Communication-avoiding micro-architecture to compute Xcorr scores for peptide identification

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Database search algorithms are vital for protein identification in systems biology using mass spectrometry (MS) data.
• High-resolution MS generates large datasets, demanding efficient computational methods for spectrum analysis.
• Current accelerators often overlook communication costs, leading to inefficient memory access and poor performance.

Purpose of the Study:

• To propose a novel communication-avoiding micro-architecture for accelerating similarity score computation in MS data analysis.
• To address the computational and communication bottlenecks in current database search algorithms.
• To improve the efficiency of protein identification from mass spectrometry data.

Main Methods:

• Developed a communication-avoiding micro-architecture utilizing efficient local cache and peptide pre-fetching.
• Implemented a custom peptide broadcast bus for input reuse and optimized data handling.
• Designed an efficient bus arbitration scheme to minimize synchronization costs and maximize parallelism.

Main Results:

• The proposed micro-architecture significantly minimizes Direct Random-Access Memory (DRAM) accesses.
• Achieved an average performance improvement of 24x compared to a high-performance CPU implementation.
• Demonstrated efficient data utilization and reduced synchronization costs.

Conclusions:

• The novel micro-architecture offers a substantial performance gain for protein identification from mass spectrometry data.
• Addressing communication costs in accelerator design is crucial for efficient biological data analysis.
• This approach provides a scalable solution for handling large-scale, high-resolution mass spectrometry datasets.