Number of autoregulation and FFL motifs in a network

Number of autoregulation and FFL motifs in a network

We are searching data for your request:

Forums and discussions:
Manuals and reference books:
Data from registers:
Wait the end of the search in all databases.
Upon completion, a link will appear to access the found materials.

Suppose we have a network with N* nodes (N* is the number of internal nodes). Every directed link in this network exists with probability p. What would be the number of:

  1. motifs that are auto-regulatory?
  2. feed-forward loop motifs
  3. feed-back loop motifs

  1. The average number of auto-regulatory motifs (self-edges), is equal to the number of edges E times the probability that an edge is a self-edge which is $p_{self}=1/N$, with N being the total number of nodes. Therefore, $_{rand} approx E/N$
  2. & 3. According to U.Alon's book ("Introduction to Systems Biology"), the mean number of times that a subgraph G occurs in a random network is given by the following formula: $approx a^{-1}N^n p^g$.

Explaining what each term of the formula means:

  • α: is a number that includes combinatorial factors related to the structure and symmetry of each subgraph, equal to 1 for feed-forward loop (FFL) and equal to 3 for feed-back loop (FBL)
  • $N^n$: is the number of ways of choosing a set of n nodes out of N. Because there are N ways of choosing the first one, times N-1$approx$N ways of choosing the second one, and so on… (this approximation is true for large networks)
  • $p^g$ : is the probability to get the g edges in the appropriate places (each with probability p)

BioFNet: biological functional network database for analysis and synthesis of biological systems

Hiroyuki Kurata is the Director of the Biomedical Informatics R&D Center at Kyushu Institute of Technology. His group aims to develop a computer-aided design system of biochemical networks (CADLIVE).

Kazuhiro Maeda is a postdoctoral research fellow in the Department of Bioscience and Bioinformatics at Kyushu Institute of Technology. He works on dynamic modeling for metabolic and gene regulatory networks.

Toshikazu Onaka is a master course student in the Department of Bioscience and Bioinformatics at Kyushu Institute of Technology. He made records and focused on dynamic simulation and system analysis of biochemical networks.

Takenori Takata is a master course student in the Department of Bioscience and Bioinformatics at Kyushu Institute of Technology. He made records and focused on dynamic simulation and system analysis of biochemical networks.

Hiroyuki Kurata, Kazuhiro Maeda, Toshikazu Onaka, Takenori Takata, BioFNet: biological functional network database for analysis and synthesis of biological systems, Briefings in Bioinformatics, Volume 15, Issue 5, September 2014, Pages 699–709,

MOTIF: Functional Unit of An Interaction Network

In a network, integration of elements and interacting components enables identification of conserved modules and
motifs. The topological analysis, however, reveals much about the nature and functions of a network and provides sufficient statistics for any further study. Supported by several data types viz interaction data, expression data, Boolean data, and raw sequence data, modules and motifs provide an easy way to understand the specific function of a gene and protein. Basically, network motifs are characteristic network patterns comprising of both transcription regulation and protein-protein interaction that recur more often than in a random network.

The idea of network motif (sub-graph) was presented by Uri Alon and his group in 2002 [1] as they were discovered in a gene regulation network of E. coli and then in a large set of a neural network. According to their occurrence and behavior in a network, “motifs are subgraph recurring repeatedly, defined by a particular pattern of interaction between vertices that reflect a framework in which particular functions are achieved”. They are of vital importance largely because they may display functional properties and may also provide deep insight into network’s functional abilities. Significant studies have been done from perspective of the biological application as well as computational theory. Biological analysis mainly endeavors to interpret the functions of network motifs associated with genetic regulation as the first motif was found in the transcription unit of E. coli as well as Yeast and other higher organisms. Apart from those of genetic regulation, some distinct motifs were also discovered from the neural network and protein interaction network (fig-1).

Fig: 1. Different types of motifs in the biological network. (courtesy: Google image)

Statically a motif is identified as a pattern that occurs at least five times and is more significant than in a random network. With only two or at least three nodes, we may randomize to get a maximum pattern in a network. It is up to analyses that one has to perform. Patterns with two, three, four and five nodes are significant as their occurrence is more frequent in a network than any other pattern. Based on directivity, connectivity, pattern, regulation, and the number of nodes, they are classified into various categories as below:

1. Negative auto-regulation (NAR)
One of the simplest and most abundant network motifs is negative autoregulation in which a transcription factor represses its own transcription (fig. 2-a). Its generalized function is in response regulation and SOS DNA repair system response. NAR was observed to speed-up the response to signals in a synthetic transcription network. It also increases the stability of the auto-regulated gene product concentration against stochastic noises, reducing variations in protein levels between different cells.

Fig: 2. Different types of loops and motif in biological networks a. auto-regulation, b. feedforward motif, c. coherent and incoherent loops, d. different types of patterns (motifs), commonly occurring in biological networks. They all occur in almost every biological system and represent a specific regulatory functional unit. (Courtesy- Google Images)

2. Positive auto-regulation (PAR)
It is characterized by the enhancement of transcription by its own gene product (fig 2-a). Comparatively, it shows a slower response than NAR. In a case where rapid regulation is required, PAR leads to a bimodal distribution of protein levels in cell populations.
3. Feed-forward loops (FFL)
This motif commonly occurs in many genetic regulatory networks and consists of three genes and three regulatory interactions (fig 2-b). In the diagram, target gene C is regulated by 2 TFs (transcription factor) A and B and in addition TF B is also regulated by TF A. Since each of the regulatory interaction may either be positive or negative, there are eight possible types of FFL motifs. Computationally, in most of the cases, FFL represent an AND & OR gate but other circuitry inputs are also possible.
4. Coherent type 1 FFL (C1-FFL)
This is one of the sub-types of FFL, characterized by giving a pulse filtration in which a short pulse of a signal will not generate a response but a persistent response will generate a short delay. Importantly, the signal response is fastened after one shut off. Such a vital mode of signal transduction in genetic or cellular regulatory system is observed in metabolic pathways and protein-gene interaction network.
5. Incoherent type 1 FFL (I1-FFL)
It is known to be a pulse generator and response accelerator. Two signal pathways function in two opposite ways, one signal activates and the other represses. After repression, a pulse dynamics is generated. Importantly, it speeds up the activation of any gene, not necessarily a gene of a transcription factor. Feedforward regulation shows better regulation than negative feedback.
6. Multi-output FFLs
The same regulator controls (regulates) multiple genes of the same system.
7. Single-input modules (SIM)

This motif occurs when a single gene regulates a single set of a gene with no additional regulation. This is significant when genes are carrying out a specific function and therefore need to be activated in the synchronized manner.
A possible confirmation of motif importance is motif conservation. In evolution conservation implies importance. The conservation of a protein in the network may be taken as an indication of the biological importance of that motif. The conservation of a motif shows the evolutionary pressure that can be followed to find ortholog in other organism. Wuchty et al, 2003 [2] tested this hypothesis for the correlation between the protein evolutionary rate and the structure of the motif it is embedded in. The conservation of motif component was found to be tens to thousands of times higher than expected at random, suggesting conservation of motif constituents. Motifs representing small functional unit or sub-graph in a network are found using different software like M-finder ( Alon/groupNetworkMotifSW.htm), MODIS, FANMOD (http://theinf1.informatik.uni-jena.dewernicke/motifs/index.htm), MAVisto ( iGRAPH: (, HOMER & Motif-X etc. Online tools like Amadeus, Web Motif, and MEME suite are also used for the same purpose.


1. U Alon, Network Motif: theory and experimental approaches, Nature Reviews Genetics 8, 450-461,(June 2007) doi :10.1038

2. Wuchty S, et al. (2003) Evolutionary conservation of motif constituents in the yeast protein interaction network. Nat Genet 35(2):176-92.


Some FFL Types Occur More Often than Others in E. coli and Yeast. We enumerated the appearances of each FFL type in databases of E. coli (6) and S. cerevisiae (11) transcription interactions (Tables 1 and 2). In both E. coli and yeast, we found that the type 1 coherent FFL, which has three activation interactions, is by far the most common coherent configuration. Incoherent FFLs are more common in yeast than in E. coli. In both organisms, the type 1 incoherent FFL, in which X activates a repressor of Z, is the most common incoherent configuration. The next most common configuration in yeast is type 2, which is somewhat more abundant than types 3 and 4.

The rest of the article is organized as follows. We analyze the behavior of the eight types of FFL by using kinetic equations for the protein levels (see Materials and Methods). For simplicity, we first assume that X and Y act in an AND-gate fashion to regulate Z and later discuss the OR-gate case. We begin with the four coherent FFL types and discuss their steady-state and kinetic behavior. Then we analyze the four incoherent FFLs. The results are summarized in Table 3.

Steady-State Behavior of Coherent FFL with AND-Gate: Only Types 1 and 2 Respond to Sy.Table 1 lists the steady-state behavior of the four coherent FFLs as a function of the two input stimuli, Sx and Sy. The steady-state of Z is evaluated for all four combinations of Sx = <1, 0>and Sy = <1, 0>, where 1 means saturating stimulus. We find that only types 1 and 2 coherent FFL (with AND-gate regulation) respond strongly to both Sx and Sy inputs (Table 1). Types 3 and 4 respond to Sx but not to Sy. To understand this, consider the type 4 coherent FFL. Here, X activates Z directly and also represses a repressor of Z. Z can be expressed only when Sx is present, because it requires active X. However, when Sx is present X acts to repress Y. As a result, the protein Y is not significantly expressed and therefore cannot interact with Sy to affect Z. When Sx is absent, Z cannot be expressed because the activator X is inactive, and even though protein Y is present, Sy has no effect on Z. Thus, Sy has no effect on Z in either the presence or absence of Sx. In contrast, in the type 1 coherent FFL X activates Y, and thus the protein Y is expressed when Sx is present and can interact with Sy to modulate Z expression.

