patterns—key drivers of cellular function—remain underexplored at single-cell resolution. Because gene-gene correlations arise from coordinated activity within individual cells, co-expression network structure provides a natural representation of cellular state and its variation across conditions.
Here, we present SCENE (Single-cell Co-Expression Network analysis), a workflow for identifying and comparing cell type-specific gene co-expression networks. Building on the idea that gene co-expression structure defines cell state and may be cell type–specific, SCENE constructs donor- and cell type-specific co-expression networks from scRNA-seq data. These networks are compared across cell types or analyzed within each cell type to assess associations between network changes and biological conditions. Differences in gene–gene interactions are reported at gene and pathway levels.
We applied SCENE to the Chinese Immune Multi-Omics Atlas (CIMA), comprising peripheral blood mononuclear cells (PBMCs) from 421 donors with diverse demographic profiles. Using this dataset, we investigate gene co-expression rewiring across immune cell types and its association with age and sex. Our analysis demonstrates the potential of SCENE to uncover pathway-level changes in gene interaction structure.
SCENE complements existing single-cell analysis approaches by enabling systematic investigation of gene interaction dynamics. This framework provides new insights into regulatory mechanisms underlying cellular heterogeneity and their modulation across biological conditions.

Viral circular RNAs (vcircRNAs) are covalently closed RNA molecules generated by back-splicing events and have been primarily described in DNA viruses, where they contribute to the modulation of host cellular processes. Their presence in RNA viruses remains largely unexplored, partly due to methodological limitations. Current circRNA detection tools are mainly developed for eukaryotic systems, potentially limiting their applicability to viral genomes.

Here, we present a computational framework for unbiased detection of back-splicing events from RNA sequencing data, independent of predefined splicing signals and applicable to both short paired-end and long-read sequencing.

Its application to RNA-seq data from Zika virus (ZIKV) infected neural stem cells revealed a large population of putative vcircRNAs across sequencing platforms (Illumina: n=4167; Nanopore: n=234). Candidate numbers were markedly lower using a widely adopted circRNA detection tool (CIRI3) under comparable conditions (Illumina: n=164; Nanopore: n=0); for consistency, the same minimum support threshold (>= 2 BSJ-spanning reads) was applied.

Validation using rolling-circle reverse transcription followed by linear amplification confirmed the circular structure of a subset of candidates (Illumina, n=12; Nanopore, n=5). None of these were exactly reported by CIRI3 and only a limited subset (Illumina, n=7) could be recovered under relaxed matching criteria (within 50 nts for both coordinates), indicating that canonical splicing-based approaches may exclude bona fide viral back-splicing events.

Overall, our results demonstrate that unbiased detection strategies expand the detectable space of circular RNAs, highlighting the need for flexible computational approaches capable of handling sequencing data in non-canonical systems.

Alternative mRNA splicing generates a vast diversity of protein isoforms, many of which remain structurally and functionally uncharacterized. In cancer, aberrant splicing events can produce isoforms with altered protein domains, disrupted interaction interfaces, and novel immunogenic peptides, yet systematic resources linking splice variation to three-dimensional structural consequences are currently lacking.

We present SPLISOFORMS, a web-based knowledge platform that integrates long-read transcriptomics, protein structure prediction, and functional annotation to enable comprehensive exploration of cancer-associated splice isoforms and identification of "splice-enabled" neoantigens.
The database establishes a comprehensive structural baseline by including AlphaFold 3 predictions for all human GENCODE isoforms. Building on this, we integrated PacBio long-read and 10x single-cell RNA sequencing from clear cell renal cell carcinoma (ccRCC) organoids to generate over 30,000 novel, disease-specific isoform structures. Every sequence is compared against the canonical isoform, and systematically annotated across multiple dimensions, including domain architecture changes, nonsense-mediated decay (NMD) susceptibility, post-translational modification (PTM) site accessibility, and predicted neoantigen peptides.

