Overview

The shielding phenomenon leads the fluctuations associated with directly observable transitions to dominate the variance of the observable states in many networks, but this rule does not hold universally. The edge importance measure (see Eq 42 in Methods) provides an exact means to evaluate the applicability of stochastic shielding to any model (Markovian, with first-order transitions) by quantifying the effect of suppressing noise along each edge separately. This measure considers the pathwise mean square error between two trajectories: the full stochastic process with all fluctuations included, and an approximate process with a subset of fluctuations excluded. We use this measure to rank the edges in order of importance with respect to the stationary variance of the observable states. Moreover, we show that the stationary variance decomposes into a sum of contributions from each edge. This decomposition is unique and follows from a straightforward calculation that we describe and prove in Theorem 1 in the last subsection of Results. We apply the stochastic shielding method and compute the edge importance measure for the acetylcholine receptor model introduced above and for a set of simple networks (3-state chains) with timescale separation.

Biological example of a bursty process: Nicotinic acetylcholine receptor

The nicotinic acetylcholine receptor is a ligand-gated ion channel and the opening of the channel pore requires the binding of acetylcholine. For low acetylcholine concentration, the nAChR is a classic example of a bursty ion channel. This channel has been described many times in the literature, and we will follow the formulation from Colquhoun and Hawkes [30]. Following Figure 4.1 in their paper, the channel has five states with ten possible transitions between states. The states form a graph with vertices and edges (see Tables 1 and 2). Fig 2A shows the transition state diagram. The channel can be bound to zero, one, or two ACh molecules. When singly or doubly bound the channel may be open or closed, whereas the unbound state is always closed. Table 1 gives the definition of the states and labels each state as open (observable) or closed (unobservable). State 5 (T) is the unbound state (closed), state 4 (AT) is singly bound (with 1 molecule of ACh) and closed, state 3 (A₂T) is doubly bound and closed, state 2 (A₂R) is doubly bound and open, and state 1 (AR) is singly bound and open. The measurement vector M specifies which states are open and which are closed by labeling each state with a 1 or 0, respectively. In this case, M is given by (3) meaning that states 1 and 2 are open/conducting states and states 3, 4, and 5 are closed/non-conducting states. Table 2 gives the definition of the edges and the transition rates. Note that the ten transitions are numbered starting with the pair of transitions connecting states 1 (AR) and 2 (A₂R) and moving clockwise back to state 1; these are reactions 1-8. The last pair of transitions (9 and 10) connect states 4 (AT) and 5 (T). We will write the per capita transition rate for the k^th reaction, with source node i and destination node j, either with a single index denoting the reaction (α_k) or with a double index denoting the source followed by the destination (α_ij). Thus, α₁ and α₂₁ are synonymous.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

(A) Colquhoun & Hawkes’ five-state model for the nicotinic acetylcholine (ACh) receptor [30]. White disks: closed (non-conducting) states (M_j = 0). Gray disks: open (conducting) states (M_j = 1). Nodes 1-5 (large black numbers) are defined in Table 1. Transitions 1-10 (small blue numbers) are defined in Table 2. The opening of the channel requires the binding of acetylcholine. Transitions 2, 6, and 10 are driven by ACh concentration. Transitions 3, 4, 7, 8 are directly observable through a conductance change. (B) Timescale separation (ratio of non-zero eigenvalues of the graph Laplacian) as a function of ACh concentration. (C) Edge importance R_k for k ∈ {1, …, 10} for each edge in the graph as a function of [ACh], see Eq 42. (D) Sample trace of the model exhibiting burstiness of the channel for low agonist concentration, here [ACh] = 0.5μM. (E) Zoomed in version of the burst in (D) labeled by the red arrow.

https://doi.org/10.1371/journal.pcbi.1006206.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Colquhoun & Hawkes’ five-state model for the nicotinic acetylcholine receptor [30], cf. Fig 2A.

Definition of the states and the measurement vector M (normalized conductance). M_i = 1 means that state i is open (conducting/observable) and M_i = 0 means that state i is closed (non-conducting/non-observable).

https://doi.org/10.1371/journal.pcbi.1006206.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Colquhoun & Hawkes’ five-state model for the nicotinic acetylcholine receptor [30], cf. Fig 2A.

Definition of the edges and the transition rates α_ij. The acetylcholine concentration is c ≥ 0. Bold font denotes edges with [ACh]-dependent transition rates.

https://doi.org/10.1371/journal.pcbi.1006206.t002

Burstiness is defined by the observation of isolated single channels opening and closing in bouts [29–31]. Fig 2D shows a sample trace of our model simulation exhibiting burstiness of the channel for low agonist concentration ([ACh] = 0.5μ Mol). (For details on the model simulation, see Numerical Methods, in Methods.) Fig 2E zooms in on the burst in panel D labeled by the red arrow. The distribution of closed intervals shows a mixture of slow and fast timescales, requiring combinations of two or more exponentials with widely separated time constants. These time constants are related to the eigenvalues of the graph Laplacians (see Eq 4, and see Methods for details). The ratio of eigenvalues will be used as a measure of timescale separation. Fig 2B shows the presence of timescale separation at low [ACh] concentrations by plotting the ratios of the eigenvalues {λ₂/λ_j}_{j = 3,4,5}. Significant timescale separation occurs when λ₂/λ_j << 1, or in words, when the two eigenvalues differ by at least one order of magnitude. The graph Laplacian has leading eigenvalue λ₁ = 0. For the acetylcholine receptor, and for the systems we study here, the remaining eigenvalues are real and negative, and are ordered so that , where is the number of states.

We apply the stochastic shielding method to the nAChR model and show that it works well for high acetylcholine concentration, but not in the bursty regime characterized by low ACh concentration. In fact, we see a reversal of edge importance at low agonist levels (see Fig 2C and discussion below). In light of the network filtering effect underlying stochastic shielding, we might naïvely expect that the edges connecting states 2 and 3, and states 1 and 4, should contribute the most to the stationary variance of the observable states (1 and 2), but this is not the case. There is even a regime where the observable edge pair (edges 3 and 4) is only the third most important edge, as defined by our edge importance measure.

