Deep Symbolic Learning Architecture for Variant Calling in NGS

Abstract

[EN]The Variant Detection process (Variant Calling) is fundamental in bioinformatics, demanding maximum precision and reliability. This study examines an innovative integration strategy between a traditional pipeline developed in-house and an advanced Intelligent System (IS). Although the original pipeline already had tools based on traditional algorithms, it had limitations, particularly in the detection of rare or unknown variants. Therefore, SI was introduced with the aim of providing an additional layer of analysis, capitalizing on deep and symbolic learning techniques to improve and enhance previous detections. The main technical challenge lay in interoperability. To overcome this, NextFlow, a scripting language designed to manage complex bioinformatics workflows, was employed. Through NextFlow, communication and efficient data transfer between the original pipeline and the SI were facilitated, thus guaranteeing compatibility and reproducibility. After the Variant Calling process of the original system, the results were transmitted to the SI, where a meticulous sequence of analysis was implemented, from preprocessing to data fusion. As a result, an optimized set of variants was generated that was integrated with previous results. Variants corroborated by both tools were considered to be of high reliability, while discrepancies indicated areas for detailed investigations. The product of this integration advanced to subsequent stages of the pipeline, usually annotation or interpretation, contextualizing the variants from biological and clinical perspectives. This adaptation not only maintained the original functionalities of the pipeline, but was also enhanced with the SI, establishing a new standard in the Variant Calling process. This research offers a robust and efficient model for the detection and analysis of genomic variants, highlighting the promise and applicability of blended learning in bioinformatics