Biology subjects

Ioannidis, N. M.

Publications and source records attributed to Ioannidis, N. M..

At least 19 recordsLinked to original sources

bioRxiv2026

Extracting interpretable single-cell metabolic states with graph-guided representation learning

Metabolism shapes cellular function and state, yet measuring single-cell metabolic states at scale remains a challenge. We present Metabolic Representation Net (MeRN), a graph-guided variational autoencoder that leverages prior metabolic knowledge as a topology graph to learn latent representations of metabolic state and reaction activity from single-cell transcriptomes. MeRN's scalable estimation of reaction activity enables the definition of data-driven pathways (DDPs): context-specific metabolic modules supported by transcriptomic evidence and agnostic of standard pathway definitions. Using DDPs, we introduce the weakest link analysis to identify metabolic network rewiring. MeRN recovers metabolic zonation in the mouse intestine, links a folate deficiency-induced break in de novo purine synthesis to embryonic neural tube defects, shows cytokines with similar non-metabolic effects can elicit divergent T cell metabolism, and identifies metabolic drivers of T cell exhaustion and therapy response in human cancers. Our results establish MeRN as a unified method for metabolic analysis of single-cell transcriptomes.

bioinformatics↗

bioRxiv2026

The evolution of structural variation across 500 million years of vertebrate evolution

Structural variants (SVs) contribute substantially to genetic variation and play vital roles in adaptation and disease1,2. Nonetheless, SVs are poorly captured by short reads and thus remain understudied, especially in non-model organisms3,4. Here, using haplotype-resolved genome assemblies from >600 vertebrate species, we comprehensively survey the landscape of SVs across >500 million years of evolution. We identify 35.3 million SVs and 3.12 billion single nucleotide variants (SNVs) segregating between two representative haplotypes across species, with SVs impacting [~]12-fold more base pairs. SV and SNV heterozygosity are correlated across species, with endangered and threatened species exhibiting reduced genetic diversity. However, the contribution of SVs relative to SNVs fundamentally differs across major vertebrate clades: given the same number of SNVs, fishes, amphibians, and reptiles have 4.3-to-9.1 times the number of SVs than birds, and 1.7-to-3.6 times more than mammals. This reduction in the relative contribution of SVs in mammals and birds is linked to fewer non-repeat-associated SVs as well as lower transposable element (TE) abundance and diversity. We identify features underlying genomic instability across vertebrates, finding that SVs frequently occur in repetitive and SNV-rich regions and are mediated by both homology and non-canonical DNA structures. Notably, G-quadruplex structures are enriched 11.5-fold around SV breakpoints in birds, while Z-DNA structures are enriched 2.2-fold in cartilaginous fishes. TEs uniquely contribute to SVs both directly through transposition and indirectly by mediating ectopic recombination, with the proportion of TE-mediated SVs influenced by both genomic TE density and diversity. We identify >10,000 instances of recent TE turnover including extinction of LINE-2 in therian mammals and slowing of CR1 activity in passerine birds. Finally, we show that SVs have an outsized role in functional genetic variation and are >70 times more likely to strongly impact protein-coding sequences than SNVs. While SVs are on average deleterious, we identify extensive recurrent structural variation across multiple taxa in genes involved in sensory, immune, and metabolic systems. Together, this study highlights extraordinary variation in the abundance, composition, mechanism, and functional impact of SVs across vertebrates.

evolutionary biology↗

bioRxiv2026

Epigenetic conditioning improves sequence-based modeling of gene regulation across cell types and alleles

Epigenetic state modulates gene regulation in a manner not always predictable from DNA sequence alone, yet current genomic deep learning models do not leverage epigenetic state as input. We present MethylSeqNet, a model that conditions pretrained sequence embeddings on CpG methylation, a stable epigenetic mark increasingly available from long-read sequencing data. Using a novel conditioning mechanism enabling scalability and interpretability, MethylSeqNet improves predictions in cases where differential epigenetic state drives regulatory variation. We show improvements over a sequence-only baseline for cell-type-specific chromatin accessibility and transcription. Epigenetic conditioning enables prediction of phenomena not encoded in allele sequence, including parent-of-origin imprinting, random monoallelic activity, and X-inactivation. We highlight a promising application of methylation conditioning by predicting the effects of a structural rearrangement in one rare disease patient case study. In silico motif insertion analysis confirms that MethylSeqNet learns methylation-dependent regulatory grammar, establishing a paradigm for integrating epigenetic information into genomic deep learning with immediate applications in rare disease interpretation.

