# What is the difference between structure assignment and structure prediction?

I am absolutely new to structural bioinformatics (only started last week). I am working on the secondary structure assignment/prediction (actually I am not sure) of proteins using machine learning.

When it comes to protein analysis, most of the books or research papers talk about protein structure prediction. Only a handful of books or articles talk about protein structure assignments.

What is the difference between structure assignment and structure prediction?

• I suppose these resources do not speak about these terms out of context. Is this context the same or different for both terms ? Jul 5, 2021 at 5:59
• My comment was a hint for you to finish your homework of elaboration of the question. // Your own reasoning – based on searching, reading and thinking – is supposed to be present to avoid the question closure for lack of own effort. Jul 5, 2021 at 6:13
• rostlab.org/papers/2008_rev_assignment2/paper.html#DSSP Aug 24, 2021 at 11:46
• pubmed.ncbi.nlm.nih.gov/8749853 Aug 24, 2021 at 11:48

Structure assignment is relating e.g., a measured absorbance frequency (like a triplet in $$\pu{^1H NMR}$$), or a structural feature (like $$\Phi$$ and $$\Psi$$ in a Ramachandran plot) to a structure already known. It equally may be based on a putative / postulated structure; then, you compare e.g., how a simulated NMR spectrum is similar to the experimentally recorded one. If the two spectra match 1:1, the putative structure likely is a correct one for this NMR experiment.

Structure prediction departs knowing e.g., only a sequence of $$\alpha$$-amino acids, e.g. Ala-Leu-Cys, and offering a guess about their spatial arrangement (e.g., folding). Its foundation may be more or less educated / reasonable, which is why one typically infers from already known relationships of sequences of amino acids, and their 3D structure already established, e.g., by NMR, or X-ray crystal diffraction. Machine learning (ML) calls this stage of establishing a model as training. Aside from comparing with other algorithms, you eventually benchmark your ML algorithm in multiple runs with different training sets (e.g., sequences) to watch the outcome of the predicted structures; for one, comparison with experimentally determined data intentionally excluded from the training set now offer a check about the correctness of the results. For two, this equally offers to estimate the likelihood a folding predicted in silico to equally occur in vitro e.g., reporting a range for a particular $$\Psi$$ between two adjacent amino acids.

I am working on the secondary structure assignment/prediction (actually I am not sure) of proteins using machine learning.

Secondary structure assignment is an automated method that defines the secondary structure (helix, sheet, loop) based on a known three-dimensional structure. Secondary structure prediction refers to taking the primary structure (the sequence of amino acid residues in a protein) and predicting the secondary structure.

There are also some experimental methods that help to characterize secondary structure in the absence of a 3D structure. Circular dichroism spectroscopy gives an estimate of the secondary structure content. Assigned chemical shifts of C-alpha and C-beta atoms from NMR experiments give a residue-by-residue prediction (or characterization) of secondary structure.

What is the difference between structure assignment and structure prediction?

Without the qualifier "secondary", I am unfamiliar with structure assignment. If you google "protein structure assignment", most hits will talk about secondary structure assignment. NMR resonances are assigned to protein primary structure (i.e. which residue gives rise to which signal).

Structure prediction most often refers to tertiary structure prediction. However, you can also predict secondary structure, protein-ligand structures (docking studies) and quaternary structure.

I need a layman's explanation of "protein secondary structure assignment". Adding a diagram or picture would be much appreciated.

Secondary structure refers to the conformation of the main chain in a protein, and the hydrogen bonds between the carbonyl oxygen and the amide hydrogen that are observed. The example below shows that the two secondary structure elements called beta sheet and alpha helix differ in the conformation of the main chain and in their hydrogen bonding pattern.

To assign a secondary structure based on an atomic model, you would measure the main chain torsion angle and deduce hydrogen bonds from distances of hydrogen bond donors and acceptors. Often, the secondary structure assignment is shown juxtaposed with the primary sequence and sometimes with other annotations. Here is an example screen shot. The highlighted Cys30 is part of a helix (red shade) and makes a disulfide bridge with Cys115 (see additional annotation).

Different software packages usually agree on the presence of helices and sheets, but sometime have small differences in the exact start and end of the secondary structure elements (see e.g. here). It is also possible to assign secondary structure from electron density or cryo-EM maps, see e.g. here.

• Is "Assignment" a part of the "Prediction"-process?
– user93950
Aug 28, 2021 at 10:34

Assignment refers to a statement that a particular amino acid residue

(1) has a given backbone conformation described by dihedral angles falling within certain limits

(2) based on the local conformation of a series that includes adjacent amino acids, the residue can be said to be part of a secondary structure such as one of various types of helix, sheet or turn (or simply lacking specific structure, ie intrinsically disordered).

The basis for the determination of dihedral angles in NMR is to pool together various spectroscopic variables and use these as constraints during molecular dynamics simulations of the protein, in order to generate an ensemble of possible structures, then to evaluate the dihedral angles in those structures. The data usually includes NOEs, which are signal enhancements observed when two nuclei are close and provides a sensitive constraint on the magnitude of their separation. NOEs are labelled based on the atoms involved, including the distance between the residues (in the primary sequence) to which each atom belongs. Certain patterns of NOEs are consistent with particular secondary structures and can provide qualitative evidence of such structure before applying the more elaborate analysis previously described. Other valuable data include heteronuclear chemical shifts and various J-couplings. A high-quality set of NOEs and complementary data can in principle constrain the conformation of the protein down to a non-redundant choice.

It is possible for (1) and (2) to be inconsistent, that is, to assign an amino acid to particular secondary structure, but for that particular amino acid to reveal, on the basis of spectroscopic data and/or simulations, conformations inconsistent with that secondary structure. In that case the assignment is said to violate the data (or vice-versa), making the assignment suspect if not falsifying it, and calling details of the analysis into question. The quality of an ensemble is assessed on the basis of a statistical analysis of the agreement between the data and the assigned secondary structures.

Secondary structure prediction is a term used to refer to the use of known correlations between secondary structures and observables, and can be said to include using spectroscopic or crystallographic data to estimate what the structure of a protein is like, including how flexible. Assigning residues to particular secondary structure is a form of prediction, but when the input is experimental data and the degree of confidence is sufficiently high the process is not usually called prediction. Prediction involves an input of data, possibly sparse and usually including parts or all of the primary sequence, obtained for a protein of interest, into a prediction algorithm which takes such inputs and returns a list of possible secondary structures ranked by likelihood of being accurate. The algorithm is developed (a step in some cases called training) by optimizing the agreement of the output obtained for proteins of known structure.

• Is "Assignment" a part of the "Prediction"-process?
– user93950
Aug 28, 2021 at 10:34