The vast amount of sequence data collected over the past two

Transcription

1 Local grah alignment and motif search in biological networks Johannes Berg and Michael Lässig Institut für Theoretische Physik, Universität zu Köln, Zülicherstrasse 77, Cologne, Germany Edited by Richard M. Kar, International Comuter Science Institute, Berkeley, CA, and aroved July 8, 2004 (received for review August 13, 2003) Interaction networks are of central imortance in ostgenomic molecular biology, with increasing amounts of data becoming available by high-throughut methods. Examles are gene regulatory networks or rotein interaction mas. The main challenge in the analysis of these data is to read off biological functions from the toology of the network. Toological motifs, i.e., atterns occurring reeatedly at different ositions in the network, have recently been identified as basic modules of molecular information rocessing. In this article, we discuss motifs derived from families of mutually similar but not necessarily identical atterns. We establish a statistical model for the occurrence of such motifs, from which we derive a scoring function for their statistical significance. Based on this scoring function, we develo a search algorithm for toological motifs called grah alignment, a rocedure with some analogies to sequence alignment. The algorithm is alied to the gene regulation network of Escherichia coli. The vast amount of sequence data collected over the ast two decades is at the heart of quantitative molecular biology. Biological information is extracted from these data mainly by analyzing similarities between sequences. This aroach is based on efficient sequence alignment algorithms and a statistical theory to assess the significance of the results (see ref. 1). Its ultimate goal is to infer functional relationshis from correlations between sequences. Over the last few years, however, it has become clear that functions in many cases cannot be identified at the level of single genes. A given function may require the cooerative action of several genes, and conversely, a given gene may lay a role in quite different functional contexts. The genome is thus a highly interactive system and the exression of a gene deends on the activity of other genes. The athways of these interactions are encoded in so-called regulatory networks. Similarly comlex networks govern signal transduction, that is, the influence of external signals on gene exression, or rotein interactions, that is, the ability of two or more roteins to enter a bound state in a living cell. A few exemlary cases of gene networks have been studied in much detail, such as the regulation of early develoment in the sea urchin Strongylocentrotus ururatus (2) or in Drosohila (3). In some aroximation, these structures can be understood as logical networks: the exression level of a gene is reduced to a binary variable (on or off) and is secified in terms of binary inut data, i.e., the exression levels of its ustream genes. On the other hand, a large amount of data on molecular interaction networks is now obtained by high-throughut exeriments, for examle rotein interaction mas in yeast (4) or gene exression arrays (5). In these arrays, one robes the activity of an entire genome, rather than of just a few genes. However, the detailed logical connection of interaction athways is tyically lost. The information is reduced to a toological network, that is, a set of nodes (reresenting, e.g., genes or roteins) and links reresenting their airwise interactions. These links can be directed as in the case of regulatory interactions or undirected as for rotein rotein binding. The amount of toological data on molecular networks is exected to increase raidly in the next few years, aralleling the earlier exlosion of sequence data. What can be learned from these data? Using the network toology alone, can we distinguish atterns of biological function from random background? The urose of this article is to develo a bioinformatics aroach to the search for local modules in networks. We discuss a heuristic search algorithm and its statistical grounding in a stochastic model of network evolution. This aroach is designed to comlement exeriments in secific organisms by large-scale database searches. Two seminal studies (6, 7) recently have shown that toological networks indeed contain statistically significant atterns indicative of biological functions. These motifs are atterns that occur more frequently in the observed network than exected in a suitable null ensemble. The motifs found so far have been identified because they occur identically at different ositions in a network. If network evolution is a stochastic rocess, however, functionally related motifs do not need to be toologically identical. Hence, the notion of a motif has to be generalized to a stochastic one as well. Variations arise because of uncertainties