Abstract
RNA sequencing (RNA-seq) is a pivotal tool for transcriptomic analysis, providing comprehensive exploration of gene expression across diverse biological contexts. However, RNA-seq data are susceptible to various biases that can significantly compromise the accuracy and reliability of transcript quantification. This study investigates the influence of high-dimensional RNA structures on local sequencing efficiency using an innovative unsupervised variational autoencoder-Gaussian mixture model (VAE-GMM). The VAE-GMM effectively captures intricate high-dimensional k-mer structural similarities by learning compact latent representations, which reduces dimensionality while meticulously preserving essential structural features crucial for bias identification. This sophisticated modeling allows precise tracking of local RNA-read conversion dynamics and the identification of complex, often overlooked bias sources. We rigorously validate the VAE-GMM model's performance and robustness against conventional machine learning techniques, including Gaussian mixture models (GMM-only), principal component analysis-based GMMs, k-means clustering, and hierarchical clustering. These validations, using an extensive and diverse array of data sets, including synthetic RNA constructs, various human cell lines, and authentic tissue samples, consistently demonstrate the model's superior versatility and accuracy across different biological systems. Furthermore, in silico simulations of the sequencing process closely align with actual sequencing data, strongly reinforcing the critical role of high-dimensional RNA structures in determining sequencing efficiency and their impact on data quality. Our findings offer valuable insights into the underlying mechanisms of RNA structure-mediated sequencing bias. This deeper understanding enables more accurate and reliable RNA-seq analyses and is expected to improve the interpretation of transcriptomic data in future genomic studies.