FiCOPS: Hardware/Software Co-Design of FPGA Computational Framework for Mass Spectrometry-Based Peptide Database Search

This paper presents FiCOPS, a hardware/software co-designed FPGA framework that accelerates mass spectrometry-based peptide database searches by exploiting parallelism to achieve significant speed-ups and power reductions compared to existing CPU and GPU solutions.

The Gist

Imagine you are a detective trying to solve a massive mystery. You have a bag full of broken glass shards (these are the Mass Spectrometry data). Your goal is to figure out exactly which vase each shard came from by comparing them against a library containing blueprints for billions of different vases (the Protein Database).

For decades, detectives have been doing this by hand, one shard at a time. It's slow, tedious, and when the library gets huge (like when studying complex environments or rare diseases), the process takes days or even weeks.

This paper introduces FiCOPS, a new "super-detective" built using a special type of computer chip called an FPGA. Think of FiCOPS not as a single super-smart detective, but as a factory assembly line of thousands of tiny, specialized workers who can solve the puzzle simultaneously.

Here is the breakdown of how they did it, using simple analogies:

1. The Problem: The Library is Too Big

In the old way (using standard computers), the detective would:

1. Take a glass shard.
2. Walk to the library.
3. Look at every single blueprint to see if it matches.
4. Walk back, write down the score, and repeat.

If the library has billions of blueprints, this "walking back and forth" (moving data) takes forever. Also, trying to account for every possible way a vase could be chipped (called Post-Translational Modifications) makes the library explode in size, making the search impossible for standard computers to handle quickly.

2. The Solution: The "FiCOPS" Factory

The authors built a custom machine (FiCOPS) that changes the game entirely. Instead of one detective walking around, they built a factory floor right next to the library.

• The Assembly Line (Pipelining): Imagine a conveyor belt. As soon as one worker finishes checking a shard, the next one grabs it immediately. There is no waiting time.
• The Swarm of Workers (Parallelism): Instead of one person, FiCOPS has hundreds of tiny workers (Processing Elements) all working at the exact same time.
• The Local Desk (On-Chip Memory): In the old way, workers had to run to the main library to get a blueprint. In FiCOPS, every worker has their own small desk with the blueprints they need right next to them. They never have to leave their station. This saves a massive amount of time.

3. The "Co-Design" Secret Sauce

The authors didn't just buy a faster computer; they redesigned the workflow to fit the hardware.

• The Analogy: Imagine you are trying to move a mountain of sand.
  • The CPU (Old Way): Uses a giant dump truck. It's powerful, but it has to drive back and forth to the dump site constantly.
  • The GPU (Graphics Card): Uses 1,000 small scoops. It's fast, but the scoops are heavy, and the energy cost to run 1,000 scoops is huge.
  • FiCOPS (The New Way): They built a custom conveyor belt system specifically for sand. It uses a tiny motor (low power) but moves the sand continuously without stopping.

They used a mathematical model (a "crystal ball") to figure out exactly how many workers they needed and how big the desks should be before they even built the machine. This prevented them from wasting time building a factory that was too big or too small.

4. The Results: Fast, Cheap, and Green

When they tested FiCOPS against the best existing methods:

• Speed: It was 3.5 times faster than the best standard computers and 3 to 5 times faster than the best graphics cards (GPUs).
• Energy: This is the biggest win. FiCOPS used 3 times less electricity than a standard computer and 5 times less than a graphics card.
  • Analogy: It's like getting a Ferrari's speed but with the fuel efficiency of a bicycle.

Why Does This Matter?

Currently, scientists studying things like the human microbiome or rare diseases have to wait days for results. With FiCOPS, they could potentially get results in minutes.

Even better, because it uses so little power, this technology could eventually be built directly inside the mass spectrometer machines used in hospitals and labs. Imagine a machine that analyzes a patient's blood sample and gives the doctor the diagnosis while the machine is still running, rather than sending the sample to a lab and waiting days for a report.

In summary: The authors took a slow, energy-hungry process of matching broken glass to blueprints, and built a custom, energy-efficient assembly line that does it in the blink of an eye.

Technical Summary

1. Problem Statement

Mass spectrometry (MS)-based proteomics relies on database search algorithms to deduce peptide sequences from experimental spectra. However, current methods face critical scalability issues:

• Search Space Explosion: The inclusion of Post-Translational Modifications (PTMs) and the study of non-model organisms (meta-proteomics) create combinatorial search spaces that are exponentially larger than the original protein databases.
• Serial Algorithm Limitations: Traditional serial algorithms (e.g., X!Tandem, Crux) suffer from poor scalability due to excessive paging and data movement, leading to search times ranging from days to weeks for large datasets.
• Hardware Inefficiency: While High-Performance Computing (HPC) solutions (CPU clusters, GPUs) exist, they often fail to provide real-time, on-instrument solutions due to high power consumption, cost, and communication bottlenecks. Existing GPU implementations (e.g., GPU-Tide) often fail to outperform optimized serial CPU tools due to unoptimized data transfer overheads.
• Bias: Current filtering mechanisms to manage search space often result in the identification of only abundant peptides, missing rare or modified variants.

