Biology subjects

Lee, S.-I.

Publications and source records attributed to Lee, S.-I..

At least 19 recordsLinked to original sources

bioRxiv2026

Dissecting and directing pathology foundation models

Foundation models (FMs) are central to digital pathology, encoding histology images into dense embeddings for facilitating diagnostic classification, molecular alteration prediction, and clinical outcome modeling. However, the opacity of these embeddings renders FM-based systems "black boxes," limiting their trustworthiness for clinical translation and utility for scientific discovery. Here, we introduce PICASSO (Pathology Image Concept Atlas built via SparSe dictiOnary learning), a framework that makes pathology FMs interpretable and controllable. PICASSO decomposes FM embeddings into human-interpretable visual concepts using a sparse autoencoder. It is trained on more than 120 million tissue patches across 32 cancer types, producing the first pan-cancer atlas of histomorphological concepts. We demonstrate that PICASSO enables diverse downstream applications of FM embeddings by exposing interpretable structure within learned representations and supporting concept-level intervention. It enables auditing of clinical model behavior by revealing the morphological features driving predictions. Beyond transparency and validation, PICASSO enables the discovery of new biological insights; for example, it identified hobnailing epithelial morphology as a previously unrecognized biomarker of EGFR mutations in lung adenocarcinoma. By linking PICASSO-derived concepts with spatial transcriptomics, we uncover associations between morphological patterns and gene expression programs. Furthermore, PICASSO allows suppression of concepts associated with technical artifacts, thereby reducing model reliance on spurious signals. Finally, PICASSO enables controlled manipulation of learned concepts to generate counterfactual embeddings for exploratory therapeutic analysis, such as modulating tumour-infiltrating lymphocyte density to assess impacts on predict survival outcomes. Together, PICASSO provides a principled framework for transforming pathology FMs into platforms for mechanistic insight and discovery.

pathology↗

bioRxiv2026

Interactome mapping in human excitatory neurons reveals novel risk genes and pathways in Alzheimer's disease

Alzheimers disease (AD) is an irreversible neurodegenerative disease defined by its molecular hallmarks - amyloid beta peptide plaques and neurofibrillary Tau tangles. Despite significant progress that has been made in uncovering a large number of genetic risk factors through extensive genomic sequencing and genetic studies, the molecular mechanisms driving AD-associated pathology and cognitive decline remain poorly understood. Therefore, alongside the identification of more risk genes, it is also paramount to study how these genes function and influence each other within the cellular pathways and overall molecular networks in AD-relevant brain cell types. However, current human protein-protein interactome datasets were all generated in either yeast or generic human cell lines. Consequently, many important neuronal interactions, especially neuron-specific ones, have yet been discovered. To address this critical gap, we developed a highly scalable, high-quality interactome mapping pipeline in human excitatory neurons derived from induced pluripotent stem cells (iPSC), and generated a comprehensive, neuron-specific interactome map, named ADNeuronNet, for key AD risk genes. ADNeuronNet consists of 1,767 high-confidence interactions among 1,189 proteins and is the only dataset enriched with neuron-specific genes when compared to known protein interactions, including previous large-scale interactome maps, for the same baits in the literature. Within ADNeuronNet, we identified 1,375 novel interactions, many of which are likely neuron specific. For example, we identified a neuron-specific interactor, RIN2, for major AD risk factor BIN1 and confirmed RIN2s function in recruiting BIN1 to RAB5 positive early endosomes, a process that has been well-associated with AD etiology. Additionally, we performed quantitative interaction perturbation analyses on AD risk genes with AD-associated mutations or isoforms and identified significant changes in 99 protein interactions among 11 different protein variants. Finally, we found that subunits from the anaphase-promoting complex/cyclosome (APC/C), another novel BIN1 interactors identified by ADNeuronNet, mediated modulation of Tau-aggregation in neurons via regulation of APOE expression, uncovering a previously unrecognized BIN1-APC/C-APOE regulatory axis in AD pathobiology. In summary, these findings illustrate how our neuron-specific ADNeuronNet can be leveraged to uncover new risk gene candidates and cellular pathways that help advance our understanding of molecular mechanisms underlying AD etiology.