SPLISOFORMS provides a scalable framework to move beyond sequence-based splicing analysis, offering a structural lens to prioritize functional isoforms and immunotherapeutic targets in cancer.

Background: Unsupervised patient stratification using single-cell RNA-sequencing (scRNA-seq) has the potential to reveal clinically relevant disease subtypes. While complex computational methods have emerged to represent sample-level data, they often ignore the compositional nature of cell-type proportions and lack interpretability.

Methods: We benchmarked seven state-of-the-art sample representation methods across eleven diverse scRNA-seq cohorts for their ability to recover known biological groupings in an unsupervised setting. We compared them against simple baseline approaches, including pseudobulk gene expression and Centered Log-Ratio (CLR) transformed cell-type compositions.

Results: Surprisingly, we found that simple compositional baselines consistently match or outperform complex methods in recovering biological groupings. CLR-transformed cell-type proportions achieved the highest stratification performance while being robust to batch effects and requiring orders of magnitude less computational power. Our analysis revealed that stratification signals are often concentrated in a small subset of highly variable cell types, making the results highly interpretable.

Conclusions: Our results suggest that cell-type composition is the primary driver of clinically relevant inter-sample variation in scRNA-seq cohorts. To facilitate these analyses we introduce scECODA, an open-source R Bioconductor package for scalable and interpretable cohort-level Exploratory COmpositional Data Analysis.

Podocyte injury is an early and critical event in the onset of chronic kidney disease (CKD), yet the molecular mechanisms underlying distinct injury etiologies remain poorly defined. Here, we present a human induced pluripotent stem cell (iPSC)-derived podocyte platform to model early, sublethal injury across multiple disease-relevant conditions. By integrating transcriptomic profiling with comparative computational analysis, we systematically characterized both shared and injury-specific molecular responses.

Differentiated podocytes were exposed to distinct stressors representing diverse pathogenic mechanisms, followed by bulk RNA sequencing and downstream pathway analysis. Our results reveal a core transcriptional injury program conserved across conditions, enriched for pathways related to cytoskeletal remodeling, metabolic dysregulation, and stress response signaling. In parallel, each injury modality exhibited unique gene expression signatures, reflecting mechanistic heterogeneity at early stages of disease progression.

Importantly, cross-condition integration identified a subset of consistently dysregulated genes and pathways that may serve as candidate biomarkers for early podocyte injury. These shared molecular features provide a foundation for developing diagnostic strategies that are independent of disease etiology. Ongoing validation using advanced platforms, including organ-on-chip systems, aims to further assess the translational potential of these candidates.

Together, this study establishes a scalable stem cell-based framework for modeling podocyte injury and highlights the power of integrative transcriptomics to uncover convergent and divergent disease mechanisms. These findings have implications for early detection and therapeutic targeting in CKD.

Sequence models are increasingly used to predict the effects of regulatory variants, but systematic comparisons against experimental results from Massively Parallel Reporter Assays (MPRAs) in different cellular settings remain limited. We present a benchmarking study evaluating sequence-to-function models (e.g., Enformer, Basenji2, and AlphaGenome) against MPRA-measured allelic regulatory activity using our adjustable, scalable, and reproducible Snakmake-based pipeline (https://github.com/kircherlab/Model-MPRA-Benchmark-Project). As benchmark data, we use IGVF-released MPRA data in three cell lines (https://data.igvf.org): HepG2 (147,485 SNVs), HEK293T (147,532 SNVs), and NGN2-differentiated neurons (38,968 SNVs). We aggregated SNP-Allelic-Difference (SAD) scores across multiple model outputs or use cell-type-specific subsets to assess performance on statistically significant variant effects (FDR threshold ≤ 0.1)).