in the network data, or, more imortantly, because some of the interactions can change without affecting the functionality of the motif. This noise is an imortant characteristic of biological systems, familiar from sequence analysis, where one searches for local sequence similarities blurred by mutations and insertionsdeletions, rather than for identical subsequences. It leads us to the notion of a robabilistic motif in which each link occurs with a certain likelihood. Probabilistic motifs arise as consensus from finding a family of sufficiently similar subgrahs in a network. The search for mutually similar subgrahs and their robabilistic motifs is the central issue of this article. The motifs of interest here are nonrandom in two ways: they have an enhanced number of internal links, associated, e.g., with feedback, and they aear in a significant number of subgrahs. Identifying these local deviations from randomness in networks requires a statistical theory of local grah structure, which we establish in this article. This is a comlementary aroach to the global statistics measured by the connectivity distribution (8) or connectivity correlations (9, 10) of a network. Our aroach leads to an algorithmic rocedure termed local grah alignment, which is concetually similar to sequence alignment. It is based on a scoring function measuring the statistical significance for families of mutually similar subgrahs. This scoring involves quantifying the significance of the individual subgrahs as well as their mutual similarity, and is thus considerably more comlicated than for families of identical motifs. Our scoring function is derived from a stochastic model for network evolution. There is indeed evidence that network evolution can be described as a stochastic rocess. For examle, the comarison of the regulatory networks for early develoment in several Drosohila secies has revealed the continuous buildu and loss of gene interactions following an aroximate molecular clock (11). Yet little is known about the secific athways of network evolution. Our scoring function is comatible with divergent evolution of subgrahs but also with convergent evolution toward a common functional motif. These ath- This aer was submitted directly (Track II) to the PNAS office. To whom corresondence should be addressed. berg@th.uni-koeln.de by The National Academy of Sciences of the USA PHYSICS GENETICS PNAS October 12, 2004 vol. 101 no

2 ways can be illustrated by a comarison with sequence evolution. An examle of convergent evolution is the formation of sequence motifs serving as binding sites of secific enzymes (12, 13). An examle of divergent evolution is a set of sequences stemming from a common ancestor undergoing mutations indeendently. The robabilistic grounding of grah alignment allows us to infer otimal scoring arameters by a maximumlikelihood rocedure (14). As a comutational roblem, grah alignment is more challenging than sequence alignment. Sequences can be aligned in olynomial time by using dynamic rogramming algorithms. For grah alignment, a olynomial-time algorithm robably does not exist. Already simler grah matching roblems such as the subgrah isomorhism roblem (deciding whether a grah contains a given subgrah) (15, 16) or finding the largest common subgrah of two grahs (17) are NP-comlete and NP-hard, resectively. Thus, an imortant issue for grah alignment is the construction of efficient heuristic search algorithms. Here we solve this roblem by maing grah alignment onto a sin model familiar in statistical hysics, which can be treated by simulated annealing. This article is structured as follows. In the first art, we discuss the statistics of local subgrahs based on a robabilistic model. This is done in three stes: (i) an individual subgrah with an enhanced number of internal links, (ii) a subgrah in the resence of a temlate motif secifying the functional imortance of each link, and (iii) correlated subgrahs, whose common attern is to be inferred from the data instead of being given as a temlate. We then construct a scoring function designed to distinguish sets of statistically significant network motifs with an enhanced number of links from a background of other atterns. High-scoring motifs are found by an alignment algorithm, details of which are described in Suorting Text, which is ublished as suorting information on the PNAS web site. In the second art of the aer, we aly this method to the regulatory network of Escherichia coli and discuss the robabilistic motifs found. The statistics of these motifs is used to test the assumtions of our robabilistic model. Grahs and Patterns A toological network or grah is a set of nodes and links. Labeling the nodes by an index r 1,...,N, the network is described by the adjacency matrix C, which has entries C rr 1 if there is a directed link from node r to node r and C rr 0 otherwise. Grahs with a generic