Uploaded February 2026 | Updated November 2026, 3 days ago
Speaker:
Mohammed Alser, Georgia State University
Abstract:
Genomic data formats such as FASTA, FASTQ, BAM, and VCF were designed for early sequencing technologies with low throughput and short reads. Today, genomics is entering an AI-driven decade, where large-scale machine learning models increasingly consume raw and processed data directly. This talk revisits the assumptions behind our most widely used genomic file formats and asks whether they remain fit for purpose—a question often overlooked, even when building high-performance hardware accelerators for genomics.
We examine how current formats encode sequence, quality, alignment, and metadata, highlighting structural constraints that limit downstream analysis, accuracy, compression, and random access. While these formats have enabled remarkable progress, they often prioritize human-centric pipelines and linear workflows over the needs of modern AI systems: efficient batching, rich contextual metadata, and scalable access across distributed computing.
We also share ongoing efforts to enable intelligent genome analysis through novel algorithms, hardware accelerators, and emerging technologies such as processing-in-memory. Current formats unfortunately impede these parallelism efforts, as I/O and memory bandwidth remain significant bottlenecks. Finally, we present projects exploring new genomic data formats and outline future challenges, benefits, and research directions. These efforts aim to lay a foundation for AI-ready, scalable, and intelligent genome analysis.
Speaker:
Mohammed Alser, Georgia State University
Abstract:
Genomic data formats such as FASTA, FASTQ, BAM, and VCF were designed for early sequencing technologies with low throughput and short reads. Today, genomics is entering an AI-driven decade, where large-scale machine learning models increasingly consume raw and processed data directly. This talk revisits the assumptions behind our most widely used genomic file formats and asks whether they remain fit for purpose—a question often overlooked, even when building high-performance hardware accelerators for genomics.
We examine how current formats encode sequence, quality, alignment, and metadata, highlighting structural constraints that limit downstream analysis, accuracy, compression, and random access. While these formats have enabled remarkable progress, they often prioritize human-centric pipelines and linear workflows over the needs of modern AI systems: efficient batching, rich contextual metadata, and scalable access across distributed computing.
We also share ongoing efforts to enable intelligent genome analysis through novel algorithms, hardware accelerators, and emerging technologies such as processing-in-memory. Current formats unfortunately impede these parallelism efforts, as I/O and memory bandwidth remain significant bottlenecks. Finally, we present projects exploring new genomic data formats and outline future challenges, benefits, and research directions. These efforts aim to lay a foundation for AI-ready, scalable, and intelligent genome analysis.