systems biology↗

bioRxiv2026

cGAS inhibition delays TDP-43-driven ALS Pathogenesis

Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disorder marked by motor neuron loss and cytoplasmic mislocalization of TAR DNA-binding protein 43 (TDP-43), a key regulator of RNA splicing. However, the upstream modulators of this process remain poorly defined. Here we identify cyclic GMP-AMP synthase (cGAS) as a central mediator of TDP-43 pathology and associated mis-splicing. cGAS expression was elevated in ALS patient brains and enriched across activated microglia. In human iPSC-derived microglia-motor neuron co-cultures, neuronal TDP-43 pathology triggered microglial cGAS activation, whereas pharmacological inhibition with a potent human cGAS inhibitor reduced phosphorylated TDP-43, restored lysosomal and phagocytic programs, normalized microglial reactivity, and reversed TDP-43-associated RNA splicing defects. In vivo, cGAS inhibition in TDP-43 Q331K mice reversed widespread RNA splicing abnormalities across neurons and oligodendrocyte lineage cells, attenuated neurodegenerative pathology, and preserved motor function. Together, these findings identify cGAS as a druggable upstream regulator linking innate immune signaling to TDP-43-dependent RNA mis-splicing and neurodegeneration, and establish cGAS inhibition as a promising therapeutic strategy for ALS.

neuroscience↗

bioRxiv2025

An Explainable AI Framework for Identifying Universal Aging Signatures in Cell Embeddings

Aging is a complex biological process marked by progressive physiological decline and increased disease vulnerability. Single-cell RNA sequencing offers unprecedented resolution for studying aging, yet isolating aging-related signatures remains challenging because gene expression is primarily shaped by other factors such as cell type, tissue, and sex. We present ACE (Aging Cell Embeddings), an explainable deep generative framework that disentangles aging-related gene expression changes from background biological variation. ACE employs two latent representations: one capturing aging-related signatures and another representing non-aging-related variation in the data. Through explainable AI, ACE identifies key genes and pathways associated with aging amid dominant non-aging-related variations. Applied to large-scale mouse, fly, and human datasets, ACE uncovers aging signatures both within specific tissue-cell-type contexts and across all tissues and cell types, enabling accurate prediction of biological age. Moreover, ACE identifies aging genes conserved across species, highlighting its ability to reveal shared biological mechanisms of aging. Experimental RNAi knockdowns in C. elegans validate ACEs findings, confirming its ability to prioritize novel aging genes affecting lifespan. ACE reveals key pathways involved in proteostasis, immune regulation, and extracellular matrix remodeling, and identifies Uba52 through the cross-species model as an important aging gene, whose knockdown in C. elegans significantly shortens lifespan. By providing interpretable and generalizable aging embeddings, ACE establishes a foundation for cross-species single-cell aging studies and translational geroscience.