Computing the edge importance measure (Eq 42 in Methods), the fraction of the stationary variance contributed by edge k, requires the graph Laplacian L (and its corresponding eigenvalues and eigenvectors), the noise coefficient matrix B (defined below), the stationary mean flux J_k, and the measurement vector M. The graph Laplacian L as a function of ACh concentration c is (4) and matrix B is (5) where J_k = N_tot α_ij π_i(k) is the stationary flux across edge k for a total population of N_tot ion channels, α_ij is the appropriate transition rate of reaction k (Table 2) and ζ_k is the stoichoimetry vector for reaction k. The k^th stoichoimetry vector describes how an individual moves from node i to node j in reaction k. For instance, the first two stoichoimetry vectors are (6) (7) which correspond to transition 1 (an individual moves from state 2 to state 1) and transition 2 (an individual moves from state 1 to state 2), respectively, in Fig 2A. Note that ζ₁ = −ζ₂, and this relationship holds for each edge pair in the ACh transition graph.

The matrix B depends on the equilibrium population distribution . Since is the leading eigenvector of the graph Laplacian L, the equilibrium fraction π_i of the population in state i will change as a function of c (ACh concentration). Lastly, recall that the measurement vector M = (1 1 0 0 0)^⊺ as described in Table 1.

Fig 2C plots the relative edge importance R_k (fraction of the stationary variance contributed by edge k) for each edge k ∈ {1, …, 10} as a function of acetylcholine concentration over the range [ACh] ∈ [10⁻¹, 10²] μMol. At high concentrations, the most important edges are those connecting the doubly bound closed state to the doubly bound open state (edges 3 and 4), that is, the edges along which transitions are directly observable. This situation is consistent with results for Hodgkin-Huxley ion channels and generic Erdos-Renyi random graphs with randomly assigned binary measurement vector [9, 10]. In contrast, the most important edges at low concentrations are those connecting the singly bound state to the doubly bound closed state (edges 5 and 6 in Fig 2A). Although transitions along this edge are not directly observable, they make a greater contribution to the stationary variance of the open state than the opening/closing transitions.

Moreover, we find that edges 5 and 6 have the highest relative importance for low and intermediate concentrations, followed by edges 3,4 and 9,10. Just below a concentration of 10 μMol, the relative importance switches so that edges 3 and 4 become the most important for higher concentrations (≥ 10 μMol). To begin to understand why the edge importance ranking changes for low [ACh], we note that the relative importance depends heavily on state occupancy probability.

As has been previously observed, one of the nodes in the 5-state nAChR model has very low occupancy probability across all agonist concentrations [32]. In particular, states 2 (A₂R) and 5 (T) are the most likely states to be occupied over the range of [ACh] considered. However, state 1 (AR, one of the open states) has very low occupancy probability and hence is rarely visited by the process. As a result, the most likely path between the unbound/closed state 5 (T) and the doubly bound/open state 2 (A₂R) is 5 → 4 → 3 → 2. This means that transitions 7,8 and 1,2 do not happen very often. The stochastic shielding method predicts that these reactions should be important, but if they rarely happen, they contribute little to the stationary variance. Thus, their relative importance as computed by our edge importance measure is very small. Indeed, for all values of [ACh], the equilibrium occupancy probability of state 1, π₁ is ≪ 1. The variance of the open states for a population of N_tot channels at equilibrium is Although the goal of the stochastic shielding approximation is not to change the network topology by eliminating nodes as other authors have suggested [23, 32, 33], when edges are “unimportant” it is natural to consider eliminating them. If all the edges to a node are unimportant, eliminating them would eliminate the node, and in this case the change in stationary variance of the open states would be approximately N_totπ₁(1 − π₁) − 2N_totπ₁π₂, if π₁ is small. (Compare to [32], “Scheme 1”.)

The edge importance measure R_k (for each edge k) provides an intrinsic idea of how many edges could be suppressed in an approximation (whether by suppressing the fluctuations generated by that edge, which is the focus here, or by removing the edge entirely). For the typical operating range of the nAChR, roughly 1-10 μM [ACh], there are three transition pairs with similar edge importance (edge pairs 3,4, 5,6, and 9,10), suggesting that accurate simulations of stochastic effects would require keeping the fluctuations generated by all three of these edge pairs.

The acetylcholine receptor example suggests that the inversion of edge importance is related to timescale separation. In the next subsection, we investigate the edge importance measure in the presence of timescale separation, as well as a combination of sparsely and abundantly populated vertices. We show that edge importance ranking is preserved despite the introduction of arbitrary timescale separation in simple graphs (3-state chains) with per capita transition rates at two distinct timescales. As we will see, a system needs at least three distinct timescales in order to see the method break down. Nevertheless, the edge importance measure remains exact, and informative, for arbitrary networks, and can be used to extend the original stochastic shielding method to systems with timescale separation and bursty behavior.

3-state model with timescale separation

Motivated by the example of the acetylcholine receptor, we systematically study the effects of introducing timescale separation into the simplest nontrivial model to which stochastic shielding applies: the 3-state chain with one observable state (or one pair of observable transitions into and out of the observable state). Specifically, we consider a discrete state, continuous time Markov jump process with N_tot random walkers moving independently on a graph with three nodes. See Fig 3 for an illustration of the graph, and see Methods for general notation and see S1 Supporting Information for a detailed description of the 3-state model. Here we assume that state 3 (black disk) is the observable state, which yields the following measurement vector: M = (0 0 1)^⊺.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. An illustration of the general 3-state chain with per capita transition rates α_k for k = 1, 2, 3, 4.

State 3 (black) is the observable state (or open/conducting state) of the system, and all other states are not observable (or closed/non-conducting states). By convention, we identify α₁ ≡ α₁₂, α₂ ≡ α₂₁, α₃ ≡ α₂₃, and α₄ ≡ α₃₂.

https://doi.org/10.1371/journal.pcbi.1006206.g003

If we think of this model as a simplified ion channel with three states, then the observable state is the open or conducting state of the system, and all other states are closed or non-conducting. There are four directed edges in the graph, and edge k represents a transition from source node i(k) to destination node j(k) which happens at rate α_k (or α_ij, see Methods for details on notation). We focus on the observed process M^⊺N(t) which describes the evolution of the open state, and approximate processes that suppress noise along a subset of the four edges. In particular, we use the following two approximate processes to illustrate how stochastic shielding “usually” works: (i) suppress noise along edge pair 1,2 (and preserve noise along edge pair 3,4) and (ii) suppress noise along edge pair 3,4 (and preserve noise along edge pair 1,2). In most cases (i) is the best approximation; we investigate here whether or not this heuristic holds universally.

