Deep learning, epitomized by models like AlphaFold2, has achieved unparalleled accuracy in protein structure prediction. However, the depth of multiple sequence alignment (MSA) remains a bottleneck, especially for proteins lacking extensive homologous families. Addressing this, we present MSA-Generator, a self-supervised generative protein language model, pre-trained on a sequences-to-sequences task with an automatically constructed dataset. Equipped with protein-specific attention mechanisms, MSA-Generator harnesses large-scale protein databases to generate virtual, informative MSAs, enriching subpar MSAs and amplifying prediction accuracy. Our experiments with CASP14 and CASP15 benchmarks showcase marked LDDT improvements, especially for challenging sequences, enhancing both AlphaFold2 and RoseTTAFold's performance.