Schematic illustration of ESMPair that builds interologs as the input to AlphaFold-Multimer. Given a pair of query sequences as input: (1) we first search the UniProt database [35] with JackHMMER [36] to generate the MSA for each query sequence, (2) sequences of the same taxonomy rank are grouped into the same cluster, (3) ESM-MSA-1b is applied to estimate the column attention score (ColAttn_score) between each sequence homolog of MSA with the query sequence. (4) One interolog is obtained by directly concatenating two matched sequence homologs. We match two sequence homologs of the same taxonomy group with similar attention scores from the two query sequences (5) AlphaFold-Multimer takes the interolog MSA as input to predict the complex structure.