Predicting 3D structure and domains

There are three main ways to predict a protein's 3-D structure, and which one to use depends on what is already known.

• Homology modelling: line up the unknown sequence with proteins whose structures are already known. If they are very similar, the unknown protein probably has the same overall fold. Tools: MODELLER and SWISS-MODEL.
• Fold prediction (threading): if no close relative is known, try fitting ("threading") the unknown sequence onto the backbones of known structures, and check which fits physically. It needs more computing than homology modelling. Tools: LIBELLULA and Threader.
• De novo (ab initio) prediction: build the 3-D structure from the amino acid sequence alone, with no template at all, as QUARK does.

The decision (shown in the book's Figure 10.1) is simple: is there a similar protein with a known structure in the PDB? If yes, use homology modelling. If no, use fold prediction or ab initio prediction.

Predicted and experimentally solved structures are stored as lists of atomic coordinates (the position of every atom) in text files ending .pdb, in the Protein Data Bank, alongside structures solved by X-ray crystallography and NMR.

Many proteins are built from domains: distinct parts that fold on their own, have their own function or structure, and turn up again and again in different proteins, like standard building blocks. Domains are found with InterProScan (from EMBL) and CDD search (from NCBI).