AlphaGenome consistently outperformed the other two models. For significant variants, it achieved Spearman ρ=0.65 (HepG2) and ρ=0.60 (HEK293T), compared to Basenji2 (ρ=0.56/0.53) and Enformer (ρ=0.55/0.53). On the data of NGN2-differentiated neurons, all models performed noticeably worse (Enformer ρ=0.17, Basenji2 ρ=0.19, AlphaGenome ρ=0.16). This may reflect the representation of neuronal and other cell-type-specific contexts in model training, or the small proportion of large effect-size variants within this dataset.

Filtering tracks by (closest) cell-type improved performance for Basenji2 and AlphaGenome but worsened Enformer performance outside the liver context. AlphaGenome represented our cellular contexts the best, suggesting that the composition of training data is crucial for performance across cell-types. At assay level, DNase tracks yielded the highest per-assay correlations across all models, while CAGE performance varied substantially between models. We will present cross-model comparisons by effect size and contrast those with genomic LM performance.

Hypertrophic cardiomyopathy (HCM) is a common cardiac disease in cats and humans, characterized by left ventricular (LV) hypertrophy and diastolic dysfunction. Despite its prevalence, molecular mechanisms involved in the disease pathogenesis remain incompletely understood.
To improve the understanding of HCM associated cardiac remodeling processes, we performed bulk RNA sequencing of LV and left atrial (LA) tissues donated from pet cats with naturally occurring HCM (n=7 LV, n=7 LA) and healthy control cats of similar age (n=5 LV, n=5 LA). Differential gene expression analysis was complemented by two pathway enrichment approaches using distinct statistical frameworks: Ingenuity Pathway Analysis on a filtered set of differentially expressed genes and Gene Set Enrichment Analysis on the entire ranked gene list. Obtained pathways were compared for similarities and differences.
Chamber-specific pathway profiles were identified: the HCM LV showed enrichment of inflammatory activation, fibrofatty remodeling, and cell death pathways, whereas in the HCM LA antihypertrophic, reparative, and cardioprotective pathways were observed, with Rho signaling emerging as a potential regulatory hub. For both chambers, gene expression patterns associated with epigenetic regulation, metabolic, proteostasis-related, and electrophysiological alterations were detected, features suspected to be involved in cardiac disease, but not described in naturally occurring HCM. HCM hearts further showed activation of inflammation, immune cell migration, fibrosis, and extracellular matrix remodeling pathways consistent with previously described disease processes. These findings characterize chamber specific molecular disease processes in naturally occurring HCM that have not previously been reported, identifying pathways for targeted mechanistic investigations.

Amyloid precursor protein (APP) is a transmembrane protein expressed in neurons and is involved in synapse formation and cell signaling. In the amyloidogenic pathway associated with Alzheimer's disease, APP contributes to the generation of amyloid-β (Aβ) peptides, which aggregate into plaques, disrupt synaptic function, activate astrocytes, and induce metabolic dysfunction.

In Alzheimer's disease, astrocytes responsible for Aβ clearance become reactive, reducing their clearance capacity and transitioning to a pro-inflammatory phenotype. Epoxyeicosatrienoic acids (EETs) counteract this reactivity, but their degradation by soluble epoxide hydrolase (sEH), which is upregulated near plaques, limits this effect. Genetic deletion of sEH (sEH-KO) has been linked to reduced astrocyte reactivity and improved metabolic function.

We analyze 10x Genomics Visium spatial transcriptomics data from APPF1 mouse brain tissue, comparing sEH-KO and control conditions. Our focus is on spatial gene expression changes related to astrocyte signatures and metabolic pathways, assessing how sEH-KO alters astrocyte metabolism.

While sEH-KO has been studied in Alzheimer's pathology, less is known about how its effects are spatially organized within brain tissue. By applying spatially resolved analysis, we investigate the topological structure of cell-cell communication networks associated with astrocyte transcriptomic and metabolic shifts.

Our approach includes quality control, cell type deconvolution, and spatial clustering. We further explore cell-cell communication to identify ligand-receptor interactions between astrocyte and neighboring cell types to identify spatial communication patterns and localize functional hotspots within the brain.

