HNSC filtered datasets from GSE139324 and GSE164690

The filtered HNSC datasets include GSE139324 and GSE164690, comprising scRNA-seq samples from healthy donors, as well as HPV-negative and HPV-positive cases.<br>This project uses deep learning models to improve clustering accuracy in large-scale single-cell atlases. By integrating <b>contrastive learning</b> and transformer based <b>variational autoencoders (VAEs)</b>, these models learn biologically meaningful low-dimensional representations, enabling precise identification of cell types, including rare and transitional populations.Key innovations include:Robust handling of batch effects and technical noiseScalable architecture for millions of cellsImproved clustering metrics (ARI, NMI) on benchmark datasetsEnhanced interpretability by linking latent features to gene programsThese models offer a powerful tool for building accurate, scalable cell atlases and advancing our understanding of cellular diversity.

经筛选的头颈部鳞状细胞癌（HNSC）数据集包含GSE139324与GSE164690，涵盖来自健康供体的单细胞RNA测序（single-cell RNA sequencing）样本，以及人乳头瘤病毒（Human Papillomavirus, HPV）阴性与阳性病例。本项目采用深度学习模型，以提升大规模单细胞图谱的聚类精度。通过融合对比学习（contrastive learning）与基于Transformer的变分自编码器（variational autoencoders, VAEs），这些模型可学习具有生物学意义的低维表征，从而精准识别细胞类型，包括稀有细胞群与过渡态细胞群。核心创新点包括：可稳健处理批次效应与技术噪声；支持百万级细胞的可扩展架构；在基准数据集上优化了聚类指标（调整兰德指数（Adjusted Rand Index, ARI）与归一化互信息（Normalized Mutual Information, NMI））；通过将潜在特征与基因程序关联，提升了模型可解释性。这些模型可为构建精准且可扩展的细胞图谱，以及深化我们对细胞多样性的认知提供强有力的工具。

搜集汇总

背景与挑战

背景概述

该数据集包含来自GSE139324和GSE164690的过滤后的HNSC数据，涵盖健康捐赠者、HPV阴性和HPV阳性病例的单细胞RNA测序样本。数据集采用深度学习模型，结合对比学习和基于变压器的变分自编码器（VAEs），以提高大规模单细胞图谱的聚类准确性，并能够识别细胞类型，包括稀有和过渡群体。