Simone G. Riva

Motivation: High-throughput genomic technologies have transformed our understanding of biological systems, yet direct comparison and visualisation of these complex datasets remains challenging. Existing normalisation methods often fail to align genomic signal across samples due to sensitivity to sequencing depth differences and localised high-signal artefacts, leading to inconsistent replicate behaviour and increased downstream variability. Results: We introduce Zone Equalisation Normalisation (ZEN), a novel approach designed to improve cross-sample signal alignment of genomic data. ZEN rescales genomic signal based on variance estimated within biologically enriched regions, reducing the influence of extreme outliers while preserving underlying biological structure. Using a diverse collection of data and our new genome-wide benchmarking approach, we reveal that ZEN improves biological and technical replicate alignment across the majority of tested conditions and experimental platforms. We further show that this improved signal comparability is associated with fewer differential accessibility calls between technical replicates and a more conservative set of biological differences. Together, these results demonstrate that ZEN provides a complementary framework to improve the accuracy and reliability of genomic data analysis and that normalisation choice can affect downstream analyses and biological interpretation. Availability and implementation: ZEN is available as an open-source Python package via conda and PyPI. Source code, documentation, tutorials, and code to reproduce the analyses are available at https://github.com/Genome-Function-Initiative-Oxford/Zone-Equalisation-Normalisation and Zenodo (https://doi.org/10.5281/zenodo.21067751).

High-throughput chromatin accessibility assays such as bulk and single-cell ATAC-seq have generated large collections of processed signal tracks in bigWig format, which are widely used for visualisation, data integration, and Machine Learning (ML)–based analyses. Despite their central role, systematic quality control (QC) frameworks operating directly at the level of bigWig signal tracks remain underdeveloped. This gap limits the ability to assess data reliability and hampers robust downstream analyses. Here, we present a biologically grounded QC framework for chromatin accessibility bigWig files that integrates peak-level information, background noise estimation, and recovery of stable genomic reference features. Using an ML-based peak caller (LanceOtron), we derive complementary quality metrics capturing signal structure and signal-to-noise properties. We further define constant promoter and CTCF regions as internal biological controls and show that their recovery provides a sensitive measure of data quality across diverse cellular contexts. We apply this framework to a collection of 502 human chromatin accessibility bigWig tracks spanning a wide range of tissues and cell types. The proposed metrics capture related but non-redundant aspects of signal quality and motivate the use of constant promoter and CTCF recovery as biologically meaningful targets. An XGBoost model trained on LanceOtron-derived features accurately predicts recovery of these stable genomic elements on held-out data (R² = 0.97), yielding a continuous and interpretable quality score. Feature importance analysis using SHAP values highlights that model decisions are driven by biologically relevant signal properties rather than arbitrary heuristics. Quantile-based stratification of the quality score is further supported by clear qualitative differences in genome browser visualisations. Together, this work provides a principled and extensible frame-work for assessing the quality of chromatin accessibility bigWig tracks, enabling more reliable data integration and supporting downstream ML applications in regulatory genomics.

A substantial fraction of disease-associated genetic variants reside in non-coding regions of the genome, where they act by perturbing cis -regulatory elements (CREs) such as enhancers, promoters, and insulators. While recent sequence-based deep learning models, such as Enformer, accurately predict continuous epigenomic signals from DNA sequence, they do not directly provide discrete and interpretable CRE annotations. Here, we present CREP ( Cis -Regulatory Element Predictor), a fine-tuned version of Enformer trained to predict regulatory element identity from sequence using REgulamentary-derived annotations across multiple human cell-types. Through a controlled experimental framework, we show that incorporating diverse cell-types improves model performance. CREP leverages cell-type-specific training data to learn regulatory representations while producing a unified prediction of CRE identity from sequence. This is demonstrated by the Vanuatu SNP, a non-coding variant that creates a de novo erythroid regulatory element, which is correctly detected only when erythroid data are included during training. Error analysis further reveals that apparent misclassifications between enhancers and promoters reflect their shared regulatory architecture, supporting the view of CREs as a functional continuum rather than strictly discrete classes. Together, these results demonstrate that CREP enables interpretable prediction of regulatory element identity from sequence and provides a framework for the functional interpretation of non-coding genetic variation.

