Three methods are used to perform the final screen: the RMS deviation screen , the accessibility profile screen, or energy only (backbone and sidechains of the final model).

a) RMS deviation (RMSD) screen

As has been seen here, many of the kinked crystal structure H3 loops are fairly similar to each other. This has been exploited in a screen for the final model, which is (see table below) the most effective screen at present.
The RMSD screen works by essentially comparing each model with a weighted average of all the kinked crystal structures of that length, the weighting being based on sequence similarity to the model using the Dayhoff mutation matrix.

where E is meant to be a sigma (sum), dx is the RMS deviation between the model and structure x and mx the sequence similarity between the model and structure x.

Advantages: Works very well for kinked loops with the typical "hairpin" structure. Also, structures which can be predicted from sequence to be either atypical or part of a particularly well-defined subgroup of the "typical" group , (see here ) can be screened using an average of known structures with that feature, rather than the whole set. This particularly applies to 5-residue H3s. It also now applies to 11-residue loops with certain sequence motifs .

Disadvantages: For obvious reasons works less well for loops with atypical structure, where no sequence-structure rules have been proposed. 1kem and 1kb5 in the table , and 2jel to some extent, are examples of this. However,it still works reasonably for 11 residue loops, where the crystal structures examples are fairly well scattered (RMSD 2.0-3.0 angstroms) around the mean (e.g. 2h1p, 1mcp and 1fpt in the table; note that, although 1fpt belongs to the group of 11-residue H3s with a well-defined sequence-structure relationship, this experiment was done before that was discovered, and thus the RMSD screen for 1fpt was done comparing the models to all other 11-residue kinked H3s).

Important note: The method can only be used for 5 and 7-12 residue loops as for other lengths only two or three examples exist, rendering the average structure unrepresentative.

b) Accessibility profile screen

Like the RMSD screen, the accessibility profile screen is a knowledge-based method. It is based on the fact that kinked CDR-H3 loops have a characteristic accessibility profile for certain residues, partly as a result of the kink, but also resulting from the fact that residues at the apex of the loop (typically residues 3-5, counting from the N terminus) are usually exposed. The accessibility profile method scores conformations according to their deviation from the average accessibility in known structures of that length at those key residues:

where O is the observed accessibility in the model for the current key residue, M is the mean accessibility in the known structures for this length, s is the standard deviation and f is a power that the s.d. is raised to (2 was found to be best while testing).

An important point regarding this screen is that the sidechains must be built first; therefore, rather than selecting the final model and building sidechains on it, one builds the sidechains on the whole set of clustered CDR-H3 conformations and then screens.

Advantages: Picks out kinked structures with exposed apices (no "crumpled-in" loops), and is not restricted by the mean observed CDR-H3s as much as the RMSD screen is, so can potentially work well for atypical kinked H3s (e.g. 1kem). Works well with 7-9 residue loops (e.g. 1vfa, 1tet, 2fbj in the table).

Disadvantages: For longer (10+) loops, the apex of the loop can vary in conformation a good deal while retaining the accessibility profile, and therefore an accurate prediction is not guaranteed. It will therefore perform considerably poorer than the RMSD screen on 10+ loops with a conformation of, or close to, the typical hairpin shape (see 1for, 1dba, 1clz, 2h1p, 1mcp in the table).

Important note: This method can only be used for 5-14 residue loops, as shorter loops do not have enough residues to show a well defined profile, and longer loops consist almost all of flexible (and therefore variable accessibility) apex. In any case, there are barely any examples of known structures outside this range, making it difficult to determine a profile in any case.

c) Energy-based screen

The final method, and the only one that can be used if the loop is extended, is to simply select the lowest-energy clustered conformation using the VFF, the same forcefield as is used to energy-minimise each backbone. Sidechains are built on all the clustered backbones and the lowest is taken.

Advantages: It can be used on extended loops, for which there are not enough examples in the database to develop a knowledge-based screen at present. It is also the safest option for longer (12+) residue loops, where the apex comprises a very high percentage (50% or more) of the loop.

Disadvantages: Experience has shown that energy-based methods that do not express the environment with 100% accuracy (for example here, the solvent effect is ignored), are highly prone to false positives. Hence, although an energy screen is a good idea in the general sense to screen out the very worst conformations, they are less good at discriminating between reasonably good and moderately poor conformations than the knowledge based methods.

Comparison of screens

Pictures of the final models superimposed on the crystal structure H3 for the RMSD screen can be viewed here.

The table shows the RMS deviation of the selected model with respect to the crystal structure for each screen.
Note that the RMSD screen was done comparing the models to an average of all other kinked H3s of that length; the sequence-structure rules for certain 8 and 11 residue CDR-H3s were not known at the time.
Note also that the "Closest by MDM" column is the RMSD of the model from the closest match by the Dayhoff mutation matrix (excluding the structure being modelled) in the database of known CDR-H3 structures; it is to show that the weighted average approach of the RMSD screen is preferable to doing this simple screen.
Finally, note that 1mlb to 1igm have 5 canonical loops, whereas 1tet to 1kb5 had at least one other non-canonical loop at the time of this analysis (although our reclassification has subsequently placed the majority of these into canonical classes). This shows that the CDR-H3 can be modelled accurately even in this situation; however, the other non-canonical loop(s) can also be modelled reasonably well using the database-only search , typically having RMSD in the range 1.0 to 2.5 angstroms. Canonical loops, of course, are well modelled with RMSD below 2.0 angstroms in all instances.

Latest developments

Several of the latest developments in H3 structural knowledge have led to refinement of these screens for certain length H3 loops. See here for more details...

Which screen should I use?

As a guideline when modelling loops predicted to be kinked, either the RMSD or the accessibility screen is suitable for 7-9 residue loops, while the RMSD screen is the preferred option for 10 and 11 residue loops, as well as for 5-residue loops, for which a tentative series of sequence-structure rules have been drawn up (see here ). For 12 residue loops or longer, apices vary considerably, and the VFF (energy) screen is the method of choice. More on selecting a screen can be found here .

Last updated 22/11/01