genomics↗

bioRxiv2026

Point cloud local ancestry inference (PCLAI): continuous coordinate-based ancestry along the genome

Local ancestry inference (LAI) predicts a discrete ancestry label for each segment of an individuals genome and has become integral to studying population history, genetic variation, and polygenic trait association. We present a new local ancestry paradigm that eschews discrete categorical labels and instead performs inference in a continuous coordinate space. We call this method "point cloud local ancestry inference" (PCLAI), since it represents an individuals genetic ancestry as a point cloud with each point corresponding to a small haplotypic segment in their genome. This formulation works in any co-ordinate space (for instance, geographic or principal components) permitting the representation of continuous genetic variation at the haplotypic-segment level without resorting to artificially constructed discrete labels. We illustrate PCLAI by training on ancient samples from multiple time periods separately, yielding chromosome paintings based on geography that are time-stratified and provide insight into how individuals genomic segments moved across space and time.

genomics↗

bioRxiv2026

Sequence models conditioned on splicing factor expression predict splicing in unseen tissues

Predicting how RNA splicing varies across tissues is important for understanding the impact of genetic variation and identifying splicing-based disease mechanisms. Although many sequence-based deep learning models have been developed to predict splicing, most predict splice sites rather than full splicing events, are restricted to tissues seen during training, or do not account for trans-regulatory variation such as differences in splicing factor expression. Here, we present Splice Ninja, a sequence-based deep learning model that predicts percent spliced-in (PSI) values for individual splicing events across tissues by conditioning on the expression levels of 301 splicing factors. Trained on PSI measurements from many different human tissues and cell types, Splice Ninja is evaluated on three entirely held-out tissues. Despite not seeing these tissues during training, it makes accurate PSI predictions and can identify a substantial fraction of splicing events with high tissue-specificity. Its performance is comparable to Pangolin [1], which is trained directly on the test tissues, but falls short of TrASPr [2], a substantially larger model also trained on the test tissues. Splice Ninja demonstrates that integrating trans-regulatory context into sequence-based splicing models enables generalization to new cellular environments. This framework offers a promising direction for building robust, context-aware predictors of alternative splicing. Our code is available at https://github.com/anikethjr/splice_ninja.

genomics↗

bioRxiv2025

Machine learning-based prediction of human structural variation and characterization of associated sequence determinants

Structural variants (SVs) represent a major source of genetic diversity and play key roles in human disease and evolution. Yet, the extent to which local sequence context shapes the likelihood of structural variant formation remains poorly quantified. Here, we develop machine learning models to predict the occurrence of SVs across the human genome and characterize genomic determinants associated with their formation. We developed both a sequence only-based convolutional neural network (CNN) model as well as a random forest approach integrating diverse genomic annotations. Both models achieve high predictive performance individually (>90% AUROC) which can be further improved in an ensemble. The predictive ability of these models demonstrates that SV-prone regions can be accurately inferred from sequence context. Model interpretability techniques reveal key genomic contributors to SVs, including effects of sequence motifs such as microhomology and non-canonical DNA structures, as well as the presence of SV hotspots. We find that different classes of SVs exhibit distinct sequence determinants, with transposable elements and inversions displaying particularly unique signatures. Moreover, predicted SV probability correlates with allele frequency and gene functional constraint, indicating the potential utility of the model for variant effect prediction. These findings demonstrate that machine learning models trained on local sequence features can identify unstable genomic regions and provide a framework for quantifying SV susceptibility and SV variant effects in personalized genomics.

