Omar Abdelwahab is a dedicated researcher, bioinformatics consultant, and aspiring entrepreneur in genomics, with a strong interdisciplinary background bridging both wet lab and computational biology. His academic journey began with a bachelor's degree in biomedical sciences, with a concentration in Genomics and Computational Biology from the University of Science and Technology at Zewail City in Giza, Egypt. This program provided him with extensive knowledge and hands-on experience in both experimental and computational methodologies, equipping him with a unique skill set that enables him to tackle complex biological questions from multiple angles. Omar further honed his expertise by pursuing a master's degree and subsequently a PhD in Computational Genomics at Laval University, where his research focuses on building new computational pipelines for the identification and analysis of genomic variations. His work aims to enhance the understanding of genetic diversity and its implications in various biological contexts. By leveraging his expertise in both laboratory and computational methods, he is developing innovative approaches to improve genomic analysis and interpretation. In addition to his wet lab skills, Omar possesses a robust background in bioinformatics and computational biology. He has mastered multiple programming languages, including Python, R, and Bash, and has extensive experience working with artificial intelligence and machine learning frameworks such as TensorFlow and PyTorch. His proficiency in these tools has enabled him to develop sophisticated computational models and pipelines for genomic data analysis. His research has been published in peer-reviewed journals, reflecting his contributions to the field. Beyond academia, Omar is a seasoned entrepreneur with significant industrial experience. He has worked as a bioinformatics consultant for multiple companies in the genomics and biotechnology sectors. His work involves providing data-driven solutions, optimizing genomic workflows, and developing innovative tools for real-world applications. Additionally, he has served as an international subject matter expert in data science and machine learning, mentoring professionals and collaborating on research-driven projects. Omar’s goal is to become a leading entrepreneur in the genomics sector by establishing his own company. Through his extensive expertise in bioinformatics and AI-driven genomic analysis, he aims to develop innovative solutions that bridge the gap between research and industry. His vision is to leverage innovative technology to revolutionize genomic data interpretation, ensuring that advancements in genomics translate into impactful real-world applications. With a solid foundation in both experimental and computational genomics, Omar continues to drive innovation in the field, contributing equally to academic research and industrial advancements. His combined expertise allows him to bridge the gap between scientific discovery and entrepreneurial impact, ensuring that genomic innovations are effectively translated into meaningful applications. A deep learning-based pipeline for detecting high-quality variants from genomic data Direction de recherche : Davoud Torkamaneh By Omar Abdelwahab Despite its importance in genomic analyses for detecting mutations, accurate identification and classification of genomic variants is still an exhaustive process. To acquire high-quality variants, there are specific procedures: 1) Use a high-quality sequencing instrument that produces high coverage and low error rates; 2) Alignment: Align the sequencing reads to the appropriate reference genome or transcriptome using a high-quality alignment algorithm. This involves correcting for sequencing errors, accounting for insertions and deletions, and minimizing false positives and false negatives; and 3) Variant calling: Use a high-quality variant calling algorithm that considers the quality scores of each variant call, the depth of coverage, and the frequency of the variant across the sample. Filter out low-quality variants based on appropriate thresholds for quality, depth of coverage, and allele frequency. In the last two decades, short reads (mainly generated by Illumina sequencing technologies) have been the reference-based data for various variant calling methods (Altshuler et al. 2012). However, with the emergence of other sequencing technologies like Pacific Biosciences high-fidelity (PacBio HiFi) and Oxford Nanopore Technologies (ONT), a new question emerged of whether they can accurately detect genomic variations. Moreover, up to date, there is no clear advantage of a specific variant calling tool over others using a different type of sequencing data (i.e., short vs. long reads). Furthermore, most of the variant calling pipelines produce a large number of candidate variants that need to be filtered as part of a very important step known as “Variant refinement”, in which the variants are filtered from false positives in the produced list of candidate variants (Kim et al. 2019). This process can mainly be performed by two approaches: heuristic filtering and/or manual review. Although heuristic filtering can be relatively fast, it fails most of the time to find the optimal solution. As for the manual review, it is an exhaustive time-consuming process that requires direct visual analysis of variants. With the advancements in computational power and artificial intelligence (AI), I hypothesize that the process of variant calling and variant calling can be completely automated. The deep learning (DL) algorithm, known as TRANSFORMERS (Vaswani et al. 2017), was initially made for natural language processing (NLP). One of its main applications is to understand languages and perform analyses on that languages including translation, detecting grammatical and spelling mistakes, and editing them. Accordingly, I hypothesize that the inner workings of TRANSFORMERS can be tweaked into treating the genomic sequences as languages from which we can detect mistakes (genomic variations). To test this hypothesis, the proposed project has three objectives: I) To test the performance of conventional and AI-based variant callers using different types of sequencing data (i.e., short vs. long reads); II) To develop and evaluate a DL model, using TRANSFORMERS, for variant calling refinement; and III) To establish and integrate the proposed DL-based model for detecting high-quality variants from genomic data. For objective I, three conventional variant calling tools (BCFTools, GATK v4, and Platypus), and two AI-based (DNAscope and DeepVariant) will be used to call variants from different sequencing technologies data types (Illumina, PacBio HiFi, and ONT). Three samples (HG003, HG006, and HG007) will be acquired from the Genome In A Bottle (GIAB) project (Zook et al. 2016), and the results will be compared to truth sets of high-quality variants made by GIAB. Three accuracy metrics will be used to report the results: precision, recall, and F1-score. This will lead to exploring the strengths and weaknesses of each tool and finding room for improvement. For objectives II and III, we propose using a deep learning technique (TRANSFORMERS) for variant calling and also to create a model to refine and classify previously called variants with different tools. Different GIAB samples (HG001, HG002, HG004, and HG005) collected from various sequencing technologies (Illumina, PacBio HiFi, and ONT) and another dataset of 41K variants from (Ainscough et al. 2018) will be used to train and evaluate the proposed model. Then, the model will be optimized to function on various datasets from different species with different genome sizes and ploidy levels. Finally, multiple datasets from different species with different ploidy will be used to test the proposed model. Overall, the final pipeline should be able to automate the whole variant calling process starting from FASTQ files to filtered high-quality VCF files. Furthermore, the model will have the ability to be integrated into other variant calling pipelines (e.g. Fast-GBS). Ideally, this model will be capable of replacing the old variant calling pipelines and manual refinement techniques by introducing a new variant calling pipeline. The proposed variant calling tool will represent a deep learning-based pipeline for detecting high-quality variants from genomic data from various sequencing technologies.