From: Subject: Science -- Jansen et al. 302 (5644): 449 Date: Sat, 18 Oct 2003 20:10:02 -0400 MIME-Version: 1.0 Content-Type: multipart/related; boundary="----=_NextPart_000_0000_01C395B3.D0924F80"; type="text/html" X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2800.1165 This is a multi-part message in MIME format. ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: text/html; charset="Windows-1252" Content-Transfer-Encoding: quoted-printable Content-Location: http://www.sciencemag.org/cgi/content/full/302/5644/449 Science -- Jansen et al. 302 (5644): 449
3DElsevier=20 3DAACR=20 =
Institution: YALE UNIVERSITY |=20 Sign In as = Individual | FAQ | Access Rights | Join AAAS
3DHelp=203DFeedback=20

A= bstract=20 of this Article
PDF=20 Version of this Article
Supp= orting=20 Online Material
 
Down= load=20 to Citation Manager
Alert me when:
new articles cite this=20 article
 
Search for similar articles in:=20
  Science Online=20
  PubMed
Search Medline for articles by:
Jansen,=20 R. || Gerstein,=20 M.
 
This=20 article appears in the following Subject=20 Collections:
Cell=20 Biology =

A Bayesian Networks Approach for Predicting = Protein-Protein=20 Interactions from Genomic Data =

Ronald=20 Jansen,1*=20 Haiyuan Yu,1 Dov=20 Greenbaum,1 Yuval = Kluger,1=20 Nevan J. Krogan,4 Sambath=20 Chung,1,2 Andrew=20 Emili,4 Michael = Snyder,2=20 Jack F. Greenblatt,4 Mark=20 Gerstein1,3

We have developed an approach using Bayesian networks to=20 predict protein-protein interactions genome-wide = in=20 yeast. Our method naturally weights and combines = into=20 reliable predictions genomic features only weakly = associated with interaction (e.g., messenger=20 RNAcoexpression, coessentiality, and colocalization). = In=20 addition to de novo predictions, it can integrate = often=20 noisy, experimental interaction data sets. We = observe=20 that at given levels of sensitivity, our = predictions are=20 more accurate than the existing high-throughput=20 experimental data sets. We validate our predictions = with=20 TAP (tandem affinity purification) tagging = experiments.=20 Our analysis, which gives a comprehensive view of = yeast=20 interactions, is available at genecensus.org/intint.=20

1 Department of Molecular Biophysics and = Biochemistry,=20 Yale University, 266 Whitney Avenue, Post Office Box 208114, = New=20 Haven, CT 06520, USA.
2 Department of = Molecular,=20 Cellular and Developmental Biology, Yale University, 266 = Whitney=20 Avenue, Post Office Box 208114, New Haven, CT 06520,=20 USA.
3 Department of Computer Science, Yale=20 University, 266 Whitney Avenue, Post Office Box 208114, New = Haven,=20 CT 06520, USA.
4 Banting and Best Department = of=20 Medical Research, Department of Molecular and Medical = Research,=20 University of Toronto, Toronto, M5G 1L6, Ontario, Canada.=20


* Present = address:=20 Computational Biology Center, Memorial Sloan-Kettering=20 Cancer Center, 307 West 63rd Street, New York, NY = 10021,=20 USA.

3D{dagger}=20 To=20 whom correspondence should be addressed. E-mail: mark.gerstein@yale.edu=20

Many fundamental biological processes involve = protein-protein=20 interactions, and comprehensively identifying them is=20 important to systematically defining their = cellular role.=20 New experimental and computational methods have = vastly=20 increased the number of known or putative = interactions,=20 cataloged in databases (= 1=96= 7).=20 Much genomic information also relates to interactions=20 indirectly: Interacting proteins are often = significantly=20 coexpressed (as shown by microarrays) and = colocalized to=20 the same subcellular compartment (= 8,=20 = 9).=20

Unfortunately, interaction data sets are often incomplete = and contradictory (10=9612).=20 In the context of genome-wide analyses, these=20 inaccuracies are greatly magnified because the protein=20 pairs that do not interact (negatives) far outnumber = those=20 that do (positives). For instance, in yeast, the = 6000=20 proteins allow for 3D~=2018 = million=20 potential interactions, but the estimated number = of=20 actual interactions is <100,000 (10,=20 13,=20 14).=20 Thus, even reliable techniques can generate many = false=20 positives when applied genome-wide. This is = similar to a=20 diagnostic with a 1% false-positive rate for a = rare=20 disease occurring in 0.1% of the population, = which would=20 roughly produce one true positive for every 10 false=20 ones. Further information is necessary.

Consequently, when evaluating protein-protein = interactions,=20 one needs to integrate evidence from many different=20 sources (15=9617).=20 Here, we propose a Bayesian approach for integrating=20 interaction information that allows for the = probabilistic=20 combination of multiple data sets and demonstrate = its=20 application to yeast (18).=20 Our approach can be used for combining noisy = interaction=20 data sets and for predicting interactions de novo, = from=20 other genomic information. The basic idea is to = assess=20 each source of evidence for interactions by = comparing it=20 against samples of known positives and negatives=20 ("gold-standards"), yielding a statistical = reliability.=20 Then, extrapolating genome-wide, we predict the = chance of=20 possible interactions for every protein pair by = combining=20 each independent evidence source according to its = reliability. We verified our predictions by comparing=20 them against existing experimental interaction data = (not=20 in the gold-standard) as well as new TAP (tandem = affinity=20 purification) tagging experiments.

Among the many possible machine-learning approaches that=20 could be applied to predicting interactions = (ranging from=20 simple unions and intersections of data sets to = neural=20 networks, decision trees, and support-vector = machines),=20 Bayesian networks have several advantages (19):=20 They allow for combining highly dissimilar types = of data=20 (i.e., numerical and categorical), converting = them to a=20 common probabilistic framework, without unnecessary=20 simplification; they readily accommodate missing data; = and=20 they naturally weight each information source = according=20 to its reliability. In contrast to "black-box"=20 predictors, Bayesian networks are readily = interpretable=20 as they represent conditional probability = relationships=20 among information sources.

