Flexible fitting in 3D-EM using superfamily information

We are working in a new method for flexible fitting in three-dimensional electron microscopy (3D-EM). The method explores the variability among the protein domains of a given superfamily, according to the definition given in databases as CATH or SCOP. For this purpose, a structural alignment of protein domains belonging to the superfamily, followed by a principal components analysis is done. The first three principal components of the decomposition are explored. Then, using rigid body transformations for the secondary structure elements (SSE) of the domain, plus a robotics algorithm to close the loops, a bio-chemically correct model is built for each structure. Finally, all the structures are fitted into the 3D-EM map using the program SITUS-COLORES, and the best one is chosen based on cross-correlation measures.

We have applied the method both simulated and experimental cases. In all types of experiments, the flexible fitting was able to produce better results than only applying rigid body fitting.

Results

We have performed a first initial assessment of the flexible fitting approach considering high resolution structures from several superfamilies. The test was to force the fitting of two different members of the same superfamily, one into the simulated low resolution (8 Å) map of the other, checking which were the deformations applied to make them fit. A variable number of member of the superfamily were used for each case. In all cases the three first principal components were explored considering 10 values for each component. The fittings at 8 Å were done with COLORES.

Nor surprisingly, we have found that the final results are highly dependent of the variational space of each superfamily rather than on the number of elements in the superfamily. Also, and quite logically too, the performance depends on how close the two structures being fitted are.

An example of a successful experiment is shown in Figure 4, where 20 members of 1.10.238.10 superfamily (1wdcB2, 5tnc02, 4tnc02, 4cln02, 4cln01, 3cln02, 3cln01, 2tn400, 2bbnA1, 2bbmA2, 2bbmA1, 1vrkA1, 1top02, 1tnq00, 1tn400, 1tcoB0, 1tcf00, 1qiwB1, 1qiwA1 and 1cll02) were used.

The cross correlation coefficients changes from 0.60 to 0.64 and the RMSD (computed for the backbone atoms belonging to the core) descends from 3.68 to 2.60 Å. If the RMSD is computed for the SSE’s atoms only, the value is reduced from 3.52 to 2.54 Å . Figure 4 shows the initial and final structures of 1wdcB2 together with 1cll02.

 

Figure 4 Flexible fitting at 8 Å of domain 1wdcB2 (regulatory domain of myosin molecule for Aequipecten irradians), belonging to CATH superfamily 1.10.238.10 (EF-hand), in the 3D-EM map of domain 1cll02 (domain #2 of calmodulin of Homo sapiens). Green: structure of 1cll02 that generated the map. Purple: structure of 1wdcB2. Red: Final structure of 1wdcB2 after flexible fitting in the map (represented in the image on the right).

Another experiment with a more complex fold, belonging to superfamily 2.80.10.50 and being of b class, is shown in Figure 5. In this case the variational information comes from 15 structures (1jlxA1, 1jlyA1, 1jlyB1, 1jlxB1, 1itbA0, 5i1b00, 1dfqA2, 1dllA2, 1d0hA2, 1avaD0, 1a8d02, 1avaC0, 1diwA2, 1af902, 1avxB0).The initial structure of 1jlxA1 is deformed to better fit into de map of 1a8d02. The LCCC value increases from 0.542 to 0.569 and the RMSD for the core atoms decreases from  3.90 to 3.53 Å. If only the atoms of the SSEs are considered, the RMSD changes from  3.66 to 2.46  Å.

Figure 5. Flexible fitting at 8 Å of domain 1jlxA1 (domain #1 of aglutinin of Amaranthus caudatus), belonging to superfamily 2.80.10.50 (lectin, trefoil topology), in the 3D-EM map of 1a8d02 (domain #1  of tetanic neurotoxin of Clostridium tetani). Green: structure of 1a8d02 that generated the density map. Purple: Structure of 1jlxA1. Red: Final structure of 1jlxA1 after flexible fitting on the map  (represented in the image on the right).

Discussion

The approach has worked well with the test cases of simulated maps of isolated domains with plenty of variational information coming from well populated superfamilies, and also in a complicated multi-domain experimental case, where not all the domains belonged to big superfamilies. In both types of experiments  the LCCC is improved, and in the simulated cases, where correct coordinates are available, that lead to reduced values of RMSD.

In the experiments where little variational information is available, or it is not well covered by the three first components, almost no change in the initial structures are to be expected (no matter how many structures are used in the alignment), whereas in the favourable cases changes occur and the fitting is improved. The required conditions to do so are that the variational information is adequately captured by the three first components, and that the initial structure to fit is similar enough to the actual one present in the 3D-EM map.