Differential expression analysis uses well-established tools (limma, edgeR, DESeq2) to compare conditions. However, it is unclear how these tools perform in difficult scenarios such as strong sample imbalance. Here, we study differences between cancer subtypes comparing genetic ancestries, where existing omics datasets are heavily biased towards Europeans. While the portability of genotype-phenotype relationships in genomics data has been analyzed in depth using polygenetic risk scores, similar analyses are lacking in transcriptomics or epigenomics.
We developed a differential comparison approach that accounts for ancestry-related sample imbalance. This approach uses meta-analysis of randomly drawn European subsets to obtain interaction effects between cancer subtypes and genetic ancestry. Based on simulated and permuted data, we show that this approach resulted in reliable FDR control under strong sample imbalance, while retaining sensitivity compared to standard approaches. Applied to TCGA data, our approach revealed robust ancestry-specific cancer pathways but to a smaller extent than expected from baseline differences between ancestries. Our approach further enabled us to assess portability of predictive models from Europeans to other ancestries. This revealed that, while model loss was increased in other ancestries, classification performance remained largely unchanged, suggesting that these models relied on features without ancestry-specific effects.
In summary, we provide a computational framework to robustly identify ancestry-specific cancer effects and to evaluate prediction performance across ancestries, applicable to various multi-omics data.

The majority of disease-associated variants are located in the non-coding genome, and their functional effects remain challenging to decipher. In genome-wide association studies (GWAS), more than hundred non-coding genomic loci have been linked to Parkinson's disease (PD) risk. To study effects of non-coding genetic variation on gene regulation, we generated single nuclei multi-omic atlases of human cingulate cortex and substantia nigra with matched long-read whole-genome sequencing data for 190 donors (115 controls, 75 PD). The atlases consist of 1.1M snRNA-seq and 3.1M snATAC-seq high-quality nuclei, which allowed us to profile gene expression and chromatin accessibility in all major cell types for both brain regions. By integrating chromatin accessibility quantitative trait loci (caQTL), DNA methylation QTL (meQTL), and allele-specific chromatin accessibility (ASCA), we identified 53,841 high-confidence cis-acting genetic variants that modulate cell type-specific enhancer accessibility in one or both brain regions. We further demonstrate that sequence-to-function models can accurately predict the impact of these variants directly from the genomic sequence. Novel explainability approaches allowed stratifying these variants according to their regulatory function, with the majority disrupting specific transcription factor binding sites in a cell type specific manner. Integrating these "enhancer variants" (EV) with eQTL mapping and gene locus modeling linked an EV subset to their target genes. Finally, we applied these models to prioritize regulatory variants at known PD GWAS loci, bypassing statistical limitations in rare disease-relevant populations like dopaminergic neurons. We establish a unique resource and new sequence modeling strategies to interpret functional non-coding variation in the human brain.

Single-nucleus DGE analyses suffer from p-value inflation when treating cells as independent replicates. Multiple testing corrections are overly conservative when feeding into downstream analyses. Bulking mitigates pseudo replication but discards single cell information (heterogeneity, distributions). We evaluate metacells - inspired by high dimensional weighted gene correlation network analyses - to preserve within population structure while stabilizing variance.
We compared results from Wilcoxon tests on simulated data and an atlas, contrasting one against other cell types. We applied (i) pseudobulk, (ii) metacell, and (iii) random down sampling. Metacells were constructed via KNN graphs while tuning hyperparameters. Consistency of DGEs and biological relevance were evaluated. We further recorded runtime and memory.
Tuning strongly affected DGE stability and biological interpretability. Intermediate metacell resolutions maximized gene set concordance while limiting false positives. Other hyperparameters than size most influenced the metacell outcome. Down sampling reduced computation but at too low fractions degraded sensitivity. Pseudobulk improved type I error control but missed cell type signals. Metacell aggregation dominated runtime.
Metacells offer a practical compromise between robustness and resolution. We recommend multivariate desirability optimization for tuning. Overall, metacell based DGEA yields reproducible gene sets at competitive cost, reducing loss of information compared to down sampling or pseudobulk. MetaDGEA is a promising tool for discovery and assessment of therapeutic concepts to deliver innovation, safety, and quality for our patients.

