The authors have declared that no competing interests exist.
Conceived and designed the experiments: JHP BLdG. Performed the experiments: JHP. Analyzed the data: JHP. Contributed reagents/materials/analysis tools: JHP BLdG. Wrote the paper: JHP BLdG.
Protein-protein interactions play an important role in all biological processes. However, the principles underlying these interactions are only beginning to be understood. Ubiquitin is a small signalling protein that is covalently attached to different proteins to mark them for degradation, regulate transport and other functions. As such, it interacts with and is recognised by a multitude of other proteins. We have conducted molecular dynamics simulations of ubiquitin in complex with 11 different binding partners on a microsecond timescale and compared them with ensembles of unbound ubiquitin to investigate the principles of their interaction and determine the influence of complex formation on the dynamic properties of this protein. Along the main mode of fluctuation of ubiquitin, binding in most cases reduces the conformational space available to ubiquitin to a subspace of that covered by unbound ubiquitin. This behaviour can be well explained using the model of conformational selection. For lower amplitude collective modes, a spectrum of zero to almost complete coverage of bound by unbound ensembles was observed. The significant differences between bound and unbound structures are exclusively situated at the binding interface. Overall, the findings correspond neither to a complete conformational selection nor induced fit scenario. Instead, we introduce a model of conformational restriction, extension and shift, which describes the full range of observed effects.
Due to their importance in biological processes, the investigation of protein-protein interactions is of great interest. Experimental structures of protein complexes provide a wealth of information but are limited to a static picture of bound proteins. Ubiquitin is a signalling protein that interacts with a wide variety of different binding partners. We have used molecular dynamics simulations to compare the dynamic behaviour of bound and unbound ubiquitin in complex with different binding partners. Our observations suggest that the conformations accessible to bound ubiquitin, while often restricted in comparison to unbound ubiquitin, still occupy a subspace of the conformational space as those of unbound ubiquitin. This corresponds to the “conformational selection” binding model. Only on a local level near the binding interface, differences between bound and unbound structures were found in specific regions of the bound ensemble. To account for the different types of behaviour observed, we extend the currently available binding models by distinguishing conformational restriction, extension and shift in the description of binding effects on conformational ensembles.
Protein-protein interactions are crucial in most biological processes, yet the principles governing the conformational effects of these interactions are still poorly understood. X-ray structures of protein complexes provide a wealth of high resolution structural information but reflect a static snapshot of the structure, leaving the mechanism of complex formation and dynamics in the complex unaddressed. In addition, compared with the growing number of experimentally determined structures of unbound proteins, there is only a small number of known structures of protein complexes. Computational methods are being developed to derive complex conformations from unbound structures, but this remains a challenging and highly non-trivial task
Two different models have been suggested to explain the conformational differences observed experimentally between bound and unbound proteins. The induced fit model
A good model system to investigate the conformational effects of complex formation is ubiquitin with its binding partners. Ubiquitin is a 76 residue protein that plays an important role in metabolic pathways, as the ubiquitination (covalent attachment of ubiquitin to a lysine side chain of a protein) can, among other functions, control the degradation or regulate transport of this protein. In this function, ubiquitin is recognised by and interacts with a multitude of other proteins. Lange et al.
Thus far most studies have focused on static snapshots of ubiquitin complexes in comparison to solution ensemble of unbound ubiquitin. Here, based on several experimental structures of ubiquitin in different complexes
Statistical evaluation of simulations of ubiquitin both in the presence and the absence of a binding partner indicates conformational selection to be the appropriate model for complex formation when considering the dominant backbone dynamics, while some localised differences between bound and unbound ensembles can be found near the binding interface.