adjacency matrix are called directed. The secial case of a symmetric adjacency matrix can be used to describe undirected grahs. The in and out connectivities of a node, k r r C rr and k r r C rr, are defined as the number of in- and outgoing links, resectively. The total number of links is denoted by K r,r C rr. The networks considered here are sarse, i.e., their average connectivity KN is of order 1. A subgrah G is given by a subset of n vertices {r 1,...,r n } and the resulting restriction of the adjacency matrix. More recisely, we define the matrix c(g, A) with the entries c ij C ri r j (i, j 1,...,n) secifying the internal links of the subgrah for a given order A of the nodes. This matrix c is called a attern, which is contained in the subgrah. The definition of a attern used here imlies that two atterns are counted as searate if the matrices c and c are different. This assumes that nodes are distinguishable by their biochemical identity and their functional role even if they are at symmetric ositions, i.e., if c and c differ only by the labeling of the nodes. An alternative definition would count two matrices c and c related by a relabeling as defining an identical attern. Which definition is more aroriate deends on the articular biological alication. The most imortant characteristic of atterns for what follows is their number of internal links, Lc c ij. [1] i, j Fig. 1. Motifs and alignment in toological networks. (a) A randomly chosen connected subgrah is likely to be a tree; i.e., it has the number of internal links equal to its number of nodes minus 1. (b) Putatively functional subgrahs are distinguished by internal loos, i.e., by a higher number of internal links. (c) An alignment of three subgrahs with four nodes each. Each node carries an index 1, 2, 3 labeling its subgrah and an index i 1, 2, 3, 4 given by the order of nodes within the subgrah. Nodes with the same index i are joined by dashed lines, defining a one-to-one maing between any two subgrahs. Network links are shown as solid lines (with their arrows suressed for clarity). (d) The consensus attern of this alignment. Each link occurs with a likelihood c ij indicated by the gray scale. Fig. 1 shows two subgrahs that differ in the values of L. Grah Alignments and Motifs A grah alignment is defined by a set of several subgrahs G ( 1,..., ) and a secific order of the nodes {r 1,..., r n } in each subgrah; this joint order is again denoted by A. For simlicity, we assume here that the subgrahs are of the same size n, but it is not difficult to generalize our aroach to include subgrahs of different size. For a given set of mutually disjoint subgrahs, there are (n!) different alignments. An alignment associates each node in a subgrah with exactly one node in each of the other subgrahs. This association can be visualized by n strings, each connecting the nodes with the same index i as shown in Fig. 1c. A given alignment A secifies a attern in each subgrah; we write c c(g, A). The consensus attern of this alignment is given by the matrix c 1 c. [2] Thisisarobabilistic attern, the entry c ij denoting the likelihood that a given link is resent in the aligned subgrahs. For any two aligned subgrahs G and G, we can define the airwise mismatch n Mc, c c ij 1 c ij 1 c ij c ij. [3] i, j1 The mismatch is 0 if and only if the matrices c and c are equal, and is ositive otherwise. It can be considered as a Hamming distance for aligned subgrahs. The average mismatch over all airs of aligned subgrahs, M M(c, c), is termed the fuzziness of the consensus attern c. Analogously, the average number of internal links is denoted by L L(c). We now define network motifs as statistically significant consensus atterns of grah alignments, which are distinguished by a high number of internal links and low fuzziness. Clearly, this definition Berg and Lässig

3 is mathematically loose before we quantify the statistical significance. This quantification will be done in the next three sections. Guided by the results of refs. 6 and 7, we take an enhanced number of internal links as a toological indicator of ossible functional modules in networks. The additional links beyond a treelike toology can be associated with feedback or feed-forward loos in transcrition networks, or clusters in rotein interaction networks. For examle, the triangle shown in Fig. 1b can be interreted (6) as a low-frequency bandass filter: the central node is activated if both to nodes are active. However, the right-hand node is activated by that on the left with a small delay, so the central node is activated rovided the left node is active for a time longer than this delay. The nontreelike nature of this motif is crucial for its function. On the other hand, most randomly chosen connected subgrahs would be treelike. Clearly, an enhanced number of internal links is but the simlest toological indicator of utative functionality, and