The gold-standard data set on which we train=20 ("parameterize") the Bayesian network should = ideally be=20 (i) independent from the data sources serving as=20 evidence, (ii) sufficiently large for reliable=20 statistics, and (iii) free of systematic bias. We = used=20 the MIPS (Munich Information Center for Protein = Sequences)=20 complexes catalog as the gold-standard for positives = (= 6).=20 This hand-curated list of proteincomplexes is = based on=20 the literature [8250 pairs in our filtered = version (19)].=20 A negatives gold-standard is harder to define, = but=20 essential for successful training. Thus, we = synthesized=20 negatives from lists of proteins in separate = subcellular=20 compartments (= 9).=20 Our positive and negative gold-standards satisfy = the=20 first two criteria and provide a good practical = solution=20 for the third. Hence, our goal, precisely defined, = was to=20 predict whether two proteins are in the same complex, = not=20 whether they necessarily had direct physical contact. =

As a measure of reliability, the overlap of information=20 sources (i.e., "interaction data sets," which = could=20 either be noisy experimental data or sets of = genomic=20 features) with the gold-standards can be = expressed in=20 terms of a "likelihood ratio." For example, = consider a=20 genomic feature f expressed in binary terms = (i.e.,=20 "present" or "absent"). The likelihood ratio=20 L(f) is then defined as the = fraction of=20 gold-standard positives having feature f = divided=20 by the fraction of negatives having f. For two = features=20 f1 and f2 with=20 uncorrelated evidence, the likelihood ratio of = the=20 combined evidence is simply the product=20 L(f1, f2) =3D=20 = L(f1)L(f2).=20 For correlated evidence, = L(f1,=20 f2) cannot be factorized in this = way.=20 Bayesian networks are a formal representation of such=20 relationships between features. The combined = likelihood=20 ratio is proportional to the estimated odds that = two=20 proteins are in the same complex, given multiple = sources=20 of information.

We predict a protein pair as positive if its combined=20 likelihood ratio exceeds a particular cutoff = (L=20 > Lcut) (negative otherwise). = To get=20 an overall assessment of how the prediction performs, = we=20 segmented the gold-standard into separate training and = testing=20 sets (using a sevenfold cross-validation protocol). = Then=20 we evaluated the number of true- (TP) and=20 false-positive (FP) predictions in the = testing=20 set. Finally, we applied the Bayesian network = beyond the=20 testing set, computing likelihood ratios for all = possible=20 protein pairs in the genome.

Fig= ure=20 1 schematically shows the information sources and = results=20 of our calculations. We term the results = "probabilistic=20 interactomes" (PIs), in which each protein pair = is=20 associated with a probability measure for being = in the=20 same complex (i.e., likelihood ratio L). = Our=20 procedure not only allows combining existing = experimental=20 interaction data sets (resulting in a PI-experimental = or=20 "PIE"), but also the de novo prediction of = protein=20 complexes from genomic data sets (when the input = data are=20 not interaction data sets per se, resulting in a=20 PI-predicted or "PIP").

= Fig. 1. The=20 information sources integrated in our analysis and = their=20 comparison with each other. (A) The three = different=20 types of data used: (i) Interaction data from = high-throughput=20 experiments. These comprise large-scale two-hybrid = screens=20 (Y2H) (= 1,=20 = 2)=20 and in vivo pull-down experiments (= 3,=20 = 4).=20 (ii) Other genomic features. We considered expression = data,=20 biological function of proteins (from Gene Ontology = biological=20 process and the MIPS functional catalog), and data = about=20 whether proteins are essential (= 6,=20 19=9622).=20 (iii) Gold-standards of known interactions and = noninteracting=20 protein pairs. (The MIPS functional catalog differs = from the=20 MIPS complexes catalog used for the gold-standard.) = (B)=20 Combination of data sets into probabilistic = interactomes.=20 (C) Comparison of the probabilistic = interactomes with=20 the gold-standards and our new experimental data. = Numbers next=20 to the arrows indicate which figures refer to these = various=20 comparisons. [Vi= ew=20 Larger Version of this Image (25K GIF = file)]=20

We combined four interaction data sets from = high-throughput=20 experiments into the PIE (= 1=96= 4)=20 (Fig= .=20 1B). The PIE represents a transformation of = the=20 individual binary-valued interaction sets into a = data set=20 where every protein pair is weighted according to = the=20 likelihood that it exists within a complex.

We computed the PIP from several genomic data sources: = the=20 correlation of mRNA amounts in two expression = data sets=20 (one with temporal profiles during the cell = cycle, one of=20 expression levels under 300 cellular conditions), = two=20 sets of information on biological function, and=20 information about whether proteins are essential = for=20 survival (= 6,=20 20=9622).=20 Although none of these information sources are=20 interaction data per se, they contain information = weakly=20 associated with interaction: Two subunits of the same=20 protein complex often have coregulated mRNA expression = and=20 similar biological functions and are more likely = to be=20 both essential or nonessential (= 8).=20

For computing the PIE and the PIP, we used two different=20 types of Bayesian networks: a "na=EFve" network = for the PIP=20 and a fully connected one for the PIE (19).=20 The na=EFve network is simpler to compute but = requires=20 information sources with essentially uncorrelated = evidence. In contrast, the fully connected = Bayesian=20 network accommodates correlated evidence, which is = the=20 case for the four experimental interaction data sets. =

Finally, we combined the PIP, PIE, and gold-standard into = a=20 total PI (PIT), which represents our most = comprehensive=20 view of the known and putative protein complexes = in yeast=20 (23).=20 Because the PIP and PIE data provide essentially=20 uncorrelated evidence for protein-protein = interactions,=20 we chose a na=EFve network to construct the = PIT.=20

Fig= ure=20 1C gives an overview of how we compared the PIP, = PIE,=20 gold-standard, and our new experiments. In particular, = Fig= .=20 2 shows the performance of the integration = resulting=20 in the PIP and PIE. When tested against the=20 gold-standard, we observed that the ratio of true = to=20 false positives (TP/FP) increases=20 monotonically with Lcut, confirming = L=20 as an appropriate measure of the odds of a real=20 interaction. Conservatively estimated, protein = pairs with=20 L > 600 have a better than 50% chance = of being=20 in the same complex, suggesting Lcut =3D = 600 as a=20 useful threshold (19).=20 Unless otherwise noted, we use this throughout = our=20 analysis. It gives 9897 predicted interactions from the = PIP and 163 from the PIE. In contrast, likelihood = ratios=20 derived from single genomic features (e.g., mRNA=20 coexpression) or from individual interaction = experiments=20 (e.g., the Ho data set) did not exceed the cutoff = when=20 used alone, with TP/FP values far = below 1.=20 This demonstrates that information sources that, taken=20 alone, are only weak predictors of interactions can = yield=20 reliable predictions when combined.

