Abstract
The HIV1 virus can enter a dormant state and become inactive, which reduces accessibility by antiviral drugs. We approach this latency problem from an unconventional point of view, with the focus on understanding how intrinsic chemical noise (copy number fluctuations of the Tat protein) can be used to assist the activation process of the latent virus. Several phase diagrams have been constructed in order to visualize in which regions of the parameter space noise can drive the activation process. Essential to the study is the use of a hyperbolic coordinate system, which greatly facilitates quantification of how the various reaction rate combinations shape the noise behavior of the Tat protein feedback system. We have designed a mathematical manual of how to approach the problem of activation quantitatively, and introduce the notion of an “operating point” of the virus. For both noisefree and noisebased strategies we show how operating point offsets induce changes in the number of Tat molecules. The major result of the analysis is that for every noisefree strategy there is a noisebased strategy that requires lower dosage, but achieves the same antilatency effect. It appears that the noisebased activation is advantageous for every operating point.
Introduction
HIV1 latency is a serious problem that prevents eradication of the virus
Human immunodeficiency virus 1 (HIV1), initially reported nearly 30 years ago, represents a major global health problem with millions of people infected [1,2]. One of the biggest problems with HIV1 is that the virus can enter a dormant state and effectively “hide” from drug cocktails that are therapeutically administered. Several reviews have been written on the subject [313]. Latent viral reservoirs can persist for many years. Once a therapy is interrupted, the latent reservoirs remain as the source of eventually renewed infection. This behaviour has been identified as the key problem in eradicating HIV1.
There are many reservoirs, e.g., cell types, in the body that can harbor the latent virus. CD4+ Tcells have been identified as one of the largest pools of the viral DNA. This is also one of the best characterized reservoirs [5]. Once a Tcell is infected, it can either become activated and produce new virus particles that will infect other cells, or it can enter an inactive state. In the inactive (latent) state the transcription of the viral DNA is silenced, despite the fact that the viral RNA has been reversely transcribed and inserted into the host DNA.
The process of entry into the latent state is rather complex since it is controlled by a sizeable number of processes which need to occur nearly at the same time [810,13]. There is a relatively low percentage of latently infected Tcells, being roughly one per 10^{6}, though the frequency can be lower [13]. This also suggests that once the latent state is established it remains very stable, and spontaneous activation events are very rare. In order to activate a latent cell, the inactive cellular processes need to be reactivated, which is not likely to occur simultaneously. A large number of studies have been performed in the past with the goal to find a way of reliably activating the latent virus, as reviewed in [313].
The mechanisms responsible for the maintenance of the HIV1 latency work mostly at the molecular level. Processes such as chromatin control, a shortage of host transcription factors that initiate transcription, the presence of molecules that slow down (or even block) polymerase elongation, transcriptional interference, DNA methylation, and insufficient transport of viral mRNA from the nucleus to the cytoplasm are some prominent examples. All known mechanisms have been reviewed in detail, for example in [810,13].
A small number of studies focused on the role of chemical (intrinsic) noise in establishing the latent state [1417]. A generic conclusion extracted from these studies is that noise is detrimental for the latency decision. Noise comes from fluctuations in copy numbers of the proteins that are involved in the latency control. The level of noise in the system is partially modulated by host factors (proteins that are present in healthy cells) and partially with virus specific factors (proteins encoded by viral genetic machinery). Noise makes the latency decision stochastic (unpredictable) and the literature suggests that the noise driven inactivation occurs spontaneously.
In the present study the focus is placed on understanding how noise can be manipulated to reverse the latency decision, i.e., to drive the activation process. This idea is fully in line with several previous findings and suggestions in the literature. For example, it has been clearly appreciated that noise plays a role in various cellular decision making processes [18,19], but the idea to use noise to steer cell fate decisions has never been seriously explored. It has been argued, merely on a general basis, that one should strive to control noise better in order to steer cell fate decisions [19]. In the context of the HIV activation it has been found that altering noise of the HIV genetic machinery can bias virus decision making towards productive replication or latency [20].
In our study the focus is on finding ways to manipulate noise in order to specifically facilitate the productive activation, which could become the foundation of medical treatment strategies in the future. To our knowledge, this way of approaching the eradication of the latent virus has not yet been much discussed in the literature.
The first goal of this study is to identify regions in parameter space where noise greatly influences the dynamics. We investigate fluctuations of copy numbers of several key proteins that are involved in the maintenance of the latent state of infected cells. Based on these insights, the second goal is to identify suitable medical strategies that can be used to harvest noise in order to achieve more efficient activation. These intuitive considerations will be addressed in a quantitative way by means of a rigorous mathematics.
The Tat feedback loop is a main source of noise driving the activation of the latent virus
The Tat feedback loop has been identified as an important part of the HIV gene expression machinery, reviewed in [35,7] and others. The main biological function of the Tat loop is to accelerate viral RNA production which further increases the number of viral particles in the infected cell. The presence of Tat molecules in the cell increases the transcription rate by roughly two orders of magnitude. This normally leads to cell lysis and continued infection of other cells. A lack of Tat molecules in latently infected cells has been identified as a strong barrier to activation of the latent cell [810,13].
Interestingly, the loop appears to be important for noise driven entry into the latent state [1417]. If viral DNA integrates into regions of high basal transcription then noise does not play such a big role. However, if it integrates in regions of low basal transcription, then the positive feedback loop can amplify noise effectively and produce bursts of activity. In such a way the fate of the cell is determined in a stochastic manner, as shown in [14,15]. Such behavior splits the typical low transcription isogenic cell population into two groups. In the first group the Tat feedback loop produces new Tat particles (the “on” state). In the second group the Tat feedback loop is inactivated (the “off” state). This phenomenon was termed the phenotypic bifurcation (PheB). A series of carefully designed experiments were performed [14] to show that such behavior is indeed a product of intracellular noise, and is independent from extracellular perturbations. The possibility of a spontaneous latency decision caused by uncontrollable fluctuations of the Tat copy number has been clearly demonstrated.
Current antilatency treatments need to be improved
Even though several antilatency drugs have been investigated, it is not clear how to administer them efficiently. For example, there is no consensus whether they should be administered aggressively (all at once), by repeated injection of smaller amounts, or constantly administered over a longer time period [4,5,21,22].
It is generally agreed upon that due to potential side effects of the treatment (e.g. toxicity), minimal dosages are preferable. A convincing argument for administering Minimal dosages is that too large amounts of the activation drug might release the virus beyond control, such that the usual highly active antiretroviral therapy (HAART) cannot contain it any longer [21]. Moreover, there are viral reservoirs, e.g., the brain, that are not easily accessible by HAART. In such organs it is very important to activate the virus gradually.
