Last Updated on 2026-10-07
PEZY Computing, Inc. (hereinafter “PEZY Computing”) and QuettaAI, Inc. (hereinafter “QuettaAI”) are pleased to announce that they have successfully implemented an accelerated version of HMMER [1,2], a sequence search tool widely used in the field of bioinformatics, for the “PEZY-SC3” — an MIMD processor for HPC developed and manufactured by PEZY Computing — and for general-purpose CPUs.
The project was led by software engineer Toshiyuki Oda (QuettaAI). By making full use of a state-of-the-art generative AI coding agent (Claude [3] Code / Opus 5), code modification for a complex architecture was accomplished rapidly.
Project Background and the Use of AI
With the dramatic improvements in generative AI in recent years, AI is now able to provide powerful support for hardware-specific code modification and optimization, work that has traditionally required a high degree of specialized expertise. In this project, the implementation process was entrusted to Claude Code, and the complex HMMER codebase was successfully optimized for the PEZY-SC3 in a short period of time.
To ensure reliability in scientific computing, development was carried out under the requirement that computational results match the original version 100%. Parts of the computation in which the ordering of floating-point operations would be problematic were avoided, and numerical values such as E-values and bit scores were confirmed to match exactly in the final output.
Dramatic Performance Gains
Benchmark tests using a range of datasets confirmed dramatic improvements in processing speed.
● Up to approximately 70x faster: For hmmscan, a speedup of up to approximately 70x over the original version was achieved (processing 5,000 fragment sequences predicted by prodigal [5] from high-throughput sequencing (HTS) reads ERR056999 [4]; Table 1-1).
● Substantially shorter analysis times in practice: Using an environment equipped with four PEZY-SC3 units, Pfam [6] annotation of one HTS run (all 262,574 reads of ERR056999 translated by prodigal) can be completed in approximately 11 minutes, and one bacterial proteome (the E. coli proteome, 4,403 sequences) in approximately 27 seconds (roughly a 40x speedup). Even compared with the inversion technique [7], which uses hmmsearch in place of hmmscan, this represents a speedup of approximately 10x (Table 2-1).
● Substantial speedup on CPUs as well: Thanks to our own optimization of the host-side code, speedups of several-fold to several tens of fold for hmmscan are achieved even in CPU-only environments that do not use the PEZY-SC3.
Future Developments
The optimized source code developed in this project is scheduled for public release as open source. Using accelerators such as the PEZY-SC3 efficiently has until now demanded deep specialist knowledge, but recent advances in coding AI agents have made it possible to implement accelerator code with far greater ease. This is expected to remove bottlenecks in the processing of the vast volumes of life-science data generated by metagenomic analysis and other applications, and to significantly accelerate research and development in the life sciences.
PEZY Computing and QuettaAI will continue to combine cutting-edge hardware technology with artificial intelligence technology in order to open up new possibilities in high-performance computing (HPC).
[1] http://hmmer.org/ (accessed 2026/09/24)
[2] Eddy, Sean R. "Accelerated profile HMM searches." PLoS computational biology 7.10 (2011): e1002195.
[3] https://claude.ai/ (accessed 2026/09/24)
[4] https://trace.ncbi.nlm.nih.gov/Traces/?view=run_browser&acc=ERR056999, retrieved on 2026-08-31 using sratoolkit [9]
[5] Hyatt, Doug, et al. "Prodigal: prokaryotic gene recognition and translation initiation site identification." BMC bioinformatics 11.1 (2010): 119.
[6] Paysan-Lafosse, Typhaine, et al. "The Pfam protein families database: embracing AI/ML." Nucleic acids research 53.D1 (2025): D523-D534.
[7] http://cryptogenomicon.org/hmmscan-vs-hmmsearch-speed-the-numerology.html (accessed 2026/09/24)
[9] https://github.com/ncbi/sra-tools version 3.4.1 (accessed 2026/09/24)
For details, such as specific processing times, please refer to the PDF linked below.