I've been seeing using sequence alignment to help with prediction of protein folding, but am unable to understand why. Google returns other research papers for these search terms, hence I'm hoping the stackoverflow community can help thank you.
Protein sequences from the same family form a sequence alignment. Some positions in the alignment will be conserved, i.e. the same in all sequences. These are often associated with function, e.g. a catalytic residue at position 50.
However there is a more subtle signal in sequence alignments called covariation. Two positions, say residue 30 and residue 60, may appear to change together across the alignment. We infer that they are close in 3D space in the structure:
Initially statistical techniques (DCA, PSICOV etc.) were used to separate direct from indirect (chained) couplings and extract predicted contacts from sequence alignments. Now deep neural networks have improved performance and even allowed the prediction of residue-residue distances and orientations.
These constraints on the structure allow model building, for example using the CNS software. A good set of constraints effectively guarantees the correct fold. The depth and quality of the sequence alignment is the major factor in the accuracy of the predicted constraints.
This approach to protein structure prediction is the current state of the art. Use of covariation from sequence alignments is the only approach that has reliably given accurate models for larger proteins without templates. See this paper and this paper for the best current methods, and tune into the CASP conference this December to see if anyone can improve on them.