more detailed ways of identifying network motifs are likely to emerge in the future. Statistics of Individual Subgrahs To quantify the statistical significance of a given number of internal links, we first comute the relevant robability distribution in a suitable random grah ensemble, which is generated by an unbiased sum over all grahs with the same number of nodes and the same connectivities k r, k r (r 1,...,N) as in the data set but randomly chosen links (10, 18). This null ensemble is aroriate for biological networks, whose connectivity distribution generally differs markedly from that of a random grah with uniformly distributed links. In the null ensemble, the robability of finding a directed link from node r i to node r j is in good aroximation given by w ij k ri k rj K (19). Hence, a given subset of nodes {r 1,...,r n } forms a subgrah G with robability n P 0 G 1 w ij 1cij c w ij ij. [4] i, j1 This exression neglects double links, which can be included as in ref. 19. The robability P 0 (G) deends on the attern c(g) ofthe subgrah, as well as on its environment given by the connectivities k i, k i (i 1,...,n). In this ensemble, the exected number of internal links er node is small, L 0 n nn, where we denote the average over a given ensemble by. Hence, most random subgrahs in a large and sarse grah are disconnected. Within the subset of connected subgrahs, most are treelike. (Later we will be interested in the subset of nontreelike subgrahs, and this will require a modification of the null ensemble.) We now assume that subgrahs containing network motifs are generated by a different ensemble P (G). The robability that a given air of nodes carries a link is enhanced by a factor e relative to the null ensemble 4, leading to P G/P 0 G Z 1 exlc. [5] Again the robability P (G) that a given subset of nodes {r 1,..., r n } forms a subgrah G deends on the matrix c(g). We have introduced the normalization factor Z ij cij 0,1 ex[l(c)]p 0 (G), which ensures that P (G) summed over all matrices c gives unity. The quantity, called the link reward, is multilied by the total number L of internal links given by Eq. 1. The ensemble 5 is a statistically unbiased way to describe that functional motifs are distinguished by a large number of internal links. (Technically, it is the ensemble of maximal information entroy with a given average link number L, which is determined by the value of.) This ensemble may be thought of as resulting from an evolutionary rocess favoring the formation of links due to selection ressure; such a rocess has recently been studied for regulatory networks (20). Here we focus on the detection of evolved motifs rather than on the reconstruction of evolutionary histories. Hence, we do not need to make assumtions on dynamical details of motif formation but only on its outcome, which is described by the ensemble P (G). We have tested the form of this ensemble for the regulatory network of E. coli as discussed in Results and Discussion below. Moreover, the value of the link reward can be inferred from the data. One finds e Nn, which results in a finite exected number Ln of internal links er node within a motif. Statistics in the Presence of a Temlate The distribution 5 describes an ensemble with an enhanced number of links, which is aroriate for scoring individual subgrahs in the absence of further knowledge. Consider now an evolutionary rocess directed toward a given network motif reresented by a temlate adjacency matrix t. An alignment A between the motif t and the subgrah G is secified by a given ordering of the nodes {r 1,..., r n }ing. The outcome of this evolutionary rocess can be modeled by an ensemble Q t (G, A) with a bias against links that do not occur in the temlate, Q t G,A/P 0 G Z t 1 ex Lc 2 Mc, t. [6] This exression denotes the robability that a given subset of nodes {r 1,...,r n } forms an aligned subgrah (G, A), with the definition 3 of the airwise mismatch of the subgrah G and the temlate t in a given alignment A. Again, Z t is given by normalization. This is a hidden Markov model: the outcome of the stochastic rocess is an aligned subgrah (G, A), whereas only G is observed. The likelihood of observing G is then a sum over all alignments, Q t G Q t G,A. [7] A This ensemble has two free arameters, the link reward and the mismatch enalty (with a factor 12 introduced for later convenience). It is concetually similar to hidden Markov models for the alignment of sequences with gas. Statistics of Correlated Subgrahs Now we turn to the case where a network motif is not given as a temlate but has to be inferred from a family of suitably aligned subgrahs. The underlying evolutionary rocess can be regarded as a biased link formation as in the revious section, with the consensus attern c as temlate. Assuming the link formation is indeendent for each subgrah, we obtain an ensemble given by Q, G 1,...,G, A P 0 G Z 1, ex Lc Mc 2,c Z 1, ex Lc Mc,c 2,, [8] where A secifies an alignment of all subgrahs and we have used the definition 2 of the consensus attern. The normalization is given by Z, A c 1,...,c ex Lc Mc,c 2, P 0 G. [9] PHYSICS GENETICS Berg and Lässig PNAS October 12, 2004 vol. 101 no