= Fig. 2.=20 Comparison of PIP and PIE with each other and with the = individual information sources. (A) The=20 TP/FP ratio as a function of=20 Lcut for the PIP and the individual = data=20 from which it was computed. The ratio is computed as = follows:=20

where pos(L) and neg(L) are the = number of=20 positives and negatives in the gold-standard with a = given=20 likelihood ratio L. The vertical line indicates = our=20 standard threshold Lcut =3D 600. = (B)=20 The same plot as in (A), but for the PIE. (C)=20 Comparison of TP/FP ratios between the = PIP and=20 PIE. The abscissa represents the sensitivity of the=20 probabilistic interactomes. The gray area indicates = the gain=20 of sensitivity of the PIP over the PIE for equal=20 TP/FP ratios. The arrow shows the = difference in=20 sensitivity at TP/FP =3D 0.3. At this = level, the=20 PIP contains 183,295 protein pairs, of which 6179 are=20 gold-standard positives (75% sensitivity), whereas the = PIE=20 contains 31,511 protein pairs and 1758 gold-standard = positives=20 among these (21% sensitivity). This difference in = sensitivity=20 between PIE and PIP illustrates the value of the de = novo=20 prediction. It also reflects, to some degree, that the = experiments were done only on subsets of the genome = and may=20 have been measuring different types of interactions = than the=20 complexes' gold-standard, which we used to = parameterize the=20 PIP. The white circles show the performance of a = voting=20 procedure in which each of the four genomic features = (from=20 which we computed the PIP) contributed an additive = vote. There=20 are four possible outcomes in the additive voting = procedure,=20 depending on how many data sets contribute a positive = vote (19).=20 [Vi= ew=20 Larger Version of this Image (21K GIF = file)]=20

The PIP had a higher sensitivity than the PIE for = comparable=20 TP/FP ratios (Fig= .=20 2C). ("Sensitivity" measures coverage and is = defined=20 as TP/P, where P is the number of = gold-standard=20 positives.) Specifically, the sensitivity of the = PIP is=20 3D~27%=20 at our cutoff. This may seem low, but compares = favorably=20 with the PIE, which had a sensitivity of less = than 1%.=20 This means that we can predict, at comparable = error=20 levels, more complex interactions de novo than = are=20 present in the high-throughput experimental interaction = data sets.

One might ask whether simpler voting procedures can match = the performance of more complicated = machine-learning=20 methods such as Bayesian networks. To test this=20 hypothesis, we compared the PIP with a voting = procedure=20 where each of the four genomic features = contributes an=20 additive vote toward positive classification. We = found=20 that the Bayesian network achieved greater sensitivity=20 for comparable TP/FP ratios (Fig= .=20 2C) (19).=20

Fig= ure=20 3 shows parts of the PIP and PIE graphs and how = these=20 compare with the gold-standard and our new = experiments.=20 First, to test whether the thresholded PIP was = biased=20 toward certain complexes, we looked at the = distribution=20 of predictions among gold-standard positives (Fig= .=20 3A); they were roughly equally apportioned = among the=20 different complexes, suggesting a lack of = bias.=20

= Fig. 3.=20 Representations of the thresholded PIP (de novo = prediction)=20 compared with different data sets. (A) The = complete set=20 of gold-standard positives and their overlap with the = PIP. The=20 PIP (green) covers 27% of the gold-standard positives=20 (yellow). (B) A graph of the largest complexes = in the=20 PIP, i.e., only those proteins in the thresholded PIP = having=20 3D"=3D" = src=3D"http://www.sciencemag.org/math/ge.gif"=20 border=3D0>20 links. (Left) Overlapping gold-standard = positives=20 are shown in green, PIE links in blue, and overlaps = with both=20 the PIE and gold-standard positives in black. (Right)=20 Overlapping gold-standard negatives are shown in red. = Regions=20 with many red links indicate potential false-positive=20 predictions. (C) Three PIP complexes that we = partially=20 verified by TAP-tagging. Each complex contains the = proteins=20 linked to a central protein (gray) after thresholding = the PIP=20 at Lcut =3D 300. Interactions = verified by our=20 TAP-tagging are shown in dark blue and PIE links in = light=20 blue; gray links indicate where TAP-tagging overlapped = with=20 PIE links. [Vi= ew=20 Larger Version of this Image (36K GIF = file)]=20

We have thus far treated all interactions as independent. = However, the joint distribution of interactions = in the=20 PIs can help identify large complexes: An ideal = complex=20 should be a "clique" in an interaction graph = (i.e., a=20 subgraph with N(N =96 1)/2 links = between=20 N proteins). Although this rarely happens in = practice,=20 because of incorrect or missing links, large complexes = tend to have many interconnections within them, = whereas=20 false-positive links to outside proteins tend to = occur=20 randomly, without a coherent pattern (Fig= .=20 4).

= Fig. 4.=20 TP/FP for subsets of the thresholded PIP = that=20 only include proteins with a minimum number of links.=20 Requiring a minimum number of links isolates large = complexes=20 in the thresholded PIP graph (Fig= .=20 3B). Increasing the minimum number of links raises = TP/FP by preserving the interactions = among=20 proteins in large complexes, while filtering out=20 false-positive interactions with heterogeneous groups = of=20 proteins outside the complexes. [Vi= ew=20 Larger Version of this Image (8K GIF file)] =

Fig= ure=20 3B shows parts of the thresholded PIP that are = restricted=20 to proteins with 3D"=3D"=20 src=3D"http://www.sciencemag.org/math/ge.gif" border=3D0>20 = links (23),=20 highlighting large complexes. Some predicted = complexes=20 overlap with the gold-standard positives = (cytoplasmic=20 ribosome) or the PIE (exosome, RNA polymerase I,=20 26S proteasome). Comparison with the gold-standard=20 negatives showed where the PIP likely produced = false=20 complexes. Many protein associations only appear = in the=20 PIP and thus potentially represent new = interactions and=20 complexes. An interesting example is the = mitochondrial=20 ribosome; it has appreciable overlap with both=20 gold-standard positives and the PIE and contains=20 plausible, newly predicted interactions with = three=20 proteins (19).=20

To further test the predictions in the PIP, we conducted=20 TAP-tagging experiments, in which a protein = expressed at=20 its normal intracellular concentration ("bait") = is tagged=20 and used to "pull down" endogenous protein = complexes. We=20 picked 98 proteins as TAP-tagging baits. These = produced=20 424 experimental interactions overlapping with = the PIP=20 thresholded at Lcut =3D 300. (Of these, = 185, in=20 turn, overlapped with gold-standard positives, = and 16=20 with negatives, highlighting the reliability of = our=20 experiments.)

Fig= ure=20 3C shows three examples of the overlap between the = PIP=20 and TAP-tagging. We predicted that the putative = DEAD-box=20 RNA helicase Dbp3 interacts with three other RNA=20 helicases (Hca4, Mak5, and Dbp7), with proteins=20 implicated in ribosomal RNA (rRNA) metabolism = (e.g.,=20 Nop2, Rrp5, Mak5, and components of RNA polymerase = I),=20 and with Nsr1, the yeast homolog of mammalian Nucleolin = and a GAR domain=96containing protein (24).=20 When Dbp3 was TAP-tagged and purified, we found=20 previously unknown interactions with Nsr1, Hca4, = and=20 Nop1, connecting Dbp3 with known rRNA-processing=20 proteins. Further purifications with TAP-tagged = versions=20 of Mak5, Rrp5, Dbp7, Dbp3, Nsr1, Hca4, and Nop2 = verified=20 the physical association.

