DeepMSA (version 2) is a composite approach to generate high quality multiple sequence alignments for protein monomers or protein multimers based on huge genomics and metagenomics databases with a structure model-based multi-MSA ranking system. For protein monomer, the MSAs are produced by three iterative MSA generation pipelines with large alignment depth and diverse sequence sources by merging sequences from whole-genome sequence databases (Uniclust30 and UniRef90) and from metagenome databases (Metaclust, BFD, Mgnify, TaraDB, MetaSourceDB and JGIclust). For protein multimer, the top N ranked MSAs for each constituent protein are selected for generating potential paired MSAs. Each selected MSA for one constituent protein can be paired with the MSA of another constituent. Large-scale benchmark data show that the performance of several important tasks in protein research, including protein monomer and complex structure prediction, template detection, and contact/distance prediction, can be significantly improved after utilizing DeepMSA2-derived MSAs.

DeepMSA (version 2) consists of two separate pipelines for monomer and multimer MSA constructions respectively. For monomer MSA construction, it utilizes three parallel blocks (dMSA, qMSA, and mMSA) built on different searching strategies to obtain raw MSAs from a diversity set of databases from whole-genome and metagenome sequence libraries. In each of the three MSA generation blocks, a similar logic is followed, in which an initial query is searched against a sequence database, and if a sufficient number of effective sequences is not achieved, iterative searches into larger databases are attempted. Finally, up to 10 raw MSAs are scanned and ranked through a rapid deep learning folding process to select the optimal candidate MSA. For multimeric MSA construction, multiple composite sequences are created by linking the monomeric sequences from different component chains that have the same orthologous origins. Here, a set of M top ranked monomeric MSAs from each chain are paired those of all other chains, which result in M^m hybrid multimeric MSAs with m being the number of distinct monomer chains in the complex. The optimal multimer MSAs are then selected based on a combined score of the depth of the MSAs and folding score of the monomer chains.