The mechanism of stochastic shielding can be readily understood by considering the power spectrum of the observed process M^⊺N(t). The relationship between the power spectrum and the covariance matrix of a stochastic process is well known; the power spectrum is the Fourier transform of its covariance [34]. The stationary covariance C of a discrete state Markov process (such as N described above) is given by Gadgil, et al. [8], and satisfies the Lyapunov Eq 46 (see Methods).

The stationary variance R of the full and approximate observed processes has the following connection to the power spectrum: integrating over the power spectral density (PSD) S(ω) gives the stationary variance. Moreover, since the stationary variance decomposes into a sum of contributions from each edge in the graph (R = ∑_kR_k where R_k is the edge importance measure of edge k given in Eq 42), the power spectrum decomposes as well (S(ω) = ∑_k S_k(ω), see Eq 66). We provide more details on how the power spectrum is obtained in Methods §Numerical Methods.

Fig 4B shows sample trajectories for the full process (denoted by X, black trace) and the two approximations (i) and (ii) described above (denoted by X_3,4 (red trace) and X_1,2 (blue trace), respectively) in the Gaussian (OUP) version of the model. Fig 4A shows the corresponding power spectral contributions for the three processes: S(ω) is the total PSD (shown in black), S_3,4(ω) is the PSD for approximation X_3,4 (red), and S_1,2(ω) is the PSD for approximation X_1,2 (blue). See Methods §Numerical Methods for details on model simulation and calculation of the power spectra.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Stochastic shielding and the power spectrum when all transition rates equal unity.

Panel A shows that the majority of the power comes from the observable edges (red dashed line), as expected from the edge importance measure and the stochastic shielding method. Black line is the total power spectral density (S) for the observed process X, red dashed line is the PSD (S_3,4) for the approximate process X_3,4 with noise from observable edges preserved, blue dashed line is PSD (S_1,2) for the approximation X_1,2 with noise from hidden edges preserved. Panel B shows trajectories (Gaussian version of the model) of the full observed process with all noise sources included (black trace), the approximate process with noise preserved on the observable edges (red), and the approximate process with noise preserved on the hidden edges (blue). The red trace closely follows the black trace, whereas the blue trace only tracks mean behavior and misses most fluctuations; X_3,4 is the best approximation of X, in agreement with the stochastic shielding method.

https://doi.org/10.1371/journal.pcbi.1006206.g004

At all frequencies, the power from the observable edge pair 3,4 predominates, as shown by the red dashed line (S_3,4(ω)) closely following the black line (total PSD). This spectral decomposition agrees with our edge ranking based on edge importance (i.e. edge pair 3,4 contributes the most to the stationary variance), and illustrates why the stochastic shielding method says that the best approximation of the full process is to preserve the noise along edge pair 3,4 and to suppress the noise along edge pair 1,2. Fig 4B illustrates the consequence in the time domain: the red trajectory closely follows the black trajectory, but the blue trajectory only captures a rough approximation of the full process.

However, this situation breaks down and leads to edge importance reversal for certain bursty systems, which we aim to understand in the rest of the paper. In the remainder of this section, we show that edge importance inversion cannot be obtained by taking a 3-state chain and accelerating or decelerating any single edge, pair, or trio of edges with a single parameter (i.e. by introducing two distinct timescales). As we shall see, in order to invert the edge importance as we did in the nAChR example for low agonist concentration, we need to introduce a third timescale. This will be addressed in §Generalized 3-State Model with Timescale Separation.

Two distinct timescales.

Starting with two distinct timescales, we systematically survey all 3-state chains with one, two, or three out of four edge transition rates (α_k) accelerated or decelerated relative to the remaining edges. The graph Laplacian L and matrix B for the 3-state model are given by (8) recalling that the stationary flux along the k^th edge is given by J_k = N_totα_kπ_i(k). The systems we consider satisfy detailed balance, which means that J₁ ≡ J₂ and J₃ ≡ J₄. Matrices L and B will be used to compute the relative edge importance R_k for each edge in all cases described below.

We introduce a parameter α (ranging from 10⁻⁴ to 10⁴) and consider seven different cases: three cases where two transition rates are set to α and the other two rates are 1, and four cases where one rate is set to α and the other three rates are 1. See Table 3 for a detailed description of these cases. In Case 1, transition rates between closed states differ from transition rates between the open and closed states; in Case 2, transition rates into the middle (closed) state differ from transition rates out of the middle state; in Case 3, upward transition rates differ from downward transition rates; in Cases 4-7, one of the four transition rates differs from the other three. We use Eq 42 (see Methods) to compute the relative edge importance R_k for each edge k as a function of α. The relative edge importance is computed with respect to the total variance of the third state.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Detailed description of each case for the 3-state model.

The first seven cases correspond to the chain that has the third state observable (see Fig 3). The last five cases have the middle state as the observable state. Transition rates α_k for k = 1, 2, 3, 4 are given by columns 3-6. The final column shows the characterization of each case into one of six different types determined by their edge importance graph as a function of parameter α (see right most column in Figs 5 and 6).

https://doi.org/10.1371/journal.pcbi.1006206.t003

Fig 5 shows all possible cases of the 3-state chain with state 3 as the open/conducting state and α ranging from 10⁻⁴ to 10⁴. The left column gives the 3-state diagram for each case with transition rates α_k equal to 1 or α, as described in Table 3. The middle column shows the “timescale separation” for each case, defined to be the logarithm of the ratio of the two non-zero eigenvalues (λ₂/λ₃) versus α. As in the nAChR model, significant timescale separation occurs when λ₂/λ₃ ≪ 1, or when the two eigenvalues differ by at least one order of magnitude. The right column shows the relative edge importance R_k versus α for each case.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. All possible cases for the 3-state chain with state 3 conducting.