Coherent FFL Kinetics: All Types Serve as Sign-Sensitive Delay Elements. We now consider the kinetic response of Z to step-like addition of the inducer Sx, in the presence of Sy. In the type 1 coherent FFL (with AND-gate regulation), for example, upon a step addition of Sx, Z expression begins only when the activator Y builds up to a sufficient concentration and crosses the activation threshold for Z (Fig. 2Left). The speed of the response is characterized by the response time, the time that it takes Z to reach half of its steady-state level (17, 22). The response time after an on step of Sx is longer in the coherent FFL than for a corresponding simple regulation design (Fig. 1b) that has the same steady-state Z levels (Fig. 2, compare thick and medium curves to the thin curves). The magnitude of the delay can be tuned by the relationships between four biochemical parameters: lifetime of Y, lifetime of Z, the threshold Kyz, and the basal Y level (Fig. 2).

Kinetics of coherent type 1 (Left) and type 4 (Right) FFLs with AND regulatory logic, in response to on and off steps of Sx. Note that the delayed response to on steps of the FFLs (thick, medium lines) compared to a corresponding simple system (thin line). Note that FFLs can behave as simple regulation for nonfunctional parameter domains (see Materials and Methods). Simulation parameters: Kxz = Kxy = 0.1 for type 1, Kyz = <0.5, 5> for type 4, Kyz = <0.6, 0.3> all others are as stated in Materials and Methods.

The delay in the response is sign sensitive: the response to on steps is delayed, but the response to off steps of Sx is not delayed (Fig. 2). We term this behavior sign-sensitive delay. It is carried out by all four types of coherent FFLs. Type 2 and 3 coherent FFLs have reversed sign sensitivity: the response to off, but not on steps is delayed. The delay response is summarized in Table 1.

Steady-State Behavior of AND-Gate Incoherent FFL with No Basal Activity: Only Types 1 and 2 Respond to Sy. As in the case of AND-gate coherent FFLs, we find that only type 1 and 2 incoherent FFLs with AND-gate Z regulation are able to respond in steady state to both of their input stimuli, Sx and Sy. Types 3 and 4 have a constant steady state, which does not depend on either Sx or Sy values (Table 2).

Kinetics of Incoherent FFL with No Basal Activity: Only Types 3 and 4 Are Good Pulsers. We now consider the kinetic response of the incoherent FFL to steps of Sx, in the extreme case where Y modulation by X leads to a strong effect on Z. In this case, the incoherent FFL functions as a pulser. For example, in the type 4 incoherent FFL, when Sx turns on, Z is first induced by the joint action of X and Y. Meanwhile, Y production is repressed by X and its levels drop, until Z production begins to decrease. Thus, in type 4 upon an on step of Sx, in the presence of Sy, Z levels first rise and then drop (Fig. 3Right). A similar scenario holds for type 3. We find that type 1 and 2 incoherent FFLs (with AND gate Z regulation) are generally poor pulsers (Table 2 and Fig. 3a). The pulse amplitude is much smaller than the maximal level that can be reached by the circuit (the maximum level is reached in the absence of Sy, Fig. 3Left Bottom). We find that type 1 and 2 incoherent FFLs are poor pulsers for all parameter values. In contrast, types 3 and 4 are good pulsers: For some biochemical parameters, the pulse reaches high amplitude relative to the maximal circuit response (Table 2 and Fig. 3Right). The pulse occurs in the absence but not in the presence of Sy in the case of type 3 and 4 FFLs. Thus Sy is an enabling signal that can be used to allow or block the pulse (Fig. 3Right).

Kinetics of incoherent type 1 (Left) and type 4 (Right) FFLs with AND regulatory logic and no basal activity of Y, in response to on and off steps of Sx. Note that type 4 FFLs can produce a strong pulse that is enabled by Sy. Type 1 can produce only a weak pulse when Sy = 1, and the pulse-like nature of the response is lost when Sy = 0. Simulation parameters: Kxz = Kxy = 0.1 for type 1, Kyz = <0.01, 0.1, 0.3> for type 4, Kyz = <1, 0.3, 0.1>(thick, medium, thin lines) all others are as stated in Materials and Methods.

Kinetics of Incoherent FFL with Basal Y Activity: All Four Types Are Sign-Sensitive Accelerators. We now consider the incoherent FFL where the affect on Z by the indirect path through Y is not complete. In the type 1 incoherent FFL-AND, for example, upon a step addition of Sx, Z expression first rises, and then when Y levels build up, Z expression decreases to a nonzero level (Fig. 4Left).

Kinetics of incoherent type 1 (Left) and type 4 (Right) FFLs with basal Y activity and AND regulatory logic, in response to on and off steps of Sx. Note that the response of the FFL to on steps (thick, medium lines) is faster than that of a corresponding simple system (thin line). Simulation parameters: for type 1, Kxz = 1, Kxy = 1, Kyz = 0.5, By = <0.5, 0.3> for type 4, Kxz = 1, Kxy = 0.1, Kyz = 0.5, By = <0.45, 0.35> all others are as stated in Materials and Methods.

We find that the response time of the incoherent FFL is smaller than the response time of a simple regulation system (Figs. 1b and 4). To make a mathematically controlled comparison (9), we compare a simple regulation system and an FFL that have the same steady-state Z expression upon addition of Sx (that is, with a Z promoter in the type 1 incoherent FFL that is stronger than in the corresponding simple regulation design, to compensate for the repressing effect of Y on the steady state). The simple regulation design has a response time of one lifetime (17) of protein Z (Fig. 4, thin line). The response time of the incoherent FFL is shorter (Fig. 4, thick and medium lines). The accelerated response occurs because Z initially rises quickly because of its relatively strong promoter and is then stopped by the repressor Y. Thus, in cases where speedy responses are needed, an incoherent FFL has an advantage over simple regulation with the same steady state.

The acceleration of the response is sign sensitive. We find that all four types of incoherent FFLs show sign-sensitive acceleration. In types 1 and 4, for example, the response is accelerated for on steps of Sx, but not for off steps (Fig. 4). Types 2 and 3 show sign-sensitive acceleration for off but not on steps of Sx (Table 1). The acceleration is tunable and controlled by the same parameters that control the delay in the coherent FFLs. For example, decreasing Y basal activity enhances the acceleration.

FFLs with OR-Gates Have the Same Functions but with Reversed Sign Sensitivity Relative to FFLs with AND-Gates. The discussion thus far considered FFLs in which X and Y act as an AND-gate to regulate gene Z. We now consider the effect of an OR-gate, where either X or Y is sufficient to express Z. We find that the FFLs with OR-gate regulation have the same sign-sensitive acceleration or delay functions, but with the sign sensitivity reversed relative to FFLs with AND gates. For example, the type 1 coherent FFL with an OR-gate shows a delayed response to off steps of Sx and a rapid response to on steps (Fig. 5 and Table 1).

Kinetics of coherent type 1 with AND (Left) and OR (Right) regulatory logic at Z promoter. Note that the AND FFL has delayed response to on steps, whereas OR FFL has delayed response to off steps. FFL: thick, medium lines simple system: thin line. Simulation parameters: Kxz = 0.1, Kxy = 0.5 for AND, Kyz = <0.5,5> for OR, Kyz = <0.7,0.3> all others are as stated in Materials and Methods.

Incoherent FFLs that are poor pulsers with AND-gates, namely types 1 and 2, are better pulsers with OR-gates. Conversely, the good pulsers with AND-gates, types 3 and 4, are poor pulsers with OR-gates. The steady-state behavior of OR-gate FFLs is more intricate than that of AND-gate FFLs, because more intermediate states of expression are generally found.

Only Coherent Type 1 AND-Gate FFL Shows Increased Apparent Cooperativity. We checked the effect of the FFL on the cooperativity of Z induction as a function of Sx (12), both analytically and by using simulations (Fig. 6). We found that only the type 1 AND-gate FFL shows a non-negligible increase of the apparent cooperativity. This effect occurs at low Sx levels, where the effective Hill coefficient for type 1 AND FFL is proportional to Hxz + HyzHxy, where Hij is the Hill coefficient for gene j by transcription factor i. The other FFL types, including coherent type 1 OR-gate FFLs, showed no significant increase in cooperativity (some types even reduce apparent cooperativity). We note that simple transcription cascades are known to increase cooperativity (23).

Apparent cooperativity of steady-state Z response as a function of X activity. The graph shows the z(x) response curve for type 1 (thick line), type 4 (thin line) FFLs, and a simple regulation system (○), for Hzx = Hyx = Hzy = 2. Simulation parameters: αi = 1, βi = 1, Kij = 1, Bi = 0. Type 1 coherent FFL (thick line) has an effective cooperativity of Heff = Hxz + Hxy*Hyz, where Hij is the Hill coefficient of the regulation reaction of protein j by protein i. Other coherent FFL types have Heff = Hzx.


Tasks involved in finding network motifs typically include the definition of frequency concepts, random graph generation, determining statistical significance of the frequency of a sub-graph and deciding sub-graph isomorphism, etc. A few definitions follow.