Clinical studies show that a global Tcell activation cannot eradicate the virus. Instead, it causes unwanted side effects [22]. An additional problem is that many of the activating agents are generic to a wide range of gene regulation processes; administering them is expected to show severe toxicity in the host. Thus the reduction of the dosage of antilatency drugs appears desirable. In this context, noisedriven activation could be highly beneficial.
Intrinsic fluctuations in protein copy numbers could be used to achieve more efficient activation
A therapy can be envisioned where several types of drugs work in synergy with the specific aim of moving the operating point of the virus into a noisy region, where the frequency of spontaneous virus activation would increase. Such therapy could be sustained for a longer time since, presumably, lower dosages of antilatency drugs could be used. Targeting fluctuations in Tat copy number is very natural in this context, since the Tat protein is an essential part of the gene transcription machinery which produces new viral particles.
Fluctuations in Tat copy number have the potential to drive the activation. The up to date understanding of the feedback loop is that Tat controls transcriptional elongation rather than initiation. However, it seems that these two processes cannot be clearly separated, as it has been shown that an exogenous injection of Tat can activate the latent cell [14,2327]. We note that a study on mice [25] showed no obvious side effects.
It is somewhat surprising that exogenous administration of Tat can in fact have a positive effect on the activation of infected cells. The lack of Tat molecules is just one of the many barriers to activation. If the transcription of the viral DNA is also blocked by other mechanisms that are not controlled by the Tat protein, the injection of exogenous Tat should not speed up the transcription process automatically. In fact, there are also cases where Tat by itself cannot reactivate the virus [28].
This supports a notion that in latently infected cells the transcription system is in a rather labile state, and that the factors affecting the transcription process do not work strictly in a binary onoff fashion. Such a system could be activated by spontaneous fluctuations in Tat copy number. Based on this insight, in the next section we will construct a mathematical machinery to identify useful strategies to achieve noiseassisted activation of the virus.
A mathematical manual for the design of noisebased activation strategies
The central idea in the subsequent discussions is the concept of the “operating point” of the virus. The term will refer to a particular choice of the parameters that define the dynamics of the system, i.e., the virus and the cell it infects. In fact, one of the difficulties in combating HIV is that the integration of the virus into different sites gives rise to proviruses that have variable gene expression properties. It is conceivable that this results in different operating points of the virus. Therefore it is not immediately clear how a drug designed to achieve noisebased activation at a specific operation point would work consistently at other operating points. It is important to address this concern.
The manual we envision must contain a description of how (i) each antiviral treatment affects an operating point, (ii) how the amounts of administered agents affect the magnitude of the offset, and (iii) which antilatency effect the induced offset has. If the operating point of the virus is moved to regions where the latent state is less stable, the effect of antiviral drugs might be enhanced, which implies reduced dosage and shortened treatment. Below, each of these key considerations is mathematically formalized.
Mathematical description of therapies and dosage
Assumption I. The main function of each antilatency agent is to alter the operating
point of a virus: Assume that the virus is at an operating point
In this abstract sense a treatment of the HIV latency is a vector mapping
where the vector
should be always satisfied, i.e., with nothing administered no operating point offset is induced.
Assumption II. The size of the operating point offset depends on the dosage: The size of the offset
where the size of the offset is defined as
being the standard Cartesian norm of the vector
A possibility is to define
To avoid working with complicated metric spaces and projection procedures, the simplest possible relationship in the form of a proportionality law will be assumed.
Probability of activation and related observables
The central quantity of interest is the probability of the activation of a latent cell in an observation interval t_{0}<t^{′}<t, to be denoted by P(t_{0},t).
Assumption III. Offsets should be induced to reduce the survival probability of a latent virus. In strict mathematical terms P(t_{0},t) is a complex functional on the space of trajectories (histories). A history of the latent cell is defined as a set of individual trajectories for each particle type that define the biochemical content of the cell
where n_{i}(t^{′}) for i=1,2,⋯,N are copy numbers of relevant reactants (e.g. proteins) at time t^{′}.
Intuitively, one expects that not all details of trajectories are important and that the activation probability depends strongly on a distinct feature of a history quantified by an observable (quantifier) Φ,
The observable Φ is a functional on the space of histories and Π(Φ) is a real valued function. Two examples of Φ will be discussed later. If the observable Φ is chosen well, the activation probability should depend on Φ in a threshold like manner,
where Φ_{∗} is a constant.
The noisefree and noisebased activation concepts
Since trajectories are stochastic, the variable Φ is also stochastic and can be described by some probability distribution function
Γ(Φ). Assume that two types of treatments have been designed which adjust the operating
point of the stable latent virus (OP_{0}) so that the respective distributions for Φ are obtained as depicted in Figure 1 (OP_{1}, OP_{2}). The two distinct scenarios will be referred to as the “noisefree” (NFA) and the
“noisedriven (based)” (NBA) activations. The distributions are characterized by their
respective means (
Figure 1. The meaning of the noisefree and the noisebased activations. The figure illustrates the meaning behind the noisefree and noisebased activations.
All graphs were drawn by hand. Γ_{i}(Φ) with i=0,1,2 are distribution functions for the observable Φ for three systems in three operating points: the latent cell i=0, the cell that has been activated using the noisebased therapy (i=1), and the cell that has been activated using noisefree therapy (i=2).
Workings and effects of noisefree therapies
The purpose of a noisefree (NF) therapy
as much as possible. Under the influence of the treatment with a small dosage the expected change of this quantity
can be approximated by
where
is the gradient computed at the operating point
Very likely, due to experimental constraints, not all offsets can be realized. In an ideal situation where all offsets can be induced, there is a class of treatments which are the most efficient. Such therapies should induce the offset in the direction of the gradient
as
Such a treatment will be referred to as optimal noisefree.
Workings and effects of noisebased treatments
For a noisebased treatment
where variable κ quantifies the typical size of a fluctuation. For a small dosage, the expected change of this quantity
can be approximated by
where
It is useful to partition this gradient further into a noisefree part and a noiserelated part
where the noise relatedpart is given by
The precise value for κ depends on the character of the particle number distribution function and the confidence interval, but in here it will be taken as κ∼1 for simplicity reasons. Note that the effects of noise can be shutoff by taking κ=0.
Among all noisebased (NB) treatments a class of most efficient treatments exist, which induce the offset in the direction
where
Such a treatment will be referred to as optimal noisebased.
The dose reduction coefficient can be used to quantify at which operating points the noisebased treatment is useful
Of particular interest is to find a proper combination of drugs that can achieve a
maximal effect with a minimal dosage. We compare two arbitrary procedures
implying that both therapies induce the same change of the quantities they respectively
try to maximize. The quantities are given in (9) and (15) respectively, and
If the size of the vector
