AlphaFast: High-throughput AlphaFold 3 via GPU-accelerated MSA construction

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• Accurate biomolecular modeling is crucial for biological research.
• AlphaFold 3 (AF3) provides high-quality predictions but is hindered by slow multiple sequence alignment (MSA) generation.
• The CPU-bound nature of MSA limits the speed and scalability of AF3.

Purpose of the Study:

• To develop a faster method for MSA generation to overcome the limitations of AlphaFold 3.
• To reduce the computational bottleneck in AlphaFold 3's workflow.
• To enable rapid and cost-effective protein structure prediction.

Main Methods:

• Integration of GPU-accelerated MMseqs2 sequence search into a drop-in framework named AlphaFast.
• Utilizing a single GPU for MSA construction and end-to-end runtime analysis.
• Deployment on multiple GPUs for high-throughput structure prediction.

Main Results:

• Achieved a 68.5x speedup in MSA construction.
• Reduced end-to-end runtime by 22.8x on a single GPU.
• Enabled structure prediction in 8 seconds per input using four GPUs with maintained accuracy.
• Demonstrated cost-effective serverless deployment at $0.035 per input.

Conclusions:

• AlphaFast effectively removes the MSA bottleneck in AlphaFold 3.
• The framework offers substantial speedups and cost reductions for biomolecular modeling.
• AlphaFast makes accurate protein structure prediction more accessible and efficient.