2. Methodology

The authors propose FiCOPS, a hardware/software co-design framework implemented on an FPGA to accelerate peptide database searching. The methodology involves three main steps:

A. Algorithmic Analysis & Parallelism Identification

The authors analyzed the standard database search workflow (specifically the "closed-search" algorithm) to identify parallelism:

• Inner Loop (Dot-Product): The core computation involves matching experimental spectrum peaks against theoretical peptide ions. This was identified as highly parallelizable via loop unrolling.
• Middle/Outer Loops: While candidate peptides and spectra can be processed in parallel, memory bandwidth is a primary bottleneck.
• Indexing Strategy: Unlike serial tools that use memory-intensive "fragment-ion indexing" (which duplicates peptide references and consumes massive disk space), FiCOPS utilizes peptide-indexing. This decision was made to fit within the constrained on-chip memory of FPGAs, trading some parallelism granularity for memory efficiency.

B. Hardware Architecture Design

FiCOPS employs a configurable, parameterized, multi-core processing architecture on an Intel Stratix 10 FPGA:

• System Topology: A host CPU communicates with the FPGA via PCIe. The FPGA contains multiple Processing Units (PUs), each containing a pipeline of Processing Elements (PEs).
• Processing Element (PE): Each PE contains:
  • On-chip RAM: Stores experimental spectra and candidate peptides to minimize off-chip access.
  • Dot-Scorer Module: A configurable unit that performs parallel dot-product calculations. It features a Theoretical Ion Generator (computing b/y ions on-the-fly) and an Ion Matching Unit (comparing theoretical ions against experimental spectra using parallel comparators).
  • Double Buffering: Used to prevent computation stalls during data loading.
• Dataflow: The system uses a producer-consumer pipeline where spectra and peptides are streamed into PEs, and scores are streamed out, overlapping computation with communication to hide latency.

C. Design Space Exploration (DSE)

To find the optimal configuration without exhaustive trial-and-error (which is time-consuming for FPGA compilation), the authors developed an Analytical Performance Model:

• Model Inputs: Workload parameters (number of spectra, peptides, PTMs) and architectural parameters (number of PUs, PEs per PU, loop unrolling factor).
• Optimization: The model predicts total execution time, resource utilization (ALMs, RAM), and power consumption. It formulates the design problem as a constrained optimization to minimize execution time subject to FPGA resource budgets.
• Outcome: The DSE revealed that using simpler PEs arranged in deeper pipelines yields better performance than complex PEs, as it reduces memory bus contention and enables better data reuse.

3. Key Contributions

1. FiCOPS Framework: The first end-to-end FPGA-based computational framework specifically designed for MS peptide database searching, utilizing a hardware/software co-design approach.
2. Analytical Performance Model: A novel model that enables rapid design space exploration, allowing researchers to predict resource usage and performance without full FPGA synthesis for every configuration.
3. Memory-Efficient Architecture: A shift from memory-heavy fragment-ion indexing to a peptide-indexing approach optimized for on-chip FPGA memory constraints, utilizing on-the-fly ion generation.
4. Scalable Pipeline Design: A highly pipelined architecture that overlaps computation and data transfer, effectively mitigating the PCIe communication bottleneck.

4. Experimental Results

The system was implemented on an Intel Stratix 10 FPGA and evaluated against state-of-the-art tools (X!Tandem, Crux, MSFragger, HICOPS, GPU-Tide, GiCOPS) using six real-world datasets (ranging from 138K to 3.7M spectra).

• Speedup Performance:

  • vs. Serial CPU: FiCOPS is 101× faster than X!Tandem and 82× faster than Crux.
  • vs. Optimized Serial (MSFragger): FiCOPS achieves a 3.5× speedup in closed-search and 4× speedup in open-search.
  • vs. Parallel CPU (HICOPS): FiCOPS outperforms a 2-node HICOPS cluster and is comparable to a 4-node cluster, despite running on a single FPGA.
  • vs. GPU (GPU-Tide): FiCOPS is 5× faster than GPU-Tide in closed-search. Notably, GPU-Tide was found to be slower than its serial counterpart in some cases due to I/O overhead.

• Power Efficiency:

  • Power Consumption: FiCOPS consumes an average of 32.8W (Open Search) and 34.5W (Closed Search).
  • Comparison: This represents a 3× reduction compared to CPU solutions (e.g., MSFragger at ~106W) and a 5× reduction compared to GPU solutions (e.g., GPU-Tide at ~132W).
  • Efficiency: FiCOPS achieves a power efficiency of 228.6 kOPs/W (kilo-operations per watt) for open search, which is 100× higher than existing GPU solutions.

5. Significance

• Real-Time Potential: The high speed and low power consumption make FiCOPS a viable candidate for on-instrument, real-time analysis, a critical requirement for clinical proteomics and personalized medicine.
• Energy Efficiency: The framework demonstrates that specialized hardware (FPGA) can outperform general-purpose accelerators (GPUs) and massive CPU clusters in terms of both speed and energy efficiency for specific scientific workloads.
• Scalability: The analytical model proves that the architecture scales linearly with resources, offering a path for future integration into larger System-on-Chip (SoC) designs for mass spectrometry machines.
• Paradigm Shift: The paper challenges the notion that simply porting serial code to GPUs yields performance gains, highlighting the necessity of hardware-aware algorithm redesign and co-design to overcome I/O and memory bottlenecks.