Maria Chikina

Allele-specific quantification of sequencing data allows for a systematic investigation of how DNA sequence variations influence cis gene regulation. Current methods for analyzing allele-specific measurements for causal analysis rely on statistical associations between genetic variation across individuals and allelic imbalance. Instead, we propose DeepAllele, a novel deep learning sequence-to-function model using paired allele-specific input, designed to learn sequence features that predict subtle changes in gene regulation between alleles. Our approach is suited for datasets with unambiguous phasing, such as F1 hybrids and other controlled genetic crosses, or long-read sequencing technologies used in Fiber-seq, in which reads can be assigned to complete allele sequences. We apply our framework to allele-specific measurements in immune cells from F1 hybrid mice, and show that the model's additionally learned cis -regulatory grammar aligns with known biological mechanisms across a significantly larger number of genomic regions compared to baseline models. In summary, our work presents a computational framework to leverage genetic variation to uncover functionally-relevant regulatory motifs, enhancing discovery in genomics.

Motivation: Convergent evolution, or the independent acquisition of similar phenotypes in distinct lineages, provides a powerful framework for investigating genomic changes associated with a phenotype. This paper details an update to RERconverge, a powerful R package that tests for associations between gene relative evolutionary rates (RERs) and convergent phenotypes to infer genomic regions associated with traits or selective pressures. We introduce new customizable analysis choices and scalable and efficient algorithms that can process larger genomic datasets, a critical improvement as genomic data become available for more species. Results: Modifications to core functions in the RERconverge pipeline resulted in an immense speedup (by a factor of up to 28.6). The function that tests for associations between phenotypes and RERs has been expanded to include two new analytical methods for outlier control; we also provide here a summary of the statistical tests users can perform, along with their use cases. Availability and implementation: The code and walkthrough vignettes for the package are available at https://github.com/nclark-lab/RERconverge . Contact: Nathan L. Clark nclark@pitt.edu ; Maria Chikina mchikina@pitt.edu

Sequence-to-function (S2F) models can evaluate arbitrary DNA sequences, yet they struggle to fully capture inter-individual variation in gene expression. We introduce SAGE-net, a scalable framework for training and evaluating S2F models using personal genomes. While personal genome training improves gene expression prediction accuracy for held-out individuals, performance gains arise primarily from identifying predictive variants rather than learning a cis -regulatory grammar that generalizes across loci. Scalable software will be critical to advancing S2F models for personal genomics.

Sequence-to-function (S2F) models trained on reference genomes have achieved strong performance on regulatory prediction and variant-effect benchmarks, yet they still struggle to predict inter-individual variation in gene expression from personal genomes. We evaluated AlphaGenome on personal genome prediction in two molecular modalities--gene expression and chromatin accessibility--and observed a striking dichotomy: AlphaGenome approaches the heritability ceiling for chromatin accessibility variation, but remains far below baseline for gene-expression variation, despite improving over Borzoi. Context truncation and fine-mapped QTL analyses indicate that accessibility is governed by local regulatory grammar captured by current architectures, whereas gene-expression variation requires long-range regulatory integration that remains challenging.

Convergent evolution offers a natural test for adaptive predictability, yet pinpointing its molecular basis remains difficult. Current methods often grapple with distinguishing adaptive convergence from background noise, either by demanding overly stringent identical substitutions or relying on low-resolution evolutionary rate shifts. Here we introduce BUSTED-PH (Branch-site Unrestricted Statistical Test for Episodic Diversification – Phenotype), a branch-site codon model that detects phenotype-associated episodic diversifying selection. Unlike standard approaches, BUSTED-PH explicitly contrasts selective regimes between phenotype-positive (foreground) and phenotype-negative (background) lineages, effectively winnowing out spurious associations driven by pervasive background adaptation. BUSTED-PH has already been applied in numerous independent studies, and here we rigorously validate it using canonical positive controls and simulations to confirm high power and strict false positive control. A genome-wide scan of 120 mammalian species using strict statistical criteria ( FDR ≤ 0.01) identified 72 genes associated with echolocation; while recovering paradigmatic auditory drivers (e.g., Prestin, TMC1 ), we also uncover novel candidates in physiological support systems ranging from lipid homeostasis to neural development. A parallel analysis of mammalian gigantism identifies 91 genes linked to musculoskeletal reinforcement, organ size governance, and genomic integrity, characterizing the molecular adaptations required to support massive body size. As projects like Zoonomia and the Vertebrate Genomes Project expand comparative datasets, BUSTED-PH provides a robust framework for dissecting the genetic architecture of complex convergent traits.

