MCtandem: an efficient tool for large-scale peptide identification on many integrated core (MIC) architecture

Chuang Li, Kenli Li, Keqin Li, Feng LinBMC Bioinformatics 2019 · DOI

TL;DR

This paper introduces MCtandem, a novel tool designed to accelerate large-scale peptide identification from mass spectrometry data by leveraging the Intel Many Integrated Core (MIC) architecture. By developing a parallel scoring algorithm and optimizing its implementation, MCtandem achieves significant speedups, making it a crucial advancement for computational proteomics.

What Problem Does This Paper Address?

The field of proteomics relies heavily on tandem mass spectrometry (MS/MS) to identify peptides, which are the building blocks of proteins. This identification process typically involves searching experimental MS/MS spectra against vast databases of theoretical peptide sequences. However, the rapid advancements in mass spectrometry technology have led to an explosion in the volume of spectral data generated. Existing peptide database search algorithms struggle to keep pace with this data deluge, creating computational bottlenecks that hinder large-scale analyses, such as identifying entire proteomes. This paper addresses the critical need for more efficient and scalable peptide identification methods to handle the ever-increasing demands of modern proteomics research.

Methods

The core innovation of MCtandem lies in its novel parallel scoring algorithm, MIC-SDP (spectrum dot product), specifically designed for the MIC architecture. This algorithm employs a two-level parallelization strategy. To further boost performance, MCtandem incorporates several optimization techniques:

  • MIC-SDP Algorithm: This algorithm is designed to be highly parallelizable on the MIC's many cores. It dispenses with an initial loop found in standard SDP algorithms, allowing threads to work in parallel on experimental spectra. It also merges the SDP calculation and peak matching steps, reducing computational complexity.
  • Two-Level Parallelization: This involves task-level parallelism between the host CPU and the MIC coprocessor (using MPI) and thread-level parallelism within the MIC coprocessor (using OpenMP) through peptide sequence decomposition.
  • Optimization Strategies:
    • Pre-fetching: Implemented using a double-buffering mechanism to hide memory access latency by preparing data in advance.
  • Multithreading and Hyper-threading: Leverages the MIC's multiple cores and hyper-threading capabilities to increase throughput.
  • Vectorization: Exploits the MIC's vector processing units (VPUs) for efficient numerical computations.
  • Communication Overlapping: Uses an improved proxy infrastructure with memory-mapped circular queues to minimize communication latency between the CPU and MIC, ensuring data pipes are kept busy.
  • Dynamic Task Scheduling: Implemented to balance the workload across MIC coprocessors and avoid system overload.

Datasets

The paper utilized two protein sequence databases for evaluation:

  • UniProtKB/Swiss-Prot: A 5.2GB database containing 540,171 proteins.
  • UniProtKB/TrEMBL: An 18GB database containing 1,821,879 proteins.

The experimental spectra data were generated from tandem spectrometry experiments analyzing human liver samples. Three specific test datasets (Dataset 1, Dataset 2, and Dataset 3) derived from these experiments were used for performance comparisons, with varying sizes and numbers of spectra. No new benchmark datasets were introduced; rather, existing large-scale datasets were used to evaluate MCtandem's performance.

Key Insights and Ideas

  • MIC Architecture's Potential: The paper demonstrates that the Intel MIC architecture, with its numerous cores and vector processing capabilities, is well-suited for accelerating computationally intensive tasks like peptide identification.
  • MIC-SDP Algorithm's Effectiveness: The novel MIC-SDP algorithm is shown to be a significant improvement over standard SDP approaches, enabling substantial speedups on the MIC.
  • Multi-level Parallelism is Crucial: The combination of task-level parallelism (CPU-MIC) and thread-level parallelism (within MIC) is key to achieving high performance on heterogeneous systems.
  • Optimization Synergy: The effectiveness of MCtandem is not due to a single optimization but rather the synergistic application of pre-fetching, multithreading, vectorization, and communication overlapping.
  • Scalability for Big Data: MCtandem exhibits excellent scalability, handling increasingly large datasets and demonstrating significant speedups even on multi-node clusters, outperforming existing MapReduce-based solutions.

Key Results

  • Single MIC Speedup: MCtandem achieved up to a 28-fold speedup on a single MIC coprocessor compared to the original CPU-based X!Tandem implementation.
  • Cluster Performance: On a three-node MIC cluster, MCtandem achieved a 61.7-fold speedup compared to MR-Tandem, which required 50 nodes to achieve a 20.56-fold speedup.
  • Large Dataset Handling: MCtandem successfully processed datasets up to 12.11 GB (over 3 million spectra) with a linear increase in computation time, demonstrating its capability for very large-scale analyses.
  • Accuracy Preservation: The accuracy of MCtandem, measured by cosine values, was found to be consistent with that of MR-Tandem, indicating that performance gains were achieved without sacrificing identification correctness.
  • Optimization Benefits: The combined optimization techniques provided a 23.4% performance boost compared to the unoptimized version.

Limitations and Future Work

  • Limitations: While the paper highlights significant speedups, the exact computational complexity of the MIC-SDP algorithm is presented as O(lg(T) + |H|NK), which, while improved, still depends on the size of the theoretical spectra (T) and hypothesis spectra (H). The paper also notes that the relatively low bus bandwidth between system memory and the MIC coprocessor can be a bottleneck for data read-back.
  • Future Work: The authors suggest that the optimization strategies employed in MCtandem could provide insights for similar database search applications. The paper implicitly points towards further exploration of heterogeneous computing architectures for proteomics. The availability of MCtandem as open-source software encourages further development and application.

Summary updated

BibTeX
@article{li2019mctandem,
  author = {Chuang Li and Kenli Li and Keqin Li and Feng Lin},
  title = {{MCtandem: an efficient tool for large-scale peptide identification on many integrated core (MIC) architecture}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-019-2980-5},
}
Figure