Left column: 3-state diagram with accelerated/decelerated edges labeled 1 or α where α ∈ [10⁻⁴, 10⁴]. Middle column: logarithm of the ratio of the two non-zero eigenvalues (λ₂/λ₃) versus α. This shows the “timescale separation” (present when λ₂/λ₃ ≪ 1). Right column: relative edge importance R_k versus α.

https://doi.org/10.1371/journal.pcbi.1006206.g005

According to the stochastic shielding method, we expect edge pair 3,4 to have higher relative edge importance than edge pair 1,2. In Fig 5, this is shown by the blue curve (R₃, R₄) lying above the red curve (R₁, R₂) in the edge importance graphs in the right column. This holds for all cases except cases 2 and 6 which show a convergence of the two edge importance curves for large values of α. In these cases we also find significant timescale separation for large α, as shown by widely separated eigenvalues of the graph Laplacian (middle column graphs). For instance, when α = 10⁴, Cases 2 and 6 show that λ₂ differs from λ₃ by four orders of magnitude. However, there are several cases for which we find significant timescale separation but do not see a convergence or reversal of edge importance. Therefore, timescale separation is a necessary but not sufficient condition to see a breakdown of the stochastic shielding phenomenon.

Cases 2 and 6 share a feature that distinguish them from the other cases: although the importance of the hidden edges never exceeds that of the observable edges, they become equally important in the limit α ≫ 1. In both cases, this “fast” rate applies to the transition from the closed states to the open state (2 → 3). In Case 6, the 2 → 3 transition has the only accelerated rate; in Case 2 the 2 → 1 transition is also accelerated. However, in Case 3, which accelerates both the 2 → 3 and the 1 → 2 transitions, the edge importance does not converge. This conundrum will be resolved when we consider the general 3-state model (cf. Eq 22).

3-state model with middle state conducting

For completeness, we may consider the same 3-state chain as in Fig 3, except that we set the middle state (state 2) to be the open/conducting state instead of state 3. The measurement vector in this case is M = (0 1 0)^⊺. See the left column of Fig 6 for an illustration, and note that there are five possible cases to consider. In this version of the 3-state chain, all transitions are observable since each edge connects the conducting state to a closed state, and hence, all edges should be important in terms of stochastic shielding. State 2 no longer acts as a “shield” as it did when state 3 was the conducting state. We expect that the most important edges will either depend on the parameter α or all edges will be equally important in terms of the edge importance measure. We repeat the same analysis as in the previous section and the results are shown in Fig 6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. All possible cases for the 3-state chain with middle state 2 conducting.

In this case there are no hidden transitions, and hence no stochastic shielding effect. Left column: 3-state diagram with accelerated/decelerated edges labeled as 1 or α where α ∈ [10⁻⁴, 10⁴]. Middle column: logarithm of ratio of the two non-zero eigenvalues (λ₂/λ₃) versus α. This shows the timescale separation (present when λ₂/λ₃ ≪ 1). Right column: relative edge importance R_k versus α.

https://doi.org/10.1371/journal.pcbi.1006206.g006

Fig 6 has the same three column format as Fig 5. The left column shows the 3-state diagram with accelerated/decelerated transition rates (1 or α as outlined in Table 3) where again α ∈ [10⁻⁴,10⁴]. The middle column shows timescale separation as defined by the ratio of the two non-zero eigenvalues (λ₂/λ₃) versus α. The right column shows the relative edge importance R_k versus α. In contrast to the previous cases with state 3 conducting, now we see edge importance reversal or convergence in every case. This is what we expect, given that the stochastic shielding method says that all edges are important in this version of the model. However, we find edge importance reversal in Case 10 without corresponding timescale separation since λ₂ and λ₃ differ by less than one order of magnitude in that case.

Generalized 3-state model with timescale separation

We showed above that the presence of two distinct timescales was not sufficient to see an inversion of the edge importance in a 3-state network. However, as we show next, a network exhibiting three separate timescales can lead to edge importance reversal. In order to find examples of inversion, we study an ensemble of 3-state chains with observable state 3 (see Fig 2) with arbitrary transition rates {α₁₂, α₂₁, α₂₃, α₃₂}. We randomly draw the transition rates α_ij independently from a lognormal distribution with a given width w, that is, log(α_ij) is Gaussian distributed with mean zero and standard deviation w. Then we calculate the edge importance for each realization of transition rates for this general 3-state model and look at the instances for which R₁₂ = R₂₁ > R₂₃ = R₃₂. Note that R_ij refers to the importance measure for the edges connecting node i to node j.

For an ensemble of 10⁵ samples with (i.e. ), we find that inversion of the edge importance occurs approximately 9.8% of the time. This observation raises a number of questions. Which factors contribute to inversion of the usual edge importance relation (e.g. timescale separation)? Given an arbitrary set of transition rates, is there a canonical transformation leading to edge importance reversal? Can we obtain an exact expression for the relative contribution of the hidden edges to the stationary variance? The balance of this section addresses these questions.

Fig 7 illustrates the distribution, for this ensemble, of several factors that might be expected to play a role in inverting edge importance. Each panel plots the relative importance of the hidden edges (9) versus factors representing node occupancy, timescale separation, flux distribution, and local timescale difference. Inversion of edge importance occurs when R₁₂ > R₂₃, that is, when η > 1/2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Factors contributing to edge importance reversal.