Seventeen structures of ubiquitin in complex with eleven different binding partners were selected from the protein database (PDB)
PDB code | binding partner | reference |
1NBF | Ubiquitin carboxyl-terminal hydrolase 7 (HAUSP) |
|
1P3Q | CUE domain of Vacuolar protein sorting associated protein (Vps9p) |
|
1S1Q | Tumor susceptibility gene 101 protein (TSG101) |
|
1UBI | none (unbound reference) |
|
1UBQ | none (unbound reference) |
|
1UZX | UEV domain of Vps23 |
|
1XD3 | Ubiquitin Carboxyl-terminal esterase L3 (UCH-L3) |
|
2D3G | UIM from hepatocyte growth factor-regulated tyrosine kinase substrate (Hrs-UIM) |
|
2FIF | Rab5 GDP/GTP exchange factor |
|
2G45 | Ubiquitin carboxyl-terminal hydrolase 5 |
|
2HTH | Vacuolar protein sorting protein 36 |
|
2IBI | Ubiquitin carboxyl-terminal hydrolase 2 |
|
2OOB | E3 ubiquitin-protein ligase CBL-B |
|
Structures used for simulation setup.
To investigate the effect of binding on the backbone dynamics of ubiquitin, a principal component analysis (PCA) of the backbone atoms of residues 1–70 of the ubiquitin chain was performed. It reveals a functionally relevant “pincer mode” in the first eigenvector (
It corresponds to pincer mode already described in
Projection to the first two PCA-eigenvectors based on the backbone of residues 1–70 of all simulated ensembles. For comparison, the unbound reference ensemble is also plotted in blue. The original xray structures are marked in yellow. Histograms for the projection on the first eigenvectors are plotted above the corresponding plots. PDB codes for the starting structures of the simulations are in the upper left corner of each plot. Capital letters denote the chain identifier.
Like the unbound simulation ensemble, also simulations of bound ubiquitin show considerable conformational variety and in fact show a conformational entropy similar to unbound simulations (
However, while the dynamics of bound ubiquitin ensembles are considerable, specific restrictions can be observed in most of the 11 complexes when considering the main backbone dynamic modes (
In all but one of the bound ensembles, the free energy profile along the “pincer mode” appears to have changed to shift the equilibrium towards either side of the conformational range (
Detail from the xray structure (A) 1nbf (ubiquitin bound to HAUSP) and (B) 1xd3 (ubiquitin bound to UCH-L3). For each structure, the compatible ubiquitin structure is shown in blue, while an incompatible structure that has been fitted to the same position is shown in yellow. Clashes with the binding partner are marked in orange.
The C-terminal tail of ubiquitin, comprising residues 71–76, shows high flexibility in the unbound and most of the bound ensembles to a degree that some parts of it are fully resolved only in four of the eleven experimental structures used for simulation setup (PDB codes 1nbf, 1s1q, 1ubi and 2g45) with three experimental structures (PDB codes 1uzx, 1xd3 and 2ibi) missing only the last residue. Four of these structures (1nbf, 1xd3, 2g45 and 2ibi) are the only ones in this study that show a significantly stronger restriction of dynamics if the C-terminal residues are included in the analysis (
The principal component analysis indicates conformational overlap between bound and unbound ensembles on the level of the dominant collective backbone degrees of freedom. However, PCA as a method is not aimed at discrimination, especially if the amplitude of the differences is small compared to the variation within the ensembles. It is well possible that differences between the ensembles on a more local level are not detected by PCA. To determine differences between multidimensional ensembles, partial least squares discrimination analysis (PLS-DA, cf.
Indeed, using this method, models can be found to almost completely distinguish some of the bound ensembles from the unbound reference ensemble The magnitude of these differences is however significantly smaller than that of the main fluctuation modes of ubiquitin (compare length-scales in
Different bound ensembles (red) and the unbound reference ensemble (blue) have been projected onto the difference vector between these ensembles as determined by PLS-DA.
PLS-DA distinguishes between ensembles both on a global as well as on a local level. Even the systematic difference between two ensembles in e.g. a single side chain rotamer will result in a zero overlap.
While both bound and unbound control ensembles are fully covered by the unbound reference ensemble along the main mode of ubiquitin dynamics (
The histogram-coverage of bound ensembles (red) compared to coverage of unbound control ensembles (blue) after projection of the structures onto the first PCA-eigenvector (
The observed differences correlate well (
To localise differences between bound and unbound ensembles, individual PLS-DA calculations were performed on the conformations of each residue (including side chains) of ubiquitin separately after fitting the backbone of the whole chain.
Only a small number of residues for each complex ensemble show an overlap with the unbound reference ensemble which is significantly below
Overlap has been calculated with the unbound reference ensemble after projection to the difference vector found by PLS-DA on single residues after fitting to the backbone and plotted versus distance from the binding partner. Residues displaying a significant difference in the bound ensemble are labelled.