The nucleosome, a fundamental unit within chromatin,=20 provides a second example of overlap. It is = composed of=20 eight histones (two H2A, two H2B, two H3, and two = H4),=20 which can block RNA polymerase II progression. = This=20 blockage is relieved upon interaction with the = FACT=20 complex (also known as SPN or yFACT), which consists = of=20 Spt16 and Pob3 in yeast. Mammalian Pob3 has a high = mobility=20 group (HMG) domain for interaction with histones; = however,=20 yeast Pob3 lacks this domain. Instead, the HMG = protein=20 Nhp6 (with two virtually identical isoforms, = Nhp6A and=20 Nhp6B) binds histones (25=9627).=20 [Nhp6 also binds DNA in competition with the = nucleosome=20 (28).]=20 Our thresholded PIP and experimental data = document a=20 specific interaction between Nhp6A and Hhf1 (H4),=20 pinpointing the contact between the nucleosome and = Nhp6 to=20 the H3-H4 heterodimer (Hhf1 and Hht1). This is = plausible;=20 because Nhp6 has been shown not to influence = nucleosome=20 reassembly (29),=20 it is unlikely that it binds with the H2A-H2B dimer, = which=20 needs to reassociate with the nucleosome after = binding=20 FACT.

The replication complex, a third experimental validation = of=20 the PIP, assembles and dissembles from transiently=20 interacting subcomplexes (e.g., MCM proteins, = ORC, and=20 polymerases) throughout the cell cycle (= 8,=20 30).=20 Our predicted and experimentally verified = interactions=20 connect it, probably transiently, to another = subcomplex,=20 replication factor A (RFA, composed of Rfa1, Rfa2, and = Rfa3). Specifically, we predicted and verified=20 interactions between RFA and two proteins = associated with=20 other replication subcomplexes: Rfa2 with Top2 (a = component of the nuclear synaptonemal complex) = and Rfa1=20 with Pri2 (DNA polymerase 3D{alpha}=20=96primase=20 subunit).

Finally, we predicted and verified by TAP-tagging that = two=20 proteins involved in translation elongation (Tef2 = and=20 Eft2) interact. This is plausible given that = protein=20 elongation is mediated by three factors in yeast: = EF-13D{alpha} (Tef1, Tef2), EF-2 (Eft1, Eft2), and = EF-3=20 (Hef3, Yef3); most other eukaryotes lack EF-3. Previous = experimental data suggest an interaction between yeast = EF-13D{alpha} and EF-3 (31).=20 An interaction between EF-13D{alpha}=20 = and EF-2=20 had not been demonstrated, although this is = reasonable=20 given their similar roles in elongation and their = overlapping binding sites on the ribosome (32).=20

In summary, we have developed a Bayesian approach for=20 integrating weakly predictive genomic features = into=20 reliable predictions of protein-protein = interactions. Our=20 de novo prediction of complexes replicated = interactions=20 found in the gold-standard positives and PIE. In=20 addition, we confirmed several of our predictions = with=20 new experiments. The accuracy of the PIP was comparable = to that of the PIE while simultaneously achieving = greater=20 coverage.

Our procedure lends itself naturally to the addition of = more=20 features, possibly further improving results. We=20 anticipate that protein-protein interactions in = organisms=20 other than yeast can be explored in similar = ways.=20


References and=20 Notes

1. P. Uetz et al., = Nature=20 403, 623 (2000).[CrossRef][ISI][Medline]
2. T. Ito et al., Proc. = Natl. Acad.=20 Sci. U.S.A. 98, 4569 (2001).[Abstract/Free = Full Text]
3. A. C. Gavin et al., = Nature=20 415, 141 (2002).[CrossRef][ISI][Medline]
4. Y. Ho et al., = Nature=20 415, 180 (2002).[CrossRef][ISI][Medline]
5. I. Xenarios et al., = Nucleic Acids=20 Res. 30, 303 (2002).[Abstract/Free = Full Text]
6. H. W. Mewes et al., = Nucleic Acids=20 Res. 30, 31 (2002).[Abstract/Free = Full Text]
7. G. D. Bader et al., = Nucleic Acids=20 Res. 29, 242 (2001).[Abstract/Free = Full Text]
8. R. Jansen, D. Greenbaum, = M.=20 Gerstein, Genome Res. 12, 37 (2002).[Abstract/Free = Full Text]
9. A. Kumar et al., Genes = Dev.=20 16, 707 (2002).[Abstract/Free = Full Text]
10. C. von Mering et al., = Nature=20 417, 399=20 (2002).[CrossRef][ISI][Medline]
11. A. M. Deane, L. Salwinski, = I.=20 Xenarios, D. Eisenberg, Mol. Cell. Proteomics = 1,=20 349 (2002).[Abstract/Free = Full Text]
12. A. M. Edwards et al., = Trends=20 Genet. 18, 529 (2002).[CrossRef][ISI][Medline]
13. G. D. Bader, C. W. Hogue, = Nature=20 Biotechnol. 20, 991 (2002).[CrossRef][ISI][Medline]
14. A. Kumar, M. Snyder, = Nature=20 415, 123=20 (2002).[CrossRef][ISI][Medline]
15. A. M. Marcotte, M. = Pellegrini, M. J.=20 Thompson, T. O. Yeates, D. Eisenberg, Nature=20 402, 83 (1999).[CrossRef][ISI][Medline]
16. M. Steffen, A. Petti, J. = Aach, P.=20 D'Haeseleer, G. Church, BMC Bioinformatics = 3, 34=20 (2002).[CrossRef][Medline]
17. R. Jansen, N. Lan, J. = Qian, M.=20 Gerstein, J. Struct. Funct. Genomics 2, = 71=20 (2002).[CrossRef][Medline]
18. A. Drawid, M. Gerstein, = J. Mol.=20 Biol 301, 1059 (2000).[CrossRef][ISI][Medline]
19. Materials and methods are = available=20 as supporting material on Science Online.
20. T. R. Hughes et = al.,=20 Cell 102, 109 (2000).[ISI][Medline]
21. R. J. Cho et al., = Mol.=20 Cell 2, 65 (1998).[ISI][Medline]
22. M. Ashburner et = al.,=20 Nature Genet. 25, 25 (2000).[CrossRef][ISI][Medline]
23. See http://genecensus.org/intint.
24. I. P. Girard et = al., EMBO=20 J. 11, 673 (1992).[Abstract]
25. N. K. Brewster, G. C. = Johnston, R.=20 A. Singer, Mol. Cell. Biol. 21, 3491 = (2001).[Abstract/Free = Full Text]
26. A. A. Travers, EMBO = Rep.=20 4, 131 (2003).[Abstract/Free = Full Text]
27. T. Formosa et al.,=20 Genetics 162, 1557 (2002).[Abstract/Free = Full Text]
28. Y. Yu, P. Eriksson, L. T. = Bhoite, D.=20 J. Stillman, Mol. Cell. Biol. 23, 1910 = (2003).[Abstract/Free = Full Text]
29. R. C. Bash, J. M. = Vargason, S.=20 Cornejo, P. S. Ho, D. Lohr, J. Biol. Chem. = 276,=20 861 (2001).[Abstract/Free = Full Text]
30. O. M. Aparicio, D. M. = Weinstein, S.=20 P. Bell, Cell 91, 59 (1997).[CrossRef][ISI][Medline]
31. M. Anand, K. Chakraburtty, = M. J.=20 Marton, A. G. Hinnebusch, T. G. Kinzy, J. Biol. = Chem.=20 278, 6985=20 (2003).[Abstract/Free = Full Text]
32. O. Kovalchuke, R. = Kambampati, E.=20 Pladies, K. Chakraburtty, Eur. J. Biochem. = 258,=20 986 (1998).[Abstract]
33. We thank C. Sander and G. = Bader for=20 critical discussions.