4 The Scoring Function We now construct a scoring function designed to select a set of (utatively) functional subgrahs, characterized by a consensus motif with a high number of internal links and low fuzziness, from the background of random subgrahs in a large network. Based on the receding discussion, we assume that the statistics of functional motifs is described by an ensemble Q(G 1,..., G, A) Q, (G 1,...,G, A), where the scoring arameters and remain to be determined from the data. For the biological alications described above, where internal links are associated with feedback loos, it is clearly useful to restrict the motif search to the set of all connected subgrahs that contain internal loos, i.e., that are nontreelike. For connected subgrahs of size n, this set is given by the constraint L n on the internal link number. A large random grah tyically contains a number of order one of such subgrahs, and these define the relevant null ensemble for motif search. We model these subgrahs by using the ensemble P 0 with an enhanced number of links defined in Eq. 5. The arameter 0 will be adjusted such that the average number of internal links in the null ensemble equals that found in the nontreelike subgrahs of a suitable randomized grah. Comaring with the ensemble P 0 of random subgrahs introduced earlier, it is clear that the constraint L n corresonds to a link reward 0 0. Given these two ensembles, we define the log-likelihood score SG 1,...,G, A log Q,G 1,...,G, A P 0 G 1,...,G, A 0 Lc Mc,c 2, logz, /Z 0, [10] which is ositive if a set of subgrahs G 1,...,G and an alignment A between them is more likely to occur in the ensemble Q, than in the null ensemble P 0. The term log(z, Z 0 ) acts as a threshold assigning a negative score to alignments with too large fuzziness or a too small number of internal links. As is clear from the form of the scoring function, grah alignment is a nontrivial otimization roblem, the statistical weight of each subgrah G deending on the scoring arameters as well as on the other subgrahs included in the alignment. We address this roblem in two stes. First we find the maximum-score alignment(s) for given score arameters, which is essentially an algorithmic search roblem. Then we discuss the arameter deendence of highscoring alignments and obtain the otimal values of and for a given data set from a maximum-likelihood rocedure. Maximum Score Alignments and Parametric Otimization Finding the maximum score alignments involves a huge search sace of ossible alignments. The number of alignments is of order (n) N for given and the comutational exense grows further when the otimization over is erformed. Here we use a heuristic algorithm, which can be described by a maing to a discrete sin model. First we enumerate all nontreelike subgrahs of n nodes, which is feasible for modest values of n, and label them by the index 1,..., max. Next we evaluate the internal link numbers L L(c ) and the airwise mismatches M, defined as the minimum of M(c, c ) over all airwise alignments of the subgrahs G and G. High-scoring multile alignments are then found by a simulated annealing algorithm in the sace (s 1,...,s max ), where each sin s takes the value 1 if G is included in the alignment and 0 otherwise. The resulting Hamiltonian H is H 0 max L s max M s s logz 2,/Z, 0, [11] where s. The couling between s and s is given by M, which is equal to the airwise mismatch M if subgrahs and do not overla, and a large ositive constant if they do. (Two subgrahs overla if they have more than one node in common. According to this definition, links in nonoverlaing subgrahs form indeendently as assumed in Eq. 8.) The threshold term log(z, Z 0 ) is evaluated by saddle-oint integration; details are given in Suorting Text. Simulated annealing using the Hamiltonian 11 will then yield high-scoring alignments of nonoverlaing subgrahs (22). For fixed values of the scoring arameters, the algorithm is exected to roduce well defined maximum-score alignments. This can be understood as follows. For a (hyothetical) alignment of subgrahs with equal number of internal links and equal airwise mismatches, the score 10 scales linearly with, the number of aligned subgrahs. This behavior is consistent with the interretation of Eq. 10 as a log-likelihood score, because the aligned subgrahs occur indeendently. A high-scoring alignment in a realistic network may consist of a limited number of identical or very similar motifs. As we extend this alignment to include more subgrahs, subgrahs with increasing mutual mismatches are included. Hence, we exect the total mismatch to increase faster than