David R. Kelley

This paper introduces Akita Semifreddo, a framework for designing DNA sequences that produce specific 3D genome folding patterns, demonstrating the ability to create boundaries, dots, and flames with controlled strengths.

The paper introduces ecCLIP for high-resolution RBP binding maps and Pythia for structure-aware RNA-protein interaction modeling, improving RNA half-life prediction and variant effect assessment.

The paper introduces Cerberus, an improved regulatory sequence model using bidirectional state space blocks, achieving higher accuracy and efficiency than existing models like Borzoi.

The paper introduces Multi-ISM, a scalable framework for saturation mutagenesis that reduces computational cost while maintaining accuracy, and applies it to 5,000 genes to map regulatory elements and rare variant effects.

This paper introduces Shorkie, a DNA language model pretrained on fungal genomes that improves gene expression prediction in budding yeast by capturing regulatory motif dynamics and variant effects.

Sniff improves causal variant prioritization in GWAS by integrating Borzoi-predicted variant effects with functional annotations, identifying more fine-mapped variants with higher posterior inclusion probabilities.

The paper trains a model on single-cell 3-seq data to predict cell type-specific coverage profiles and improves variant interpretation across diverse cell types.

The paper introduces Variant-FlowFISH to measure the effects of DNA sequence variants on gene expression, identifies regulatory mechanisms, and benchmarks predictive models.

This study evaluates parameter-efficient fine-tuning of the Borzoi model to improve gene regulation analysis across new RNA-seq datasets with high accuracy and efficiency.

This paper improves DNA CNNs' ability to predict indel effects on gene regulation by addressing technical variance through shift augmentation and introduces an interpretation technique validated with experimental data.

This paper compares state-space models and transformers for predicting RNA-Seq read coverage, finding Mamba-based models slightly outperform transformers in accuracy but not in downstream SNP classification.

The paper uses AkitaV2 to analyze CTCF-binding site sequences and predict their role in genome folding, identifying key sequence determinants and grammatical rules for insulation strength and pattern formation.

This paper introduces Borzoi, a model that predicts RNA-seq coverage from DNA sequence, enabling variant effect scoring across multiple regulatory layers and outperforming existing methods on QTLs.

ENCODE-rE2G, a classifier trained on K562 CRISPR perturbations, maps 13,455,443 enhancer-gene interactions across 352 biosamples, matching or exceeding prior models on CRISPR, eQTL, and GWAS benchmarks.

This paper creates a single-cell atlas of aging in C. elegans, revealing cell-type specific aging patterns and identifying transcription factors that regulate aging rates.

The paper compiles mRNA decay rate datasets from humans and mice, identifies biases, corrects them, and trains a deep learning model to predict half-life from sequence with improved accuracy.

scBasset uses convolutional neural networks to model scATAC data, achieving state-of-the-art results in cell type identification and other tasks.

The paper analyzes replicative senescence in WI-38 cells using multi-omics data, finding early senescence markers, myofibroblast-like transitions, and YAP1/TEAD1 regulation.

Enformer, a sequence-based transformer integrating elements up to 100 kb away, lifts held-out gene expression prediction (correlation 0.81 to 0.85 versus Basenji2) and variant effect prediction.

scNym transfers cell identity annotations across single-cell experiments using semi-supervised adversarial neural networks, outperforming existing methods.

The paper quantifies alternative polyadenylation (APA) across 2 million single cells from developing mouse embryos, revealing dynamic changes in 3' UTR length and identifying potential regulators in neuronal and hematopoietic lineages.

The paper uses supervised learning to fine-map cis-eQTLs, identifying 20,913 additional putative causal eQTLs and 310 candidate genes.

The paper investigates age-related changes in muscle progenitors by analyzing single-cell RNA-seq data from young and aged mice, finding that differentiation amplifies age-related changes in progenitors but masks some changes.

This paper introduces a deep learning model that predicts 3D genome folding from DNA sequence and highlights the role of CTCF in genome structure.

This study evaluates how well deep learning annotations predict disease heritability and finds they provide limited unique information beyond existing functional annotations.

The paper develops a cross-species approach to predict regulatory sequence activity using deep learning, improving gene expression prediction and enabling transfer learning from mouse to human data.

This study uses single-cell RNA-seq to analyze aging in mice, finding that cell identity has a greater influence on aging trajectories than tissue environment.

Solo identifies doublets in scRNA-seq data using semi-supervised deep learning, improving accuracy over existing methods.

The paper introduces a method to detect directional effects of transcription factor binding on disease risk using GWAS data and applies it to identify new causal genes and mechanisms.

The paper develops a machine learning system using convolutional neural networks to predict cell type-specific epigenetic and transcriptional profiles from DNA sequence, showing alignment with eQTL causal variants.