Scholarlink

Advances in single-cell RNA sequencing (scRNA-seq) technologies have revolutionized cancer research by enabling detailed characterization of cellular heterogeneity [6]. However, predicting drug sensitivity at single-cell resolution remains challenging due to the scarcity of annotated data, high dimensionality, and technical noise. In this work, we reproduce and enhance the scDEAL framework, a deep transfer learning model that integrates bulk and single-cell transcriptomic data for drug response prediction. The enhancements focus on optimizing Denoising Autoencoders (DAE) and incorporating progressive training strategies. Specifically, we introduce Gene Prioritization Regularization (GPR) to emphasize biologically relevant genes, implement Curriculum Learning to gradually increase task complexity during training, and apply direct filtering of highly variable genes to reduce dimensionality. Experiments conducted on publicly available datasets from GDSC, CCLE, and GEO demonstrate that filtering the top 20% most variable genes leads to significant improvements in predictive performance, achieving an F1-score of 0.9641 and an AUC of 0.9549, while reducing computational costs by 80%. These results highlight the importance of feature selection and progressive training strategies for enhancing drug response prediction from scRNA-seq data.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimization of Denoising Autoencoders with Progressive Learning Strategies for ScRNA-Seq Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Francisco Javier Franco-Ruiz,
  • Sara Fernandez-Malvido,
  • Belén Vega-Marquez,
  • Isabel A. Nepomuceno-Chamorro

摘要

Advances in single-cell RNA sequencing (scRNA-seq) technologies have revolutionized cancer research by enabling detailed characterization of cellular heterogeneity [6]. However, predicting drug sensitivity at single-cell resolution remains challenging due to the scarcity of annotated data, high dimensionality, and technical noise. In this work, we reproduce and enhance the scDEAL framework, a deep transfer learning model that integrates bulk and single-cell transcriptomic data for drug response prediction. The enhancements focus on optimizing Denoising Autoencoders (DAE) and incorporating progressive training strategies. Specifically, we introduce Gene Prioritization Regularization (GPR) to emphasize biologically relevant genes, implement Curriculum Learning to gradually increase task complexity during training, and apply direct filtering of highly variable genes to reduce dimensionality. Experiments conducted on publicly available datasets from GDSC, CCLE, and GEO demonstrate that filtering the top 20% most variable genes leads to significant improvements in predictive performance, achieving an F1-score of 0.9641 and an AUC of 0.9549, while reducing computational costs by 80%. These results highlight the importance of feature selection and progressive training strategies for enhancing drug response prediction from scRNA-seq data.