The relative importance due to the hidden edges, η = R₁₂/(R₁₂ + R₂₃), was calculated for an ensemble of 3-state chains (100,000 samples, see text for details). Relative edge importance is inverted when η > 0.5. Left column shows η plotted versus stationary occupancy probability of node 3 (π₃, panel A), node 2 (π₂, C), and the ratio of nodes 2 to 3 (π₂/π₃, E). The corresponding plot for π₁ appears similar to that for π₃ (not shown). Edge importance can be inverted for any values of π₁ and π₃, but requires π₂ ≲ 1/6. Right column shows η plotted versus timescale separation (ν = λ₃/λ₂, B), relative fraction of flux generated by the hidden edges (ΔJ = (J₁₂ − J₂₃)/(J₁₂ + J₂₃), D), and ratio of relaxation times for isolated 2-state systems corresponding to the hidden versus observable transitions (τ₁₂/τ₂₃, F). Edge importance reversal requires timescale separation (|λ₃| ≳ 15|λ₂| or ν ≳ 15), larger mean flux along the observable edges than the hidden edges (J₂₃ > J₁₂), and faster relaxation along the visible edges than along the hidden edges (τ₁₂ > τ₂₃). None of these conditions alone are sufficient. However, panel G shows η versus the two factors F₁ and F₂ in the exact expression for η, black ‘+’ line is F₁ ⋅ F₂ (see Eq 22).

https://doi.org/10.1371/journal.pcbi.1006206.g007

Node Occupancy: The left column of Fig 7 plots η versus the stationary occupancy probability of each state: π₃ for state 3 (panel A), π₂ for state 2 (panel C), and the ratio π₂/π₃ (panel E). Panel A suggests that edge importance can be inverted for any values of π₃ (mutatis mutandis π₁), but panel C suggests that inversion requires π₂ ≲ 1/6. Moreover, panel E indicates that inversion requires π₂ < π₃ (equivalently, α₂₃ > α₃₂ since π₂/π₃ ≡ α₃₂/α₂₃). Together, these conditions suggest that sparse occupancy of the hidden state directly connected to the observable state (relative to the occupancy of the observable state) contributes to inversion of edge importance. However, this condition alone is not sufficient, as shown by Example A below (see Examples subsection), for which the relative importance due to the hidden edges is η = 0.4132 < 0.5.

We can extract several strict inequalities relating η to properties of the 3-state process. Maximizing η with π₂ fixed, we find (10) Fig 7C shows this inequality is tight (dashed red curve superimposed on the dots matches the upper boundary). In panel E, maximizing η with π₂/π₃ fixed, we observe that (11) (dashed red curve matching boundary), which shows that inversion (η > 0.5) is only possible if π₂ < π₃, or equivalently, α₂₃ > α₃₂. More extreme edge importance inversion requires a more extreme likelihood difference between the observable state and its neighbor or between the transition rates connecting these states.

Timescale Separation: We introduce two different notions of timescale separation. First, we define (12) which is the ratio of the two non-zero eigenvalues of the graph Laplacian L. This quantity is shown in Fig 7B where η is plotted versus ν. (Note ν is the reciprocal of the ratio used to define timescale separation in the previous 3-state model sections with two distinct timescales and discussed in Figs 5 and 6). Large timescale separation, defined via the eigenvalues of the graph Laplacian, occurs when ν ≫ 1. Specifically, Fig 7B shows that edge importance reversal requires timescale separation such that |λ₃| ≳ 15|λ₂| or (ν ≳ 15).

Second, we consider the relaxation time (13) for an isolated 2-state Markov processes with rates α_ij, α_ji between the nodes i, j. The ratio of two such local relaxation times gives an alternative measure of timescale separation within the network. Specifically, consider the two possible 2-state processes in our 3-state model (nodes 1-2 and nodes 2-3). In the first system (between nodes 1 and 2), the eigenvalues of the graph Laplacian are 0 and α₁₂ + α₂₁ = 1/τ₁₂. Likewise, looking at states 2 and 3 as a 2-state Markov process yields eigenvalues 0 and α₂₃ + α₃₂ = 1/τ₂₃. Fig 7F shows the dependence of η on the ratio of the non-zero eigenvalues for these two 2-state systems. Empirically, we see that (14) (dashed red curve in Fig 7F where η is plotted versus τ₁₂/τ₂₃). That is, inversion of the edge importance (η > 0.5) occurs only when equilibration along the hidden edges is slower than along the observable edges (τ₁₂ > τ₂₃).

Stationary Flux Distribution: Recall that the stationary flux along edge k is given by J_k = N_totα_kπ_i(k). We can also represent this term as J_ij, the stationary flux from node i(k) to node j(k) (see §Notation in Methods). Here we define (15) which is the relative fraction of the stationary flux generated by the hidden edges. In Fig 7D, we observe that the upper boundary is given by (16) which says that edge importance reversal (η > 0.5) requires larger mean flux along the observable edges than along the hidden edges. In other words, the system needs to satisfy ΔJ < 0 or J₁₂ < J₂₃.

Exact expression for η

Reproducing edge importance reversal in 3-state chain models is advantageous because such simple Markov models can be analyzed more completely than models with greater numbers of states [23]. Fortunately, explicit expressions may be derived for the eigenvalues and eigenvectors of the 3-state chain model which allows direct calculation of η, the fraction of the stationary variance generated by the hidden edges (see S1 Supporting Information §Explicit calculation of η for detailed derivation): (22) where Tr[L] ≡ ∑_i L_ii is the trace of L.

The fraction in Eq 22 is a product of two factors (denoted by F₁ and F₂ and shown in Fig 7G for the ensemble). The first factor F₁ is the ratio of the speed of transition from hidden state 2 to hidden state 1 (α₂₁) to the sum of the transition rates between states 1 and 2. Equivalently, this is the proportion of time spent in hidden state 1 relative to hidden state 2. F₁ approaches 1 as α₁₂ decreases, which only occurs if condition 2 holds (α₁₂ ≪ α₂₁). The second factor F₂ is the ratio of the opening transition rate (α₂₃) to the sum of the four rates. This factor is large if and only if the opening rate is much faster than the other rates, and this is exactly condition 1 (α₂₃ ≫ max{α₁₂, α₂₁, α₃₂}). Together these two conditions bring about a reversal of edge importance (η > 0.5) in this simple scenario.

While the exact formula for the relative edge importance (22) applies only for the 3-state chain model considered here, we anticipate that analogous results may hold for more general Markov processes. We consider this question further in §Discussion.