The minimization problem defined in (23) and (24) can be applied to decide whether a superior noisebased treatment exists, resulting in a lower dosage. The solution of the optimization problem is given by
where the dot denotes the scalar product. The ratio
will be used to quantify the comparison between the two strategies. This quantity
indicates the degree of the dose reduction for the treatment associated with a vector
In the case where the optimal noisefree treatment
to be referred to as the dose reduction coefficient. Ideally,
The space of operating points where a noisebased activation strategy can work
Now the mathematical manual will be used to find regions in the space of operating points of the latent virus where a noisebased activation is effective. To do this we focus on the Tat feedback loop. It is essential to model the noise of the Tat feedback loop, and to choose a relevant observable. The loop itself will be modelled in the simplest possible way as shown in the next subsection. The features of the system we wish to describe and the related observables are discussed subsequently.
The resistor model of HIV latency
The “resistor model” of the HIV dormancy control has been suggested to explain how the lack of the Tat molecules maintains (stabilizes) the latent state [14,15]. This simplified description of the transcription machinery will be used for the theoretical analysis. The biochemical model [15] consists of entities (or particle types) TatA and TatD which denote the acetylated and deacetylated form of the Tat molecule respectively. To simply the notation TatA and TatD will be abbreviated as A and D respectively.
The model consists of the following chemical reactions. Each deacetylated Tat molecule can get acetylated with rate α,
and each acetylated Tat molecules can get deacetylated with rate β,
Deacetylated Tat molecule decay with rate δ
An acetylated Tat molecule works as a transcription factor for the expression (production) of an additional TatD molecule,
It is clear that the model defined above neglects many biochemical details and is far from complete. For example, the mechanisms listed above are just a subset of all modifications of the Tat protein that occur in the cell [29]. Furthermore, a weak basal expression of TatD is continuously occurring [14] but this process will be neglected in the same way as in [15]. The model does not discriminate between mRNAs and proteins, either.
A class of noisebased activation strategies
The resistor model exhibits two states depending on the choice of the reaction rates. In the active state the feedback loop dominates the dynamics and all copy numbers increase. In the stable (dormant) state, bursts of activity last for a finite period of time during which the number of TatA molecules increases, reaches a maximum at t=t_{max}, and then decays to zero. Such bursts of activity will be referred to as pulses. A pulse lasts typically for a time t_{eq}. They rarely happen spontaneously and need to be initiated by injection of deacetylated Tat. This behavior is illustrated in Figure 2.
Figure 2. The characteristics of a typical pulse. The figure illustrates the key features of the activity pulse. All graphs were drawn by hand. The pulse lasts roughly t_{eq} and attains the maximum at roughly t≈t_{max}. Additional key quantities of interest are the height of the pulse μ_{max}, fluctuations around the peak described by the standard deviation σ_{max}, and the surface under the curve. The patterned triangle can be used to approximate the shape.
The stability condition has been formulated mathematically in [15]. The system is stable if
This condition is a fundamental property of the system. It defines the balance between the processes that produce and destroy TatA molecules. The system is stable if the production is slower than the destruction.
The key behavior of the resistor model we wish to exploit is related to pulses of activity. We investigate a hypothetical situation where Tat molecules are administered weakly at a constant rate in combination with an antiviral drug and one or more operation point offsetting agents. The injection of Tat should trigger repeated pulses of activity. The antiviral drug should reduce the copy number of viruses, while the offsetting agents move the operating point of the virus towards a nosier regime. The role of the offsetting agents is to boost fluctuations in the Tat expression levels so that the dosage of the antiviral drug can be reduced.
Two pulse characteristics as key observables
We now apply the mathematical machinery to compute the dose reduction coefficient for a wide range of reaction rate parameters (operating points) of the resistor model. It is necessary to identify an observable upon which the activation probability strongly depends. It will be assumed that the determining observable that governs the activation probability is the amount of viral mRNAs (proteins) produced by the feedback loop for the duration of a single pulse triggered by a single injection of Tat. If there are many viral particles that are ready for packaging and transport, the latent cell should activate.
The number of tatD mRNAs produced during the activation of the pulse can be computed by investigating how many molecules are produced by the A→A+D channel during the pulse duration. The expected amount of mRNAs produced in a small time interval dt is exactly given by kn_{A}(t) dt. Integrating this quantity for the whole pulse duration gives
where
is the pulse surface (PS), i.e., the surface under the n_{A}(t) curve (see Figure 2). This scenario will be referred to as the pulse surfacedependent threshold scenario, and when appropriate quantities computed in this context will be labelled by PS as in the example above.
There is no explicit experimental evidence that the surface is the one determining
factor that most strongly influences activation. However, this idea is supported by
several experimental studies. For example, in the experiment published in [15] it was argued that the duration of the activity pulse is likely to be important.
The strength of the feedback loop regulates the duration of activity bursts. For less
stable states the duration of such bursts increases dramatically, until they become
equally long, and eventually longer than the lifetime of the cell. Furthermore, in
[30] it was experimentally shown that a stronger feedback implies more likely activation.
Stronger feedback should correlate with the size of the surface
Describing the fluctuations of the area
where n_{A}(t) is the number of A at time t (see Figure 2). The quantities computed in this context will be labeled by PH as above.
While Φ_{PH} will be primarily used for illustrative purposes to gain qualitative understanding, the concentrationdependent threshold scenario does have practical significance. One expects that the number of Tat molecules in the cell correlates with the number of other viral proteins since their expression is encoded in the same gene. In that sense, high copy numbers of Tat protein should correlate with large activation probability. There are also some experimental indications in favour of the scenario.
Although there is no direct evidence that Tat on its own ensures activation, there is experimental evidence that exogenous injection of Tat can activate the latent cell [25]. Interestingly, there is also a report of the opposite [28]. Moreover, it has been suggested in several publications that Tat operates in a threshold dependent manner [6,8,31,32], but there is no direct experimental evidence for that. Interestingly, it is known that the Rev protein which controls export of viral mRNAs operates in a threshold dependent manner, and Tat drives the concentration of Rev above the threshold (See [15] for a discussion).
Computation of the dose reduction coefficients for the concentrationdependent threshold scenario (the pulse height scenario)
The average number of acetylated Tat molecules at the peak μ_{max} and the related standard deviation of it σ_{max} are the key quantities that need to be computed. Figure 2 illustrates the meaning of these quantities. Both have to depend on the parameters that define an operating point of the latent cell,
The functions
In this context the following choice for the observables of interest is the most natural:
and
This results is the following dose reduction coefficient computed at an operating
point of interest
where
and
Note that
Computation of the dose reduction coefficients for the pulsesurface dependent threshold scenario
We try to estimate fluctuations of
where t_{eq} denotes the average equilibration time of the system. This quantity has to be a welldefined