Accurate annotation of intermediate cell states remains a major challenge in single-cell RNA sequencing (scRNA-seq), particularly in continuous differentiation systems such as erythropoiesis. Existing reference-based methods often lack the resolution required to distinguish early and transitional erythroid progenitors and may generalise poorly across datasets and modalities. Here, we present a supervised framework for erythroid lineage annotation based on expert-curated training data that integrates bulk and single-cell transcriptomic information. Starting from a human bone marrow scRNA-seq atlas, we refined erythroid annotations by introducing previously unresolved progenitor stages, including burst-forming unit–erythroid (BFU-E), colony-forming unit–erythroid (CFU-E), and pro-erythroblast (ProE), guided by canonical marker genes and bulk RNA-seq references. We trained and benchmarked four classical machine learning models and identified LightGBM as the best-performing approach, achieving a validation macro F1-score of 0.821 and balanced accuracy of 0.826. On a held-out test set, the model showed strong performance across most erythroid stages, with errors largely confined to adjacent differentiation states. The classifier was further transferred to independent bulk RNA-seq samples and an external bone marrow scRNA-seq dataset, where it recovered expected erythroid progression and refined coarse-grained annotations into higher-resolution cell states. Together, these results show that expert-curated supervised learning can improve erythroid cell state annotation in scRNA-seq and provide a practical framework for studying differentiation hierarchies in settings where finely resolved public references are limited.

Enformer is a deep learning model trained on human and mouse genomes to predict regulatory activity from 196,608 bp DNA windows. Its trunk embeddings capture long-range cis-regulatory interactions, but whether this signal generalises across the tree of life has not been assessed. We embed universal single-copy orthologous groups (OGs) from OrthoDB v12 across three taxonomic scales and evaluate reconstructed trees against TimeTree5 using Mantel r and Normalised Robinson-Foulds (NRF). On 702 OGs across 34 Primate species (≤74 Mya), the consensus tree achieves Mantel r=0.902 and NRF=0.481, correctly recovering major clades. A key finding is that flanking regulatory context - not the gene locus itself - carries the phylogenetic signal: restricting pooling to central 448 bins collapses Mantel r to 0.355. Applying the same fixed configuration to Vertebrates (≤450 Mya, 83 OGs, 150 species) and Plants (≤1,500 Mya, 92 OGs, 40 species) yields consensus Mantel r of 0.752 and 0.803 respectively, with NRF worsening monotonically across tiers. Distance-ordering fidelity degrades smoothly with evolutionary distance while topological accuracy declines steadily, with no sharp taxonomic boundary. These results show that an unmodified regulatory deep learning model encodes robust phylogenetic signal well beyond its training distribution, reaching across 1,500 million years of divergence.

Genome-wide association studies have revealed that many disease-associated genetic variants lie in non-coding regions of the genome. To prioritise these variants and clarify their functional roles, accurate classification of cis-regulatory elements is essential. Early approaches relied on characteristic histone marks, while more recent methods use Hidden Markov Models to segment the genome into chromatin states. However, these models often produce abstract states that require manual interpretation to assign regulatory function. REgulamentary is introduced as a rule-based framework for de novo, genome-wide annotation of cis-regulatory elements in a cell type-specific manner. Its behaviour is compared with count-based and segmentation-based approaches to highlight differences in classification strategy and the interpretability advantages of a rule-based design. Finally, its utility in the analysis of complex disease loci is demonstrated through application to published genetic association data to prioritise likely causal variants.