Edge importance and power spectra

Additional insight into the error arising from different noise-suppressing approximations can be obtained by examining the power spectral density distributions of the true and approximate processes. Recalling Fig 4A in the case α_ij ≡ 1, the power spectra for the full process with all noise sources included (S, black curve) and the approximate process with hidden edge flux noise suppressed (S_3,4, red curve) are very similar, with an order of magnitude less power arising from the hidden edges at all frequencies. In contrast, Fig 8A shows the power spectra for the 3-timescale model. In particular, it shows that at low frequencies, the power spectrum for the approximate process with visible edge flux noise suppressed (S_1,2, blue curve) is very similar to the PSD for the full process, but that the blue and red curves cross at an intermediate frequency (ω ≈ 3) so the red curve dominates at high frequencies.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Stochastic shielding and the power spectrum in the 3-state chain for the case where edge importance is reversed (α = 10 shown here).

Panel A shows that the majority of the power comes from the hidden edges (blue) for low frequencies, but the red and blue curves cross at ω ≈ 3, so for high frequencies the majority of the power comes from the observable edges (red). This switch in dominant spectral contributions is reflected in the corresponding Gaussian model trajectories shown in Panel B with the blue trace closely following the black trace, and the red trace deviating from the black trace. This shows that X_1,2 best approximates the full process X in this case (whereas X_3,4 is the best approximation in the uniform transition rate case shown in Fig 4).

https://doi.org/10.1371/journal.pcbi.1006206.g008

The change in power spectral contributions is also reflected in model simulations (see Numerical Methods for details on simulations). Fig 8B illustrates sample trajectories for the three processes described above: full process X (black), approximation X_3,4 (red), and approximation X_1,2 (blue) where α = 10. Comparing this edge importance reversal case to the base case shown in Fig 4B, we see that the blue trajectory (instead of the red one) closely follows the black trajectory. Hence, X_1,2 is the better approximation to the full process in this case.

Thus, the edge importance reversal observed under the combined conditions α₁₂ ≪ α₂₁ and α₂₃ ≫ max(α₁₂, α₂₁, α₃₂) may be understood as resulting from enhancement of the noise power contribution from the hidden edges at low frequencies, as well as the small amplitude of the full process’ power spectrum at high frequencies.

We see a similar mechanism at work in the 5-state acetylcholine receptor model in the low-[ACh] regime (where a hidden edge becomes more important than a visible edge) as opposed to the high-[ACh] regime, in which the usual edge importance ordering is observed. Figs 9 and 10 show the power spectrum and Gaussian model trajectories in the high-[ACh] and low-[ACh] regimes, respectively. Here we have similar notation to the 3-state cases: X (black) is the full observed process (Gaussian version) with all sources of noise included and X_i,j is the approximate process that preserves noise on edge pair i, j but suppresses noise on all other edges. In particular, we focus on the red trace (X_3,4, noise preserved on visible edges 3,4) and the blue trace (X_5,6 noise preserved on hidden edges 5,6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Stochastic shielding and the power spectrum in the acetylcholine receptor model.

For a high concentration ([ACh] = 100 μM shown here), edge importance is not reversed. Panel A shows that the majority of the power comes from the visible edges at all frequencies (S_3,4, red trace), in agreement with the usual edge importance ranking. The corresponding Gaussian model trajectories shown in Panel B illustrate that X_3,4 is the best approximation to X, whereas the other approximation only captures average behavior.

https://doi.org/10.1371/journal.pcbi.1006206.g009

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. Stochastic shielding and the power spectrum in the acetylcholine receptor model.

For a low concentration ([ACh] = 0.5 μM shown here), edge importance is reversed. Panel A shows that the majority of the power comes from the hidden edges at low frequencies (S_5,6, blue trace) and from the visible edges at high frequencies (S_3,4, red trace), just as we saw in the 3-state model under the case of edge importance reversal. The corresponding Gaussian model trajectories shown in Panel B illustrate that X_5,6 (blue) is the best approximation to X, although it misses some the fluctuations captured by X_3,4 (red).

https://doi.org/10.1371/journal.pcbi.1006206.g010

The usual edge ordering via the edge importance measure for high [ACh] ranks edge pair 3,4 the most important, followed by edges 5,6, then 9,10 (the last two edge pairs have relative importance close to 0 and make the two lowest spectral contributions); See Fig 2. Fig 9A shows that for [ACh] = 100 μM, most of the power is attributable to the observable edge pair 3,4, and this agrees with the edge importance ranking. Model trajectories in panel B illustrate that X_3,4 is the best approximation of the full process X and that the other approximations at best only capture the mean behavior of the system.

In the low-[ACh] case shown in Fig 10 ([ACh] = 0.5 μM), however, we see the crossing of the top blue and red power spectral density curves at an intermediate frequency (ω ≈ 2). As in the 3-state case, this indicates a reversal of edge importance whereby now the hidden edge pair 5,6 contributes the most to the stationary variance of the observable process. Again, this change in spectral contributions is reflected in model trajectories shown in panel B. We see that the blue curve X_5,6 closely follows the full process X, and is the best approximation in this case, but the blue curve misses some of the fluctuations captured by the red curve X_3,4 even though the red curve clearly deviates from the other two processes.

Edge importance and decomposition of the stationary observed variance for a general first-order transition network

Gadgil et al. showed rigorously that the time evolution of the second moments of a discrete population evolving as a first-order reaction network system can be represented explicitly in terms of the eigenvalues and eigenvectors of the matrix that governs the evolution of the mean population dynamics [8]. We apply their general results to the specific example of a first-order transition network in two ways. First, we use the spectral decomposition of the stationary variance to establish our main stochastic shielding result. Second, their result on time varying systems allows us to obtain the decomposition of the power spectrum in terms of the eigenvalue spectral decomposition, shown in Eqs 64–66.

Consider an arbitrary first-order reaction network with graph Laplacian L and matrix B satisfying Eqs 32–36 (see Methods). The fact that the stationary covariance matrix decomposes into a sum of contributions from each edge in the graph follows from a straightforward calculation that we describe in Lemma 1 and Theorem 1. We defer the proof of Lemma 1 to §Methods, below.

Definition 1 Let X denote the set of n × n real matrices C such that for all j = 1, …, n, . Let Y = {C ∈ X | C^⊺ = C}.