Motivation: The rapid success of deep learning sequence-to-function (S2F) models has driven a trend toward ever larger architectures for regulatory genomics. While these models achieve meaningful predictive performance, their growing size and computational cost have made them increasingly opaque, difficult to deploy, and impractical for widespread use. As S2F models mature, an open question is how much model capacity is truly necessary to represent regulatory sequence function. Results: We show that, within linear layers, much of the predictive power of state-of-the-art S2F models is concentrated in low-rank structure. Using post hoc singular value decomposition, we construct low-rank approximations of the Sei model that reduce model size by up to 90% while preserving over 90% correlation with full-model predictions. Strikingly, extremely low-rank models — including rank 1 — can match or exceed full-model performance on several variant-effect benchmarks, indicating that dominant regulatory signals lie in a surprisingly low-dimensional subspace. Combining low-rank approximation with static model quantization enables practical CPU inference, yielding over 5× speedups relative to the full Sei model and more than 100× faster inference than larger S2F models on comparable tasks. Applying the same approach to Borzoi and Enformer reveals consistent behavior, demonstrating that low-rank structure is a general property of modern S2F architectures rather than a model-specific artifact. Availability and Implementation: Low-rank Sei is available at https://github.com/kostkalab/seillra . Contact: mchikina@pitt.edu , kostka@pitt.edu

Cis-regulatory elements shape gene expression by recruiting transcription factors (TFs) to DNA motifs, yet how motifs cooperate or compete across distances remains poorly understood. Existing deep learning models predict TF binding with high accuracy but rely on computationally intensive post-hoc analyses to infer regulatory syntax, limiting interpretability and scalability. Here we present InfluenceNet, a transparent convolutional architecture that directly encodes motif influence-a position– and motif-specific profile quantifying how one motif modulates binding at another. Trained on ChIP-nexus data for pluripotency TFs, InfluenceNet achieves performance comparable to state-of-the-art models while providing immediate, global interpretability. The learned influence profiles recapitulate known principles of motif syntax, including Oct4-Sox2 pioneering activity and Nanog’s striking 10.5-bp helical periodicity, and reveal a cooperative Nanog-Nanog grammar supported by de novo motif discovery. Integrating these results with structural and molecular-dynamics modeling, we propose that Nanog forms chain-like oligomers on DNA through asymmetric contacts between WRD (tryptophan repeat domain) and DBD (DNA-binding domain) that maintain helical phase alignment. Together, these findings establish InfluenceNet as a general framework for accurate prediction and mechanistic interpretation of cis-regulatory grammar, bridging deep learning and biophysical insight.

The spatial folding of the genome shapes gene regulation by controlling which loci interact, yet inferring the mechanisms behind these 3D structures from contact maps remains difficult. Cohesin-mediated loop extrusion is a key organizer of domains and loops, but existing methods either predict contacts without mechanistic insight or simulate extrusion with limited scalability. We present the differentiable loop extrusion model (dLEM), a scalable framework that reformulates extrusion as a smooth, trainable process. dLEM represents extrusion through position-specific velocity profiles for leftward and rightward cohesin movement. Fitting dLEM to chromosome conformation capture data yields a one-dimensional, interpretable description of extrusion dynamics that aligns with genomic and epigenomic features. dLEM parameters also capture architectural changes under CTCF and WAPL perturbations, enabling genome-wide prediction of extrusion disruptions. Extending our observations, we demonstrate that dLEM can be seamlessly incorporated into deep learning models to infer extrusion parameters directly from sequence and chromatin features, reducing model complexity by nearly three orders of magnitude while preserving predictive accuracy. Indeed, when incorporated into deep dLEM, dLEM acts as a biophysically-motivated layer for long range genomic communication, and together they provide a predictive, inter-pretable framework linking 1D genomic features to 3D chromatin folding and its response to sequence and chromatin state, with dLEM’s mechanistic modeling enabling prediction of trans factor perturbation effects.

Sequence variation within transcription factor (TF)-binding sites can significantly affect TF–DNA interactions, influencing gene expression and contributing to disease susceptibility or phenotypic traits. Despite recent progress in deep sequence-to-function models that predict functional output from sequence data, these methods perform inadequately on some variant effect prediction tasks, especially with common genetic variants. This limitation underscores the importance of leveraging biophysical models of TF binding to enhance interpretability of variant effect scores and facilitate mechanistic insights. We introduce motifDiff, a novel computational tool designed to quantify variant effects using mono- and dinucleotide position weight matrices. motifDiff offers several key advantages, including scalability to score millions of variants within minutes, implementation of statistically rigorous normalization strategy critical for optimal performance, and support for both dinucleotide and mononucleotide models. We demonstrate motifDiff’s efficacy by evaluating it across diverse ground truth datasets that quantify the effects of common variants in vivo, thereby establishing robust benchmarks for the predictive value of variant effect calculations. Finally, we show that our tool provides unique insights when interpreting human accelerated regions. motifDiff is available as a standalone Python application at https://github.com/rezwanhosseini/MotifDiff.