Chapter 42Analyzing Protein Structures Using an Ensemble Representation

This chapter discusses the mathematical basis and consequences of comparing an ensemble of unfolded-state molecules versus a given reference structure, and then presents the biological meaning of the structure comparison of unfolded (or intrinsically disordered) proteins at the ensemble level [1].

We address an important problem: Suppose an ensemble of structures is computed for a protein in an unfolded or disordered state (chapters 7 and 39), or even for a unimodal native state (chapters 15–18). Given a reference structure (for example, a crystal structure), or even multiple structures of exchanging states (chapter 27), how do we compare these reference structures to the ensemble? The answer will be surprising to many readers.

Mathematical Results

Terminology

Let Ã\tilde{A} and B̃\tilde{B} be the distance matrices of structures AA and BB , respectively. The distance root-mean-square deviation between distance matrices Ã\tilde{A} and B̃\tilde{B} is defined by

dRMS(Ã,B̃)=1n(n−1)∑i=1n∑j=1n(Aij−Bij)2,(42.1)dRMS(\tilde{A}, \tilde{B}) = \sqrt{\frac{1}{n(n-1)} \sum_{i=1}^n \sum_{j=1}^n (A_{ij} - B_{ij})^2}, \quad (42.1) where Aij=∥r→i−r→j∥A_{ij} = \|\vec{r}_i - \vec{r}_j\| denotes the Euclidean distance between atoms ii and jj in structure AA .

The Cartesian coordinate-based root-mean-square deviation between structures AA and BB is defined by

RMSD(A,B)=1n∑i=1n∥r→iA−r→iB∥2,(42.2)RMSD(A, B) = \sqrt{\frac{1}{n} \sum_{i=1}^n \|\vec{r}_i^A - \vec{r}_i^B\|^2}, \quad (42.2) where r→iA\vec{r}_i^A denotes the Cartesian coordinates of the ii th atom in structure AA , and ∥⋅∥\|\cdot\| denotes the Euclidean l2l_2 -norm, that is, ∥a→∥=∥(a1,a2,…,aJ)∥=∑i=1Jai2\|\vec{a}\| = \|(a_1, a_2, \dots, a_J)\| = \sqrt{\sum_{i=1}^J a_i^2} .

42.1.2 Results

Given an ensemble of structures A={A1,A2,…,Ak}A = \{A^1, A^2, \dots, A^k\} , k=1,…,Nk = 1, \dots, N , and a reference structure BB , we have the following mathematical results:

dRMS(⟨Ã⟩N,B̃)≤⟨dRMS(Ãk,B̃)⟩N,(42.3)dRMS(\langle \tilde{A} \rangle_N, \tilde{B}) \leq \langle dRMS(\tilde{A}^k, \tilde{B}) \rangle_N, \qquad(42.3)

RMSD(⟨A⟩N,B)≤⟨RMSD(Ak,B)⟩N,(42.4)RMSD(\langle A \rangle_N, B) \leq \langle RMSD(A^k, B) \rangle_N, \qquad(42.4) where AkA^k denotes the kk th structure in the ensemble AA , and Ã\tilde{A} denotes the distance matrix of structure AA (the same definitions for Ãk,B̃\tilde{A}^k, \tilde{B}), and ⟨⋅⟩N\langle \cdot \rangle_N stands for the ensemble average over all NN structures in the ensemble.

42.1.3 Brief Proof

Here we give an intuition of a proof sketch. See the Results section in Zagrovic and Pande [1] for a rigorous proof.

A metric space is a set XX with a distance function d:X×X→ℝd : X \times X \rightarrow \mathbb{R} , such that for every x,y,z∈Xx, y, z \in X ,

  1. d(x,y)≥0d(x, y) \geq 0 with equality if and only if x=yx = y (nonnegativity);
  2. d(x,y)=d(y,x)d(x, y) = d(y, x) (symmetry);
  3. d(x,z)≤d(x,y)+d(y,z)d(x, z) \leq d(x, y) + d(y, z) (triangle inequality).

By definition of dRMS, we can easily construct a metric space with distance function:

dRMS:ℝ3n×ℝ3n→ℝ.dRMS : \mathbb{R}^{3n} \times \mathbb{R}^{3n} \rightarrow \mathbb{R}.

Now consider a simple case in this metric space, as shown in figure 42.1. Suppose we have an ensemble consisting of only two points, aa and cc , and a reference point bb . Let mm be the midpoint between aa and cc , then mm stands for the mean of the ensemble. Extend bmbm to b′b' such that |bm|=|b′m||bm| = |b'm| . By triangle inequality, we obtain |b′c|+|bc|≥|bb′||b'c| + |bc| \geq |bb'| . Since |bb′|=|b′m|+|bm|=2|bm||bb'| = |b'm| + |bm| = 2|bm| and |b′c|=|ab||b'c| = |ab| , we have |bm|≤(|ab|+|bc|)/2|bm| \leq (|ab| + |bc|)/2 . Since |bm||bm| stands for the distance between the mean of the ensemble and the reference—that is, dRMS(⟨Ã⟩N,B̃)dRMS(\langle \tilde{A} \rangle_N, \tilde{B}) —and (|ab|+|bc|)/2(|ab| + |bc|)/2 stands for the average over all distances between every point in the ensemble and the reference point—that is, ⟨dRMS(Ãk,B̃)⟩N\langle dRMS(\tilde{A}^k, \tilde{B}) \rangle_N —we thus have

Figure 42.1 An illustration of the intuition. Credit: Michael Zeng and Bruce Donald.

dRMS(⟨Ã⟩N,B̃)≤⟨dRMS(Ãk,B̃)⟩N.dRMS(\langle \tilde{A} \rangle_N, \tilde{B}) \leq \langle dRMS(\tilde{A}^k, \tilde{B}) \rangle_N.

It is easy to extend this intuition to a general case with ensemble size of more than 2.

Biological Significance

It has been observed that the mean unfolded-state Cα−CαC^\alpha - C^\alpha distance matrices of an α\alpha -helix are close to its corresponding native-state distance matrix. To exclude the effect of mathematical averaging, two tests of the α\alpha -helical protein villin were performed in Zagrovic and Pande [1]:

Test 1 Use the members of the unfolded-state ensemble as reference structures instead of the native structure, and compare the dRMSdRMS (or RMSD) distance between the ensemble mean and each reference structure.

Results The native-state distance matrix is closer to the mean unfolded-state matrix than that of any member in the ensemble.

Test 2 Use native structures of other, unrelated proteins as reference structures, and perform a similar comparison.

Results The mean unfolded-state distance matrix is more similar to the corresponding native structure of the unfolded protein than any other native protein.

Reference

[1] B. Zagrovic and V. S. Pande. How does averaging affect protein structure comparison on the ensemble level? Biophysical Journal 87;4 (2004 Oct):2240–2246.