Investigating Alternative Splicing Events in Amyotrophic lateral Sclerosis with RNA-Sequence Analysis and Deep Learning.

Abstract

Amyotrophic lateral sclerosis (ALS) is a progressive chronic neurodegenerative disorder caused by the degeneration of motor neurons and associated with dysregulated RNA metabolism, specifically alternative splicing. Investigation into AS abnormalities may help illuminate the molecular mechanisms of the disease and identify new transcriptomic biomarkers. RNA-Seq data from ALS and normal individuals were compared using a computational workflow that included HISAT2 for alignment, FeatureCounts for read counting, and DESeq2 for differential gene expression analysis, and StringTie with BEDTools for transcript assembly and analysis. Extracted AS associated transcript sequences were classified using the deep learning models CNN, RNN, unidirectional LSTM, and BiLSTM. Differential expression analysis revealed 349 statistically significant dysregulated genes. The performance of the models was assessed and found to be strongest for BiLSTM which achieved an accuracy and F1-score of 81% in differentiating ALS samples from normal individuals on the basis of transcript sequences, outperforming the CNN, RNN, and unidirectional LSTM architectures. This suggests the BiLSTM model can effectively account for sequence dependencies and distinguish between features that identify AS associated transcript sequences. A deep learning approach using a BiLSTM framework offers a scalable method to identify alternatively spliced transcripts and potential transcriptomic biomarkers in ALS. The model requires external validation on large independent datasets and combination with multi-omics data to further improve prediction accuracy and biological insight.