Avanti Shrikumar

This study examines functional connectivity changes in the Default Mode Network of Australian Rules Football players using resting-state fMRI to assess potential as an early biomarker for non-concussive impacts.

This paper develops zero-inflated regression models to better analyze Coronary Artery Calcium Scores by addressing the high prevalence of zeros in the data.

The paper introduces Affinity Distillation (AD), a method to extract thermodynamic affinities from deep learning models of in vivo TF binding, showing it outperforms motif-based methods in predicting sequence variation impacts.

This study improves cross-species prediction of transcription factor binding by addressing domain shifts caused by species-specific repeats using domain adaptive neural networks.

fastISM accelerates in-silico saturation mutagenesis for convolutional neural networks by 10x, making it feasible for large-scale sequence analysis.

This paper benchmarks deep learning interpretation methods on more realistic simulated regulatory genomic data, finding that some methods perform poorly on complex datasets.

The paper introduces BPNet, a deep learning model that predicts base-resolution ChIP-nexus binding profiles of TFs and identifies soft syntax rules for cooperative TF binding.

This paper evaluates strategies for enforcing reverse-complement equivariance in deep learning models for regulatory genomics, finding post-hoc conjoined models perform best.

This paper introduces a Fourier-transform-based attribution prior to improve deep learning model stability, interpretability, and performance in genomics tasks.

The paper introduces MPRA-DragoNN, a CNN model to predict regulatory DNA activity from MPRAs, identifies TF motifs, and explores variant effects in lipid-associated loci.

gkmexplain provides a fast and accurate method to interpret gkm-SVM models of regulatory DNA sequences, outperforming existing methods in identifying predictive patterns and mutation impact scores.

Kipoi provides a framework for standardized, reusable machine learning models in genomics, enabling tasks like variant effect prediction and model benchmarking.

Deep learning rivals or beats baselines across imaging, health records, genomics and drug discovery but is not yet transformative; scarce labels, privacy and opacity remain hurdles.

This paper introduces a separable fully connected layer for deep learning in genomics, showing improved performance and biological consistency over conventional models.

Standard genomic CNNs can score a sequence and its reverse complement differently; four weight-sharing layer types remove that inconsistency and are reported to train faster.