linearly with, leading to a maximum S*(, ) of the total score at some intermediate value of *(, ). The roerties of the maximum-score alignments deend strongly on the arameters and. With increasing, the number of internal links L*(, ) er subgrah is exected to increase. With increasing, both the number of grahs *(, ) and the fuzziness M *(, ) decrease. In this way, the maximum-score alignment varies between a set of indeendent subgrahs for 0 and a set of identical subgrahs with identical motifs for 3. A maximum-likelihood aroach can be used to infer the otimal scoring arameters *, * for a given data set, which we obtain as the oint of the global score maximum S* max, S*(, ). Results and Discussion In this section, we discuss the alication of local grah alignment to motif search in the gene regulatory network of E. coli, taken from ColiNet-1.1, containing 424 nodes and 577 directed links. Each labeled node in this network reresents a gene. A directed link between two nodes signifies that the roduct of the gene reresented by the first node acts as a transcrition factor on the gene reresented by the second. Throughout we consider motifs with a fixed number of nodes n 5. First we show that our algorithm indeed roduces well defined alignments of maximal score (i.e., of maximal relative likelihood). For fixed arameters 3.8 and 4.0, this is illustrated by Fig. 2a, which shows the score S and the fuzziness M for the highestscoring alignment with a rescribed number of subgrahs, lotted against. As exected, the fuzziness increases with increasing, and the total score reaches its global maximum S*(, ) at an intermediate value *(, ). It is lower for *(, ) because the alignment contains fewer subgrahs and for *(, ) because the subgrahs have higher mutual mismatches. Fig. 2b shows the score S*(, ) as a function of and. This function has a unique global maximum S*, which defines the maximum-likelihood oint (* 3.8, * 2.25, * 24). The scoring arameter 0 of the null ensemble P 0 is determined as follows: the data set is randomized by generating a network with the same connectivities as in the data set but randomly chosen links (10, 18). Again, the nontreelike subgrahs are extracted and their Berg and Lässig

5 Fig. 2. Maximum score alignment and arametric otimization. (a) Score otimization at fixed scoring arameters 3.8 and 4.0. The total score S (thick line) and the fuzziness M (thin line) are shown for the highest-scoring alignment of subgrahs, lotted as a function of. (b) The score S*(, ) lotted against the arameters and. The unique maximum S*defines the maximum-likelihood arameters * 3.8 and * average number of internal links is determined. The value of 0 is uniquely determined by the condition that the exected number of links in the null ensemble 5 equals the average number of internal links found in the nontreelike subgrahs. One obtains As exected, 0 *, which shows that the data set has an enhanced number of internal links relative to the randomized network. At the maximum-likelihood scoring arameters, we can, moreover, verify the functional form of the ensembles used to construct the score function 10. To test the model 5 for individual subgrahs, we enumerate all subgrahs with n 5 that have nontreelike atterns (i.e., a link number L 5). All ordered airs of nodes i, j are then binned according to the robability w ij of a directed link existing between them in the ensemble P 0 (G). In Fig. 3a the fraction of these airs i, j carrying a link is lotted against w (). The exectation value of this fraction is given by Eq. 5 as e w(1 w e w), shown as a solid line with a fit value * 3.8. Our model (Eq. 8) for generic alignments can be tested in a similar way. From this ensemble, the marginal robability that a given ordered air of nodes secified by, i, j is linked can be comuted. We grou all such airs with the same exectation value c ij *,* according to Eq. 8 to build a histogram. For each grou, the average of c ij over node airs in the actual maximum-likelihood alignment is comuted and lotted against the model rediction (see Fig. 3b). The same rocedure is reeated for the two-oint correlations c ij c ij *,* between associated nodes in different subgrahs and as also shown in Fig. 3b. In both histograms, the data oints cluster well around the straight line equating exectation values in the model 8 and averages in the actual alignment. The fluctuations seen reflect the limited size of the data set and the small number of fitting arameters in the model. For such data, more detailed models can hardly be tested because they would lead to overfitting. We now turn to the robabilistic consensus motifs found in the data for different number of nodes n 4 and n 5. Fig. 4a shows the n 4 consensus motif c ij at consecutive values of * 3.6, 8, and 15. The gray scale encodes the average number of Fig. 3. Statistics of motif ensembles. (a) Testing the statistical model