Sub-graph frequency

Frequency here refers to the number of matches of a query sub-graph in a network [ 27]. Three different frequency concepts were discussed in Schreiber and Schwöbbermeyer [ 28, 29], F1, F2 and F3, where F1 allows arbitrary overlapping of nodes and edges between two sub-graphs F2 only allows node overlapping and F3 does not allow any overlapping of nodes or edges. Figure 2, Figure 3, and Table 1 illustrate variations in sub-graph frequency based on different frequency concepts.

Autoregulation as a Network Motif

Suppose that gene X regulates itself via activation or repression. This is represented symbolically as for activation and for repression.

Going back to the equation for regulation,

Here we replace Y by X. Also we replace &beta(X) with f(X) because is a fast reaction. We get the following differential equation for autoregulation,

where the type of hill function f(X) depends on whether the reaction is or . We call this an autoregulated or self-regulated circuit/network.

Q. Why is this arrangement important in transcription networks?

For this we introduce the idea of "network motif".

Network Motif

Take a transcription network .Try to spot a "motif".

By evolutionary processes, different edges are being generated or killed at random.

Q. Which patterns are "significant" or "accidental"?

We can generate a random network ( Edros-renyi) to see what a randomly generated transcription network looks like

Recipe/Algorithm for creating random network

Given N nodes ( Here proteins/genes) and E edges ( E will be the # of edges in the network),

2) Pick another node randomly.

3) Put an edge from the 1st to 2nd node.

4) Repeat the process E times.

Consider the probability of having a self-edge,

(A node has N choices to chose between)

Average number of self-nodes of E-edges

Standard Deviation (Assume Poisson Distribution)

Example: For N = 424 , &mu = 1.2

In this case, expected self-edges = 0 , 1 , 2

But let us consider a real E.Coli Network with the same number of nodes and edges,

So, there must be a reason nature keeps auto-regulation in transcription networks. Patterns such as auto-regulation that are extremely hard to explain as "evolutionary accidents" are called "Network Motifs". Examples of other network motifs are:

2) Two node feedback loops.

There are many more network motifs (see ref). Motifs can also be discovered in other types of networks.

Autoregulation, Feedback and Bistability

Download the video from iTunes U or the Internet Archive.

Description: In this lecture, Prof. Jeff Gore continues his discussion of gene expression, this time with a focus on autoregulation (when a gene regulates its own expression). He begins by discussing the network motif, then moves on to both negative and positive autoregulation.

Instructor: Prof. Jeff Gore

Introduction to the Class a.

Input Function, Michaelis-M.

Autoregulation, Feedback an.

Synthetic Biology and Stabi.

Oscillatory Genetic Networks

Graph Properties of Transcr.

Feed-forward Loop Network M.

Introduction to Stochastic .

Causes and Consequences of .

Life at Low Reynolds Number

Robustness and Bacterial Ch.

Robustness in Development a.

Microbial Evolution Experim.

Evolution in Finite Populat.

Clonal Interference and the.

Fitness Landscapes and Sequ.

Survival in Fluctuating Env.

Parasites, the Evolution of.

Ecosystem Stability, Critic.

Dynamics of Populations in .

The Neutral Theory of Ecology

The following content is provided under a Creative Commons license. Your support will help MIT OpenCourseWare continue to offer high-quality educational resources for free. To make a donation or view additional materials from hundreds of MIT courses, visit MIT OpenCourseWare at

PROFESSOR: So today, what we're going to do is we want to talk about this idea of autoregulation, so what happens when a gene regulates its own expression. And I think that this is a really nice topic because it really encapsulates several of the big themes that we're going to be seeing throughout the semester.

So first there's this idea of a network motif. So this motif, it's the simplest possible motif where a gene regulates itself. It occurs more frequently than you would expect by chance. We'll explain what we mean by that. And then possibly an evolutionary explanation is that this thing may be advantageous in one way or another.

Now, there are two basic kinds of negative autoregulation, or sorry, two basic kinds of autoregulation. There's negative and positive. So negative means that that protein represses its own expression. Positive autoregulation would mean that it is activating or upregulating its own expression. So down here is positive autoregulation. And these things have very different possible purposes.

All right, we'll see that there are basically two things that negative autoregulation may do for us. First is that it speeds the on time relative to simple no regulation. But the other thing is kind of an interesting thing, which is that the protein concentration is somehow robust. And robust means that it's comparatively non-sensitive or insensitive to something. And particularly, this means protein concentration is robust to say fluctuations in the production rate.

In various contexts, we're going to see how robustness is an important goal, that biology might have, a particular cell might have. And that you would like this protein concentration to be robust to a wide range of things. But in this case, anytime you talk about robust, you have to say all right, it's robust. What is it that's robust against what? In this case, its protein concentration is robust, the fluctuations in the production rate.

Now, positive autoregulation, on the other hand, in some ways makes these two things worse. We'll discuss what we mean by that. But it perhaps performs some other functions. In particular, it allows for the possibility of bistability, and hence memory. And we'll explain how memory appears in this process.

But before we get into autoregulation, I just want to do a quick review of this thing about getting intuition or binding of activators, for example, to the promoter. But also how we can use this thing about sequestration in order to generate what we call this ultra-sensitive response.

Now, when we talk about this, we're often thinking about a question that we'd like the rate of expression of some gene to be sensitive or be ultra-sensitive to the concentration of something. So if you have x activating y, then what you might like, in many cases, is for there to be an essentially digital response, where if you think about the production rate of y as a function of this input x, you might like it to just be very sharp, to be essentially 0. Quickly to come up, and then saturate at some level, beta maybe.

Now, the question is how is it that you can do that? One thing that we discussed that could get you something kind of like this. Does anybody remember what the base solution might be?

AUDIENCE: Cooperative regulation.

PROFESSOR: Right, cooperative regulation. So if it's the case that you have, for example, a tetramer of x that is somehow activating y, then you can get something that looks, maybe not this sharp, but much sharper than what you get, which is the simple Michaelis-Menten-type looking curve if you just had a single monomer of x activating y. So one solution is indeed cooperativity.

But I think that there's another interesting one that is this-- second solution is this idea of molecular titration. I kind of explained the basic idea at the end of the lecture on Tuesday. But then what we want to do is try to get a sense of what the requirements might be for the bindings in order to generate this effect. So what we have is a situation where we have, again, just x is regulating y. But it's a little bit more complicated, because this protein x, in addition to binding the promoter and activating expression of y, we also have the possibility that x can bind to something else. In particular, there might be some other protein, maybe w, which combined reversibly into some complex wx.

Now, for some situations-- in particular, if we describe this by some Kw and this by some Kd we can ask in what regime will this generate an ultra-sensitive response, where as a function of the total concentration of x-- So we might want to call that x total, just be clear. We'd like it that there's very little expression of y, and that all of a sudden to get maximal, if you like, expression of y. And the question is, well, what do we need in terms of these KwKd's, and so forth.

So there are going to be three different conditions that we're going to play with. First might be the relationship between this Kd, the affinity of binding of that transcription factor to the promoter, as compared to this Kw. I will write some options, but as always, you should start thinking. You don't need to watch me write.

Don't know again. So I'll give you that. Just 30 seconds to think about it. Yes.

AUDIENCE: Is Kw the dissociation constant? Like if the concentration w times concentration x over the--

PROFESSOR: Right, so it's the guy that has units of concentration. So lower in both these cases corresponds to tighter binding.

I'll let you think for 20 seconds maybe.

Do you need more time? And it's OK if you are not 100% convinced of something. It's useful to just make a guess and then we can discuss. Shall we go for it? Everybody have your tools in front of you? All right, ready. Three, two, one. So we have, I'd say, a fair amount of disagreement, actually. I'd say we got some A's, B's, C's. I don't know. There might be a one or two D's. Why don't we go ahead and turn to a neighbor there. You should definitely be able to find a neighbor that disagrees with you. And try to convince them that you--

AUDIENCE: What's the question?

PROFESSOR: Oh, yeah, sorry. No, no, all right, so the question is, what conditions do these things need in order to have an ultra-sensitive response, where the rate of production is what?

PROFESSOR: All right, why don't we go ahead and reconvene. And maybe I'll just get a sense of what the state of your thinking is right now. All right, so once again, the question is, we want to know what the relationship between all these binding affinities concentrations has to be in order to get an ultra-sensitive response, where the function of the total amount of x-- you add x. At first, you don't get really much of any expression of y, but all of a sudden, you get a lot. And first, we want to know the relationship between Kd and Kw. Let's go ahead and vote. Ready? Three, two, one.

Oh, I'm sorry. We're still on this one. So yeah, you could ignore these. It's just that it takes me time to write. So I wanted to take advantage of your discussion. We're still on this one. Do you guys understand what I'm trying to ask? Okay, all right, ready. Three, two, one. All right, I'm going to cover this so nobody gets confused.

All right, so we have a fair pocket of C's back there, but then some other distributions. So we've got some discussion over here. So maybe somebody can offer their insight?

What did your neighbor think?

AUDIENCE: My neighbor thought that at first, there should be the binding between x and to the gene. So x should bind much more equally with w. That's why Kw should be--

PROFESSOR: All right, so you're arguing that you want Kw to be much smaller than Kd here. Because as you add x, you want initially for them all to be sequestered by molecule w. Is that OK? So if somebody's neighbor thought it should instead be B, do you want to offer their argument? Everybody's neighbors now convinced that C is--

So I'm going to side with that. So the idea here is that what you'd really like is-- The principle of this is that Kw is perfect, and then that corresponds to going to 0. Then that means that if you plot the free x as a function of x total, then what does it start out being if Kw is 0? 0, right? And when thinking about the concentration of free x, do we have to think about or worry about how much x is down to the promoter? No, we'll say that so long as you have any reasonable concentration of x, then the one x that binds to that promoter, binds that DNA, is not going to affect the concentration of free x.

