Peter K. Koo

This paper introduces PICASSO, a framework that makes pathology foundation models interpretable and controllable by decomposing embeddings into human-interpretable visual concepts, enabling new biological insights and downstream applications.

This study introduces pVEP, a framework showing how genetic background affects AI predictions of variant effects, revealing variant effects vary across haplotypes.

This paper introduces DEGU, a method that improves deep learning reliability in genomics by combining ensemble learning and knowledge distillation to capture both average predictions and uncertainty.

NTv3 is a foundation model that predicts genome function, annotates genes, and designs DNA sequences across species, outperforming existing models in multiple tasks.

This paper introduces SEAM, a deep learning framework that identifies functional binding sites and mutation impacts in regulatory DNA, revealing two classes of regulatory signals in human and Drosophila loci.

This study evaluates pre-trained DNA language models' ability to predict and interpret cell-type-specific functional genomics data, finding they do not outperform conventional machine learning approaches.

GAME provides a standardized framework for benchmarking genomics models across tasks, enabling consistent evaluation and community contributions.

The paper evaluates and optimizes sequence-based gene regulatory deep learning models using a DREAM Challenge dataset, showing improvements through architectural and training strategy refinements.

The paper experimentally quantifies amyloid nucleation for >100,000 sequences, trains an interpretable deep learning model to predict it, and reveals the model's decision-making process.

This paper introduces SQUID, a framework for interpreting genomic deep neural networks using surrogate models, showing improved motif consistency and variant-effect predictions.

This paper introduces D3, a generative model for designing regulatory DNA with tunable activity, showing improved performance over baselines in sequence generation and activity prediction.

This paper extends EvoAug, an evolution-inspired data augmentation method for genomic deep learning, to TensorFlow, demonstrating comparable performance to the original PyTorch-based version.

This paper introduces CREME, a toolkit to interpret cis-regulatory interactions from deep neural networks, revealing gene regulation rules and functional sequence elements.

The paper corrects gradient-based interpretations of deep neural networks in genomics to reduce noise from off-simplex gradients, improving attribution method interpretability.

EvoAug improves genomic DNN generalization and interpretability through evolution-inspired data augmentations while preserving functional integrity.

The paper analyzes single-cell CRISPR perturbations to show enhancers act multiplicatively with limited evidence for epistatic interactions.

This study evaluates deep learning models for predicting chromatin accessibility, showing that quantitative models outperform binary ones in generalization and interpretability.

The paper benchmarks neuronal cell-type annotation in scATAC-seq data using marker genes and introduces a deep learning model to predict chromatin accessibility from DNA sequence.

This paper introduces a differentiable Smith-Waterman algorithm to jointly learn multiple sequence alignments and downstream tasks, showing improvements in contact prediction and structure prediction.

The paper shows that a single attention layer can predict protein contacts as well as Potts models and pretrained Transformers, suggesting attention models can capture meaningful protein features from unlabeled data.

This study investigates how architectural choices in convolution-based hybrid networks affect motif representations and the interpretability of their attention maps in genomics.

This paper introduces Global Importance Analysis (GIA), a method to quantify the effect size of genomic features in deep neural networks, demonstrating its application in predicting RNA-protein interactions.

This study shows that using exponential activations in the first layer of CNNs improves the interpretability and robustness of genomic sequence motif representations in both synthetic and real DNA data.

This study evaluates how well generative sequence models capture structural and fitness patterns from protein families, finding that structural and fitness information are not necessarily correlated.

This paper argues for global importance analysis to quantify population-level importance of genomic features in deep neural networks, beyond traditional attribution methods.

This study investigates how CNN architecture affects the learning of genomic sequence motifs, finding that hierarchical designs lead to distributed representations while non-hierarchical designs yield interpretable localist representations.

This study shows that training deep neural networks with robustness methods improves interpretability in genomics tasks.

This paper introduces ResidualBind, a convolutional residual network that infers RNA-binding protein sequence and structure specificities from enriched sequence data, outperforming previous methods and learning RNA structure context from sequence data.

The paper improves a deep learning method for inferring functional neural connectivity from neural activity data, achieving high precision on synthetic datasets.