function of the operating point
The standard deviation of the area is much harder to estimate. A rough estimate for a typical fluctuation of this surface is given by
where
These assumptions results in the following observable for the noisefree activation analysis,
and the related noisecorrected quantity is given by
The rescaled equilibration time
measures the ratio between the length of the pulse duration and the time needed to produce a single tat mRNA. It equals roughly the number of viral particles produced during the single pulse. The definitions above result in the following dose reduction coefficient:
where
Methods
Thus, the required key quantities are the average number of Tat molecules at this peak, its standard deviation
The mathematical description of noise: An overview of computing the mean and the standard deviation
The mean and the standard deviation can be computed from a few lowest order factorial moments of the particle number distribution function P(n_{A},n_{D},t) which specifies the probability that the system will be found in a state (n_{A},n_{D}) at time t. The factorial moments are defined as
where x and y are arbitrary positive integers, and the angular brackets denote the usual ensemble average of an arbitrary function (observable) f(n_{A},n_{D}),
The mean and the variance are computed from the related factorial moments; ρ_{x,y}(t) with x+y≤2:
Of particular interest will be the values of μ_{A}(t) and σ_{A}(t) at
The equation system for factorial moments
By using the procedure detailed in [33,34], it is possible to show that the equation system for the factorial moments is given by
In principle, such equation system forms an infinite hierarchy which for the present system decouples automatically. It is known that when all propensity functions in the stochastic formulation are linear with respect to the population count the computation of statistical moments is simple. For example, the equations for ρ_{1,0} and ρ_{0,1} form a closed system, and likewise the equations for ρ_{2,0}, ρ_{0,2}, and ρ_{1,1}. It can be seen that the equations for ρ_{x,y} with x+y≤ξ form a closed system for every ξ=1,2,3,⋯. This analysis implies that the solution to a particular equation system is exact. Also, since the equation system for factorial moments up to order ξ=2 is exact, the values for the mean and the variance will be exact.
Direct numerical integration should be avoided
In order to solve the equations it is necessary to integrate them from some initial condition from time t=0 until t=t_{max}. There are several reasons why a direct numerical integration should be avoided. First, the computational cost of a single time integration scales linearly with the inverse of a typical time step size used. A direct numerical integration is highly impractical, since to construct the phase diagrams, the noise measures have to be computed at many points in the reaction rate space.
Second, there will be a need to compute derivatives of the mean and the standard deviation with respect to the reaction rates. If obtained numerically, his would multiply the computational effort by a factor two or more, depending on which technique is used to numerically calculate the derivatives.
We have found a procedure to avoid the numerical integration and yet obtain an analytic
approximation of these quantities. The details of how μ_{max}≡μ_{A}(t_{max}) and σ_{max}≡σ_{A}(t_{max}) are computed are given in the next section. Both quantities depend on ratios of
the reaction rates. Accordingly, it is useful to rescale all rates to obtain dimensionless
parameters, e.g., by rate α. This reduces the four dimensional Cartesian space of reaction rates defined by tuples
(α,β,δ,k) into a threedimensional Cartesian space defined by tuples
where the functions f_{μ} and f_{σ} are detailed in the next section. When convenient the subscript “max” will be omitted. To simplify notation we will use only μ and σ instead of μ_{max} and σ_{max}.
Details of computing the mean and the standard deviation
In this section it will be shown how to obtain the equations of motion for factorial moments and the functional forms for the mean and the standard deviation.
A procedure for avoiding numerical integration
The numerical analysis of the equations of motion (not shown) suggests that the following parameterization is useful:
where φ_{A} and φ_{D} are the noise strengths for A and D particle types since it is trivial to see that
The variable φ_{AD}(t) is a generalization of the noise strength concept for a pair of particle types.
A numerical integration of the equations of motion shows that for large times η=σ/μ→∞ but, in the same limit, the noise strengths become constant,
This can be used to obtain approximations for the noise strengths at t=t_{max}.
First, it is possible to construct an algebraic system of equations for these quantities by studying the asymptotic limit of the ODE system for φ_{A}(t), φ_{D}(t) and φ_{AD}(t). This implies that finding the asymptotic values for the noise strengths can be reduced to an algebraic problem. Second, for the Poisson initial condition the noise strengths are given by
This can be seen from the fact that for the Poisson initial condition
A single cell can be infected by more than one virus particle. The number of viral particles that infect a given cell (the multiplicity of infection) varies randomly and is usually Poissondistributed. Accordingly, in what follows it will be assumed that the initial particle number distribution function is Poissonlike.
With the assumptions at hand, the values of φ are known at two time instances, around t=0 and t≈t_{eq}. Here and in the following t_{eq} denotes the time after which the φ variables reach their asymptotic values. This information can be used to obtain a very crude approximation for the noise strengths in the form of a linear interpolation between the points t=0 and t=t_{eq}. Once the interpolation has been carried out, the values of the noise strengths at t=t_{max} are then given by
This is the approximation that will be used to compute noise strengths at the time instance where the number of the acetylated Tat particles reaches maximum.
There is no a priori reason why the noise strengths should vary linearly with time. We have inspected several curves where noise strengths were computed numerically to see whether the time dependence is linear. Interestingly, while the time dependence is not strictly linear it seems that the approximation used is qualitatively correct. We performed more rigorous tests of such an approximation by comparing it with the results of a numerical integration for wide range of parameters and found reasonable agreement.
The time t_{eq} can be found (not shown) by computing the eigenvalues of the matrix that defines the ODE system for the first and the second order moments. The smallest eigenvalue governs the relaxation time which is given by
where
and
The equation system for noise strengths
By using the parameterization just introduced it is possible to obtain the equations of motion for the means and the noise strengths. In addition, to make the analytic analysis easier it is useful to map the means onto the Poincaré sphere:
This is a useful mathematical technique to perform asymptotic analysis (see [35] and references therein). By using the new variables the equations for the means
become
By combining (58) with (6165) it is possible to obtain the equations for the noise strengths that are given by
It will be shown later that the terms proportional to a(t) and d(t) also drop out in the asymptotic limit when the stability condition is satisfied.
Locating the peak region
The value for t_{max} can be easily found by requiring that
One can also find that
In order to see what happens when relatively few deacetylated Tat molecules are injected the equation above will be used with d_{0}=1.
Computing the asymptotic noise strengths
The equation (83) does not involve the variable z(t) and can be used to obtain the asymptotic value for the ratio d(t)/a(t) as time approaches infinity. This differential equation has one stable fixed point in the physical region u(t)>0. The fixed point value for u(t), i.e., limt→∞u(t)=u^{∗}, is given by
Note that there are no restrictions on the reaction rates, only that they are positive real numbers. This implies that in this model the ratio of the number of acetylated and deacetylated Tat molecules approaches a constant value regardless on whether the system is stable or not.
The asymptotic (fixed point) value of z(t) can be determined by considering how the term β−αu(t) on the right hand side of Eq. (82) behaves as time becomes large. From (89) one can see that