So we really just have to worry about how much is bound to w. So it's going to be 0. And this is in the limit of Kw going to 0. And so when is it that something starts happening?

AUDIENCE: 1x is the concentration of w.

PROFESSOR: Right, so it's when you get n. And is it the concentration of w, or is it-- maybe we can be a little bit more precise? What shall I actually write here? Should I just write w, or should I write-- w total, right. Because the free w is changing all the time. And indeed, at you approach here, the free w goes to--

PROFESSOR: 0, right. So over here, the free w is equal to wt. But then it decreases linearly until you get here. And now, there's no free w. And it's at that stage that this free x is going to go up with a slope of what? 1, right? Right, exactly. All the x that you add just turns into free x, so this goes up with slope 1.

So this is the idea behind this mechanism. What I'm drawing is the case of perfect sequestration. And in this case, what happens is, if you look at the production rate of y, it's a function again of x total. The production rate in here is to be equal to what? 0. Until you get to this wt, and then it's going to come up. And how is it going to come up? Does it immediately come all the way up to maximal? No, so what determines how rapidly that's going to happen?

AUDIENCE: The concentration of free x.

PROFESSOR: The concentration of free x. And we actually know what the concentration of free x is here now.

PROFESSOR: Right, so it's the Kd of the binding. So this is going to be some-- so this is again a Michaelis-Menten looking kind of curve, where the half maximal here is just going to be this wt plus that Kd.

All right, so there's a sense that this is ultra-sensitive, because initially you don't get much of any expression, then all of a sudden you start getting significant amounts.

Are there any questions about the basic-- the intuition here? We still have come back and think a little bit more carefully, kind of quantitatively, about what this mechanism means for the other comparisons? But at least, based on this idea of the perfect situation, where the sequestration of w with x is complete, then this is the idea of what's going on. Is it clear right now? Yeah.

AUDIENCE: What's the direction of Kw in this situation? Is that--

PROFESSOR: In what I'm drawing here?

AUDIENCE: What you drew there.

PROFESSOR: This one? Oh, OK, so the definition of kw is going to be that kw-- and this is a definition. It has units of concentration, so that means we have to put up-- this a conservation of w, a concentration of x, the concentration of wx.

PROFESSOR: Any other questions about this? Yeah.

AUDIENCE: Why does the flow go the free x 1, because isn't the free x now binding with the gene?

PROFESSOR: OK, right, so this is what we were saying is that there's just one copy of this gene. So given that there might be-- where in this case, this might be 1,000 of these proteins, and then it's just not a significant-- yeah, in the back?

AUDIENCE: So whenever you have a k, do have always concentration?

PROFESSOR: OK, so in this class whenever possible, at least in the lectures, we'll always try stick with the dissociation constant, which is the guy that has units of concentration. You can also talk about the association constant, and then they also use k for that. Horribly confusing, right? So whenever possible, I'll be referring to the dissociation constant.

AUDIENCE: Is that written a lower case k or a big K?

PROFESSOR: Oh, that's right. No, it's a good thing that my K's are so clearly big K's that there's no possible source for confusion there. Yes. But if you're ever confused about which K or whatnot I'm referring to, please just ask. All right, any other questions about the basic mechanism? Yeah.

AUDIENCE: Just a correct one. So if we take the other limit for Kd going to 0, would it be the same-- so free x versus x total, would it be the same slope, but starting from 0 that should be left?

PROFESSOR: Oh, OK. The other limit of Kd going to 0. Yeah, this is an interesting point. Right, so I think this plot is independent of Kd, because binding to that promoter doesn't affect the free concentration of x anyways, right? It does affect this, though, because if Kd gets smaller and smaller, then this curve actually gets more and more steep. Does that answer your question?

AUDIENCE: Isn't that more sensitive if the curve is deeper?

PROFESSOR: OK, right, OK. Yes, OK, that's a good point. But now you're taking two limits of things going to 0, so we have to be a little bit careful, because it still maybe depends on the relationship between those two as they go to 0. No, that's fair. But the idea is that if you're in this other limit, then you end up not having significant sequestration in the sense that as you start adding x, you start getting expression of that protein early on. And so then the whole mechanism doesn't work. So it's true that in principle that thing is steep, but it was never inhibited to begin with. Because the moment you start adding x, you start getting expression from the gene. Other questions about that?

Let's try out this next one. Now, this is a question of K-- so this is the binding affinity of x to that sequesterer as compared to the total amount of that w. All right, so what's going to be the requirement there? So we'll go ahead and give you 30 seconds to--

AUDIENCE: Excuse me, but don't we already have what we want? What's the more specific question?

PROFESSOR: Right, OK, so here, this is the case where w went to zero. Sorry, Kw went to 0. Right? And indeed, this works. And the question is, in general-- Kw is not actually going to be equal to 0. And it might be relevant for this question still. Yeah, because in some ways, this is the idealized version, and any real thing is going to have just some numbers for these things. The question is, will those numbers work.

Do you need more time? Let's go ahead and make our best guess. In order to get ultra-sensitivity response of y to the concentration of x, what is that you need here? All right, ready. Three, two, one. So we got some A's, some B's, some D's. C is not very popular. OK, so it seems like-- Well, why don't we just go ahead and spend 30 seconds? Turn your neighbor. You should be able to find someone who disagrees with you.

PROFESSOR: OK, why don't we go ahead and reconvene, and you can give your argument to the group? All right, does somebody want to volunteer an explanation? Yeah.

AUDIENCE: I haven't had time to test this explanation.

PROFESSOR: That's just fine. Well, your neighbor will appreciate it.

AUDIENCE: So I was thinking that the time scale of that axis really on the second graph is set by how quickly that slope, the curving slope, rises?

PROFESSOR: OK. so times scale I'm a little bit worried about.

AUDIENCE: Sorry, concentration scale.

AUDIENCE: It's set by how quickly the second curve rises, so once you go above 0. That's the only feature on that graph. And that should be in a Michaelis-Menten curve that's determined by K.

AUDIENCE: Oh, the K of the Kd.

PROFESSOR: OK, right, so I think I like everything you're saying, although I think you're about to answer the next one, although the next one, I think, is in some ways the hardest one, so you're ready.

AUDIENCE: Right, which we determined should be a lot more than Kw.

PROFESSOR: Right, so we know that Kd should be a lot a lot more than Kw.

AUDIENCE: And I think that this should be a decent amount less than-- Sorry, so I think that Kd should be a decent amount smaller than w2.

PROFESSOR: All right, so you think Kd should be a decent amount smaller than w2. OK, so you really are answering the next question.

AUDIENCE: But no, I think this is important.

PROFESSOR: Oh, no, I agree it's important. But--

AUDIENCE: Basically speaking, if you let wt become too small, then it compares into everything else, all the relevant quantities, then this feature, which is really determining the over-sensitivity strengths have 0 side. That's what I wanted to say.

PROFESSOR: OK, yeah. I very much like that explanation. And since we have it on tape, we can now play it again in a few minutes and then-- I think you're making a very nice argument, and actually, we'll go ahead and even--

Does anybody want to argue against. Yeah. So what he's arguing is that what you-- So he's actually arguing that it's this over here, which is that what sets the scale that this thing increases by is the Kd. And you want this thing to come up quickly relative to this other scale, which is the total w. So if total w is just too small here, then it's not like you get any ultra-sensitive response, because you want it to be low, low low, and then come up kind of quickly. So there's some sense that you want this scale to be of the order of this scale or maybe even a bit shorter. Is that-- and I agree that this actually one that generates a lot of argument an discussion in general, because I think that reasonable people can disagree. I think this is actually-- I will side with you on this. But what about this, because you haven't said anything about KW.

AUDIENCE I think that combining one and three.

AUDIENCE: That's why I was trying to answer--

PROFESSOR: OK, no. OK, fair. But it's just too many logical leaps. I think it's true. And there's a reason I ordered the questions in this way, so that is so this one you could-- Otherwise, I agree that if you combine these things, you do get-- and which one do you get?

PROFESSOR: Is that what other people got? OK, I agree that actually you can get to A here from the combination of C and B here. But it's a little bit crazy. Is there a more direct explanation we can get? Yeah, go ahead.

AUDIENCE: Well, if you don't have a lot of w, then your binding is too sensitive in a way. You'll add a little bit of x, and then the binding will be all used up, and you'll be making y immediately. You want a lot of w to be able to soak up a sizeable amount--

PROFESSOR: Yeah, and then maybe in the back.

AUDIENCE: So originally, I agreed, but that just makes you switch, like Evy said, you can sequester less, but that still gives you a sense of [INAUDIBLE]. You're more sensitive to fluctuations if you have a lower sequestration capacity, but it doesn't really change-- You still have that switch--

PROFESSOR: Sure, OK. So I think that this is actually [INAUDIBLE], because there are two ways of having more of this wt helps. One is that it pushes this boundary further to the right. But the other is that it makes it a better sequestering agent because there are just more of these w's to bind. And if you go and you do you do the calculations, and I encourage you to do it, what you actually find is that the concentration of free x is given approximately by x total divided by wt over K sub w. So what you see is that the free concentration of x-- and this is in the region. This is for x total less than and maybe significantly less than w total.