bioinformatics↗

bioRxiv2024

A deep generative model for capturing cell to phenotype relationships

Single-cell omics has proven to be a powerful instrument for exploring cellular diversity. With advances in sequencing protocols, single-cell studies are now routinely collected from large-scale donor cohorts consisting of samples from hundreds of donors with the goal of uncovering the molecular bases of higher-level donor phenotypes of interest. For example, to better understand the mechanisms behind Alzheimers disease, recent studies with up to hundreds of samples have investigated the relationships between single-cell omics measurements and donors neuropathological phenotypes (e.g. Braak staging) [4, 9, 3, 10]. In order to ensure the robustness of such findings, it may be desirable to aggregate data from multiple distinct donor cohorts. Unfortunately, doing so is not always straightforward, as different cohorts may be equipped with different sets of phenotype labels. Continuing the previous Alzheimers example, recent AD study cohorts have provided various subsets of neuropathological phenotypes, cognitive testing results, and APOE genotype. Thus, it is desirable to be able to infer any missing phenotype labels such that all available cell-level data in the study of a given phenotype of interest could be used. Moreover, beyond simply imputing missing phenotype information, it is often of interest to understand which groups of cells and/or molecular features may be most predictive of a given phenotype of interest. As such, there is a pressing need for computational methods that can connect cell-level measurements with donor-level labels. However, accomplishing this task is not straightforward. While a rich literature exists on learning meaningful low-dimensional representations of cells [7, 8, 1, 2] and for inferring corresponding cell-level labels (e.g. cell type) [11], the donor level prediction task introduces substantial additional complexity. For example, different numbers of cells may be recovered from each donor, and thus our prediction model must be able to handle arbitrary numbers of samples as input. Moreover, ideally our model would not a priori require any additional prior knowledge beyond our cell-level measurements, such as the importance of different cell types for a given prediction task. To resolve these issues, here we propose milVI (multiple instance learning variational inference), a deep generative modeling framework that explicitly accounts for donor-level phenotypes and enables inference of missing phenotype labels post-training. In order to handle varying numbers of cells per donor when inferring phenotype labels, milVI leverages recent advances in multiple instance learning. We validated milVI by applying to impute held-out Braak staging information from an Alzheimers disease study cohort from Mathys et al. [9], and we found that our method achieved lower error on this task compared to naive imputation methods.

bioinformatics↗

bioRxiv2024

APOE3-R136S mutation confers resilience against tau pathology via cGAS-STING-IFN inhibition

The Christchurch mutation (R136S) on the APOE3 (E3S/S) gene is associated with attenuation of tau load and cognitive decline despite the presence of a causal PSEN1 mutation and high levels of amyloid beta pathology in the carrier1. However, the specific molecular mechanisms enabling the E3S/S mutation to mitigate tau-induced neurodegeneration remain unclear. Here, we replaced mouse ApoE with wild-type human E3 or E3S/S on a tauopathy background. The R136S mutation markedly decreased tau load and protected against tau-induced synaptic loss, myelin loss, and reduction in theta and gamma powers. Additionally, the R136S mutation reduced interferon response to tau pathology in both mouse and human microglia, suppressing cGAS-STING activation. Treating tauopathy mice carrying wild-type E3 with a cGAS inhibitor protected against tau-induced synaptic loss and induced similar transcriptomic alterations to those induced by the R136S mutation across brain cell types. Thus, suppression of microglial cGAS-STING-IFN pathway plays a central role in mediating the protective effects of R136S against tauopathy. One-sentence summaryThe R136S mutation on APOE3 enhances resistance to tau-related pathology and toxicity by downregulating the cGAS-STING-IFN signaling pathway.

neuroscience↗

bioRxiv2024

A deep profile of gene expression across 18 human cancers

Clinically and biologically valuable information may reside untapped in large cancer gene expression data sets. Deep unsupervised learning has the potential to extract this information with unprecedented efficacy but has thus far been hampered by a lack of biological interpretability and robustness. Here, we present DeepProfile, a comprehensive framework that addresses current challenges in applying unsupervised deep learning to gene expression profiles. We use DeepProfile to learn low-dimensional latent spaces for 18 human cancers from 50,211 transcriptomes. DeepProfile outperforms existing dimensionality reduction methods with respect to biological interpretability. Using DeepProfile interpretability methods, we show that genes that are universally important in defining the latent spaces across all cancer types control immune cell activation, while cancer type-specific genes and pathways define molecular disease subtypes. By linking DeepProfile latent variables to secondary tumor characteristics, we discover that tumor mutation burden is closely associated with the expression of cell cycle-related genes. DNA mismatch repair and MHC class II antigen presentation pathway expression, on the other hand, are consistently associated with patient survival. We validate these results through Kaplan-Meier analyses and nominate tumor-associated macrophages as an important source of survival-correlated MHC class II transcripts. Our results illustrate the power of unsupervised deep learning for discovery of cancer biology from existing gene expression data.

bioinformatics↗

bioRxiv2023

Isolating structured salient variations in single-cell transcriptomic data with StrastiveVI