Again, in none of the cases, a complete distinction between bound and unbound ensembles could be found. Even for the residue displaying the smallest overlap between bound and unbound ensembles (residue His68 in ensemble 1nbf chain C) a small fraction of bound structures can be found in the same conformational region as the unbound ones (
Histogram of the projection of bound (red) and unbound (blue) ensemble onto the difference vector found by PLS-DA for residue His68 of ensemble 1nbf chain C. Out of all 11 complexes studied, this residue shows the smallest overlap between bound and unbound ensembles. The inset shows the corresponding structures from the simulation ensembles.
We compared ensembles of ubiquitin structures from molecular dynamics simulations with and without binding partners aimed at a detailed investigation of the conformational effects of protein binding.
The main collective mode of fluctuation found in unbound ubiquitin is the “pincer mode” which strongly influences the shape of the binding surface (
Employing the partial least squares discrimination analysis method, that specifically aims at identifying differences between ensembles, low amplitude difference modes between bound and unbound ubiquitin ensembles were identified.
The observation of the unbound protein displaying the bound state conformation is often considered indicative of conformational selection (
It is still possible that a portion of the binding events occurs according to an induced fit scenario. An alternative classification of the binding process is based on the inclusion of binding kinetics
An aspect not considered in recently discussed binding models
In general, two effects of binding on the conformational space of the protein can be expected (
The blue ensemble would be that of the unbound protein, the red that of the bound. A sketch of possible free energy profiles fitting the corresponding models is given on the right.
These two effects are not mutually exclusive and indeed in most cases we observe a combination of both effects in the binding behaviour of ubiquitin. In the most extreme cases, all conformations accessible to the unbound protein are restricted, with all the conformations in the complex being the effect of conformational extension. This “conformational shift” corresponds best to the induced fit binding model.
In the case of conformational extension, changes of the energy landscape due to binding allow the protein to access conformations that are energetically unfavourable in the absence of the binding partner. While not generally considered, conformational extension is well compatible with the conformational selection model of binding, as the binding process itself can well take place in the overlap between the bound and unbound states.
Most complexes considered in this study can be described by the scenario of conformational extension combined with conformational restriction, showing a significant overlap between bound and unbound ensembles. Interestingly, also for those complex with near-zero overall overlap, substantial overlap is found between the bound and unbound states on the level of individual residues. Hence, for these complexes, each residue samples states in the unbound state that are found in the bound state, but the probability to find all contact residues in a complex compatible state
The consideration of conformational ensembles is a common feature of modern computational protein docking approaches to account for conformational changes due to binding
From the Protein Data Bank (PDB,
Principal component analysis
Partial least squares regression (PLS) can be used to find a linear model to calculate an external parameter from protein structures. By defining a label of which structures belongs to which class (in this case
If a structural difference between the classes exist, the projection of structures onto this difference vector will make it possible to assign a structure to one or the other class. If it is not possible to completely distinguish structures belonging to the two different classes, the model will still produce the best possible distinction, allowing quantification of the remaining overlap between bound and unbound ensembles. For this, both ensembles are projected onto the difference vector and histograms of the projections are calculated (
The PLS-DA algorithm used in this study produces a model that maximised the difference of the projection of two structures from different classes (bound vs. unbound) while minimising the difference between structures from the same class. Consequently, if more than one structural mode can be used to distinguish the two classes, the resulting model will not necessarily represent both of them, especially if one would result in stronger variation within the classes. While the method can be used to determine whether or not a full distinction between bound and unbound ensembles can be found, additional steps are necessary to fully characterise the structural differences. For this, PLS-DA was performed on sub-groups of atoms (i.e. the backbone as well as each residue including side chain individually) after fitting of the ensemble on the backbone atoms.
Helland's Algorithm
For comparison, both ensembles were sorted into the same set of 100 bins spanning their combined range. The overlap of one ensemble by the other is defined as the normalised sum of the products of the number of structures for each bin. Coverage of one ensemble by another is defined as the fraction of structures from the first ensemble in bins containing a minimum number (50) of structures from the other ensemble.
The stationary bootstrap algorithm
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)