But the idea is that when you're in this sequestering regime, what we see is that the concentration of x is going to increase linearly with the total x. But it's going to be sort of suppressed by this amount here, where this ratio is much larger than 1. You want to be much more than 1, wt over Kw. And this is telling you how good of a sequesterer this guy, this w is, because you want to bind tightly and you need to have kind of enough to keep the free concentration of x from growing.

AUDIENCE: So that's what I was thinking as well, but then I was wondering if there's too much of w, then wouldn't x never bind completely, but always get sequestered?

PROFESSOR: So this is a very important point, which is that it needs to be possible for the cell to make enough x to overcome the w. If there's so much of the sequestering protein that you cannot even get beyond that, then you're never-- It's true that you might have a nice ultra-sensitive switch out there, but you just can't get there. So yeah, this is certainly relevant as well.

All right, incidentally, if you plot free x as a function of xT on a log-log scale, the question is what is it going to look like? Now we, know that on a linear scale, it looks something like that. So if it's on a log-log scale, question is-- And log-log is nice because you can really get the full dynamic range of what's going on. If you plot x and this is free x, as a function of x total in this regime for large X and x total, what is it going to end up looking like?

AUDIENCE: It's a straight line.

PROFESSOR: It's a straight line. And eventually, it's going to be straight line with slope 1, actually, because they're just linearly related to each other. Now, down in the low regime, what is it going to look like well below this region of wt?

AUDIENCE:You should also get a straight line with 1 slope.

PROFESSOR: With what slope?


PROFESSOR: OK, so why does everybody agree? No, we have a disagreeable-- Right, so what's the--

AUDIENCE: It's going to be also slope 1, but just lower.

PROFESSOR: OK, right. So it's actually also slope 1, but it has a lower-- This is dangerous. This is why I brought this up, because it's really easy to look at these things and think that-- Yeah, because they're still linearly related to each other. So when you take the logs, it just affects the level. So you still get the same slope, but it's going to be down here somewhere. And then in this regime, you get this ultra-- this thing where it goes up suddenly.

One thing that I strongly encourage everyone to do is, in these sorts of problems, it's wonderful to spend time-- Oh, sorry. This is log. Free x log of xT. It's really very valuable to plot things in multiple different ways, by hand or by the computer or both or whatnot, just to make sure that you're keeping track of what's going on. Because often what you see and what you think is very different depending on what you plot. And you'd like to be able to see your problem from as many different angles as possible.

All right, I think that we've probably spent about as much time on this as we ought to. But are there any other questions about how this is going to come about? Yes

AUDIENCE: Are there any negative autoregulation that can use sequestration to get a switch-like [INAUDIBLE]?

PROFESSOR: OK, that's an interesting question, although we have to be careful about-- If you really want it to be more switch-like, you'd probably use positive autoregulation. And I'm not aware of a case where this has been combined, although-- It's likely there are some. I just don't know them.

I'm going to switch gears to this autoregulation, which is something, of course, that you guys just read about. And it looked like your understanding of it was solid. But we want to move through these different ideas. First, this idea of a network motif. And this is just the simplest example of a network motif. And it's so simple, we often don't even call it a network motif. But the idea here is that we have some network, and it has maybe N nodes and E edges. And the example that they give, that Uri gives in his book, N was 420, and E was 520. ,

And there's a basic question that if you have a network with N nodes and then you have directed edges-- these are edges that have an arrow pointing on one end. Now, in that case, and if you allow self-directed , edges how many possible edges are there? Does anybody remember what this looks like?

AUDIENCE: More than 20 possible, right?

AUDIENCE: There'd be more than 20 possible.

PROFESSOR: Right, not even in terms of the actual number. Just in terms of N, for example.

AUDIENCE: Total or just self-directed?

PROFESSOR: Total self-directed. Or Sorry, total directed edges, total number of possible directed edges, if we included self edges, I guess. N squared. And you can think about this in multiple ways. One is just that, well, you can start at any of the N nodes. And you can end at any of the N nodes, and that gives you N squared. Right?

But there's another way that you could-- For instance, this is Emacs. It's NE N squared. You could also think about this, if you like, as just-- say, well, as they point out, there's sort of-- 1/2 N times N minus 1 is the total number of pairs in the network. And then for each of the pairs, you can have an edge pointing either direction, so that gives you a 2. Plus, you can have N self edges, right? Now, of course, these are just different ways of counting N squared. But it's useful to just think about it in different ways to make sure that you're comfortable with the combinatorics of it. In particular, because next week or the week after that, we're going to be talking about network motifs in more detail, in particular the feed-forward loop. And then we really have to keep track of these sorts of combinatorics better.

The way that Uri thinks about this is he says, all right, well, we're going to invoke this null model of a network, this Erdos-Renyi network, where we just simply say we're going to assign the E edges randomly across the Emacs possible. So of all possible edges, we're going to place the edges randomly and then I generate some random network. So this is basically what we typically call a random network. And that's going to allow us to define some null model that if everything were random, we can ask, for example, how many self edges do expect to get?

Well, one way to construct this sort Erdos-Renyi network-- Yeah.

AUDIENCE: But how do you know if you have other kind of constraints in the system, how do you know that transcription might require in some cases autoregulation--

The answer that was give that was a good answer was that evolution is the only thing that can-- if you find a network motif that has to do with evolution--

PROFESSOR: So that is the argument that Uri makes, and you're maybe saying maybe that's not a good argument. And what Uri is saying as well, if you see these networks more frequently than you would expect by chance-- and of course, you can define what you mean by chance-- then you can say oh, maybe it was select for, it was evolved. And I think that, in the case of this autoregulation case, I think the results are not very sensitive. But I think that this question of what the right null model is, is a real issue, especially when you're talking about some of these other networks.

And what we'll see for the feed-forward loop is that you have to decide-- well, one thing we're going to see is an Erdos-Renyi random network is very much not an accurate description of real transcription networks. So then you could say, well, that's not a good null model to use. And so we'll definitely spend some more time thinking about this.

In the context of the Erdos-Renyi network, though, one way that you can generate it is that for each of the Emacs possible, each of these total number of edges, there is some probability that you're going to actually place a real edge there. And that probability is just E over N squared. E is the number of edges. N squared is the number of possible edges. So if you just create a random network in that way, then this is a manifestation of a random network that has at least the basic properties, the same number of edges as our network.

So from this, you say, well, how many self edges would you expect in this world? And you'd say well, in that case-- There are two ways of thinking about this. So you can either say, we're going to take, for each of the N edges, there's one possible self-directed arrow. And for each of those cases, we can just multiply this by P. And this gives us E over N.

You could also think about it as-- There are multiple ways once again, of doing the counting. In an Erdos-Renyi network, you can say, all right, you would expect to get roughly E over N autoregulatory loops.

And this is of order 1. So this is 1.2, in the case of the network that Uri analyzes in his book and his paper. Whereas, how many were actually observing in the network that he studied?

PROFESSOR: There were 40, right? So in the observed transcriptional network-- and this is in E. coli-- he found that there were 40. And the basic statement here is that 40 is just much larger than 1.2.

And you can quantify this a little bit better, because, it's really you would expect 1.2 plus or minus the square root of this, in a random network like this. But that is, you'd expect 0, 1, 2, maybe 3. So 40 is definitely not what you would expect based on an Erdos-Renyi network. So this is the sense in which it's a network motif. It's that the observed network just doesn't look like a random network in this very particular sense.

And of these 40, does anybody remember kind the distribution between negative autoregulation and positive autoregulation?

AUDIENCE: It was 30-10 or something.

PROFESSOR: Yeah, I think it was like 34 and 6 was my recollection. I didn't write this down. So most of these guys have a form of x inhibiting x. But some had the form of x activating x. So this was something like 34 and 6. What you would say then is that negative autoregulation is a very strong network motif, whereas positive autoregulation is a weaker network motif, but still something that occurs perhaps more than you would expect by chance. Are there any questions about that where that argument came from, other concerns about it?

All right, then from this, then Uri says, OK, well, maybe these things evolved for a reason. And so what we'll do in the next half hour just argue or discuss what possible evolutionary advantages such a social network motif might have. Yeah.

AUDIENCE: Can you really propose adversarial explanations like that?

PROFESSOR: Oh, you can propose whatever you want.

AUDIENCE: I know, but it doesn't have much value. It's unquantifiable.

PROFESSOR: Yes. No, I'd say this is a major issue in a lot of evolutionary arguments. And I would say that the purpose of ideas and hypotheses is to get you to go and make new measurements. And so right now, we're the stage, OK, well, maybe these things are occurring more frequently than you would expect by chance. So now, we can just sit down and think, oh, what advantage might it give. And then we can go and try to experimentally ask whether those advantages are at least manifested in real systems. It doesn't prove that that's why they evolved, but it makes you more comfortable with the argument. feel

Ultimately, we assign some-- we have some agent probability somewhere in our brain. And the more evidence that we can accumulate that's consistent with these ideas, the more likely that we think it is. But in general, you don't prove things in this sort of evolutionary space the way you prove things in many other fields. Yeah.

AUDIENCE: I feel like it's hard to call this an argument. It feels more like just an observation.

PROFESSOR: Which thing is an argument versus--

AUDIENCE: I guess the thing is it should be evolutionarily advantageous, that's an argument, but essentially, the whole thing is an observation, and then there's a little bit of an argument in the end.