genetics↗

bioRxiv2025

GUANinE v1.1 Reveals Complementarity of Supervised and Genomic Language Models

BackgroundThere has been much debate about the benefits of supervised versus unsupervised learning on genomes. Answering the question of "which is better?" requires the development of comprehensive benchmarks spanning representative functional and evolutionary tasks. Importantly, such benchmarks need large sample sizes to enable well-powered ranking of models under equivalent evaluation, such as L2-regularized probing (linear evaluation). ResultsHaving developed such an assessment (GUANinE v1.1), we can conclude that the answer is both: each paradigm outperforms on certain tasks and offers key advantages over the other. Supervised sequence-to-function models excel at annotating functional states characterized by chromatin accessibility, histone marks, or CTCF binding, while unsupervised language models outper-form on evolutionary conservation and related tasks without being limited to data-rich organisms. Our hundreds of new evaluations since v1.0 provide evidence for a direct tradeoff between input context size and model parameter count when on a fixed compute budget, which we depict with novel metrics like parameters/base pair. We also describe two new large-scale variant interpretation tasks: cadd-snv measuring proxy deleteriousness, and clinvar-snv measuring clinical pathogenicity. We find that conservation scores, and by extension, language models, dominate deleteriousness prediction, but successfully translating deleteriousness to pathogenicity remains challenging. ConclusionsGUANinE v1.1 is a large-scale and thorough evaluation of pretrained genomic models. We identify uniquely performant models across tasks, and we conclude by suggesting hybrid-supervised language models may define the next era of genomic sequence modeling.

genomics↗

bioRxiv2025

Integrated analysis of multimodal long-read epigenetic assays

Long-read sequencing assays that detect base modifications are becoming increasingly important research tools for the study of epigenetic regulation, especially with the development of DiMeLo-seq and similar methods that deposit non-native base modifications to mark a range of epigenetic features such as protein-DNA interactions and chromatin accessibility. A main benefit of these methods is their inherent capacity for multimodality, enabling the encoding of multiple genomic signals onto single nucleic acid molecules. However, there are limited tools available for visualization and statistical analysis of this type of multimodal data. Here we introduce dimelo-toolkit, a python package built to enable flexible visualizations and easy integration into custom data processing workflows. We demonstrate the utility of dimelo-toolkits preset visualizations of multiple base modifications in long-read single-molecule sequencing data with a novel extension of the DiMeLo-seq protocol that can capture three separate aspects of chromatin state on the same single reads: target protein binding, CpG methylation, and chromatin accessibility. We apply this multimodal method to simultaneously map chromatin accessibility, CpG methylation, and LMNB1 and CTCF binding patterns, respectively, in GM12878 cells. Our flexible design allows us to investigate previously unexplored technical biases that arise when working with this type of multimodal data. Additionally, we show that dimelo-toolkit enables analysis for a wide range of other long-read sequencing methods, such as mapping endogenous patterns in RNA base modifications with direct RNA sequencing. This software tool will pave the way for developing well-optimized protocols and help unlock previously inaccessible biological insights.

genomics↗

bioRxiv2025

Improving interpretability of transcription factor binding models with DNA shape features

Deep learning models in genomics that predict molecular phenotypes from DNA sequence traditionally focus on one-hot encoded nucleotide representations. Here, we develop a novel model that extends this approach by incorporating DNA structural attributes indicative of local DNA shape alongside canonical sequence inputs. This augmentation provides an additional axis for model interpretability and aids in identifying regulatory patterns not apparent from sequence alone. Applying this approach to prediction of transcription factor binding (ChIP-seq) demonstrates that combining sequence and structural DNA information can improve the identification of regulatory elements to provide a more nuanced understanding of genomic function and regulation. Graphical abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=96 SRC="FIGDIR/small/646034v1_ufig1.gif" ALT="Figure 1"> View larger version (23K): org.highwire.dtl.DTLVardef@b648b6org.highwire.dtl.DTLVardef@15fa931org.highwire.dtl.DTLVardef@15d0d31org.highwire.dtl.DTLVardef@d212e_HPS_FORMAT_FIGEXP M_FIG C_FIG Schematic overview of the DeepShape model. One-hot encoded sequence, and five DNA shape attributes--minor groove width (MGW), helical twist (HelT), propeller twist (ProT), roll, and electrostatic potential (EP)--are input separately to sequence and shape branches of the model, each processed through two convolutional layers and subsequently concatenated for further processing through additional convolutional layers.

