Table 1.

Approaches that could utilize metagenome data properties for better protein structure prediction​​​​​

Approach	Metagenome source	Number of biomes	Strategy	Source
HMM + Rosettaa	IMG database	Multiple biomes	Combined	[63]
HMM + C‐QUARKb	Ocean microbiome	Single biomes	Single	[64]
Alphafoldc	Metagenome	Multiple biomes	Combined	[57]
DeepMSA + C‐I‐TASSERd	Mgnify	Multiple biomes	Combined	[70]
MetaSource + DeepMSA + C‐I‐TASSERe	Mgnify	Multiple biomes	Targeted	[70]

Note: Single strategy: using a single large biome as the protein source. Combined strategy: using a set of large biomes as protein sources. Targeted strategy: customized methods that select different biomes for different proteins.

^a Using IMG database [70, 82], models for 614 protein families were generated for unknown structures.

^b Using Tara Oceans data [80], proteins for 27 Pfam families were modeled with unsolved structures.

^c A deep learning algorithm, leveraging multisequence alignments were used for modeling protein structures.

^d Built on 4.25 billion microbiome sequences, 1044 Pfam families foldable by C‐I‐TASSER [70].

^e As targeted approach, MetaSource model was used to identify a set of biomes to supplement homologous sequence for specific Pfam families [70].