Evolutionarily related sequences provide information for the protein structure and function. Multiple sequence alignment, which includes homolog searching from large databases and sequence alignment, is efficient to dig out the information and assist protein structure and function prediction, whose efficiency has been proved by AlphaFold. Despite the existing tools for multiple sequence alignment, searching homologs from the entire UniProt is still time-consuming. Considering the success of

more »

... aFold, foreseeably, large- scale multiple sequence alignments against massive databases will be a trend in the field. It is very desirable to accelerate this step. Here, we propose a novel method, fastMSA, to improve the speed significantly. Our idea is orthogonal to all the previous accelerating methods. Taking advantage of the protein language model based on BERT, we propose a novel dual encoder architecture that can embed the protein sequences into a low-dimension space and filter the unrelated sequences efficiently before running BLAST. Extensive experimental results suggest that we can recall most of the homologs with a 34-fold speed-up. Moreover, our method is compatible with the downstream tasks, such as structure prediction using AlphaFold. Using multiple sequence alignments generated from our method, we have little performance compromise on the protein structure prediction with much less running time. fastMSA will effectively assist protein sequence, structure, and function analysis based on homologs and multiple sequence alignment.