Supporting Online = Material=20

www.= sciencemag.org/cgi/content/full/302/5644/449/DC1=20

Materials and Methods

Figs. S1 to S3

Tables S1 and S2

References

29 May 2003; accepted 29 August=20 2003
10.1126/science.1087361
Include this information = when=20 citing this paper.

A= bstract=20 of this Article
PDF=20 Version of this Article
Supp= orting=20 Online Material
 
Down= load=20 to Citation Manager
Alert me when:
new articles cite this=20 article
 
Search for similar articles in:=20
  Science Online=20
  PubMed
Search Medline for articles by:
Jansen,=20 R. || Gerstein,=20 M.
 
This=20 article appears in the following Subject=20 Collections:
Cell=20 Biology =
Volume 302, Number 5644, Issue of = 17 Oct 2003,=20 pp. 449-453.
Copyright =A9 2003 by The American Association for = the=20 Advancement of Science. All rights reserved.=20

3D"Functional 3D"Next =



------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: application/octet-stream Content-Transfer-Encoding: base64 Content-Location: http://oas.highwire.org/RealMedia/ads/adstream_lx.ads/www.sciencemag.org/cgi/content/full/302/5644/991696202/TopLeft/sci/SOL-Elsevier-10.15.03-10.30.03/searchalertbanner_w234xh60.gif/38323834373861323366393164363430 R0lGODlh6gA8AKIHAOTyzs/snKrVY09+AWGVC////4fFIf///yH/C05FVFNDQVBFMi4wAwEAAAAh +QQFXgEHACwAAAAA6gA8AAAD/2i63P4wykmrvTjrzbv/YCiOZGmeaKqubOu+cCzPdG3feK7vfO// wKBwSCwaj8ikcslsOp/QqHRKxQQAgWsWq+1yv9mweEwujwVZdBgtYLvb8LigKg2Dvdt8l4zlmv9i amprAW2FbgZziolzJG0KcCFoHYVWAQxsiZcreJ13fn17gKOkhWtuh5GPqyABBQAGAAUFIrMcAq8X rgWNBq5zsr0odnqfxXx5d6VlgmeHqamrisIcrlizjWhYC9qbV4IMsq8CWOSGANsKVwCKuVoN63O7 7JDi5Lzl6lwinsf9ocjG+Alkhg0zQwgzMVpIbQOuWQW8vbpCyxrFWBBtLRDHLv/jvSwRDcxCFxEX yU0LMkakSC/RyHsQYckCCavVwIGiOmlhR4CAwEq+yDRipi7NGkimECVi2FDDrpEKglG8pw1fSFco RdJ6mAaOxZDkCmVsgOuSSYzCsGK8ZGvcxVbLSgkgEFQMz7qBip5R16ySnDiQpj2iS2nkLLav0HVM jI6qL14MbJ1NJOtaR29tMmZVqxUtppDBtD5U3NLDq9OVUatOzTpYz9bpSFfeJBudWZ2ahAaWA+Js ZTS8tAE/d2+xMGyT1Vrj+nsiNm4T1YZeIB1y2+DoQqjczr07tp7cA9ANT6A4yW3kaj8iyoj3B2su c4kDO/9S6pDhsOWKj/rxS9//+6kTEzCQQRdRaJLN15QG3jXo3VwEkKcVR/RRiF5t+RD1VySEmdBQ L8Es+IgDTY1YgYkQoPjAghs46KJK4kW4HW2GHVZeZamZheFturUHWA3T0ZHBi0TyAt52cz0UE32r 6Vgbj2cI9qOQSRT5IoTcwRLhfejQxVpiNz6Zm13ZbNghlUVAtOGaG6rZUxyKzeVjnOXFEU+dXqVB lxyHQLIbixcoRGJWFAAV6CaTrEhoBYauoGZSRjUjaSHixNjVbD0lpZh4mrQRyp550lNnp4UQVmYb ZzpVoAManRigBQjS8kCrFkzGgpUuYjljeeFR1iRlTxqi25q9VUgPR/oAtc4h/5VisixllHKUnj7p aNQFUvtwU862i9aCa4O6wuilYTON96WTskFpx6lypJoBViZNhSw2M5lkT7MpWSXWkuPYgmMu/q4E 8GmrmqTkcyZ8C+6R4o6mma+soUuauqf4GMl7YHkakmTROVdOOzVxY0pJ/8U6WXZtsTMVNo0a7A5+ JajJ5swaR/QmHJvefM5ddjYGqrCVlNfeIXsuFYe7GOACnGLWVZTYOr4oaKuvLNlq8n6V6McYOzDB 7DIsnMXMskFkQ1q2mz6x8RtdmvrsC875dKoGT3+lwc1urRwImVrIucXtRwZTd1W/Ic/U2WG7HI5d tEtD1xEsSp+g8IMMZwTAeP/bQT7fSZ4Gi9dRFsOBtC68PGW4OP6N03VNUEGy5OOC88Kxmp11TSDt rsMeNgmTdxeuRwqoNtGNiZ2myI5j5kWsIyT2guLFgZE1gYoLpUgWoLf2vquMKl2uVY00thZmuslv we7RaB4hM81s2qNzejzDHf/OSdqpp9Fy391e+kiMBKkpAOyLINDWNpulLVrb8Amz3pE2YbWET6Xy U7v4lybtqeR3EBHPwTLIKwtJbFNAC4SUoGeERIkMewxolA4sKC7uacZcqtlE8WYDLPJ9zm5sGt0P