Kallenborn, F.; Chacon, A.; Hundt, C.; Sirelkhatim, H.; Didi, K.; Dallago, C.; Mirdita, M.; Schmidt, B.; Steinegger, M.

Sensitive search of rapidly growing protein sequence databases for evolutionary information requires continual acceleration. This is achieved by innovating algorithms to filter sequences, or perform gapped alignments. Here, we present GPU-optimized algorithms for gapless filtering, achieving up to 100 TCUPS across eight GPUs, and gapped alignment using protein profiles. Implemented in MMseqs2-GPU, they result in 20x faster and 71x cheaper search on a NVIDIA L40S GPU compared to MMseqs2 k-mer on a 128-core CPU. In ColabFold, they accelerate structure prediction 23x at matching accuracy to AlphaFold2. MMseqs2-GPU is open-source software available for CUDA devices at mmseqs.com.