Single-cell RNA sequencing (scRNA-seq) has provided deeper insights into biological processes by highlighting differences at the cellular level. Within these single-cell omics measurements, researchers are often interested in identifying variations associated with a specific covariate. For instance, in aging research, it becomes vital to differentiate variations related to aging. To address this, we introduce StrastiveVI (Structured Contrastive Variational Inference; https://github.com/suinleelab/StrastiveVI), which effectively separates the variations of interest from other dominant biological signals in scRNA-seq datasets. When deployed on aging and Alzheimers disease (AD) datasets, StrastiveVI efficiently isolates aging and AD-associated patterns, distinguishing them from dominant variations linked to sex, tissue, and cell type that are unrelated to aging or AD. In doing so, it underscores both well-known genes and potential novel genes related to aging or AD.

bioinformatics↗

bioRxiv2022

PROPOSE: Predictive and robust probe selection for spatial transcriptomics

A prominent trend in single-cell transcriptomics is providing spatial context alongside a characterization of each cells molecular state. This typically requires targeting an a priori selection of genes, often covering less than 1% of the genome, and a key question is how to optimally determine the small gene panel. Reference data from these methods covering the whole genome is unavailable, and using single-cell RNA sequencing (scRNA-seq) datasets as a surrogate can result in suboptimal gene panels due to the fundamentally different data distributions across technologies. We address these challenges by introducing a flexible deep learning framework, PERSIST, to identify informative gene targets for spatial transcriptomics studies by leveraging existing scRNA-seq data. Using datasets spanning different brain regions, species, and scRNA-seq technologies, we show that PERSIST reliably identifies gene panels that provide more accurate prediction of the genome-wide expression profile, thereby capturing more information with fewer genes. Furthermore, PERSIST can be adapted to meet specific biological goals, such as classifying cell types or discerning neuronal electrical properties. Finally, via a simulation study based on a recent in situ hybridization-based dataset, we demonstrate that PERSISTs binarization of gene expression levels enables models trained on scRNA-seq data to generalize with input data obtained using spatial transcriptomics, despite the complex domain shift between these technologies.

cell biology↗

bioRxiv2022

Principled Feature Attribution for Unsupervised Gene Expression Analysis

As interest in unsupervised deep learning models for the analysis of gene expression data has grown, an increasing number of methods have been developed to make these deep learning models more interpretable. These methods can be separated into two groups: (1) post hoc analyses of black box models through feature attribution methods and (2) approaches to build inherently interpretable models through biologically-constrained architectures. In this work, we argue that these approaches are not mutually exclusive, but can in fact be usefully combined. We propose a novel unsupervised pathway attribution method, which better identifies major sources of transcriptomic variation than prior methods when combined with biologically-constrained neural network models. We demonstrate how principled feature attributions aid in the analysis of a variety of single cell datasets. Finally, we apply our approach to a large dataset of post-mortem brain samples from patients with Alzheimers disease, and show that it identifies Mitochondrial Respiratory Complex I as an important factor in this disease.

bioinformatics↗

bioRxiv2021

Isolating salient variations of interest in single-cell transcriptomic data with contrastiveVI

Single-cell datasets are routinely collected to investigate changes in cellular state between control cells and corresponding cells in a treatment condition, such as exposure to a drug or infection by a pathogen. To better understand heterogeneity in treatment response, it is desirable to disentangle latent structures and variations uniquely enriched in treated cells from those shared with controls. However, standard computational models of single-cell data are not designed to explicitly separate these variations. Here, we introduce Contrastive Variational Inference (contrastiveVI; https://github.com/suinleelab/contrastiveVI), a framework for analyzing treatment-control scRNA-seq datasets that explicitly disentangles the data into shared and treatment-specific latent variables. Using four treatment-control scRNA-seq dataset pairs, we apply contrastiveVI to perform a broad set of standard analysis tasks, including visualization, clustering, and differential expression testing. In each case, we find that our method consistently achieves results that agree with known biological ground truths, while previously proposed methods often fail to do so. We conclude by generalizing our framework to multimodal measurements and applying it to analyze a single-cell dataset with joint transcriptome and surface protein measurements.