Resolving the gene targets of non-coding genetic variation is the major bottleneck in translating genome wide association studies into mechanistic understanding of complex diseases such as coronary artery disease (CAD). Combining new Transformer-based Machine Learning (ML) approaches trained on cardiovascular epigenetics with high-resolution, allele-specific genomic and transcriptomic technologies we create a highly scalable platform to simultaneously resolve causal variants, cell-type of action, output gene, and direction of effect. When applied to CAD genetics, our ML predicts causal variants from 20,747 candidate SNPs across 9 vessel cell-types and identifies disrupted transcription factor binding motifs using ML feature attributions. We investigate 94 of the top predictions in endothelial cells using Micro Capture-C, revealing the importance of fluid shear stress and TGF-β signaling pathways. We exploit allelic skew in heterozygous cells to demonstrate both variant causality and effect direction, demonstrating this platform can be used to rapidly resolve non-coding genetics in complex disease.

High-throughput genomic technologies have transformed our understanding of biological systems, but analysing these complex datasets remains challenging. Existing normalisation methods often struggle to accurately correct for technical variations across different experimental conditions, potentially leading to misleading biological interpretations. We introduce Zone Equalisation Normalisation (ZEN), a novel approach designed to improve signal normalisation across genomic datasets. By developing a flexible method that reduces technical noise whilst preserving biological signals, ZEN addresses critical limitations in current normalisation techniques. By applying ZEN to diverse genomic techniques, we reveal its robust performance in managing technical variations across different experimental platforms. ZEN provides researchers with a robust tool for more accurate and reliable genomic data analysis, potentially revealing biological insights that were previously obscured by technical variations.

Genome-Wide Association Studies (GWAS) have transformed our understanding of genetic traits and diseases since 2005 by analysing genetic variants across the human genome. Whilst they have successfully identified thousands of genetic locations linked to various conditions, converting these discoveries into practical understanding remains challenging. A key challenge lies in understanding through which cell types and tissues a specific set of genetic variants functions. This knowledge would aid researchers in designing experiments and identifying therapeutic targets. Causal variants often affect cell type specific cis -regulatory elements (CREs) and therefore the genetics of a trait are enriched within the CREs of effector cell types. Recent advances in genomics and computational methods are improving our understanding of genetic variants’ effects across cellular environments. In this work, we introduce Cell types Ranking and Enrichment Score for selecTive Genetic Variants (CREST-GV), a statistical tool that analyses genetic variants intersecting with open chromatin regions across hundreds of cell types, by performing a systematic computation of probability distributions and standardised metrics, transforming genomic data into quantitative enrichment measures with corresponding significance values. The tool demonstrates strong capability in prioritising disease-relevant cell types, for example, in diabetes studies where it correctly identified immune cells for Type 1 diabetes, and pancreatic islet cells and adipose tissue for Type 2 diabetes. CREST-GV shows robust generalisation across diverse genetic traits, successfully analysing variants associated with different conditions whilst maintaining consistent performance and adaptability through its flexible framework. This work includes a user-friendly web application connected to our public database of precomputed enrichment scores. The platform enables rapid querying across cell types and genetic variants, provides standardised results for comparative analysis, and allows the exploration of multiple cell type relationships simultaneously. This accessible interface makes complex enrichment analyses available to the broader research community, accelerating genetic research.

Genome regulatory elements are fundamental to cellular identity and cell type specific gene expression. Understanding how the underlying genetic code is differentially utilised by different cell types is central to understanding human health and disease. To better understand how DNA encodes genome regulatory elements such as promoters, enhancers, and boundary elements, we leverage the Enformer gene expression and epigenetic prediction model. We used transfer learning with high quality single cell ATAC datasets to develop REnformer, a model to predict chromatin accessibility. By introducing a bench-mark for comparing performances against Enformer model, REnformer significantly outperformed Enformer in terms of higher prediction outcomes and lower error rates in all extensive analyses shown; introducing these benchmarks allowed us, and possible future works, to fairly compare such models. We further tested REnformer by predicting the effects of a well characterised α -thalassemia variant and found that the prediction aligned with the observed change in genome regulatory element, previously validated. We conclude that REnformer is and can be a state-of-the-art tool to predict cell type specific regulatory elements, and interrogate the effect of genome variation in health and disease.