Deep learning innovations for accurate genomic analysis

Deep learning methods have driven significant advances in genomics, particularly in enhancing genomic analysis. In this talk, we will discuss a suite of technologies developed by our team to improve the quality and accuracy of insights derived from sequencing data. First, we will learn about DeepVariant, a deep learning-based germline variant caller that utilizes convolutional neural networks to analyze sequencing data and accurately identify genetic variations. We'll highlight its real-world application in rapid diagnostics for NICU patients and discuss its successful adaptation and performance across newer sequencing technologies, including those from Roche, Element, and Ultima. Next, we will cover DeepSomatic, an extension of DeepVariant specifically designed for the detection of challenging somatic mutations. This area of variant calling is particularly difficult because of limited ground truths. We will then discuss DeepConsensus, a transformer-based consensus caller, demonstrating how it significantly improves the base-level accuracy of PacBio HiFi reads. Complementing these callers, DeepPolisher leverages deep learning to enhance the quality of genome assemblies and reference sequences. Finally, we will explore the opportunities presented by newer pangenome references. We'll discuss how tools like DeepConsensus and DeepPolisher contribute to building more accurate and complete pangenome references. Furthermore, we will examine how these improved pangenomes, in turn, enhance DeepVariant's variant calling accuracy by incorporating valuable population-level information during the calling process.