bioinformatics↗

bioRxiv2021

Uncovering expression signatures of synergistic drug response using an ensemble of explainable AI models

Complex machine learning models are poised to revolutionize the treatment of diseases like acute myeloid leukemia (AML) by helping physicians choose optimal combinations of anti-cancer drugs based on molecular features. While accurate predictions are important, it is equally important to be able to learn about the underlying molecular basis of anti-cancer drug synergy. Explainable AI (XAI) offers a promising new route for data-driven cancer pharmacology, combining highly accurate models with interpretable insights into model decisions. Due to the highly correlated, high-dimensional nature of cancer transcriptomic data, however, we find that existing XAI approaches are suboptimal when applied naively to large transcriptomic datasets. We show how a novel approach based on model ensembling helps to increase the quality of explanations. We then use our method to demonstrate that a hematopoietic differentiation signature underlies synergy for a variety of anti-AML drug combinations.

cancer biology↗

bioRxiv2021

Revisiting the STRmixlikelihood ratio probability interval coverage considering multiple factors

In previously reported work a method for applying a lower bound to the variation induced by the Monte Carlo effect was trialled. This is implemented in the widely used probabilistic genotyping system, STRmix. The approach did not give the desired 99% coverage. However, the method for assigning the lower bound to the MCMC variability is only one of a number of layers of conservativism applied in a typical application. We tested all but one of these sources of variability collectively and term the result the near global coverage. The near global coverage for all tested samples was greater than 99.5% for inclusionary average LRs of known donors. This suggests that when included in the probability interval method the other layers of conservativism are more than adequate to compensate for the intermittent underperformance of the MCMC variability component. Running for extended MCMC accepts was also shown to result in improved precision.

molecular biology↗

bioRxiv2021

Transferable representations of single-cell transcriptomic data

Advances in single-cell RNA-seq (scRNA-seq) technologies are enabling the construction of large-scale, human-annotated reference cell atlases, creating unprecedented opportunities to accelerate future research. However, effectively leveraging information from these atlases, such as clustering labels or cell type annotations, remains challenging due to substantial technical noise and sparsity in scRNA-seq measurements. To address this problem, we present HD-AE, a deep autoencoder designed to extract integrated low-dimensional representations of scRNA-seq measurements across datasets from different labs and experimental conditions (https://github.com/suinleelab/HD-AE). Unlike previous approaches, HD-AEs representations successfully transfer to new query datasets without needing to retrain the model. Researchers without substantial computational resources or machine learning expertise can thus leverage the robust representations learned by pretrained HD-AE models to compare embeddings of their own data with previously generated sets of reference embeddings.

bioinformatics↗

bioRxiv2020

Unified AI framework to uncover deep interrelationships between gene expression and Alzheimer's disease neuropathologies

Deep neural networks offer a promising approach for capturing complex, non-linear relationships among variables. Because they require immense sample sizes, their potential has yet to be fully tapped for understanding complex relationships between gene expression and human phenotypes. Encouragingly, a growing number of diseases are being studied through consortium efforts. Here we introduce a new analysis framework, namely MD-AD (Multi-task Deep learning for Alzheimers Disease neuropathology), which leverages an unexpected synergy between deep neural networks and multi-cohort settings. In these settings, true joint analysis can be stymied using conventional statistical methods, which (1) require "harmonized" phenotypes (i.e., measured in a highly consistent manner) and (2) tend to capture cohort-level variations, obscuring the subtler true disease signals. Instead, MD-AD incorporates multiple related phenotypes sparsely measured across cohorts, and learns complex, non-linear interactions between genes and phenotypes not discovered using conventional expression data analysis methods (e.g., component analysis and module detection), enabling the model to capture subtler signals than cohort-level variations. Applied to the largest available collection of brain samples (N=1,758), we demonstrate that MD-AD learns a truly generalizable relationship between gene expression program and AD-related neuropathology. The learned program generalizes in several important ways, including recapitulation of the disease progress in animal models and across tissue types, and we show that such generalizability is not achieved by previous statistical paradigms. Its ability to identify genes with high non-linear relevance to neuropathology enabled us to identify a sex-specific relationship between neuropathology and immune response across microglia, providing a nuanced context for association between inflammatory genes and AD.