where Λ is a strictly positive constant that depends on the values of the reaction rates. Its exact numerical value is not relevant for the discussion and the formula for Λ will not be shown. For large times one has
When βδ−αk>0, z(t) grows exponentially fast, implying that the means approach zero. This condition is fully equivalent to the stability condition (32). On the other hand, when βδ−αk0, z(t) approaches zero exponentially fast, which implies that the average copy numbers become infinite. In both cases the ratio of the means becomes constant and is given by u^{∗}.
For a stable system one can neglect the terms proportional to a(t) or d(t) in Eqs. (8486) when t→∞. This results in a linear system of equations for the asymptotic values of the noise strengths given by
where
Finally, σ can be computed as
where a(t_{max}) is given in (88) and
Results
A hyperbolic coordinate system facilitates understanding
The space of the reaction rates is multidimensional and relatively hard to visualize.
Numerical tests showed that it is advantageous to reparameterise the space of
Accordingly, an operating point will be the triple
In fact, this set of coordinates is very intuitive which can be appreciated by analysing the meaning of the inverse transformation.
The coordinate v is given by
This quantity measures to which extent the reactions which remove acetylated Tat dominate over the reaction that produces it. The coordinate
measures the relative contribution of the reactions that remove acetylated Tat molecules. These two coordinates naturally form a hyperbolic coordinate system. The variable
is particularly important for several reasons. First, it measures how intensive the transcription process (the production of acetylated Tat molecules) is. This coordinate also measures how far in the ility region the operating point of the virus is placed. For example, ϵ=0 at the border of the stab stability region when αk=βδ. Furthermore, ϵ≈1 when the stability condition is strongly satisfied, i.e. when βδ≫kα. Thus, in the stable region ϵ attains values in the interval between zero and one.
This variable could potentially be used to experimentally quantify the degree of the latency for a given cell. For example, the literature suggests that for a latent virus its operating point has to lie in the region of stability where ϵ>0. Otherwise, for ϵ0, an infected cell would lyse relatively fast. Thus, it seems necessary that a treatment meant to activate a latent virus has to move its operating point into the unstable region where ϵ0 or at least sufficiently close to the instability boundary ϵ=0.
An example of a typical operating point can be obtained from the resistor model. From the experimental values [15] for the reaction rates α_{R}=0.5/day, β_{R}=5/day, δ_{R}=2/day, and k_{R}=5/day one can compute the corresponding operating point
The symbol “R” emphasizes the resistor model operating point.
Regions of parameter space where noise is large
Perhaps the biggest advantage with the hyperbolic coordinate system suggested is that the v dependence is very easy to visualize. Figures 3a, 3b, and 3c depict how the coefficient of variation η=σ/μ depends on the reaction rate parameters. Figures 3a and 3b demonstrate that when ϵ approaches the border of stability (ϵ→0) the coefficient of variation approaches infinity. Figures 3b and 3c show that if v is increased, the coefficient of variation always increases, no matter which values for u and ϵ are chosen. Figures 3a and 3c indicate that if noise is to be exploited for a treatment, one should design drugs that could move the operating point of the virus away from the regions around u≈0. These plots show in which regions of the parameter space the effects of noise are expected to dominate. Superficial analysis of these figures would suggest that one should choose rates β and δ as different as possible from each other, since this should increase the amount of noise relative to the mean. However, in doing so one might move the operating point such that despite the increase in fluctuations the threshold cannot be reached. A more quantitative analysis is needed in order to identify useful noisebased strategies. This is illustrated by a case study, where we suggest how the mathematical manual developed above can be used to guide experimental design.
Figure 3. Contour plots of σ / μ in three hyperplanes: (a) for ( v =v_{R}, u , ϵ ), (b) for ( v , u =u_{R}, ϵ ), and (c) for ( v , u , ϵ =ϵ_{R}). The position of the resistor model operating point is marked by the white circle. Contour lines are labelled by their respective σ/μ values (square boxes). Lighter (darker) regions indicate where noise does (does not) dominate the dynamics.
A case study: A strategy for improving noisefree histone deacetylase inhibitor treatment by moving the operating point of the resistor system to a noisier region
A typical antilatency strategy in the HIV therapy context is to increase the transcription
rate k, for example in treatments based on the use of histone deacetylase inhibitors [36]. These molecules open up the chromatin environment such that the transcription factors
needed for viral expression can attach easier to their respective binging sites. Several
molecules have been suggested as drug candidates. Valproic acid, trichostatin A, vorinostat
(SAHA), and many more have been reviewed in [36], and references therein. Some of these molecules are rather toxic. In the following
we investigate how toxicity could be reduced by lowering their dosage at a typical
operation point, e.g.
The expressions for the derivatives are not shown. Their numerical values are given by:
By administering a certain amount of an antilatent drug, e.g., SAHA, which only shifts k_{R} to k_{R}+Δk and leaves α_{R}, β_{R}, and δ_{R} unchanged, the operating point would be shifted by
where Δϵ is arbitrary but controlled by the amount of SAHA administered. From Eq. (26) one
can see that a more effective noisebased therapy with SAHA as the primary drug exists,
since
Different changes in the reaction rates have to be used in order to generate these two vastly different offsets. For example, let us investigate which changes in the transcription rate k are needed to achieve the changes in (108) and (109).
From (98100) follows that
where
In such a case the equations (110112) and (113116) imply that both Δv and Δu are zero, and the following relationship between Δk and Δϵ holds
However, in the case of the noisebased therapy we do not know a priori the values
for
By comparing (117) and (120) one sees that the dose of the primary drug can be reduced roughly twenty times (assuming the linear relationship between the dose change and the transcription rate change). The price one has to pay is that additional drugs have to be administered which reduce β and increase δ (note that Δϵ is negative). It is not unrealistic that this can be eventually verified experimentally, e.g. in the context of the kinetics experiments on LTR [14,15].
Even if the optimal noisefree treatment would be applied in this case, noisebased
treatment would still result in dose reduction. An immediate use of Eq. (27) shows
that
Note that
This case study shows that it is sufficient to inspect the value of the dosage reduction
coefficient
Regions of parameter space where noisedriven activation is possible
To identify regions in parameter space where noisebased activation is beneficial,
we determine how
Figure 4. Phase diagrams for the dose reduction coefficient. The panels (ac) and (ef) depict contour plots of the dose reduction coefficients
A striking result of this visual analysis is that regions where
The contours in panels (a) and (d) differ in the middle region, and likewise for panels (c) and (f). There are even some regions where a relatively large dose reduction is possible. Figure 4a indicates that an efficient dosage reduction can be achieved in the regions around u≈0 where ϵ is either very large or very small. This does not hold for panel (d). However, both (a) and (d) panels suggest that the noise based therapy is advantageous in the region where the absolute value of u is very large.
Figures 4b and 4e show that for fixed u the largest dose reduction can be reached for very small values of ϵ and very large values of v. Figures 4c and 4f indicate that the noisebased activation can be useful in the regions where v is large and where either u→∞ or u→−∞. The panels do not agree in the middle region.
The linear theory yields qualitative predictions when operating point offsets are not small
