Maria D. Chikina

This paper shows that low-rank approximations of DNA sequence-to-function models can achieve high performance while significantly reducing model size and computational cost.

The paper updates the RERconverge package with faster algorithms and new analysis features for studying convergent evolution at the genomic level.

This paper introduces SAGE-net, a framework for training sequence-to-function models using personal genomes, showing improved gene expression prediction but limited generalization, while DNA methylation shows better generalization.

The paper identifies tissue-specific transcriptional regulatory programs in rats undergoing endurance exercise training by integrating chromatin accessibility, DNA methylation, and gene expression data.

The paper evaluates AlphaGenome's performance in predicting gene expression and chromatin accessibility from personal genomes, finding it excels in accessibility but struggles with gene expression.

BUSTED-PH identifies genes associated with convergent phenotypes by detecting episodic diversifying selection across mammalian species.

The paper introduces MORALE, a domain adaptation framework for cross-species TF binding prediction, achieving state-of-the-art performance without adversarial training.

InfluenceNet directly encodes motif influence in cis-regulatory elements, enabling interpretable prediction of regulatory syntax with performance comparable to state-of-the-art models.

This paper introduces dLEM, a differentiable loop extrusion model that predicts genome folding mechanisms from contact maps and integrates with deep learning to infer extrusion parameters from genomic features.

The paper introduces motifDiff, a tool for variant effect prediction using biophysical TF binding models, showing it effectively scores variants and provides insights into human accelerated regions.

DALE-Eval benchmarks cell type-specific expression deconvolution methods with new metrics and performance insights.

CLAMP integrates genetic associations and drug signatures with large-scale RNA-seq data to identify interpretable gene modules linked to complex diseases.

This paper introduces DeepAllele, a deep learning model for allele-specific gene regulation analysis, showing improved detection of cis-regulatory elements in F1 hybrid mice immune cells.

The paper identifies context-dependent interactors of TDP-43 that regulate its dysfunction in ALS/FTLD using proximity labeling, mass spectrometry, and snRNA-seq integration.

ERC2.0 improves functional interaction prediction across large phylogenies by enhancing algorithm speed, normalization, and statistical power.

The paper introduces L0 segmentation, a framework for extracting locally coherent signals from diverse epigenomic data, showing it can identify subtle features missed by other methods.

BUSTED-E detects positive selection while identifying residual alignment errors in genome-scale datasets, improving accuracy and biological interpretation.

The paper expands RERconverge to analyze multi-categorical traits by introducing methods for ancestral state reconstruction, statistical tests, and phylogenetic permutations, showing improved performance in identifying diet-related genes and pathways.

This study benchmarks cell-type deconvolution methods using a new simulation approach that introduces biological heterogeneity, showing that marker-based methods and BayesPrism perform best.

InstaPrism re-implements BayesPrism with a fixed-point algorithm to improve computational efficiency while maintaining accuracy in cell-type deconvolution.

The paper introduces mousiPLIER, a mouse pathway-level information extractor model trained on brain RNA-seq data, which reduces dimensionality and identifies aging-related latent variables.

The paper identifies proteins and regulatory elements associated with vocal learning by analyzing bat neural activity and epigenomic data.

This study compares the impact of different optimization methods on model performance and sparsity in transcriptomic prediction using LASSO logistic regression.

This study benchmarks deep neural networks for predicting gene expression from DNA sequence across individuals and identifies limitations in predicting variant effects.

The paper introduces a method to approximate in silico saturation mutagenesis using first-order Taylor expansion, reducing computational cost while maintaining robustness.

The paper introduces tiSFM, an interpretable neural network for sequence-to-function learning, which outperforms existing models in predicting open chromatin and transcription factor activities.

This study profiles multi-omic changes in multiple tissues of rats during endurance exercise training, identifying molecular adaptations and sex-specific responses across time.

The paper introduces NIFA, a probabilistic factor analysis model for scRNA-seq data that disentangles discrete cell-type identity and continuous pathway activity, outperforming existing methods in biological interpretation.

The paper evaluates the effectiveness of various -omics data types in predicting cancer mutation states and finds that RNA sequencing is often the best, but other data types can be equally effective.

DataRemix is a normalization method that enhances biological signal detection in gene expression data by tuning hidden factors, outperforming existing methods without external knowledge.