Bioinformatics & AI Acceleration

Accelerating Protein Structure Prediction on Domestic GPU-like Accelerators

With the rapid development of deep learning models in protein structure prediction, models such as AlphaFold3 have achieved high-precision predictions, but long prediction times remain a bottleneck limiting their application. The Multiple Sequence Alignment (MSA) stage is the primary performance bottleneck, with native AlphaFold3 using the CPU-based Jackhmmer tool for sequence search. This study proposes an MSA acceleration scheme based on domestic GPU-like accelerators. We port MMseqs2 from CUDA to HIP to enable GPU acceleration on domestic accelerators, complete format conversion for 4 out of 9 AlphaFold3 databases, and integrate MMseqs2 into the AlphaFold3 workflow. A memory management strategy is proposed to enable parallel execution of 4 database searches on a single accelerator card. Experimental results on 20 protein sequences show that: (1) GPU acceleration reduces search time by 83.4% compared to Jackhmmer (from 21.7 minutes to 3.6 minutes, achieving 6.0x speedup); (2) memory optimization enables parallel execution, reducing time by 52.8% compared to serial GPU execution (from 3.6 minutes to 1.7 minutes, achieving 2.1× speedup); (3) overall acceleration achieves 92.2% time reduction (12.8x speedup) while maintaining comparable prediction accuracy (average ranking-score: 0.761 vs 0.760).