2026/01/01 by Benjamin Perry, Jeongyheon Kim, Philip Romero · 1 voice
paper · doi:10.6084/m9.figshare.31343287
openalex publication_date 2026/01/01 · openalex created_date 2026/02/17 · openalex updated_date 2026/07/01
AlphaFold 3 (AF3) relies on slow, CPU-bound multiple sequence alignment (MSA). We introduce AlphaFast, a replacement for AF3 that integrates MMSeqs2-GPU homology search to eliminate this bottleneck. AlphaFast reaches a 68.5 X speedup in MSA construction and a 22.8 Xspeedup in end-to-end inference on a single GPU. On four GPUs, AlphaFast achieves a throughput of 8 seconds per input while maintaining statistically indistinguishable predictive performance. AlphaFast is also made available to those without computational resources at a cost of less than 0.04 per input.