PROFESSOR: Yeah, I will let each person decide what fraction and observation. Yeah, I don't feel especially strongly about it. My guess is that it did evolve it because it provides some useful function. And therefore, I think it's valuable to explore what those useful functions might be. But for example, it's very hard to know which of these explanation-- this thing about increasing the response time, or sorry, increasing the response rate as compared to increasing robustness, how do you decide which one's more important? Then I think, once again, reasonable people can disagree about these things, yeah.

So first negative autoregulation, because this is the one that is the stronger network motif. I think that the book does a nice explanation of why it decreases the response time. OK, we can just ask. OK, response time-- and this is for a negative autoregulation. Response time goes down. And is this for turning on, off, both, or maybe neither, or E, don't know.

And I'll give you just 10 seconds to think about this. It's nice if you just remember it, but it's also maybe even better if you can figure it out. Because in a week, you're probably not going to just have it memorized, but you should be able to think through the logic of it and understand why this is going to be what it is.

All right, so the question is, negative autoregulation, maybe it does something. Maybe it decreases the response time. But does it decrease the response time for turning a gene on, for turning it off, for both, neither, or don't know

AUDIENCE:When you say turning it off, what exactly is the process you're imagining.

PROFESSOR: I'm imagining a process where the expression turns off immediately. So there's a signal that just stops--

AUDIENCE: what transcription can go ahead.

PROFESSOR: Right, so then it's just I chop up all the polymerases, and no more expression. But so a signal comes and tells the polymerases to stop making. Yeah.

All right, so do you need more time? No. Ready, three, two, one. All right, so we actually are all over the place on this. OK, turn to your neighbor. And you should be able to explain one way or the other why thi-- what is going on.

Let's go ahead and reconvene. I just want to remind everybody that when I say simple regulation, there's no autoregulation. It's just responding to a signal. That for a stable protein, the time to get to say, for example, half saturating concentration here is defined by the cell generation time. And that's true for turning on and for turning off. And what was the strategy that you could use if you wanted to decrease the response time in this situation?

AUDIENCE: Increase the degradation rate.

PROFESSOR: Right, so you could increase the degradation rate. And does that the on, off, or both?

PROFESSOR: Both. But there's a cost, which was what?

AUDIENCE: You have to make protein.

PROFESSOR: Right, you have to make a bunch of protein, and then you're just going to chop it up right after you make it. There is a reasonable-- there is a way to make things faster, but it has a significant cost.

The question is, if you have negative autoregulation-- so in this case, you have x that is repressing itself-- what is it that it's going to do? Is it going to affect the on time, the off time, or both. Let's just see where we are. Ready, three, two, one.

OK, so it's interesting. We're moving towards C, it seems. OK, so can somebody give me an explanation for C. Did we read the chapter?

AUDIENCE: Well, the chapter doesn't discuss the effect of negative autoregulation and turning off. I don't think it does.

PROFESSOR: Wow, it's a good thing we're doing that here then. All right. So first of all, can somebody give the explanation. Does T on go up, down, or sideways.

PROFESSOR: So T on-- It's the time that goes down. I always get this confused. So time is the one that goes down, so the rate goes up. Negative autoregulation is faster turning on, we decided. Right? And does somebody want to give the explanation for why this is?

AUDIENCE: Well, your equilibrium level is lower.

PROFESSOR: Yeah, right. Yeah, exactly. Yes, this is actually surprisingly difficult to explain even though it's not a deep concept. But the idea is that you start out expressing a lot, so that if you had kept on expressing that high level, you would have done some exponential-- It would have take cell generation time from way up here. But instead, what happens is that you shoot on up. But then, once you get up here, you repress expression. So then you get an effective thing, where the time it takes you get half of your equilibrium, that goes down. So Tl in here is shorter than here. Yes.

AUDIENCE: So in negative autoregulation, for decreasing what the book calls beta, to have the same steady state?

PROFESSOR: That's right. The initial beta, that rate, that maximal rate of expression, that goes up in a case of negative autoregulation. But then you start repressing expression once your concentration of x here gets to some reasonable level.

So now we're just talking about production rate of x. And that's as a function of x. And of this logic approximation is when it just is maximal until it gets to some K and then is completely repressed. So real versions will be much smoother, but this is just useful to start getting the intuition.

And the idea is that you shoot up to this K, and then you stop expressing. In this limit, actually, it's not even-- it's like a kink here. Because it just shoots up and then it turns around. But any real system will be smoother. Yes, question.

AUDIENCE: So if you get to a certain equilibrium level, then in autoregulation, you would need a stronger promoter.

PROFESSOR: Yes, you want a stronger promoter, because you really want to have high expression initially and then later repress that. So negative autoregulation allows you to speed up turning on, so T on goes down.

AUDIENCE: Without increasing the promoter, which is a good thing, because someone would die if you increase the promoter.

PROFESSOR: This is a very important point, which I was about to get to, which is that this is something that was done-- We could have done that without negative autoregulation by increasing the degradation rate. So the question the that we're bringing up here is, is there that same cost that we were referring to before of this futile expression of protein at equilibrium.

PROFESSOR: So it's actually not. In a cell, you start out expressing a lot, but then later, you actually bring down your rate of expression. And in any case, there's no degradation. in this. The only effective degradation is due to the dilution, the growth of the cell. So if you have the same concentration, then actually, you don't make any more protein than you did here, because you have the same concentration at equilibrium.

So this is neat because this speeds up the response when you're turning on, without the associated cost of making that protein then degrading it. Any questions about that statement?

So now what about off? Is the off time the same as the on time here?

AUDIENCE: Should the off time be slower because you have lots of degradation.

PROFESSOR: Right, and in principle, is there any active degradation that we've invoked on this?

PROFESSOR: Of course, we could have both negative autoregulation and active degradation. But in principle right now, you can have the negative autoregulation without any active degradation. In that case, how long does it take for that concentration to go away when you stop expressing?

AUDIENCE: The cell degeneration time.

PROFESSOR: The cell degeneration time. So this thing actually looks the exact same as this.

So these guides are the same, whereas this one is faster than that one. Because the idea is that the best that-- unless you're inactively degraded, all you can do is you can shut off expression. But then if you turn off expression on the negative autoregulation, it's the exact same thing as turning off expression in the absence of the neg-- in either case, you just stop making protein. So the concentration just goes down because it's being diluted away during cell growth. So this is saying that response time was down only when turning off in the case of negative autoregulation.

Are there any questions about that idea? Yes.

AUDIENCE: With the negative autoregulation, in order to reach the same protein levels, you'd need much greater production rates, correct?

PROFESSOR: Yeah, so the idea is that this beta might be-- so this is the beta of negative autoregulation. It could be much larger than the beta of simple regulation in order to get to the same equilibrium.

OK, so what about this idea of robustness? Well, this is production rate and then degradation rate. So this is an alpha x. And so my question here is, I told you that robustness-- something is robust-- Yeah, question.

AUDIENCE: My question is in this case, you're saying that the off means signal disappears, right?

PROFESSOR: OK, T off is this idea. It's the T 1/2. So this is the time that it takes for the protein concentration to reach half-- to go from halfway the distance from where you were to where you're going to end up.

AUDIENCE: But what if the signal not disappear, but to half of the original signal?

PROFESSOR: So the signal could do a range of different things. And it could be that the signal just changes so that instead of going down to 0, you go down to some other value. Is that what you're imagining?

PROFESSOR: In that case, you still go exponentially to this new value, so actually the T-- the response time there is still actually the cell generation time. So it doesn't matter, in the absence of any these, for example, autoregulation. The time, the characteristic timescale, is always the cell generation time if it's a stable protein. It doesn't matter whether you're going up, down, or all the way to 0 or not.

So the question here is-- OK, x equilibrium is robust to what? And this is to small changes in what? It's going to be A, alpha.

So this is going to be our first example of an advanced use of our cards. So the way that it works is that you can choose more than one. OK, now, this requires some manual dexterity. So what you have to do is if you think that the answer is more than one of these things, then what you have to do is show me more than one card. These cards are amazing, right? You can do so many different combinations. I'll give you 20 seconds to think about it.

AUDIENCE: So what's e? What do those things mean?

PROFESSOR: OK, the question is, are the equilibrium concentration of protein x is robust means it does not change in response to small changes in what quantities? So if I change the degradation rate, does it change equilibrium. If I change the beta. And I'm asking about this case here, perfect negative autoregulation, just so we can try to establish our intuition here. K is this repression threshold. None means that it's not robust to any of these things. DK always means "don't know."

I'll give you an extra 30 seconds. This might be--

So this one's the production rate. This one's the degradation rate. This figure might be useful to you.

AUDIENCE: Can you define K again?

PROFESSOR: Yes, so K is the concentration of the protein x at which this super effective repression kicks in. So we're assuming perfect negative autoregulation. Beta is the rate of expression for low concentrations. The moment you get to concentration K, you get perfect repression and no more expression.

Do you need more time? Question.

AUDIENCE: By saying that x equilibrium is robust, so you mean that when you change these perimeters, x equilibrium stays exactly the same, or will x equlibrium--

PROFESSOR: For now, what we'll mean right now is that a small change in this parameter leads to no change in x equilibrium. Now, for any real example, what we'll typically mean is 's going to be some sort of sensitivity analysis. For example, where you'll say oh, a 1% change in a parameter leads to a less than 1% change, for example. But in this case, there's going to be no change, I'll tell you, just so we can get the intuition clear here. All right, do you need more time? Let's go ahead and vote. Remember, you can vote for more than one thing if you like. Ready, three, two, one.

