Exploring artificial intelligence for variant calling: A comprehensive review of tools and approaches

With the advent of Next-generation sequencing (NGS) methods, significant development is being observed in the identification of genetic variants. The precision and exactness of variant calls, nevertheless, continue to trouble researchers. Miscalculations can arise from multiple sources, such as sequencing artefacts, misalignments, and genomic heterogeneity. Conventional variant callers typically rely on statistical models that calculate the likelihood or probability of each variant call. However, they are not adept at handling messy data and mining variants from problematic genomic regions. To address such challenges, artificial intelligence (AI) is being explored. It is presented as the future of variant calling to enhance stability, precision and accuracy and eliminate errors.
The present review attempts to give a broad overview of AI-derived and AI-assisted variant calling approaches in trend. We have mentioned the drawbacks of traditional variant callers and described how artificial intelligence is being used. Following this, we have covered in length major AI-based tools such as DeepVariant, DeepTrio, Clair/Clair3, NeuSomatic, PEPPER–Margin–DeepVariant, Medaka, Hello and DNAScope. 
We have provided an inclusive relative summary of the tools in use, highlighting their suitability, strengths, and drawbacks to aid in choosing the right tool. Additionally, we have suggested best practices to ensure the best output from AI-enabled pipelines. To give a complete, balanced outline, we detailed major constraints with AI-backed approaches, such as model transparency, overfitting and underfitting, computational requirements, and problems in structural variant detection. To sum up, we have offered a glimpse of emerging research areas such as transformer-based models, multi-modal data linkage, and GNN-based models.