genomics↗

bioRxiv2024

Fine-tuning sequence-to-expression models onpersonal genome and transcriptome data

Genomic sequence-to-expression deep learning models, which are trained to predict gene expression and other molecular phenotypes across the reference genome, have recently been shown to have poor out-of-the-box performance in predicting gene expression variation across individuals based on their personal genome sequences. Here we explore whether additional training (fine-tuning) on paired personal genome and transcriptome data improves the performance of such sequence-to-expression models. Using Enformer as a representative pretrained model, we explore various fine-tuning strategies. Our results show that fine-tuning improves cross-individual prediction performance over the baseline Enformer model for held-out individuals on genes seen during fine-tuning, with comparable performance to variant-based linear models commonly used in transcriptome-wide association studies. However, fine-tuning does not improve model generalizability on held-out genes, which contain sequences and variants unseen during fine-tuning, highlighting a remaining open challenge in the field.

genomics↗

bioRxiv2024

Enformation Theory: A Framework for Evaluating Genomic AI

The nascent field of genomic AI is rapidly expanding with new models, benchmarks, and findings. As the field diversifies, there is an increased need for a common set of measurement tools and perspectives to standardize model evaluation. Here, we present a statistically grounded framework for performance evaluation, visualization, and interpretation using the prominent sequence-based deep learning models Enformer and Borzoi as case studies. The Enformer model has been used for applications ranging from understanding regulatory mechanisms to variant effect prediction, but what makes it better or worse than precedent models? Does its follow-up, Borzoi, offer improved performance and more informative embeddings as well as finer resolution? Our goal is to propose a general blueprint for answering such questions and evaluating new models. We start by contrasting the few-shot performance of Enformer and Borzoi to precedent models on the GUANinE benchmark, which emphasizes complex genome interpretation tasks. We then examine Enformer and Borzoi intermediate embeddings in model-subjective principal component space, where we identify limiting aspects that affect model generalization. Finally, we present an interpretable decomposition of Enformer and Borzoi, which allows for global model interpretability and partial backtracking to explanatory causal features. Through this case study, we illustrate a new protocol, Enformation Theory, for analyzing and interpreting deep learning models in genomics.

genomics↗

bioRxiv2024

Current genomic deep learning models display decreased performance in cell type specific accessible regions

AbstractO_ST_ABSBackgroundC_ST_ABSA number of deep learning models have been developed to predict epigenetic features such as chromatin accessibility from DNA sequence. Model evaluations commonly report performance genome-wide; however, cis regulatory elements (CREs), which play critical roles in gene regulation, make up only a small fraction of the genome. Furthermore, cell type specific CREs contain a large proportion of complex disease heritability. ResultsWe evaluate genomic deep learning models in chromatin accessibility regions with varying degrees of cell type specificity. We assess two modeling directions in the field: general purpose models trained across thousands of outputs (cell types and epigenetic marks), and models tailored to specific tissues and tasks. We find that the accuracy of genomic deep learning models, including two state-of-the-art general purpose models - Enformer and Sei - varies across the genome and is reduced in cell type specific accessible regions. Using accessibility models trained on cell types from specific tissues, we find that increasing model capacity to learn cell type specific regulatory syntax - through single-task learning or high capacity multi-task models - can improve performance in cell type specific accessible regions. We also observe that improving reference sequence predictions does not consistently improve variant effect predictions, indicating that novel strategies are needed to improve performance on variants. ConclusionsOur results provide a new perspective on the performance of genomic deep learning models, showing that performance varies across the genome and is particularly reduced in cell type specific accessible regions. We also identify strategies to maximize performance in cell type specific accessible regions.