We provide a noninfinitesimal analysis for a particular operating point, and investigate how it differs from the corresponding linear analysis. To do this we will use the technically simpler PH scenario. This will be illustrated by studying how both μ_{max} and μ_{max}+σ_{max} depend on u and ϵ for fixed v (Figure 5). One particular hyperplane has been chosen, where the size of the u and ϵ components of the offset in Eq. (109) is somewhat larger than the vcomponent.
Figure 5. Reaction rate dependence ofμ_{max} andμ_{max}+σ_{max} in a ( u , ϵ ) hyperplane. Contour plots of μ_{max} and σ_{max}+μ_{max} in the (u,ϵ) plane with v=v_{R}. The full lines labelled NF1 and NF2 are two contours of μ_{max}, the two dashed lines labelled NB1 and NB2 are contours of σ_{max}+μ_{max}. The operating point of the resistor model is represented by the white circle. The
contour values for the NF1 and NB1 contours are chosen so that these lines pass through
the operating point of the resistor model. The contours labelled NF2 and NB2 have
been offset from NF1 and NB1 by the same ΔN. The black arrows labelled
Several contour lines are shown: (NF1) μ_{max}=μ_{1}, (NF2) μ_{max}=μ_{1}+ΔN, (NB1) μ_{max}+σ_{max}=μ_{1}+σ_{1}, and (NB2) μ_{max}+σ_{max}=μ_{1}+σ_{1}+ΔN where μ_{1}=0.05378, μ_{1}+σ_{1}=0.064536, and ΔN=0.010756.
The long black arrow denotes the linearized noisefree offset that needs to be induced in order to shift the mean from μ_{max} to μ_{max}+ΔN. The long grey arrow is the shortest noninfinitesimal noisefree offset that needs to be induced to move from the operating point to the NF2 contour line. The large grey circle (only in part visible) starts touching NF2 exactly at the point where the arrow meets the contour line. The short black arrow denotes the noisebased offset that needs to be induced in order to shift μ+σ by ΔN. The small grey circle denotes the minimal circle that touches NB2.
The length ratio between the black vectors is specified exactly by the dose reduction
coefficient
A few comments on the robustness of the results with regard to the model extension
The model used in here cannot capture all phenomena that might be important. The first feature of the model that can be clearly improved is its complexity. For example, we made no distinction between mRNAs and proteins. It has been found that after integration into the genome the HIV promoter makes mRNAs in random bursts of transcriptional activity [37,38]. Moreover, each mRNA makes proteins in translational bursts. This suggests that there might be other sources of noise in the system that were not considered in this work.
Also, a rather phenomenological model was used for the probability of activation and the related observable. There is clearly a need for refinement. It is possible that the probability of the virus activation depends on other features of the system, e.g., the amount of relative fluctuations. This is rather speculative, based upon indications in the literature that Tat operates also outside the Tat feedback loop. Tat is a multifunctional protein that is involved in other intracellular processes [32,39,40] and there is certainly the possibility that it influences cell homeostasis in a more complex manner than discussed here. For example, in addition to acetylation, Tat undergoes several other posttranslational modifications and interacts with several other proteins. It is currently not entirely clear how this might affect gene expression noise on one hand, and the activation probability on the other. Posttranslational modifications or other interactions could buffer, or propagate noise in levels of Tat into noise in gene expression. The question remains whether such effects could be beneficial in bringing the operating point closer to the stability boundary. Out of all parameters in the model, such effects will very likely exert strong influence on the decay constant δ of the deacetylated Tat and cause it to fluctuate in time. One could speculate that in such a situation both u in v will start fluctuating, though u, being essentially the natural logarithm of the ratio between β and δ, will fluctuate much less. This implies that v fluctuates towards smaller values, away from the noiserich regions, while u and ϵ are kept essentially constant (e.g. see the phase diagrams). Such changes in v could reduce the activation probability, and need to be further investigated.
We now briefly discuss possible effects of chromatin on the operating point of a virus. The transcription rate k will likely be influenced the most. This implies that instead of analysing the effect of a drug on a single operating point one should investigated a set of operating points. Such points should be distributed around the region where v≈v_{R}, u≈u_{R} and where ϵ∈[0,1]. Accordingly, a noisebased treatment designed to move the operating point into a noisier region will work uniformly on all points in this set. This would always result in a relatively large dose reduction; Figure 4, panels (a), (b), (d) and (e).
Conclusions
We have designed a generic mathematical manual of how to approach the problem of the HIV latency in a quantitative manner, accompanied by an example of how to use this manual. We formalised several concepts that are vaguely defined or understood only intuitively. The first key concept is the notion of an “operating point” of the virus and how it is affected by the action of a drug. The second key concepts of the mathematical formalism is the notion of a particular observable that strongly affects the activation probability. We suggested a mathematical way of describing how each therapy affects an operating point and how this in turn influences the observable that controls the activation probability. The third key ingredient is a dose reduction coefficient, which can be used to quantitatively compare various therapies.
We have suggested and investigated two rather general strategies for the virus activation, the noisefree and the noisebased strategy. In the first approach, antilatency agents are administered such that the activation happens with almost absolute certainty. In order to achieve such certainty, possibly unreasonable quantities of drugs need to be administered. In the second approach, drugs are administered in such a way that the activation is less certain but still happens with a relatively large probability. The idea behind the noisebased strategy is to reduce the quantity of drugs that need to be administered in order to achieve activation.
The mathematical manual is rather generic. To demonstrate how the mathematical manual can be used, we have focused on the simplest possible model of the Tat protein feedback loop, the most important part of the the HIV latency control. Based on the structure of the loop we suggested a class of noisebased activation strategies. We envision such an activation strategy in a procedure where one constantly supplies exogenous Tat at a very small rate, and adds a combination of antilatency drugs that would offset the operating point of the virus towards noisier regions.
In this context we considered two observables and could compute the dose reduction coefficient for both cases to answer the fundamental question: for which operating points the noisebase therapy is advantageous over the noisefree therapy in the sense of possible dosage reduction? This addresses the practical problem of reducing effects of toxicity during the antilatent treatment. Three phase diagrams were constructed to explain what controls the noise, and how this control can be used to battle latency. Our analysis of the phase diagrams indicates that the noisebased therapy is always advantageous, no matter which operating point the virus adopts. This results holds regardless of which observable is targeted. This is the major result of our analysis.
The mathematical manual is currently based on rather qualitative assumptions, but it is very generic and can be easily extended and refined. For example, we have used the assumption of small operating point offsets in order to linearize the theory. One can easily consider noninfinitesimal offsets, as demonstrated in the example. We discussed several possible extensions which are left for future work.
In summary, we suggest an activation principle where intrinsic noise is considered a feature benefiting treatment. We showed that such strategy should be efficient for any latent cell.
Competing interests
Both authors declare that they have no competing interests.
Authors’ contributions
ZK has defined the research theme, developed the model, and performed computations. ZK and AJ analysed the results, discussed, and prepared the manuscript. Both authors read and approved the final manuscript.
Acknowledgement
This work was partly funded by the Swedish Research Council (VR).
References