All right, some people are using our more than one. And of course, I can give you a hint. The reason that I'm letting you vote more than once is because more than one thing is going to be-- All right, so the majority of the group has got this, but not everyone. So let's discuss. Can somebody give an explanation for why both alpha and beta are going to work here?

AUDIENCE: So the equilibrium is basically when degradation involves production.

PROFESSOR: I want to make sure I'm okay. The equilibrium is when the production rate is equal to the degradation rate. So this is a very important thing to make sure we're on top of. And in this case, we have very sharp-- this production. So then what happens?

AUDIENCE: Well, [INAUDIBLE] is the intersection of--

PROFESSOR: Right, so in this case, what is the equilibrium concentration?

PROFESSOR: It's equal to K. Now, I strongly encourage you, whenever possible, to draw things out. Because this is a problem that when you have the drawing. It's reasonable to do. And if you don't have the drawing, you're going to get yourself tied up into weird knots. And indeed, we can see that if we change alpha, what happens in this spot? Right, it changes the slope. And you can see that if we change the slope by small amounts, we get no change where this crossing point is. And even for a real system, if it came around, you'd see that it's going to end up being a less than proportional change in the equilibrium.

And what about what about beta? That just raises and lowers this. And again, that doesn't change the equilibrium. Of course, if we changed K, then we get a 1 to 1 change. So a 10% change in K leads to a 10% change in the equilibrium concentration of x.

So this is the sense in which the equilibrium concentration in negative autoregulation is robust to changes in both-- in the book, they say oh, the production rate, but it's actually also in principle the degradation rate over some range. And this could be useful, because there are lots of things that are going to affect the production rate of a protein, and also the degradation rate, for that matter. The division rate, it changes it.

Whereas it may be that K is subjected to less severe changes, because that's determined by, for example, in the kinetics of binding of this protein to this promoter. And that is perhaps less subject to changes. It can still change depending upon the pH and so forth of the interior of the cell. But at least it's probably not subject to the big changes that alpha and beta are going to be a good experience.

So the argument that Uri makes for why it is we see so much negative autoregulation in the cell is because it both increases the rate that the cell can respond to changes, in the on direction, at least, but also that it makes the concentration of protein more robust to changes in several of the parameters that govern the equilibrium . Concentration And once again, you could argue about which one of these is more important, but I think they're both likely playing a significant role in different cases.

I'm going to want to move on, but I will tell you that only over some range of these -- alpha, beta, K-- will this thing be robust. So for example, if this comes up too high, we're going to lose this phenomenon of robustness. So I expect you to be able to tell me in some later date the conditions in which that might happen.

And I'm available for the next half hour after class, so if you do not know what I'm talking about right there, please hang out with me after, and I'll tell you the solution to that question on the exam. OK? All right.

But I do want to talk about positive autoregulation, because this is another interesting beast. So if negative autoregulation has those nice properties, then you can imagine that positive autoregulation will have some drawbacks in the same kind of ways. But it leads to some other very interesting, just qualitative features.

Positive autoregulation. So we have some x that is activating itself. And often we think about cases where it's activating its own expression in a cooperative fashion. In particular, we might assume that x dot is equal to, for example, some beta 0 plus some beta one of some cooperative thing here where N might be 2 3 4 and then again minus alpha x. Right? Now, if you just look at this, you might think oh, I don't know what this is going to do and so forth. But you've got to draw things out. Once you draw it, then you'll see that it's pretty straightforward.

So again, this is the production and the degradation rates. So that's production. And degradation, for example, might look like this. So this is the production. This is the degradation. So that's the alpha x term.

One question would be, how many fixed points does this system have? So a fixed point means that if you started right there, and in the absence of any noise, you would stay right there. So it's clearly both stable and unstable at these points.

Can you read that? I'll give you 15 seconds to count them.

Ready, three, two, one. All right, it seems like we have pretty good agreement. There are indeed 3 fixed points. Once again, the fixed point is where these curves cross. So we have one right here, one here, and one here,

Now, how many are stable? We're going to do this verbally. Ready, three, two, one.

PROFESSOR: 2. Let's try that again. Ready, three, two, one.

PROFESSOR: 2. Yeah, you get so used to the card, it's hard to speak. So they're the ones on the ends of the stable ones. And you see here that around this point, the production rate over here is more than the degradation rate. That means that if you leave that fixed point, you're going to get pushed away.

So it's very nice to draw these little arrows here to make one happy. So this thing here is stable, unstable, and again stable.

Now, the reason we call this bistability is because there are 2 stable fixed points. This is important because this phenomenon is the basic dynamical system's origin of memory. Now it's, not obvious how memory comes from this.

So memory is a generalist idea that the g-network or the cell can retain a memory of its past state. And we're going to see examples of this over the next few weeks. But just to be clear, if, for example, we imagine a situation where the alpha changes. And it could be division rate, for example, high-food, low-food environments.

What we do is we can plot-- Often, you can plot, for example, the equilibrium, but that's a little bit trickier. So I'm just going to plot the production rate as a function of alpha. Now, the question is, if we change alpha, what's going to happen? Now, for a fixed alpha, you can see already that there are two different production rates that are stable in this case.

But what happens if we increase alpha? So we increase the growth rate so it goes like this. Can that change the number of fixed points?

And indeed, what we can see is that as this line gets steeper here, eventually you only have a single fixed point, and it's stable. And that's known as a bifurcation of the dynamics of the system. So this is for large alpha, you end up-- And just to be clear, this is beta 0 down here. And then up here is the beta 1.

So what we do is we know that beta 0 is where we get for large alpha. Now, for small alpha, do we end up getting another-- We get another bifurcation. So actually, there's only again one stable point up here at small alpha. And what we're going to get is what's known as a full bifurcation where solid lines denote stable points, stable fixed points. Dashed lines represent unstable fixed points. So stable, and the dash is unstable. There are some regions of alpha conditions where the system is bistable. But then outside of that, it's just monostable.

Can somebody explain why this thing-- why I might make the argument that this thing displays memory? Well, one of those two is fine, but any new people want to explain my thought process? No. All right, maybe you.

AUDIENCE: All right, well depending on whether we had high degradation or low degradation rates in the past, we'll be on the lower or the upper range of that if we return to normal.

PROFESSOR: That's right. So the argument here is that-- let's say that this is some normal condition. This is where you are right now, for example. Now, depending upon whether you're sitting here or here, that's perhaps giving some information about the past state of the cell. Because if you were here, that means oh, maybe in the past you were out at high degradation rates, whereas if you're here, maybe were at low.

In particular, you could reset things. If you start here, then you can reset this memory module by coming over here. Once you get to this point here, that's the bifurcation dynamics, the full bifurcation. Then you come up here, and now you'll retain this state. In principle, until you get over here. Of course, there could be stochastic switching dynamics. We're going to talk a lot about that in the coming weeks. But at least in the limit of a low rates of stochastic switching, then this represents some sort of memory module, the simplest version of it.

I'd say that in the cell, most examples of such memory modules involve not just positive feedback of one protein activating itself, although this happens, but often through a whole network, where the one protein activates another, activates another, and then you come back. Or it could be repressing, repressing. 2 0's, is a po-- two negatives is a positive. Just like two lefts is a right.

Right, so are there any questions about the sense in which this thing can serve as a basic memory module?

And this is maybe not the most interesting example of it, because alpha is such a global parameter. But you can also get similar dynamics as a function of, for example, the galactose in the concentration of some sugar in your media.

So given that different small molecules such as food sources can act as inputs into these g-networks, you can also get these sorts of dynamics as a function of what you might call really some simple, external molecule, which is nice, because that means that you can have memory modules that are really independent of all the other memory modules that are going on in your cell. Whereas if you had a vary alpha, then this changes everything. Whereas if it's just a concentration of some sugar outside, then you can imagine that that could be very useful to retain a memory of what the cell has encountered in the past.

So today, what we've been able to do is analyze something about a possible evolutionary explanation for why autoregulation is as commonly observed as is. So negative autoregulation is the one that's observed perhaps most frequently. And that, I think, has some very clear purposes.

And this idea of the concentration being robust to other biochemical parameters I think is a big idea. We're going to see this idea of robustness crop up multiple times over the course of this semester. And I think that it's nice to think about robustness in this case, because it's perhaps the simplest example of how robustness as an approach can be useful as a way of thinking about a problem.

We're later going to be thinking about robustness in the context of perfect adaptation in chemotaxis, where bacteria try to find food. And there, I think everything's more subtle, because already the base phenomenon that is robust is a form of robustness. And so it kind of gets you mixed up. So I think that it's good to be very clear about what robustness means here, so that we can use that to think about robustness in other biological functions.

With that, have a good weekend. Good luck on the problem set, and I'll see you on Tuesday.

Department of Molecular Cell Biology, Weizmann Institute of Science, Rehovot, 76100 Israel

Department of Molecular Cell Biology, Weizmann Institute of Science, Rehovot, 76100 Israel

Department of Molecular Cell Biology, Weizmann Institute of Science, Rehovot, 76100 Israel

Department of Molecular Cell Biology, Weizmann Institute of Science, Rehovot, 76100 Israel

Eunice Kennedy Shriver Institute of Child Health and Human Development, Bethesda, Maryland

Freie Universität Berlin, Berlin, Germany

Institutional Login
Log in to Wiley Online Library

If you have previously obtained access with your personal account, please log in.

