Enhancing genome analysis pipeline with AI and ML

Abstract

As DNA sequencing technologies advance, the need for precise yet cost-effective genome analysis pipelines becomes increasingly vital. This dissertation unveils novel methodologies leveraging artificial intelligence (AI) and machine learning (ML) to enhance the two most critical steps in the genome analysis pipeline: read alignment and variant calling. Initially, we present BWA-MEME, an ML-augmented read alignment software. By employing learned indices, this software enhances the exact match search during the seeding phase—addressing a significant bottleneck in short-read alignment. Subsequently, we address challenges inherent to deep learning-based variant callers. These challenges encompass their reliance on vast labeled datasets and their susceptibility to diverse error profiles presented by different sequencing techniques. We devise a semi-supervised training approach that not only utilizes unlabeled data to learn error profiles but also incorporates a domain adaptation technique to minimize discrepancies arising from diverse error profiles. Together, these methods carve out novel pathways in read alignment and variant calling, underscoring the transformative potential of AI and ML within the genome analysis pipeline.