基于可解释深度学习神经模型PASNet的系统性红斑狼疮疾病特征基因挖掘与分析

Mining and Analysis of Systemic Lupus Erythematosus Disease Characteristic Genes Based on Interpretable Deep Learning Neural Model PASNet Combined with Bioinformatics

摘要: 目的：通过可解释深度学习神经模型(pathway-associated sparse deep neural network, PASNet)结合生物信息学方法寻找系统性红斑狼疮疾病的特征基因。方法：从NCBI的GEO数据库收集SLE外周血转录组数据，使用R语言筛选差异基因用于模型训练。基于PASNet模型构建用于SLE疾病研究的信号通路相关深度学习神经网络模型，获得SLE疾病相关的关键基因和关键通路。LASSO回归和随机森林方法进一步筛选疾病特征基因。验证SLE疾病特征基因在外部数据集上的表达情况并验证特征基因的临床诊断效果。结果：构建的神经网络模型在深度学习训练后，测试集上验证的预测效果较好，ROC曲线下面积(area under the curve, AUC)为0.82。模型获得47个疾病相关的关键基因，进一步机器学习算法分析及GSE61635数据验证筛选出ADHFE1、BCL11B、BLVRA和BOLAL四种SLE疾病特征基因。SLE单细胞数据验证结果显示，BCL11B和BLVRA基因与免疫细胞有不同程度的相关性，用其构建的临床诊断模型能很好地区分疾病。结论：通过深度学习结合生物信息学方法筛选的BCL11B和BLVRA基因可作为SLE的潜在生物标志物进一步研究。

Abstract: Objective: To identify characteristic genes of SLE using an interpretable deep learning neural model (PASNet) combined with bioinformatics methods. Methods: SLE peripheral blood transcriptome data were collected from the GEO database of NCBI, and differentially expressed genes were screened using R language for model training. A pathway-associated deep learning neural network model based on PASNet was constructed for SLE disease research to identify key genes and pathways related to SLE. LASSO regression and random forest methods were further employed to screen disease characteristic genes. The expression of these characteristic genes in external datasets was validated, and their clinical diagnostic efficacy was evaluated. Results: The constructed neural network model demonstrated good predictive performance on the test set after deep learning training, with an area under the ROC curve (AUC) of 0.82. The model identified 47 disease-related key genes, and machine learning algorithms analyze and GSE61635 data validation further screened out four SLE characteristic genes: ADHFE1, BCL11B, BLVRA, and BOLAL. SLE single-cell data validation results showed that BCL11B and BLVRA genes were correlated with immune cells to varying degrees, and the clinical diagnostic model constructed using these genes effectively distinguished the disease. Conclusion: The BCL11B and BLVRA genes screened through deep learning combined with bioinformatics methods can serve as potential biomarkers for SLE and warrant further research.