Greene WC: A history of AIDS: looking back to see ahead.
Eur J Immunol 2007, 37:S94S102. PubMed Abstract  Publisher Full Text

Le Douce V, Janossy A, Hallay H, Ali S, Riclet R, Rohr O, Schwartz C: Achieving a cure for HIV infection: do we have reasons to be optimistic?
J Antimicrob Chemother 2012. Publisher Full Text

Coiras M, LopezHuertas MR, PerezOlmeda M, Alcami J: Understanding HIV1 latency provides clues for the eradication of longterm reservoirs.
Nat Rev Microbiol 2009, 7:798812. PubMed Abstract  Publisher Full Text

Contreras X, Lenasi T, Peterlin BM: HIV latency: present knowledge and future directions.
Future Virol 2006, 1:733745. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Dahl V, Josefsson L, Palmer S: HIV reservoirs, latency, and reactivation: prospects for eradication.
Antiviral Res 2010, 85:286294. PubMed Abstract  Publisher Full Text

Graci JD, Colacino JM, Peltz SW, Dougherty JP, Gu Z: HIV type1 latency: targeted induction of proviral reservoirs.
Antivir Chem Chemother 2009, 19:17787. PubMed Abstract

Karn J: The molecular biology of HIV latency: breaking and restoring the Tatdependent transcriptional circuit.
Curr Opin HIV AIDS 2011, 6:411. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Mbonye U, Karn J: Control of HIV latency by epigenetic and nonepigenetic mechanisms.
Curr Hiv Res 2011, 9:554567. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Schiralli Lester GM, Henderson AJ: Mechanisms of HIV transcriptional regulation and their contribution to latency.
Mol Biol Int 2012, 2012:614120. PubMed Abstract  PubMed Central Full Text

Siliciano RF, Greene WC: HIV latency.
Cold Spring Harb Perspect Med 2011, 1:a007096. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Soudeyns H, Pantaleo G: The moving target: mechanisms of HIV persistence during primary infection.
Immunol Today 1999, 20:446450. PubMed Abstract  Publisher Full Text

