Su‐In Lee

ADVISE identifies context-specific Alzheimer's disease progression signatures using an explainable AI framework and validates ROBO2 as a candidate regulator in excitatory neurons.

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.

ACE identifies universal aging signatures across species by disentangling aging-related gene expression from other biological variations using explainable AI.

This study shows that genetic risk variants in endolysosomal network genes correlate with endolysosomal dysfunction in neural cells in Alzheimer's disease.

The paper introduces milVI, a deep generative model that infers missing phenotype labels and connects cell-level data to donor-level phenotypes in single-cell studies.

The paper introduces DeepProfile, a deep learning framework for analyzing gene expression in 18 cancers, revealing biological insights through latent space analysis and linking gene expression to clinical outcomes.

StrastiveVI isolates aging and Alzheimer's disease-associated patterns in scRNA-seq data by separating them from sex, tissue, and cell type variations.

The paper introduces contrastiveVI, a method to disentangle shared and treatment-specific variations in scRNA-seq data, showing improved performance over existing methods in four datasets.

This paper introduces a novel unsupervised pathway attribution method that identifies major transcriptomic variation sources when combined with biologically-constrained models, applied to Alzheimer's disease brain samples.

The paper introduces MD-AD, a deep learning framework that identifies non-linear gene expression-neuropathology relationships in Alzheimer's disease using multi-cohort data, revealing sex-specific immune response patterns.

This paper introduces HD-AE, a deep autoencoder that learns transferable representations of scRNA-seq data across datasets, enabling comparison without retraining.

The paper introduces AD-AE, an adversarial deconfounding autoencoder that generates gene expression embeddings free of confounders and capable of generalizing across domains.

AIControl improves ChIP-seq peak identification by replacing matched control experiments with machine learning, enhancing accuracy without requiring control datasets.

DECODER integrates large-scale gene expression data to identify mitochondrial Complex I as a therapeutic target for Alzheimer’s disease.

This paper introduces PIPER, a method to identify gene network perturbations in single-cell RNA-seq data, showing it outperforms existing methods on simulated data and identifies known regulators in real datasets.

MD-AD improves prediction of Alzheimer's disease hallmarks and staging systems by combining heterogeneous datasets with multi-task deep learning.

DeepProfile trains a variational autoencoder on thousands of unlabeled gene expression samples to learn low-dimensional representations that improve prediction of acute myeloid leukemia drug responses and cancer phenotypes.

The paper develops a machine learning system called Prescience to predict hypoxemia risk during surgery and improve anesthesiologists' ability to prevent it by providing interpretable explanations.

DeepATAC predicts regulatory factor binding activity from ATAC-seq signals and DNA sequence, outperforming existing methods.