gvQYFObOByz0SOU42DUifulCYhpEXkLnwwlkJyyaSMc3YiETdpRGGxfS0Uf/cEiPbmRDCzUj1LOo aI6oZYE6ttGO6ti3JvfhCX4x6hkWfoZFCNkvaAtJStH+pMMKSIVkY4nVhGD2L7GATW9jgddEQlQj mGRlJvAxDEVoshb4fCCDAjRKACPFQQVi0UhB6dxdNDXFhJjvjcJynmCKdZHn/EKQ/usF4kDWQ8P9 AivwwuXeyvEquV2HlyqThS9lZZogfseFMRkXjCCGGs0FS4mgW1MfKxCTzjxmKooj0XxkBzm9yYNk 5rFNaOLQy9lcx3VbK07iplE9CxjzmNxphOVsZJ7TaBB50OwKE0Owi28uBiPXoAWtoqIyrfWzOaVL qCkKuhhDOi8XD7GmaNJz6haueHOgE3inAbOEuc1drjioYZZ65DEsaX4gor46THxQh9GngEaS9zHL 4Fq3ueC9il8Std33KiUwlLrznRj8zupUClJMwW9iwOjRhnpjOG5Qo4nPWxE6UfKh61l1ephoIjXf GSOODnIk3juiM9NVHj5I4y/TnMByZjA1JkQKFaaMKyoM8D70JUR0TJTT8pZatEVolYI56IlgB0vY whr2sIhNrGIXS9fGOjatgNXBYicrWLpSdrCWtWxPMkuXzjbWs44NrWhHG9nSmva0qE2talfL2ta6 9rWwja1sZ0vb2tr2trjNrW53y4AEAAAh+QQFAAAHACwMAAwA1QAmAAAD/2i63P4wykmrvTjrzbv/ YCiOZMkEgimhD6u+cEwKBSAztN3kd+//kZwAEDAEACmAzjBEIY3ElOL4nEaZUSK0ydVuj1KFgFoE ms8UWmFdGBdQtYV6rR6+DWvlnQ54C+N0NXN0AWx0DHx3aIuMYnFqeERtcnluRIWSRVw1AZg8gUiC cYWFk31hTCiYjawqAimvTGE8eAUGfWs4cUY1REqkU5a+oTaYbX/Fbbm3k2J9R7ut0iGnoJRvq2pl jjpCR3Bvz95Rj2xMojY0nXmHU3er0/Ee1cpLOWxS7dwLpYK8lf8E0Xp2jhivIp2gzWLTZ0mEARAj SpxIsaLFixgzatzIEf8iByl2HAZBJWsHSQzfIO3A0LGly5cwY3r0UGfaoG0cIBLYybOnz59Agwod SrSo0aIz5YWIBULn0adQo0qdyjOp0mlOqWrdyvWp1RAuxIQFMQbCmJMesnZdy3bt1w+03KHlsMxB obk5B7Tdy3fqVzBDkjxJye8Kj5ThNHnBkgKMkiaFVZJ5hQvWt7R6+2reLPTrKU+5EhURnYMgqT7M wgCjx5CZIWbrQnHKAa2DWs64N/8NFwjOJVCxB7lbB2946kPV0uU6CwkYoVN5c0vXPcBBjeufe5HR /tiQpmNscNJIoszWcR25GpY6lyKPkioabk+fz/VtZXNCNiFBLKnN6rr/7vSnzXprrFPPObasctcY 8GUgH30QSvXWevD088QcSPCg3jvRLGBIeZUhiEst19mijR0F2pZZhCxKWB0GTMmBFz/64ICIiSvh QFKMTHzwYItAdvbiDQ1VAGAjPwapZE9vXYXkiktG6VOTTi6SpJRBfnXWVWWh8ApOL8gk5phkutTA XVedomYzMJTp5ptwDumME500qAQWmlx2SzpZOBHGY30689gWayYEkp4p3Vnloij59yED2YFjznKi 4OPhhwXStqYbKJrTD6WZMCpqBXf9t1Bth9wFqhoZXpoaMK/YIWuhB0GHJi4ijarrDn7cAd2lgJhH yqodKpCeo77podCmXbxIxuEuPO6qKymmnpBqpdo0FJerp4xnaxubaktIpSJOcqS0opZqCpsIOhPe P+LmeqwbBl63LLglKvauOIGYh+6/QRAko45wVbDjjAAnHIEnCjfs8MMQRyzxxNIkAAAh+QQFAAAH ACyOABcAMwAGAAADPGixwB4NvkinlWfow8WSYBVeoggG2daNbEm+jpAOnOG5LYwHRsb9wMMHR9R9 ZBpa8Kf4PIbO0/PZECSXCQAh+QQFAAAHACyPABcAOQAQAAADnVi63P4QmkGJvdfEzXsBBDVgZAB4 aCqII3mdacwFrGsZgaw/wFTZBMFu2Aj9bAZAwLRUMp/OaHMKpUYFxpZNsKx6pV+rGABYiYCEZ3gN blObZq0LxxbXvdNyFs21u+9uUUp7QAJgXU6IXW+Mao5YZ4VViUqUapWKlkoGIWhzjYmgi46PhJ4Z mal5qkyWAXEuLLJnI7Wdtri3urmzswkAIfkEBQAABwAsjgAXAD8AFgAAA9cot9z+MMIwpAw2ayvq IdtSFGG5GQx4mqwFqF8okAsg3Hiu7/wOMJQWBtgrGnmNICswajqf0Kg0ejB4YBnDdMvtjmBYDsBL LsdYs7KaqzSN13DqNWRgxu8jgHURlqTxeAcDgy0MPw2Hhg6JNYsHAQABkJKChCFDQIyQSZqdi5KR kB2DHhoEk6ihqqCsqa2rqJM2LaeSr7euubCwon0aAri7wcOukZEEg74RAryttqHPts64rALIA8oT 0MbR27HcrM0f2IUL07Hn0OnV19m/3fDh8bLOowRhCQAh+QQFXgEHACyOABcARgAeAAAD/0hxfCEw yknrXAyIxrv/BsZYZFmJAMEMX/sJgCvPdLrSLrHhPF8EqgOr11E0CsikcslsOo874ocQYziv2GdD I20ZqiOTuIQp2ITdhg6sGbsp7MYwrUvbj6J513i3F6J9QX1pBQY3dgRfg11mQXpSdVtvkxFgeXcq Sg+Uk5pymAdZoqNMAIaDAwSkq6uCfQassVhnr6Gyt0sYqXeGuL5Iuq5SG7+4MQOpwkQCxb4EyMo9 BoABYA7W1RzZDdnbAAHV4N9AyI9S4OLp4evj7eru7O8C5dE98ffw+e/76Rr0dwLQ8dOHb6C6agKe 7QJosKBDgt8ikltoB8bBi+vQjdMo8CxeuoTQ+mSU2I/jRpLh4IE78KyeHYwbYXaMGQ8kxUUHApbc