Lemma 1 Let L be an n × n real valued matrix with L_ij ≥ 0 for i ≠ j, and L_ii = −∑_{i,i ≠ j} L_ij (so that ) for j = 1, …, n, and satisfying dim(ker(L)) = 1, and with a null eigenvector Lv = 0 satisfying v_i ≥ 0 for i = 1, …, n. Then for any F ∈ Y, the equation (27) has a unique solution C ∈ Y.

Theorem 1 For an arbitrary first-order reaction network with graph Laplacian L and matrix B satisfying Eqs 32–36, there is a unique linear decomposition of the stationary covariance matrix C as a sum of contributions from each edge: (28) (29)

Proof 1 Proof of Theorem 1. Consider a first-order reaction network defined by graph Laplacian L and matrix B, satisfying Eqs 32–36. We want to solve the Lyapunov equation (30) for matrix C. Note that L satisfies the conditions in Lemma 1, and BB^⊺ ∈ Y since BB^⊺ is an n × n real symmetric matrix with columns that sum to zero. Then by Lemma 1, Eq 30 has a unique solution C ∈ Y. By replacing F with BB^⊺ in the proof of Lemma 1, we see that the unique solution is (31) since all eigenvalues of L have negative real part (except for the Perron-Frobenius eigenvalue λ₁ ≡ 0), , and B^⊺u₁ = 0.

Since BB^⊺ can be written as a sum of , we can repeat the calculation above to get Eq 29 for each k separately. The integral in Eq 29 holds for all k since the k^th stoichiometry vector ζ_k appearing in the k^th column of B is orthogonal to the steady state eigenvector. Therefore, C decomposes into a sum over the C_k terms, and Eq 28 holds.

The decomposition in Theorem 1 allows us to rank each edge in the network in terms of its contribution to the stationary variance of any given node, which we call its “importance” relative to that node. In the case of a single open or conducting node, we refer simply to the edge importance. Moreover, the decomposition allows us to quantify the accuracy of the stochastic shielding approximation with respect to the population process projected onto individual nodes. The decomposition given by Theorem 1 is exact regardless of timescale separation or node sparsity.

Notation

We begin with a directed graph with edge weights α_ij ≥ 0 representing a population of N_tot individuals moving randomly and independently among n states () along m edges {i(k)→j(k)}_1≤k≤m, with per capita transition rates {α_k}_1≤k≤m. We emphasize that edge k is the unique directed edge connecting source node i(k) to destination node j(k). The n × 1 stoichiometry vector ζ_k corresponding to edge k is defined such that ζ_k(i) = −1 and ζ_k(j) = +1, otherwise ζ_k(l) = 0; these vectors represent the effect of a transition along edge k. We use this notation to be consistent with the edge importance formula in the next subsection which is a sum of contributions to the variance of the observable state coming from each edge. Also, note that we will write the per capita transition rates either with double indexing denoting the source and destination nodes (α_ij) or with a single index denoting the reaction (α_k).

We represent the population state at time t with an integer-valued vector N(t) = (N₁(t), …, N_n(t))^⊺, where N_i(t) ≥ 0 and for all t. In other words, N(t) is a discrete state continuous time Markov process. Such processes are ubiquitous in biology [1].

We denote by M a measurement vector indicating a direction in the state space along which there is an observable of interest. For instance, M_i ∈ {0, 1} could denote the conducting state ({closed, open}) in a multi-state ion channel model. We denote the observed process by Y(t) = M^⊺N(t). The remainder of our set up follows standard nomenclature for representing a population process on a graph [8, 40–42].

Let L be the Laplacian of graph which is the n × n matrix defined by L = (A − D)^⊺ where A is the weighted adjacency matrix and D is the diagonal matrix of node out-degrees. Specifically, the entries in A are A_ij = α_ij(k) = α_k ≥ 0 and the diagonal entries in D are for i ∈ {1, …, n}. Note that L = Q^⊺ where Q is the standard generator matrix of the Markov process. It follows that, for any vector , L satisfies the following equation (32) The stoichiometry vector ζ_k is a difference of two standard unit vectors, ζ_k = e_j(k) − e_i(k). Although we do not assume that the graph Laplacian L must be a symmetric matrix, we do assume that the stationary system satisfies detailed balance, and that L has only real eigenvalues. Moreover, we assume that L has an expansion into real-valued biorthogonal eigentriples (w_λ, λ, v_λ) such that (33) (34) (35) We further assume that is connected and the process is irreducible. The Perron-Frobenius theory guarantees the existence of a unique null eigenvector with nonnegative components summing to unity, corresponding to the stationary distribution on the graph. We denote the stationary probability vector and the stationary mean flux along edge k by J_k = N_totα_kπ_i(k).

Let B be the n × m matrix defined such that (36) In other words, the k^th column of B is given by the square root of the stationary flux J_k multiplied by the stoichoimetry vector ζ_k. We can express B as a sum of matrices (37) where all the entries of B_k are zero except for the k^th column. Moreover, we will exploit the fact that the product BB^⊺ can be represented with a similar sum (38) This product appears on the right hand side of the Lyapunov equation (see Eq 46 below) and its decomposition into the above sum is a key factor in establishing the decomposition of the stationary variance into a sum over the edges.

Computations were done either by hand, or using Matlab or Mathematica.

Summary of stochastic shielding in the Langevin case

In the Langevin approximation for a time homogeneous first-order transition network, the population fraction occupying states 1, …, n is a vector satisfying (39) where L, the graph Laplacian, captures the mean flux along each directed edge . The matrix B_k gives the effects of fluctuations ξ_k around the mean flux along the k^th edge. The noise terms are independent, white and Gaussian, with 〈ξ_k(t)ξ_k′(t′)〉 = δ_kk′ δ(t − t′), one for each directed edge. Given an observable of interest, represented by a vector , the stochastic shielding approximation consists in finding a partition of the edge set, , into edges of primary importance () and secondary importance () such that and, at the same time (stationary variances), where Y is the approximate population vector satisfying (40) The noise samples ξ_k for are identical in the full and approximate models. Neglecting the noise forcing along the edges of secondary importance causes a pathwise discrepancy U(t) = Y(t) − X(t) that satisfies (41) The stochastic shielding effect consists in suppression of the resulting fluctuations in the observable process M^⊺U(t) due to the filtering effects of the network—hence “stochastic shielding”.