Recent advancements in long-read single-cell and spatial RNA-sequencing enable the profiling of RNA isoform expression and alternative splicing at unprecedented resolution. However, the computational landscape remains highly fragmented. Existing pipelines are often tied to specific protocols, require matched short-read data, or lack complete end-to-end processing. These constraints severely limit analytical flexibility, reproducibility, and the ability to compare findings across studies utilizing different methodologies.
To alleviate these computational limitations, we introduce FLAMESv2, a highly modular and protocol-agnostic R/Bioconductor package for the comprehensive analysis of long-read single-cell and spatial RNA-seq data. Building upon our previous framework, FLAMESv2 has been extensively enhanced to offer superior flexibility, processing speed, and usability. It integrates diverse data types and supports a wide array of single-cell workflows as well as emerging spatial transcriptomics protocols. The pipeline is highly configurable, scales efficiently to accommodate multi-sample analyses, and can be executed using long reads alone or in combination with short reads. Furthermore, comprehensive benchmarking confirms that FLAMESv2 achieves field-leading performance across key analysis tasks, from accurate isoform identification to precise quantification. Beyond data processing, FLAMESv2 equips users with versatile built-in functions for publication-ready data visualization and downstream analysis. By transforming a disjointed set of bioinformatic tools into a unified, robust pipeline, FLAMESv2 provides the community with a powerful approach to long-read transcriptomics. It unlocks the full potential of single-cell and spatial long-read sequencing, empowering researchers to deeply characterize the hidden layers of RNA isoform regulation in health and disease.

The gastrointestinal tract contains diverse conventional and unconventional T cell populations, including CD4+ cytotoxic T lymphocytes (CTLs), whose regulatory programs remain incompletely understood. To characterize transcriptional and regulatory changes underlying intestinal T cell heterogeneity, focusing on the epigenetic regulators HDAC1 and HDAC2, we performed single-cell RNA sequencing (scRNA-seq) combined with transcription factor activity inference across small intestinal intraepithelial lymphocytes (SI-IEL), lamina propria (SI-LP), and colonic T cells from mice lacking both Hdac1 alleles and one Hdac2 allele (HDAC1cKO-HDAC2HET). Single-cell transcriptomic analysis revealed compartment-specific effects of HDAC1/2 loss, with SI-IELs showing the strongest transcriptional alterations. Clustering and differential expression analyses identified an expansion of RUNX3high CD4-CD8a+b low subsets, including populations with mixed CD4+ lineage transcriptional features. TF activity inference supported a shift toward cytotoxic regulatory programs in HDAC1/2-deficient cells. Across intestinal tissues, CD4+ lineage cells displayed increased cytotoxic gene signatures, but reduced IL-17A-associated expression in the colon. During Citrobacter rodentium infection, HDAC1cKO-HDAC2HET mice showed impaired induction of protective Th1/Th17 programs, accompanied by increased representation of cytotoxic CD4+ CTLs. These results indicate that HDAC1 and HDAC2 regulate transcriptional programs restraining CD4+ T cell differentiation to CTLs, maintaining T cell homeostasis and effector function during bacterial infection. Additionally, we highlight the utility of single-cell transcriptomics and TF activity analysis for resolving functional heterogeneity in intestinal immune cells.

Estimating cellular composition from bulk RNA-sequencing data provides valuable insight into tissue heterogeneity and remains central to translational and clinical genomics. Most deconvolution approaches reconstruct bulk expression profiles from single-cell-derived reference signatures. Our recent work demonstrated that explicitly incorporating hierarchical relationships between hierarchically related cell types substantially improves accuracy and robustness of such estimates. Despite well benchmarked models, practical application remains challenging. Typical workflows require interaction with Python or R code, extensive preprocessing and additional tools for downstream statistical analysis. Hierarchical models further increase complexity by requiring explicit specification of cell-type groupings, creating barriers for non-expert users.