systems biology↗

bioRxiv2020

Adversarial Deconfounding Autoencoder for Learning Robust Gene Expression Embeddings

MotivationIncreasing number of gene expression profiles has enabled the use of complex models, such as deep unsupervised neural networks, to extract a latent space from these profiles. However, expression profiles, especially when collected in large numbers, inherently contain variations introduced by technical artifacts (e.g., batch effects) and uninteresting biological variables (e.g., age) in addition to the true signals of interest. These sources of variations, called confounders, produce embeddings that fail to transfer to different domains, i.e., an embedding learned from one dataset with a specific confounder distribution does not generalize to different distributions. To remedy this problem, we attempt to disentangle confounders from true signals to generate biologically informative embeddings. ResultsIn this paper, we introduce the AD-AE (Adversarial Deconfounding AutoEncoder) approach to deconfounding gene expression latent spaces. The AD-AE model consists of two neural networks: (i) an autoencoder to generate an embedding that can reconstruct original measurements, and (ii) an adversary trained to predict the confounder from that embedding. We jointly train the networks to generate embeddings that can encode as much information as possible without encoding any confounding signal. By applying AD-AE to two distinct gene expression datasets, we show that our model can (1) generate embeddings that do not encode confounder information, (2) conserve the biological signals present in the original space, and (3) generalize successfully across different confounder domains. We demonstrate that AD-AE outperforms standard autoencoder and other deconfounding approaches. AvailabilityOur code and data are available at https://gitlab.cs.washington.edu/abdincer/ad-ae. Contactabdincer@cs.washington.edu; suinlee@cs.washington.edu

bioinformatics↗

bioRxiv2020

Automated detection of glaucoma with interpretable machine learning using clinical data and multi-modal retinal images

Glaucoma, the leading cause of irreversible blindness worldwide, is a disease that damages the optic nerve. Current machine learning (ML) approaches for glaucoma detection rely on features such as retinal thickness maps; however, the high rate of segmentation errors when creating these maps increase the likelihood of faulty diagnoses. This paper proposes a new, comprehensive, and more accurate ML-based approach for population-level glaucoma screening. Our contributions include: (1) a multi-modal model built upon a large data set that includes demographic, systemic and ocular data as well as raw image data taken from color fundus photos (CFPs) and macular Optical Coherence Tomography (OCT) scans, (2) model interpretation to identify and explain data features that lead to accurate model performance, and (3) model validation via comparison of model output with clinician interpretation of CFPs. We also validated the model on a cohort that was not diagnosed with glaucoma at the time of imaging but eventually received a glaucoma diagnosis. Results show that our model is highly accurate (AUC 0.97) and interpretable. It validated biological features known to be related to the disease, such as age, intraocular pressure and optic disc morphology. Our model also points to previously unknown or disputed features, such as pulmonary capacity and retinal outer layers.

bioengineering↗

bioRxiv2018

Explainable machine learning prediction of synergistic drug combinations for precision cancer medicine

Although combination therapy has been a mainstay of cancer treatment for decades, it remains challenging, both to identify novel effective combinations of drugs and to determine the optimal combination for a particular patients tumor. While there have been several recent efforts to test drug combinations in vitro, examining the immense space of possible combinations is far from being feasible. Thus, it is crucial to develop datadriven techniques to computationally identify the optimal drug combination for a patient. We introduce TreeCombo, an extreme gradient boosted tree-based approach to predict synergy of novel drug combinations, using chemical and physical properties of drugs and gene expression levels of cell lines as features. We find that TreeCombo significantly outperforms three other state-of-theart approaches, including the recently developed DeepSynergy, which uses the same set of features to predict synergy using deep neural networks. Moreover, we found that the predictions from our approach were interpretable, with genes having well-established links to cancer serving as important features for prediction of drug synergy.

cancer biology↗