mdKkz4zzknVIAAAh+QQFXgEHACwIAAoA3AArAAAD/2i63P4wytkAKLcIyrv/oCRsYdmMKBp4aZsq blwGQqaZeK5Pxb2Dtl7gQqIEesgasgcQLDHKZRHkxPyuRlpxRGMIAl3YJvxdKXpcLGfEHF2GTcPX Yl4cm6j3F41O6Uc4RwV1aoVnUgaCTAqKPYkYRFVJBk8+hhBvChkXAAZvdwx3foOPTUJgG59hJRlT l1eiqE5xRysWqYONQ1CClBiChK8MmZ4as8CCRY0aQaRPnUGdFC5yrcJqkryVK3OtwGw+gGjFrtfF ZpFQTHDKUHm/nFUo1YPzEWUWTPWc5ti7Qr/KgKOhx8qhDTU6jSPWTxMpcpSaRAJACNQCInLU/ULF Cf/MKi//6IChI6Thj2SghCTcY0ylwUcYFi18aJKOw4ydotDMSAjjLJ1ompWT9NFOSZM7giBcktPG p5fF2vjCVQ7pvarCvkXoGMdqDkBeXIGdMNar2RlHT8DouPOs27dwPwA7YUGkSoxxS5SdhvXH3rw4 lHSlB2/JCouAPzAk6+jV4sRoLUVpN+4n5HtaHJrhsmUMy6GoQo0F9GU0ZzEfS1ft67WKNMJTBCPE e7kCU4iNOt0hguii4YxSqtwp+SZTbodICJVZ1U0ZieVdayCs09zOyDpzGaWFcVQJiSFDtmCnuMoj xQXdxoN/XhS9hYMFadUznJ27sS7jamWjaCVoy///YJDiHXfkbXZLeGdslgtYoAyxVhOhkVPeOLA1 kN2AnkD40AUEYqhHHbdIB4N+dpwHwUitELFMLn2A48ATNFSigUoDOfKUNCpsY51H3G2hiYI0jEfR e488dxGERVgD02u+hQNFNehI42A1Q7524HPheDGYHO1pR5E88QUZ5EtVPDASH/XI0phDmcTHiEQB ZcYIQTfUoEUXdnInZyLQGOSECiSE+Fw82kFFT3QvhWgikU4Q1JWgep7Wo1onasMMgCshJpwXp9iY 5pdrwnRDOpD0gQY7JYKXqomJGDjSd9ywCoesBQZKYX32WUkTeLJuhmdX4EWY0JfusXYCEikKeBti /75YIqqATrnIwCLnAHeULg6sxhpCHWjr3pMwBXMhuDDIoddQxna7rbF/yYFuYO/WZuGt1K5F4WPy 5qsvLN0lx4VhSgSz78AEh+CaUfoMCWPBDDcsV1rDhhFsug5XzLBsXlis8cZrSMVBhByHXDDGV+G7 74oypqzyyiy3vLLA+0bxJWlcEHpxMy7nrPPOLW9JsE75MEVxYjjzbPTRRzM58qwiDQ0YyoWpHKOG Lp+JtNEwizywygR0DXUBXWNAQDYrG0BA0Ve3rHXFKQdAgNmuzVaAAm/nhMJ9xlSzQd318AkwIHdw 0ygbG9RD9txrN5yyAG+/PUjXjBNAA+QAjO02JeRhw834CF17snnkUnQt+tyij1055mdvXkniisvY deWSM37E3WGfDrbkY0NBUdiMc2i57MnlbvbljpztNeRm48w6wzJevvfeolzguNt/Nu7g2KAnfzt7 SUiugfFlpt443c0usXzBMkbeOO6JQH67BdObTYDYlKwwv+qbyyeE8MZjH7bblnPf6s63tUp4DQxv s58A5Ue9Hiywc7x7XSdElw3gyU6BpXhbDbwmIwIWUGdBEwLakNUM/iALhCNUmQdjlsK0ufCFHluh vg4HwxqmzWkyhMvXbMhDlmUth/pqYQ+HGCoFJAAAOw== ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: application/octet-stream Content-Transfer-Encoding: base64 Content-Location: http://oas.highwire.org/RealMedia/ads/adstream_lx.ads/www.sciencemag.org/cgi/content/full/302/5644/1695603079/TopRight/sci/SOL-AACR-7.23.03-8.21.03/targets_science2.gif/38323834373861323366393164363430 R0lGODlh6gA8APcAAP//////zP//mf//Zv//M///AP/M///MzP/Mmf/MZv/MM//MAP+Z//+ZzP+Z mf+ZZv+ZM/+ZAP9m//9mzP9mmf9mZv9mM/9mAP8z//8zzP8zmf8zZv8zM/8zAP8A//8AzP8Amf8A Zv8AM/8AAMz//8z/zMz/mcz/Zsz/M8z/AMzM/8zMzMzMmczMZszMM8zMAMyZ/8yZzMyZmcyZZsyZ M8yZAMxm/8xmzMxmmcxmZsxmM8xmAMwz/8wzzMwzmcwzZswzM8wzAMwA/8wAzMwAmcwAZswAM8wA AJn//5n/zJn/mZn/Zpn/M5n/AJnM/5nMzJnMmZnMZpnMM5nMAJmZ/5mZzJmZmZmZZpmZM5mZAJlm /5lmzJlmmZlmZplmM5lmAJkz/5kzzJkzmZkzZpkzM5kzAJkA/5kAzJkAmZkAZpkAM5kAAGb//2b/ zGb/mWb/Zmb/M2b/AGbM/2bMzGbMmWbMZmbMM2bMAGaZ/2aZzGaZmWaZZmaZM2aZAGZm/2ZmzGZm mWZmZmZmM2ZmAGYz/2YzzGYzmWYzZmYzM2YzAGYA/2YAzGYAmWYAZmYAM2YAADP//zP/zDP/mTP/ ZjP/MzP/ADPM/zPMzDPMmTPMZjPMMzPMADOZ/zOZzDOZmTOZZjOZMzOZADNm/zNmzDNmmTNmZjNm MzNmADMz/zMzzDMzmTMzZjMzMzMzADMA/zMAzDMAmTMAZjMAMzMAAAD//wD/zAD/mQD/ZgD/MwD/ AADM/wDMzADMmQDMZgDMMwDMAACZ/wCZzACZmQCZZgCZMwCZAABm/wBmzABmmQBmZgBmMwBmAAAz /wAzzAAzmQAzZgAzMwAzAAAA/wAAzAAAmQAAZgAAMwAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACwAAAAA6gA8AAAI/wCfCRxI