Deep learning pipeline for detecting high-quality genomic variants using transformer models

Overview

This research presents a deep learning-based pipeline for detecting high-quality genomic variants from next-generation sequencing (NGS) data. The solution addresses the uncertainty surrounding optimal variant calling pipelines by leveraging transformer-based artificial intelligence models to improve the accuracy and efficiency of genomic variant analysis. The approach conceptualizes genomic sequences as a language, with the reference genome serving as the lexicon, effectively transforming the variant calling problem into a natural language processing task. The pipeline is designed to work across multiple sequencing data types, including Illumina short reads, PacBio HiFi, and Oxford Nanopore Technologies (ONT) long reads.

The technology offers value to genomics researchers, clinical laboratories, pharmaceutical companies, and biotechnology firms seeking more reliable variant detection for applications such as disease research, personalized medicine, and population genomics studies. By improving variant calling accuracy, the pipeline can reduce false positives and false negatives, leading to more confident downstream analyses and clinical interpretations.

Technical specifications

Key features and methodology:

• Transformer-based deep learning model for variant calling refinement, achieving approximately 92% accuracy in initial testing
• Multi-platform compatibility with Illumina short reads, PacBio HiFi long reads, and Oxford Nanopore Technologies data
• Language-modeling approach that interprets genomic sequences from FASTQ files as a language, using the reference genome as the lexicon
• Comparative evaluation framework benchmarking against conventional variant callers such as GATK v4 and BCFTools, as well as AI-based tools including DeepVariant and DNAscope
• Quality classification system that categorizes variants into Pass or Failed categories
• Performance metrics including AUC, F1-score, recall, and precision for comprehensive accuracy assessment
• Training and validation using Genome in a Bottle (GIAB) Consortium reference samples, including the HG003 sample

Technology readiness level

The research has progressed through significant validation milestones. The initial phase involved a comparative study of conventional and AI-based variant calling tools using GIAB sequencing data, with findings published as a preprint. The subsequent phase successfully developed and evaluated a transformer-based deep learning model for variant refinement, demonstrating approximately 92% accuracy. The final phase, currently underway, involves developing the novel AI-based variant calling pipeline. The coding phase is in progress, with training and testing on GIAB samples planned for comparison against existing tools. The complete study is expected to take two years to finalize, and the research team is seeking partnership support to advance this work toward broader deployment.