Biology subjects

Choi, Y.-L.

Publications and source records attributed to Choi, Y.-L..

5 recordsLinked to original sources

bioRxiv2026

Agentic AI enables cell segmentation-free cell-type inference and cell-type-guided cell segmentation in spatial transcriptomics

Cell-type identification in spatial transcriptomics often relies on cell segmentation, which can introduce inaccuracies in cell boundaries and transcript assignment. Segmentation-free or cell-free methods have been developed to completely avoid cell segmentation, but inaccurate cell typing hinder accurate downstream analysis. Here, we introduce SSAF-T, a segmentation-free agentic framework that identifies cell types directly from spatial gene expression without predefined reference datasets. For the reference-free cell typing, SSAF-T incorporates CeLLMan, which employs expert-level agents for each organ to annotate cell types using a literature-grounded knowledge graph. We benchmark SSAF-T on both array-based and imaging-based spatial transcriptomics datasets in terms of accuracy, reproducibility, and scalability. We further show that SSAF-T provides an accurate cell segmentation based on the segmentation-free mapping, which ensure compatibility to existing downstream analysis ecosystem, and eventually enable quantitative analysis at single-cell level.

bioinformatics↗

bioRxiv2025

SHEST: Single-cell-level artificial intelligence from haematoxylin and eosin morphology for cell type prediction and spatial transcriptomics reconstruction

A comprehensive understanding of cancer progression requires integrating tissue morphol-ogy with spatial molecular profiles. We present SHEST, a multi-task profiling framework that leverages haematoxylin and eosin morphology to predict cellular composition and re-construct spatial gene expression at single-cell resolution. SHEST employs a quadruple-tile input capturing nuclear and contextual information, combined with a neighbourhood-informed clustering algorithm to filter ambiguous cellular signals. It comprises a shared morphological encoder with two task-specific heads: a classifier for cell type prediction and a reconstruc-tor for gene expression. Multi-task optimisation uses cross-entropy and zero-inflated negative binomial losses, specifically addressing the sparsity of spatial transcriptomic data. Evalua-tion on human lung adenocarcinoma datasets demonstrated high accuracy for the principal reciprocal constituents of the tumour-immune axis (F1: 0.97 for tumour cells and 0.91 for lym-phocytes). External validation confirmed its generalisability, revealing alveolar cells and their early neoplastic transitions. Reconstructed gene expression reproduced spatially resolved, cell-type-specific marker patterns--EPCAM in tumour cells, LTBP2 in fibroblasts, and CD3E in lymphocytes--recovering biologically coherent transcriptional architecture. SHEST also pre-served distance-dependent spatial relationships and gene-level autocorrelation, reflecting the multicellular niche structure of the tumour microenvironment. By unifying cell type iden-tification, gene expression reconstruction, and spatial mapping within a single interpretable framework, SHEST provides a synergistic and cost-efficient bridge between histopathology and spatial transcriptomics. This approach facilitates comprehensive tissue characterisation and forms a foundation for precision oncology through spatially informed, cell-level insights into tumour-immune ecosystems.

bioinformatics↗

bioRxiv2025

CRISPR-Induced NUT Suppression Promotes Differentiation and Enhances Trop2-Targeted Immunocytokine Response in NUT Carcinoma

NUT carcinoma (NC) is an aggressive malignancy driven by NUTM1 gene rearrangements with limited therapeutic options. Here, we show that direct suppression of NUTM1 using CRISPR/Cas9 induces squamous-like differentiation and upregulates TROP2 expression in NC cells. Building on this finding, we developed a TROP2-interferon beta (IFN-{beta}) mutein immunocytokine that selectively targets TROP2-expressing tumors. Combined NUTM1 suppression and TROP2-targeted immunotherapy synergistically enhanced cytotoxic and immune-mediated responses in vitro. Transcriptomic and spatial analyses of NC patient tumors revealed that differentiation status correlates with TROP2 expression, upregulated immune pathways, and favorable clinical outcomes. Our results suggest that overcoming differentiation blockade not only alters tumor phenotype but also creates a more immune-permissive microenvironment. These findings highlight the therapeutic potential of sequential tumor reprogramming followed by targeted immunotherapy in treating NC and propose a broader strategy for overcoming differentiation blockade in fusion-driven cancers.

cancer biology↗

bioRxiv2025

GraphTME: a framework for predicting response to immune checkpoint inhibitors by interpreting cell-cell interactions in the tumour microenvironment using spatial transcriptomics of tumour tissue

Immune checkpoint inhibitors (ICIs) have been used to treat cancer by reactivating T-cell responses against tumours, yet clinical efficacy remains limited due to low response rates and the lack of robust predictive biomarkers. The tumour microenvironment critically shapes ICI responsiveness by modulating antitumour immunity, necessitating a deeper understanding of spatially organised cell-cell interactions. Imaging-based spatial transcriptomics (ST) enables such analysis at single-cell resolution. We present GraphTME, a spatially informed and biologically interpretable framework that models pathway-specific directional cell-cell interactions to predict anti-PD-1 response. Ligand-receptor interactions are organised by signalling pathways and represented as a multi-relational directed graph, with edge weights inversely scaled by spatial distance. Using CD8+ T cells from single-cell RNA-seq data of ICI-treated NSCLC patients, we trained a relational graph convolutional network to infer immune responsiveness. GraphTME achieved an F1 score exceeding 0.83 in predicting ICI response and was validated using MERFISH data from an NSCLC patient with a partial response. Predicted responder CD8+ T cells exhibited higher abundance and stronger directional signalling to tumour cells. They also expressed genes associated with antitumour activity, while non-responders showed expression patterns linked to poor prognosis. GraphTME is the first framework to quantitatively model single-cell-level interactions within the spatial architecture of tumour tissues and predict ICI responses from these interactions. It offers a spatially resolved, biologically grounded biomarker for immunotherapy and a tool for dissecting immune dynamics in situ.

bioinformatics↗

bioRxiv2019

JuLI: accurate detection of DNA fusions in clinical sequencing for precision oncology

Accurate detection of genomic fusions by high-throughput sequencing in clinical samples with inadequate tumor purity and formalin-fixed paraffin embedded (FFPE) tissue is an essential task in precise oncology. We developed the fusion detection algorithm Junction Location Identifier (JuLI) for optimization of high-depth clinical sequencing. We implemented novel filtering steps to minimize false positives and a joint calling function to increase sensitivity in clinical setting. We comprehensively validated the algorithm using high-depth sequencing data from cancer cell lines and clinical samples and whole genome sequencing data from NA12878. We showed that JuLI outperformed state-of-the-art fusion callers in cases with high-depth clinical sequencing and rescued a driver fusion from false negative in plasma cell-free DNA. JuLI is freely available via GitHub (https://github.com/sgilab/JuLI).

bioinformatics↗