ESM-2 outperforms all tested single-sequence protein language models across a range of structure prediction tasks. The MSA Transformer (ESM-MSA-1) can improve performance on some proteins by leveraging MSA information