SimMS: GPU-Accelerated Mass Spectrometry Similarity Engine

Client and Scientific Context

Pangea Bio operates at the intersection of computational biology and metabolomics research. Their platform enables researchers to identify unknown compounds by comparing tandem mass spectrometry (MS/MS) fragmentation patterns against reference spectral databases. This capability is foundational for drug discovery, natural product research, and clinical diagnostics. The challenge: as reference databases grow to millions of spectra, the computational cost of similarity matching becomes prohibitive.

The Computational Challenge

Untargeted metabolomics relies on cosine similarity algorithms to match experimental spectra against reference libraries. Established tools like MatchMS provide accurate results but execute sequentially on CPU, making large-scale repository searches impractical. For Pangea Bio's workloads—comparing tens of thousands of query spectra against databases containing millions of reference entries—traditional approaches would require tens of CPU-days per analysis run. This bottleneck was blocking their ability to scale.

• Reference database sizes growing into millions of spectra
• Traditional CPU methods requiring tens of CPU-days for large workloads
• Need for both greedy and modified cosine similarity algorithms
• Accuracy requirements: results must match reference implementations exactly

Engineering Approach

Scalexa's approach focused on GPU parallelization without sacrificing algorithmic correctness. We mapped the cosine similarity computation—including both greedy and modified variants—onto GPU architecture, designing kernels that exploit the inherent parallelism of spectral comparison. The implementation preserves tolerance handling and peak matching logic from MatchMS while executing thousands of comparisons simultaneously.

• Designed parallel GPU kernels for greedy and modified cosine similarity
• Preserved algorithmic equivalence with MatchMS across all tolerance ranges
• Implemented efficient memory management for large spectral datasets
• Built benchmarking infrastructure for validation against reference results

Solution Architecture

SimMS consists of a GPU compute layer that accepts batched spectral data, executes parallelized similarity computations, and returns scored matches. The engine integrates into Pangea Bio's existing computational metabolomics pipelines through a clean API. Deployment is containerized for reproducibility, with CI/CD infrastructure ensuring that every release passes regression tests against known reference outputs. The architecture supports both single-GPU workstations and cloud GPU instances for elastic scaling.

Technologies

Why It Matters

SimMS demonstrates what's possible when senior engineering meets scientific computing: a 1000× speedup without compromising accuracy. For Pangea Bio, this unlocks repository-scale metabolomics searches that were previously impractical, accelerating their research pipeline and reducing infrastructure costs. The peer-reviewed publication is available in Bioinformatics (https://academic.oup.com/bioinformatics/article/41/3/btaf081/8026685). The work exemplifies Scalexa's approach—pragmatic, maintainable, and grounded in measurable outcomes. Experience includes work for Walmart, Coca-Cola, ESPN, UNICEF, and FIFA.

Services Behind This Project

The same senior teams that delivered this work:

• DevOps & CloudCloud architecture, CI/CD and infrastructure
• AI & Machine LearningProduction ML, LLM and computer-vision systems