for single subgrahs (Eq. 5). Nontreelike subgrahs are enumerated and node airs i, j are binned according to w ij. The fraction of such airs carrying a link is shown against w ij. The solid line results from fitting the model with enhanced number of links (Eq. 5) to these data, giving 3.8. (b) Testing the statistical model for alignments (Eq. 8). (Uer) The average value of c ij over all, i, j with a given exectation value of c ij according to Eq. 8 at * 3.8 and * 2.25 against the corresonding exectation value (). For a erfect fit between model and data a straight line is exected (shown solid). (Lower) The same rocedure is used averaging the two-oint function c ij c ij over all,, i, j with a given exectation value c ij c ij. links c ij between a given air of nodes. As exected, the fuzziness decreases with increasing values of the mismatch enalty and c ij tends either to zero (no link resent) or one (link resent with certainty) as 3. The consensus motif is a layered structure, in this case with two inut and two outut nodes. A similar motif is found for n 5. Fig. 4b shows the n 5 consensus motif at consecutive values of 2.25, 5, and 12. As in the case of n 4, a layered structure is clearly discernible: the motif consistsof2 3 nodes forming an inut and an outut layer, with links largely going from the inut to the outut layer. The left node of the inut layer has an average number of about 30 outgoing links. These connectivities are excetional because the average outconnectivity of the network is Comaring the alignments of subgrahs of n 4 nodes with those of n 5 nodes in Fig. 4, one finds that many of the subgrahs found in the n 4 alignments also are a art of the subgrahs found in the n 5 alignments. This finding immediately leads to the question of how to identify larger atterns in the network from which the subgrahs at a given value of n are taken. Obviously any scoring scheme oerating at a fixed number of nodes n will be blind to the combinatorial ossibilities of selecting subgrahs from a larger attern. The henomenon is exemlified in Fig. 4c. From the 3-by-4 attern two nonoverlaing layered subgrahs with n 4 and n 5 can be generated (nonoverlaing subgrahs have at most one node in common, see above). Larger atterns generate corresondingly more nonoverlaing subgrahs. In the suorting information, we discuss a simle scheme that allows one to identify larger atterns as in Fig. 4c from smaller subgrahs. The attern of Fig. 4c is found twice in the data, contributing in total four nonoverlaing subgrahs to the alignments with n 4 and n 5. The statistics of these atterns at the level of identical atterns has PHYSICS GENETICS Berg and Lässig PNAS October 12, 2004 vol. 101 no

6 Fig. 4. Probabilistic motifs in the E. coli transcrition network. (a) Consensus motifs with n 4 nodes at different values of. From left to right, * 3.6, 8, and 15. The gray scale of the links indicates the likelihood that a given link is resent in the aligned subgrahs; the five gray values corresond to c in the ranges , , , , and , and links with c 0.9 are shown black. The link reward is ket fixed at * 3.6 and 0 takes on the value (b) Consensus motifs with n 5 for different * 2.25, 5, and 12 (left to right) at * 3.8. (c) This attern with n 7 is found twice in the data set. From each such subgrah two nonoverlaing layered subgrahs with n 4 and n 5 can be generated. (d) The number *(*, ) of subgrahs in the maximum score alignment (thick line) and the fuzziness M *(*, ) (thin line) as a function of for n 5. recently been analyzed in ref. 23, their treatment using robabilistic atterns remains a future develoment. Fig. 4d shows details of the alignments roducing the consensus motifs at n 5, namely, the number of subgrahs *( *, ) and the fuzziness M *( *, ) lotted as a function of. For 12 the fuzziness reaches zero and the alignment contains 10 identical nonoverlaing motifs. This layered attern has been found by the aroach of ref. 6, which is based on counting identical motifs. However, the maximum-likelihood alignment occurs at * 2.25 and contains a much larger number of * 24 nonoverlaing subgrahs, leading to the robabilistic consensus motif shown on the left in Fig. 4a. The same effect is found in the consensus motif of size n 4. Furthermore, at arbitrary nonzero fuzziness the robability that a given air of subgrahs have identical motifs decreases with subgrah size. As a result, counting identical motifs, rather than following a robabilistic aroach as the one resented here, will miss a fraction of relevant subgrahs resent in the data that increases with the size of the subgrah. The robabilistic grounding of motif search is also indisensable for estimating the quantitative significance of the results obtained. Here we comare the maximum-likelihood alignment in the E. coli data set with suitable random grah ensembles. We do this in two stes, to disentangle the significance of the