To address these challenges, we introduce HIDE-deconv, a Python package based on an improved and extensible version of our hierarchical deconvolution model HIDE. HIDE-deconv provides an intuitive command-line interface enabling users to run a complete workflow, from a preprocessed single-cell AnnData object and bulk RNA-seq input to cell-type composition estimates, statistical postprocessing and visualization, without requiring programming experience. Advanced users can access all core components through a Python API for custom workflow integration.
HIDE-deconv includes built-in downstream analyses such as group comparisons using Mann-Whitney U and Kruskal-Wallis tests with Dunn post-hoc correction, survival analysis using coxph and PCA to explore latent cell-type composition patterns beyond known clinical covariates. The software runs entirely locally, supporting secure analysis of sensitive clinical datasets.

HIDE-deconv is available as open-source software on GitHub https://github.com/dvoelkl/HIDE-deconv

Ribosomes typically commence translation at a methionine-encoding AUG codon flanked by a so-called Kozak region, a short nucleic acid motif that serves as an initiation site in humans. Though, the characteristic AUG start codon of an mRNA is not always effective in initiating translation. Seldomly, near-cognate codon sequences may also be recognized as start sites. Several types of ribosomal profiling techniques have been developed that elucidate active translation initiation sites (TIS). Based on data gathered from these techniques, machine learning models have been developed to predict translation start sites using mRNA sequence features. Here, a meta-model termed MetaTIS was implemented by combining outputs of genomic and protein language models fine-tuned on five different TIS datasets plus the Ensembl annotations. The model proficiently differentiates between spurious and true TIS in four distinct test sets, for both canonical and noncanonical instances. While analysing one of the base models with integrated gradients, it was found that the model considered the importance of the Kozak sequence context and the presence of an upstream open reading frame (uORF) for classifying a position as an actual TIS. We further demonstrated how MetaTIS predictions can be used to detect important uORFs in three cancer types.

Glioblastoma is an aggressive brain cancer with dismal prognosis and limited therapeutic options. Adoptive cell transfer using tumor-infiltrating lymphocytes (TILs) represents a promising strategy. We previously demonstrated that TILs enriched for tumor-reactive lymphocytes and expanded ex vivo retain specific antitumor activity, providing the rationale for the upcoming ReacTIL clinical trial evaluating this approach in patients with glioblastoma. However, successful TIL expansion is achieved in only 59% of cases, and the determinants of expansion efficiency remain elusive.
To dissect TIL-intrinsic and extrinsic drivers of expansion, we generated single-cell RNA-seq (scRNA-seq) data from 19 pre-expansion TIL-enriched samples. We then integrated these with 18 additional datasets to reconstruct a glioblastoma single-cell atlas (>1M cells, 227 patients), capturing tumor, glial, neuronal, vascular, lymphoid, and myeloid compartments. Our single-cell analysis of TIL-enriched samples with successful expansion (sExp) showed higher abundances of CD8+ effector T cells, regulatory T cells, and CD16+ NK cells, whereas neutrophils were enriched in non-expanded (nExp) samples. Notably, analysis of whole-tumor data using single-cell-informed deconvolution (n=15) revealed the opposite trend for CD8+ effector T and CD16+ NK cells, suggesting the effectiveness of our strategy for the enrichment of tumor-reactive TILs. Single-cell analysis further uncovered distinct metabolic programs in T/NK subsets from sExp and nExp samples, suggesting opportunities for metabolic reprogramming.
Altogether, this integrative analysis combining single-cell atlas construction, functional characterization, and in silico deconvolution identifies key determinants and actionable levers of TIL reprogrammability, providing a basis to improve expansion success and the clinical efficacy of TIL-therapy in glioblastoma.