genomics↗

bioRxiv2024

Designing Cell-Type-Specific Promoter Sequences Using Conservative Model-Based Optimization

Gene therapies have the potential to treat disease by delivering therapeutic genetic cargo to disease-associated cells. One limitation to their widespread use is the lack of short regulatory sequences, or promoters, that differentially induce the expression of delivered genetic cargo in target cells, minimizing side effects in other cell types. Such cell-type-specific promoters are difficult to discover using existing methods, requiring either manual curation or access to large datasets of promoter-driven expression from both targeted and untargeted cells. Model-based optimization (MBO) has emerged as an effective method to design biological sequences in an automated manner, and has recently been used in promoter design methods. However, these methods have only been tested using large training datasets that are expensive to collect, and focus on designing promoters for markedly different cell types, overlooking the complexities associated with designing promoters for closely related cell types that share similar regulatory features. Therefore, we introduce a comprehensive framework for utilizing MBO to design promoters in a data-efficient manner, with an emphasis on discovering promoters for similar cell types. We use conservative objective models (COMs) for MBO and highlight practical considerations such as best practices for improving sequence diversity, getting estimates of model uncertainty, and choosing the optimal set of sequences for experimental validation. Using three relatively similar blood cancer cell lines (Jurkat, K562, and THP1), we show that our approach discovers many novel cell-type-specific promoters after experimentally validating the designed sequences. For K562 cells, in particular, we discover a promoter that has 75.85% higher cell-type-specificity than the best promoter from the initial dataset used to train our models.

synthetic biology↗

bioRxiv2024

Variants in tubule epithelial regulatory elements mediate most heritable differences in human kidney function

Kidney disease is highly heritable; however, the causal genetic variants, the cell types in which these variants function, and the molecular mechanisms underlying kidney disease remain largely unknown. To identify genetic loci affecting kidney function, we performed a GWAS using multiple kidney function biomarkers and identified 462 loci. To begin to investigate how these loci affect kidney function, we generated single-cell chromatin accessibility (scATAC-seq) maps of the human kidney and identified candidate cis-regulatory elements (cCREs) for kidney podocytes, tubule epithelial cells, and kidney endothelial, stromal, and immune cells. Kidney tubule epithelial cCREs explained 58% of kidney function SNP-heritability and kidney podocyte cCREs explained an additional 6.5% of SNP-heritability. In contrast, little kidney function heritability was explained by kidney endothelial, stromal, or immune cell-specific cCREs. Through functionally informed fine-mapping, we identified putative causal kidney function variants and their corresponding cCREs. Using kidney scATAC-seq data, we created a deep learning model (which we named ChromKid) to predict kidney cell type-specific chromatin accessibility from sequence. ChromKid and allele specific kidney scATAC-seq revealed that many fine-mapped kidney function variants locally change chromatin accessibility in tubule epithelial cells. Enhancer assays confirmed that fine-mapped kidney function variants alter tubule epithelial regulatory element function. To map the genes which these regulatory elements control, we used CRISPR interference (CRISPRi) to target these regulatory elements in tubule epithelial cells and assessed changes in gene expression. CRISPRi of enhancers harboring kidney function variants regulated NDRG1 and RBPMS expression. Thus, inherited differences in tubule epithelial NDRG1 and RBPMS expression may predispose to kidney disease in humans. We conclude that genetic variants affecting tubule epithelial regulatory element function account for most SNP-heritability of human kidney function. This work provides an experimental approach to identify the variants, regulatory elements, and genes involved in polygenic disease.

genetics↗

bioRxiv2024

Critical assessment of missense variant effect predictors on disease-relevant variant data