Williams SA, Greene WC: Regulation of HIV1 latency by Tcell activation.
Cytokine 2007, 39:6374. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Richman DD, Margolis DM, Delaney M, Greene WC, Hazuda D, Pomerantz RJ: The challenge of finding a cure for HIV infection.
Science 2009, 323:13041307. PubMed Abstract  Publisher Full Text

Weinberger LS, Burnett JC, Toettcher JE, Arkin AP, Schaffer DV: Stochastic gene expression in a lentiviral positivefeedback loop: HIV1 Tat fluctuations drive phenotypic diversity.
Cell 2005, 122:169182. PubMed Abstract  Publisher Full Text

Weinberger LS, Shenk T: An HIV feedback resistor: autoregulatory circuit deactivator and noise buffer.
Plos Biol 2007, 5:6781. Publisher Full Text

Burnett JC, MillerJensen K, Shah PS, Arkin AP, Schaffer DV: Control of stochastic gene expression by host factors at the HIV promoter.
Plos Pathog 2009, 5:e1000260. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Razooky BS, Weinberger LS: Mapping the architecture of the HIV1 Tat circuit: a decisionmaking circuit that lacks bistability and exploits stochastic noise.
Methods 2011, 53:6877. PubMed Abstract  Publisher Full Text

Raj A, van Oudenaarden A: Nature, Nurture, or Chance: stochastic gene expression and its consequences.
Cell 2008, 135:216226. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Balazsi G, van Oudenaarden A, Collins JJ: Cellular decision making and biological noise: from microbes to mammals.
Cell 2011, 144:910925. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Singh A, Weinberger LS: Stochastic gene expression as a molecular switch for viral latency.
Curr Opin Microbiol 2009, 12:460466. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Margolis DM: Eradication therapies for HIV infection: time to begin again.
AIDS Res Hum Retroviruses 2011, 27:347353. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Shan L, Deng K, Shroff NS, Durand CM, Rabi SA, Yang HC, Zhang H, Margolick JB, Blankson JN, Siliciano RF: Stimulation of HIV1Specific cytolytic T lymphocytes facilitates elimination of latent viral reservoir after virus reactivation.
Immunity 2012, 36:491501. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Donahue DA, Kuhl BD, Sloan RD, Wainberg MA: The viral protein Tat can inhibit the establishment of HIV1 latency.
J Virol 2012, 86:32533263. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

GonzalezHernandez MJ, Swanson MD, ContrerasGalindo R, Cookinham S, King SR, Noel JRJ, Kaplan MH, Markovitz DM: Expression of human endogenous retrovirus typeK (HML2) is activated by the Tat protein of HIV1.
J Virol 2012, 86(15):77907805. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Lin X, Irwin D, Kanazawa S, Huang L, Romeo J, Yen TS, Peterlin BM: Transcriptional profiles of latent human immunodeficiency virus in infected individuals: effects of Tat on the host and reservoir.
J Virol 2003, 77:822736. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Emiliani S, Fischle W, Ott M, Van Lint C, Amella CA, Verdin E: Mutations in the tat gene are responsible for human immunodeficiency virus type 1 postintegration latency in the U1 cell line.
J Virol 1998, 72:16661670. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Bettaccini AA, Baj A, Accolla RS, Basolo F, Toniolo AQ: Proliferative activity of extracellular HIV1 Tat protein in human epithelial cells: expression profile of pathogenetically relevant genes.
Bmc Microbiol 2005, 5:20. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Pearson R, Kim YK, Hokello J, Lassen K, Friedman J, Tyagi M, Karn J: Epigenetic silencing of human immunodeficiency virus (HIV) transcription by formation of restrictive chromatin structures at the viral long terminal repeat drives the progressive entry of HIV into latency.
J Virol 2008, 82:1229112303. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Ott M, Geyer M, Zhou Q: The control of HIV transcription: keeping RNA polymerase II on track.
Cell Host Microbe 2011, 10:426435. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Weinberger LS, Dar RD, Simpson ML: Transientmediated fate determination in a transcriptional circuit of HIV.
Nat Genet 2008, 40:466470. PubMed Abstract  Publisher Full Text

Leonard JN, Schaffer DV: Computational design of antiviral RNA interference strategies that resist human immunodeficiency virus escape.
J Virol 2005, 79:16451654. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Romani B, Engelbrecht S, Glashoff RH: Functions of Tat: the versatile protein of human immunodeficiency virus type 1.
J Gen Virol 2010, 91:112. PubMed Abstract  Publisher Full Text

Konkoli Z: Mathematical explanation of the predictive power of the Xlevel approach reaction noise estimator method.
Theor Biol Med Model 2012, 9:12. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Konkoli Z: Modeling reaction noise with a desired accuracy by using the X level approach reaction noise estimator (XARNES) method.
J Theor Biol 2012, 305:114. PubMed Abstract  Publisher Full Text

Konkoli Z, Johannesson H, Lee BP: Fluctuation effects in steric reactiondiffusion systems.
Phys Rev E 1999, 59:R3787R3790. Publisher Full Text

Wightman F, Ellenberg P, Churchill M, Lewin SR: HDAC inhibitors in HIV.
Immunol Cell Biol 2012, 90:4754. PubMed Abstract  Publisher Full Text

Skupsky R, Burnett JC, Foley JE, Schaffer DV, Arkin AP: HIV promoter integration site primarily modulates transcriptional burst size rather than frequency.
Plos Comput Biol 2010, 6:e1000952. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Dar RD, Razooky BS, Singh A, Trimeloni TV, McCollum JM, Cox CD, Simpson ML, Weinberger LS: Transcriptional burst frequency and burst size are equally modulated across the human genome.
Proc Natl Acad Sci U S A 2012, 109:1745417459. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Das AT, Harwig A, Berkhout B: The HIV1 Tat protein has a versatile role in activating viral transcription.
J Virol 2011, 85:95069516. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Peruzzi F: The multiple functions of HIV1 Tat: proliferation versus apoptosis.
Front Biosci 2006, 11:708717. PubMed Abstract  Publisher Full Text