The (stationary) mean squared pathwise approximation error can be written exactly as a sum of contributions R_k from each directed edge neglected in the approximation, . This error is small compared to . We call R_k the importance of the k^th directed edge (defined in the next section). As we show below, the decomposition holds exactly not only for the Langevin process but for the discrete population process as well.

Edge importance measure

The general formula for the edge importance measure is as follows. For an arbitrary stationary population process N(t) satisfying detailed balance on a graph with n nodes, m edges, and measurement vector M (defining the observable states), is the stationary variance of the observable states where (42) In this formula, λ_n ≤ λ_n−1 ≤ ⋯ ≤ λ₂ < 0 are the nontrivial eigenvalues of the graph Laplacian L (which always has λ₁ ≡ 0); v_i and w_i are the right and left eigenvectors of L, respectively. Here and henceforth, R_k is normalized to the variance due to a single random walker by dividing out N_tot.

The stationary variance R is related to the power spectral density (PSD) S(ω) of the observed process M^⊺N. From the Wiener-Khinchin theorem, integrating the PSD gives the stationary variance: . Moreover, since the stationary variance decomposes into a sum of contributions from each edge in the graph, the power spectral density decomposes as well. By introducing (43) we define a power-spectral edge importance such that the integral of S_k(ω), the power spectral density for the observed process with noise suppressed everywhere except edge k, gives the edge importance corresponding to edge k.

To see this, note that the power spectral density of the observed process is (44) (45) provided ω > 0. For more details, see §Numerical Methods: Calculation of power spectra, below. We can use this power spectral decomposition to explore how the spectral contributions differ between the typical cases (where edge importance ranking agrees with the stochastic shielding method) and in the edge importance reversal cases.

Lyapunov equation

The Perron-Frobenius null eigenvector, suitably normalized, gives the stationary probability vector of Markov process N(t). Snapshots of the process N(t), taken under stationary conditions, are multinomial with parameters , so the covariance matrix C is known. In particular, each diagonal entry in C is the variance of state i, C_ii = N_totπ_i(1 − π_i), and each off-diagonal entry in C is the covariance of states i and j, C_ij = −N_totπ_iπ_j for i ≠ j.

The stationary covariance matrix C satisfies Lyapunov’s equation (a special case of Sylvester’s equation) [43] (46) The fact that C satisfies Eq 46 above is widely known for linear Gaussian processes such as multivariate Ornstein-Uhlenbeck processes [34], but it also holds for discrete state population processes in which the transition rates are linear functions of the population at each node, i.e. first-order transition networks, such as those we consider here (see [8, 44]).

Our system is an important special case of the general first-order reaction network presented in [8]; we only consider conversion type reactions (denoted by k^con in [8]). For our system P_i represents , summed over all identical λ if they occur with multiplicity (we both assume semisimple eigenvalue spectra). The following parameters in [8] are zero for our system: C(i, k, l), k^cat, k^s, and k^d. This simplifies Equation 50 in [8] (representing the variance of the l^th reactant in the network) and is equivalent to our edge importance measure (Eq 42). However, to our knowledge, we are the first to describe the unique decomposition of the stationary variance into a sum of contributions from each edge in the network, and [9, 10] were the first to propose the stochastic shielding approximation and justify it based on this decomposition.

The Lyapunov equation has also been used in the context of stochastic gene networks under the name of “linear noise approximation” [45, 46]; in particular [45] (pg. 1, ¶5) further cites Eqs 3.46 and 6.115 in Risken [47] for additional details. See also [48] Supporting Information §4. For the linear networks we consider here, the equation is exact.

Proof of Lemma 1

We restate the lemma for the reader’s convenience. Recall from Definition 1 that Y is the space of n × n symmetric matrices with columns (and rows) summing to zero.

Lemma 1 (restated) Let L be an n × n real valued matrix with L_ij ≥ 0 for i ≠ j, and L_ii = −∑_{i,i ≠ j} L_ij (so that ) for j = 1, …, n, and satisfying dim(ker(L)) = 1, and with a null eigenvector Lv = 0 satisfying v_i ≥ 0 for i = 1, …, n. Then for any F ∈ Y, the equation has a unique solution C ∈ Y.

Proof 2 Proof of Lemma 1. Given L ∈ X, define the linear operator A by A: C → LC + CL^⊺. First, we show that A: Y → Y. If C ∈ Y then for all j = 1, …, n, (47) because each sum over i is zero, by assumption. Moreover, (LC + CL^⊺)^⊺ = LC + CL^⊺. Therefore LC + CL^⊺ ∈ Y whenever C ∈ Y, so A maps Y into itself.

By the Fredholm alternative(cf. [49], Theorem 2.27), A(C) = F has a unique inverse for F ∈ Y provided F is in the range of A and the homogeneous equation A(C) = 0 has only the trivial solution C = 0.

Let C₀ ∈ Y be a solution of the homogeneous equation, LC₀ + C₀L^⊺ = 0. Because C₀ ∈ Y is symmetric and the nullspace of L is one dimensional, C₀ must have the form C₀ = (c₁v|⋯|c_nv) for constants c₁, …, c_n. However, the columns of C₀ must sum to zero, and , therefore c₁ = … = c_n = 0, hence C₀ = 0.

To see that F is in the range of A, we construct an explicit solution as follows: (48) and we show that this integral is well defined whenever F ∈ Y. To see this, first note that if all eigenvalues of L have negative real part, then (49) where (50) (51) and the solution in Eq 48 follows from the fundamental theorem of calculus.

It remains to show that the integral in Eq 48 is well defined whenever F ∈ Y. Assuming detailed balance, a unique null space, and that L is diagonalizable, we have that all eigenvalues of L are negative (and real) except λ₁ ≡ 0, and we can write (52) (53) Then (54) (55) (56) (57) (58) (59) (60) The underlined expressions in parentheses are all zero because the columns (and rows since F is a symmetric matrix) of F sum to zero by assumption; is orthogonal to every column of F and u₁ is orthogonal to every row of F and so and Fu₁ = 0. Thus, the integral in Eq 54 is finite and Eq 60 gives an explicit expression for it.

Numerical methods