Regular, systematic, and independent assessment of computational tools used to predict the pathogenicity of missense variants is necessary to evaluate their clinical and research utility and suggest directions for future improvement. Here, as part of the sixth edition of the Critical Assessment of Genome Interpretation (CAGI) challenge, we assess missense variant effect predictors (or variant impact predictors) on an evaluation dataset of rare missense variants from disease-relevant databases. Our assessment evaluates predictors submitted to the CAGI6 Annotate-All-Missense challenge, predictors commonly used by the clinical genetics community, and recently developed deep learning methods for variant effect prediction. To explore a variety of settings that are relevant for different clinical and research applications, we assess performance within different subsets of the evaluation data and within high-specificity and high-sensitivity regimes. We find strong performance of many predictors across multiple settings. Meta-predictors tend to outperform their constituent individual predictors; however, several individual predictors have performance similar to that of commonly used meta-predictors. The relative performance of predictors differs in high-specificity and high-sensitivity regimes, suggesting that different methods may be best suited to different use cases. We also characterize two potential sources of bias. Predictors that incorporate allele frequency as a predictive feature tend to have reduced performance when distinguishing pathogenic variants from very rare benign variants, and predictors supervised on pathogenicity labels from curated variant databases often learn label imbalances within genes. Overall, we find notable advances over the oldest and most cited missense variant effect predictors and continued improvements among the most recently developed tools, and the CAGI Annotate-All-Missense challenge (also termed the Missense Marathon) will continue to assess state-of-the-art methods as the field progresses. Together, our results help illuminate the current clinical and research utility of missense variant effect predictors and identify potential areas for future development.

genetics↗

bioRxiv2023

Characterizing uncertainty in predictions of genomic sequence-to-activity models

Genomic sequence-to-activity models are increasingly utilized to understand gene regulatory syntax and probe the functional consequences of regulatory variation. Current models make accurate predictions of relative activity levels across the human reference genome, but their performance is more limited for predicting the effects of genetic variants, such as explaining gene expression variation across individuals. To better understand the causes of these shortcomings, we examine the uncertainty in predictions of genomic sequence-to-activity models using an ensemble of Basenji2 model replicates. We characterize prediction consistency on four types of sequences: reference genome sequences, reference genome sequences perturbed with TF motifs, eQTLs, and personal genome sequences. We observe that models tend to make high-confidence predictions on reference sequences, even when incorrect, and low-confidence predictions on sequences with variants. For eQTLs and personal genome sequences, we find that model replicates make inconsistent predictions in >50% of cases. Our findings suggest strategies to improve performance of these models.

genomics↗

bioRxiv2023

GUANinE v0.9: Benchmark Datasets for Genomic AISequence-to-Function Models

Computational genomics increasingly relies on machine learning methods for genome interpretation, and the recent adoption of neural sequence-to-function models highlights the need for rigorous model specification and controlled evaluation, problems familiar to other fields of AI. Research strategies that have greatly benefited other fields -- including benchmarking, auditing, and algorithmic fairness -- are also needed to advance the field of genomic AI and to facilitate model development. Here we propose a genomic AI benchmark, GUANinE, for evaluating model generalization across a number of distinct genomic tasks. Compared to existing task formulations in computational genomics, GUANinE is large-scale, de-noised, and suitable for evaluating pretrained models. GUANinE v1.0 primarily focuses on functional genomics tasks such as functional element annotation and gene expression prediction, and it also draws upon connections to evolutionary biology through sequence conservation tasks. The current GUANinE tasks provide insight into the performance of existing genomic AI models and non-neural baselines, with opportunities to be refined, revisited, and broadened as the field matures. Finally, the GUANinE benchmark allows us to evaluate new self-supervised T5 models and explore the tradeoffs between tokenization and model performance, while showcasing the potential for self-supervision to complement existing pretraining procedures.

genomics↗

bioRxiv2023

Personal transcriptome variation is poorly explained by current genomic deep learning models

Genomic deep learning models can predict genome-wide epigenetic features and gene expression levels directly from DNA sequence. While current models perform well at predicting gene expression levels across genes in different cell types from the reference genome, their ability to explain expression variation between individuals due to cis-regulatory genetic variants remains largely unexplored. Here we evaluate four state-of-the-art models on paired personal genome and transcriptome data and find limited performance when explaining variation in expression across individuals.

genomics↗