AlphaFast: High-throughput AlphaFold 3 via GPU-accelerated homology search

AlphaFold 3 (AF3) enables accurate biomolecular modeling, but CPU-bound multiple sequence alignment (MSA) construction can limit end-to-end throughput. We introduce AlphaFast, an integrated framework that combines batched MMseqs2-GPU homology search with AF3-compatible feature generation, model weights, and outputs. AlphaFast achieves a 68.5-fold speedup in data generation and a 22.8-fold end-to-end speedup on a single H200 GPU. On four H200 GPUs, it achieves an amortized end-to-end wall time of 8.1 seconds per input, a 71.2-fold throughput improvement relative to the single-H200 AF3 baseline. Across protein-length cohorts extending to 1,500 residues, structural scores were generally close between pipelines, with greater uncertainty in the longer-protein cohorts. On 32 complete-context CASP targets, mean structural scores favored AF3, with target-specific differences between pipelines. For 32 protein--RNA targets, combining GPU-accelerated protein search with MMseqs2-CPU RNA search reduced end-to-end wall time by 7.73-fold, with small mean changes in RNA structural and interface lDDT relative to their uncertainty intervals. AlphaFast supports local, HPC, and multi-GPU execution, with measured serverless compute costs of approximately $0.039 per input in a 512-input batch. Code is available at https://github.com/RomeroLab/alphafast.