Purchase single chapter
  • Unlimited viewing of the article/chapter PDF and any associated supplements and figures.
  • Article/chapter can be printed.
  • Article/chapter can be downloaded.
  • Article/chapter can not be redistributed.


This chapter discusses the progress of the last few years in our understanding of the architecture of bacterial transcriptional networks (TRNs) and the functions provided by this architecture. The rest of this chapter concerns such local structural analysis, focusing on elementary circuits that make up the network. Other sections of the chapter are devoted to network motifs in the TRN of Escherichia coli, their structure and function, aiming to covey the notion that each motif can carry out specific information processing functions. The analysis of the E. coli TRN revealed four main recurring patterns: (i) autoregulation, (ii) feedforward loop (FFL), (iii) single input module (SIM), and (iv) dense overlapping regulon (DOR). NAR of a TRN linearizes gene response and increases the input dynamic range of its downstream genes. The dynamical functions as well as other properties of the two common FFLs: the coherent type-1 FFL and the incoherent type-1 FFL are discussed. As described in the chapter, different functions are assigned to network motifs based on theory and experiments, with new functions continuously emerging. It is likely that additional studies on other systems in both E. coli, as well as other bacteria, will result in the identification of additional functions of network motifs in isolation and in the context of the entire network. A future challenge is to view network motif behavior within the global dynamics of gene networks, and assign certain functions of the network based on network architecture.

Author Summary

Biological modules are inherently context-dependent as the input/output behavior of a module often changes upon connection with other modules. One source of context-dependence is retroactivity, a loading phenomenon by which a downstream system affects the behavior of an upstream system upon interconnection. This fact renders it difficult to predict how modules will behave once connected to each other. In this paper, we propose a general modeling framework for gene transcription networks to accurately predict how retroactivity affects the dynamic behavior of interconnected modules, based on salient physical properties of the same modules in isolation. We illustrate how our framework predicts surprising and counter-intuitive dynamic properties of naturally occurring network structures, which cannot be captured by existing models of the same dimension. We describe implications of our findings on the bottom-up approach to designing synthetic circuits, and on the top-down approach to identifying functional modules in natural networks, revealing trade-offs between robustness to interconnection and dynamic performance. Our framework carries substantial conceptual analogies with electrical network theory based on equivalent representations. We believe that the framework we have proposed, also based on equivalent network representations, can be similarly useful for the analysis and design of biological networks.

Citation: Gyorgy A, Del Vecchio D (2014) Modular Composition of Gene Transcription Networks. PLoS Comput Biol 10(3): e1003486.

Editor: Stanislav Shvartsman, Princeton University, United States of America

Received: May 17, 2013 Accepted: January 10, 2014 Published: March 13, 2014

Copyright: © 2014 Gyorgy, Del Vecchio. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Funding: Grant #FA9550-12-1-0129. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.


The observed discrepancy in occurrence frequency of FFLs and 3-CYCs is a natural consequence of topological properties of networks

Occurrences of FFLs and 3-CYCs in various biological networks (see Table 1) show patterns: there are a relatively large number of FFLs and relatively small number of 3-CYCs. In this section we explain the topological basis for these differences in their frequencies.

First we note that random connectivity within three-node subgraphs itself favours FFLs. Consider a directed, complete – there is an edge between every pair of nodes – three node graph (3-graph). Excluding bidirectional edges, for any set of 3 nodes there are 2 3 = 8 possible directed 3-graphs. Each of these configurations is isomorphic to either a FFL or a 3-CYC – any directed complete 3-graph is either a FFL or 3-CYC. Out of 8 possibilities, 6 form FFLs, and 2 form 3-CYCs. Allowing bidirectional edges, there are an extra 19 possible configurations containing at least one bidirectional edge. Each of these possibilities gives multiple FFLs or 3-CYCs or both. With or without bidirectional edges, there is a natural 3:1 bias towards forming an FFL over a 3-CYC in a 3-graph.

Global properties of biological networks also favour FFLs over 3-CYCs. Most biological networks, such as those used in our study, are scale-free [15]. In scale-free networks, the connectivity of nodes follows the power law: the probability of a node having k neighbours is P(k)

k -γ . Only a few nodes in such a network are highly-connected (and form hubs), while most nodes are sparsely connected [15].

We asked how many of the FFLs in various networks contain hubs among their nodes. (We consider as hubs the top 10% of nodes in the network that are highly-connected, having more than 10 neighbours.) Table 2 contains the percentages of FFLs enumerated in various networks, having n = <0, 1, 2, 3>nodes as hubs. A large majority of the FFLs contain at least one hub most common being the FFLs with hubs at two of their nodes. In the Yeast composite network, 961 of 997 FFLs have at least one common source-intermediate edge between them. These 961 FFLs can be grouped into 114 clusters (containing distinct source-intermediate edges) revealing that connected hubs often share many common children, automatically giving rise to FFLs. We believe that the principle of preferential attachment predisposes a biological network to have connected hubs that have shared children. This gives a network its robustness to random node failure [15].

We also observe that there is an imbalance between indegree and outdegree around hubs – there are significantly more outgoing edges than incoming edges. We have seen above that FFLs are naturally favoured over 3-CYCs in 3-graphs. The imbalances between in- and out-degree around the hubs further enhances the formation of FFLs. Consider a hub with m incoming edges and n outgoing edges. With a random addition of an edge between any pair of (m + n) nodes adjacent to this hub, the probability of forming an FFL in this system is: P FFL = 2 ( m C 2 + n C 2 ) + m n 2 ( m C 2 + n C 2 + m n ) [email protected]@[email protected]@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemiuaa1aaSbaaSqaaiabbAeagjabbAeagjabbYeambqabaGccqGH9aqpjuaGdaWcaaqaaiabikdaYiabcIcaOmaaCaaabeqaaiabd2gaTbaacqWGdbWqdaWgaaqaaiabikdaYaqabaGaey4kaSYaaWbaaeqabaGaemOBa4gaaiabdoeadnaaBaaabaGaeGOmaidabeaacqGGPaqkcqGHRaWkcqWGTbqBcqWGUbGBaeaacqaIYaGmcqGGOaakdaahaaqabeaacqWGTbqBaaGaem4qam0aaSbaaeaacqaIYaGmaeqaaiabgUcaRmaaCaaabeqaaiabd6gaUbaacqWGd[email protected][email protected] while that of forming a cycle is: P 3 − CYC = m n 2 ( m C 2 + n C 2 + m n ) [email protected]@[email protected]@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemiuaa1aaSbaaSqaaiabiodaZiabgkHiTiabboeadjabbMfazjabboeadbqabaGccqGH9aqpjuaGdaWcaaqaaiabd2gaTjabd6gaUbqaaiabikdaYiabcIcaOmaaCaaabeqaaiabd2gaTbaacqWGdbWqdaWgaaqaaiabikdaYaqabaGaey4kaSYaaWbaaeqabaGaemOBa4gaaiabdoe[email protected][email protected] . Then, P FFL P 3 − CYC = 1 + ( m − 1 ) n + ( n − 1 ) m [email protected]@[email protected]@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaqcfa4aaSaaaeaacqWGqbaudaWgaaqaaiabbAeagjabbAeagjabbYeambqabaaabaGaemiuaa1aaSbaaeaacqaIZaWmcqGHsislcqqGdbWqcqqGzbqwcqqGdbWqaeqaaaaacqGH9aqpcqaIXaqmcqGHRaWkdaWcaaqaaiabcIcaOiabd2gaTjabgkHiTiabigdaXiabcMcaPaqaaiabd6gaUbaacqGHRaWkdaWc[email protected][email protected] , which is symmetric in m and n. If there is a large disparity between m and n (i.e., mn, or mn), then one of the terms ( m n ) [email protected]@[email protected]@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaWaa[email protected][email protected] or ( n m ) [email protected]@[email protected]@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaWaa[email protected][email protected] dominates, resulting in P FFL P 3 − CYC

max ( ( m n ) , ( n m ) ) [email protected]@[email protected]@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaqcfa4aaSaaaeaacqWGqbaudaWgaaqaaiabbAeagjabbAeagjabbYeambqabaaabaGaemiuaa1aaSbaaeaacqaIZaWmcqGHsislcqqGdbWqcqqGzbqwcqqGdbWqaeqaaaaakiabc6ha+jGbc2gaTjabcggaHjabcIha4naabmaabaWaaeWaaKqbagaadaWcaaqaaiabd2gaTbqaaiabd6gaUbaaaOGaayjkaiaawMcaaiabcYcaSmaabmaajuaGbaWaaSaa[email protected][email protected] . For example, when m = 2 and n = 20, PFFL = 0.91, and P3-CYC = 0.09. This shows the odds against the formation of a 3-CYC in networks with structures typical of biological networks.

There have been suggestions that 3-CYC is an "anti-motif" – a motif that is selected against in many biological networks [14]. But, as described above, the suppression of 3-CYCs is an expected consequence of topological properties of biological networks.

These properties are sufficient to account for the observed profiles of FFLs and 3-CYCs.

Assemblies of motifs

Kashtan and colleagues [16] observed that regulatory networks contain multi-output FFL generalizations (see Figure 2(a)) in frequencies much higher than multi-input (Figure 2(d)) and multi-intermediate (Figure 2(f)) generalisations. (These authors also suggested that multi-output FFLs were selected to achieve some information processing role [16].)