GEMA: A Genome Exact Mapping Accelerator Based on Learned Indexes.

Area of Science:

• Bioinformatics
• Computer Architecture
• Machine Learning

Background:

• Genome sequencing generates vast amounts of data requiring efficient processing.
• Existing exact mapping accelerators face challenges with speed and memory bandwidth, especially for long reads.

Purpose of the Study:

• Introduce GEMA, a novel FPGA-based accelerator for exact genome mapping.
• Improve the speed and efficiency of DNA sequence mapping using machine learning.

Main Methods:

• Developed GEMA using learned indexes and a machine learning algorithm for precise read localization.
• Implemented data augmentation and distribution-aware partitioning to enhance ML model accuracy.
• Proposed speculative prefetching and an optimized FPGA architecture to reduce memory bandwidth requirements.
• Integrated an efficient error recovery technique.

Main Results:

• GEMA achieves up to 1.36x higher speed for short reads compared to recent accelerators.
• GEMA demonstrates up to ~22x faster mapping for long reads compared to existing methods.
• Optimized FPGA architecture reduces off-chip memory access overhead.

Conclusions:

• GEMA offers a significant performance improvement for exact genome mapping on FPGAs.
• The ML-based approach and architectural optimizations make GEMA highly effective for both short and long reads.
• GEMA represents a substantial advancement in accelerating genomic data analysis.