number of internal links and of the mutual similarity of atterns found in the data. (i) To assess the significance of the number of internal links, we consider the ensemble of grahs with the same in- and outconnectivities as the data set but randomly chosen neighbors (18, 10) and comute the distribution of the score with scoring arameters *, 0. The null distribution of scores from the randomized grah has the average and standard deviation given by S* The score S* 73.1 found from the data is thus significantly higher, indicating an enhanced link number with resect to the random grah ensemble. (ii) The significance of the mutual similarity of the aligned atterns is assessed by comaring the data to mutually indeendent random subgrahs with the same average density of links. (This null ensemble is generated by randomizing the internal links of each subgrah indeendently.) We then comute the score with arameters 0 * and * (thereby focusing only on the fuzziness of the data relative to that found in the ensemble of uncorrelated subgrahs). This null distribution of scores has average and standard deviation given by S* ; the corresonding score S* 50.1 found from the data is thus significantly higher. We note that the assessment of subgrah similarity is quite subtle. Subgrahs taken from a large but finite random grah may show a surious mutual similarity with resect to indeendent random subgrahs because of a revalence of internal loos. The statistical significance of the results can be formulated more recisely by using so-called values, which involve the tail of the score distribution in the random grah ensemble. Fast and reliable -value estimates are crucial for searching large databases, as is well known for sequence alignment (21). This aroach can be carried over to the grah alignments discussed here. The statistical framework resented is very flexible. For examle, as large-scale data on the logic of gene regulation become available, the definition of the airwise mismatch, Eq. 3, can be extended to reward aligning sets of nodes erforming the same logical function. In this way, features of motifs going beyond their toology can be exlored. Similarly, simle modifications of the mismatch score allow the analysis of undirected networks, networks whose links have a secific function (reressive or enhancing) or whose interaction strength is quantified by a real number. The rosect of a sizable amount of new data on biological networks becoming available over the next few years through high-throughut methods oens exciting oortunities to identify the building blocks of molecular information rocessing in a wide range of organisms, and even build hylogenetic histories of regulation from transcrition network data. We thank T. Hwa for fruitful discussions and U. Alon and P. Arndt for comments on the manuscrit. We thank the Kavli Institute for Theoretical Physics at Santa Barbara, the Abdus Salam International Centre for Theoretical Physics, Trieste, the Centro di Ricerca Matematica Ennio De Giorgi, Pisa, and the Asen Center for Physics for hositality during various stages of this work. This work has been suorted through Deutsche Forschungsgemeinschaft Grant LA Durbin, R., Eddy, S. R., Krogh, A. & Mitchison, G. (1998) Biological Sequence Analysis (Cambridge Univ. Press, Cambridge, U.K.). 2. Davidson, E. H. (2001) Genomic Regulatory Systems: Develoment and Evolution (Academic, San Diego). 3. Tautz, D. (2000) Curr. Oin. Genet. Dev. 1, Uetz, P., Giot, L., Cagney, G., Mansfield, T. A., Judson, R. S., Knight, J. R., Lockshon, D., Narayan, V., Srinivasan, M., Pochart, P., et al. (2000) Nature 403, Lockhart, D. J. & Winzeler, E. A. (2000) Nature 405, Shen-Orr, S., Milo, R., Mangan, S. & Alon, U. (2002) Nat. Genet. 31, Milo, R., Shen-Orr, S., Itzkovitz, S., Kashtan, N., Chklovskii, D. & Alon, U. (2002) Science 298, Newman, M. E. J., Strogatz, S. H. & Watts, D. J. (2001) Phys. Rev. E 64, Newman, M. (2002) Phys. Rev. Lett. 89, Berg, J. & Lässig, M. (2002) Phys. Rev. Lett. 89, Costas, J., Casares, F. & Viera, J. (2003) Gene 310, Stormo, G. D. & Hartzell, G. W. (1989) Proc. Natl. Acad. Sci. USA 86, Bussemaker, H. J., Li, H. & Siggia, E. D. (2000) Proc. Natl. Acad. Sci. USA 97, Felsenstein, J. (1981) J. Mol. Evol. 17, Ullmann, J. R. (1976) J. Assoc. Comut. Mach. 23, Messmer, B. T. & Bunke, H. (1998) IEEE Trans. Pattern Anal. Mach. Intell. 20, Garey, M. R. & Johnson, D. S. (1979) Comuters and Intractability: A Guide to the Theory of N-Comleteness (Freeman, NY). 18. Maslov, S. & Sneen, K. (2002) Science 296, Itzkovitz, S., Milo, R., Kashtan, N., Ziv, G. & Alon, U. (2003) Phys. Rev. E 68, Berg, J., Lässig, M. & Radic, S. (2003) htt:arxiv.orgabscond-mat Karlin, S. & Altschul, S. F. (1990) Proc. Natl. Acad. Sci. USA 87, Blat, M., Wiseman, S. & Domany, E. (1996) Phys. Rev. Lett. 76, Kashtan, N., Itzkovitz, S., Milo, R. & Alon, U. (2003) htt:arxiv.orgabsq-bio Berg and Lässig