Leveraging autoencoder models and data augmentation to uncover transcriptomic diversity of gingival keratinocytes in single cell analysis

Periodontitis, a chronic inflammatory condition of the periodontium, is associated with over 60 systemic diseases. Despite advancements, precision medicine approaches have had limited success, emphasizing the need for deeper insights into cellular subpopulations and structural immunity, particularly gingival keratinocytes. This study employs autoencoder models and data augmentation techniques to explore the transcriptomic diversity of gingival keratinocytes at the single-cell level. Single-cell RNA sequencing data from GSE266897 were processed using the Scanpy library, with quality control implemented to filter cells based on predefined metrics. Clustering was performed using principal component analysis (PCA) and k-nearest neighbor (KNN) algorithms. Marker gene identification and differential expression analysis were used to characterize cell clusters. Visualization techniques, including UMAP, heatmaps, dot plots, and violin plots, provided insights into gene expression patterns. The autoencoder architecture featured an encoder reducing input size to 256 units with ReLU activation, a bottleneck layer, and a decoder restoring data dimensions. The basic Autoencoder (AE) demonstrated superior performance, achieving the lowest loss (0.725), the highest accuracy (0.695), and minimal false positives. The Test-Time Augmentation AE also performed robustly, achieving an F1 score of 0.642 and an AUC-ROC of 0.800. The Basic AE effectively modeled RNA-seq data complexity compared to Variational and Denoising Autoencoders. This study highlights advanced computational techniques to investigate gingival keratinocytes’ transcriptomic diversity, revealing distinct subpopulations and differential gene expression profiles. These findings underscore the active role of keratinocytes in periodontal health and inflammatory responses, contributing to precision medicine approaches in periodontology.