A comprehensive review of variant calling tools for RNA-seq: Challenges and advances

RNA sequencing (RNA-seq) has become a powerful technology for capturing transcriptomic variation, enabling the detection of expressed genetic variants that influence gene regulation, disease mechanisms, and therapeutic response. Unlike DNA sequencing, variant calling from RNA-seq presents unique challenges, including alternative splicing, allele-specific expression, RNA editing, and variable transcript abundance. To address these complexities, diverse computational strategies have been developed, ranging from classical statistical models to advanced and deep learning approaches. Early methods such as SAMtools, VarScan2, and bcftools relied on pileup-based binomial and Bayesian likelihood models, offering computational efficiency but limited sensitivity to RNA-specific artifacts. More sophisticated tools, such as GATK HaplotypeCaller and Octopus, incorporate pair-Hidden Markov Models and de Bruijn graph assemblies, enabling haplotype-aware detection with improved accuracy across spliced regions. Machine learning-based frameworks, including RNA-SNPhunter (random forest) and RVboost (gradient boosting), leverage alignment and sequence features to enhance variant classification, while deep learning methods such as DeepVariant employ convolutional neural networks to learn complex error signatures directly from sequencing data. Complementary advances in splice-aware alignment, variant filtering, and functional annotation further strengthen accuracy and biological interpretation. Benchmarking studies highlight that the best-performing pipelines often integrate statistical rigor with AI-driven adaptability, balancing precision and recall across variant types and expression levels. Collectively, these methodological innovations are reshaping RNA-seq variant calling, advancing its role from a specialized challenge to a core component of multi-omic integration and precision medicine.