sKDBgwgTKlyI0NozZBAjSpxIsaLFixgzatzIsaPHjxMZihxJsiRBhyBTqlzJsqXLjCZjypyJ8qXN mzhzspzJs+fCmjqDCh2q06fRowKBEl3KtOnGntaiOiQITerUglWtRoUG9aHTr2DB8nQIAMCKq1XL Anh2NalDK2oB6FkFhAOQu0DIIBoJVKAzigI1BnY5+CPCsDgLQxxbpeyTqVWfqGU7sOqzFXGfVLF2 CJFnMnhB7/3p9dmTPHm8LtasByOpsqpVPoP9kVQV1LjzVCGl01nsoLPXSqQJSO3mpKXiUmZrTXJZ PQKhWeZaGTRoINQblp4s8TWAKr+7A/8gEd4wbY+e4sb19PemszzkiQaPLZOscYdR1VN2iNks863P ZEcQK3ndpVBNwQGQ2mL9LYiMb4W9Rt5DA7W32ECLReQbMvNRiOFizvhmIYfOVFNWNSEWFFFgg2Go 2IcdQgRheRWumOKLA8GmWn2YqVDWZlVh1hhsUT3jo1kAimQdEKMdhCAAzm2Xh4IUTvmdVxIONKQn pQ35xDN6PNGaM1U8QUqHzkhGAnjIkHLaa3kMd40zZTlzzTXWWMmmaVXQ+eUzjZFXJoWk+Pinc3ms 4Ft6UI74IGYrcNnmmz92OV6MF5ZkTXFVFPddVMWpkCBz/QFQZEl5gZbQk6cBwOFlAMD/yVVZeXj3 jDNZ0hnrlGcGp5mojYEX3Jk60koClM94Ryt983HlYx5RJgiACiZC6dxaYJqFzImIivqamaJKROcK VnKprJWp+XptYSblZ2pynwY3K5F6qAVIWyORoS8Qq26nwrZ/7qpgc8hCk96XElrTGCnuVoNZugJN me6xxGKLTDXVbLuWdxeGpOOr1QwbHHjvXQmNlQKdaM1relQboGlQThWSQ695SySjz0wJ3skfZypS WrFu+uM1mOlhzSpquRtvTEv2u9h41pw4Zc4D+8jwvMlCbahmJ8J2jXRUTxwrmqVuPO0zeP4WI9Vq VYy2kUFfM593ZUoGXp1TlaWCQ8ms/1ioWuRlufJ41Vhtjdw9v8oXXGcJ/V1yK+CJdFnWwFXWvTLp q6p2T5OQZ6zTfp5aWaRQB1vCjj2h2ZdEVqmgs2PTRqvLWXIFmI6+wlqxM/YxnLLZ36l+W7MklrWC as4wSmzgl2btOel5W+zzT1GbCqp+UU1uan/H4ztSqvxyzmHz47kqOsEL6uo8wb5bRrpXZA4sMmyS 5WGN+lnSqGPC1bp9a/1eKZapuHI49T0EQqRbEcU+RwJo5C9hAIRI4kYCtFJYoyqeKosVLng0va3i SGcR0EgQUaAmGQQlo5uQmiL2Oo2FTGLXyFVZOISMPPAMW6mpl8Bit5YpnWUy+atIjP9ONCT/tUmA 2JLMoX73IECRolo05NDDgrO3BxLOVqPaEUOqUq/GsSWDIazK5FYAiGPJxXtbBN+BtmMxAHTCIUoU CKNoFZVO6GiOZ2MbEtXCJdjoCnSuwqIQ6ySQ/jxBSMHZkB6xmKb16LGRy4pN22CDs2dU0kqA3ND0 tFMWPEnFU42TSnBWMCQAbNAkiLhO+Jx2Ieo4AxrXSMatAiTLZKXra03MDik6MSOBdOJMyZAlKXr1 Sr8MJE5+4YpvbCdECHGIN8k0ZmmcwUuXVagTr5IlMnjpF2zCckfvMaYyX5bLbpLImZs8YfXmUop2 PsNyK1jFKtq5ClBeC3MIYYU+9Zn/SlUiQoQn3NEF+2YZiBQUGdJJKDT6hjZoGPSCX6NQVgYaoMNZ VCB3is7XYCm3hvZNiHhiKCwh2tG0rSgqU1KB7f5zwWckoyoRrajcQgJL6cQUlxzC5UsPZ9O+KKQq pVSPUIW6gnui8RkEusu+7oIIfYpEKdK81YqG86GO+eyjiiuNiiB0o8AosjwrUk0tg6lViHwLWnWa KTJqORi2rhVGVHVpi7QYVrkya1WTq8IK9kpK9fCVr8YrpdHyqTml6gsRIXtqbEJkJTM5ailkyg1u 7iacnOThSFVAqEM7MiJXAWdV1VuFVqxRqlBmz3iWM+VRB2Qd0fBlR5g0Y1ikpR6X/2HVJgUxKUdI sQJoRqkoCAHqpxRKMLU0TjodnFYpQHiqn7YWCKxQ7GJ0qDhgKm44ihsRQeqqIQ/FtV7QQSC2FDMj G4koUxbyUIDA1kSpZkhGFHrf+KK43faCFSPBdVxLkWMWHYawotojrVosKBICFRagAX2aq9AmFa78 7QkPQkaYkNEYFUiKQ14KUSP/FaIqZHYF/4rIS3lmvzup71hfWoyeHuKmWlHJGZCSVJhIkdIIY7gs MhbTg8pUw9TxBjUOBZRjFjVkwzgpOAR2C+Xg9V8xJi21j0GwQQx8HYagxBk6fBtBpsQ1APhmSH31 8vigVIVwzRAzvimLGZklseyopf+xhVTQt7L25tmQC8dCptUMx0zjsQXrd3221hGx5ZgyZw1c97XI QYKk2pNg5l6gbKmTKVe9soh2JFQ20BpvTJ7NdjdkdqPaCqpBMPBMcSA+ogyyjhUu8kqMjWxRH8BQ CrwL8S5rorJSyGomZIZVL2wCoVjhguY8h1FpIEqUmUecFKr9fpFIkebKpNdSOeN2VCQk3JcJEyxk UakNk1SS2AVf7TWw0apVAKjGsZDBU/q8mqFpVV9jSPkEivE6o39Tc9iKdKKtDQnYz6BYwIktIZdt 9Hd7c+myC5LcSyuZ2o5rcnIhXqonxJQhz932SVSTOA45B1cDazOwWzdmaJ3mCer/JvZtOfTuMb8M NvNWneqO2Oln2DF2nhP376ol8zKJDnZnopiDCUfy4vnHyJUR8Aapw+h7YdB4kp74f8wYs9VSJVWg ie6RFfylv4iXLfX7edj6B2uHWEboQmx5h+ZTr8ckJYbNc8YCp5RzKNkOapJp3wUbYz+RCR3kzxtb d4dl5IRao4tJEpoXIx71BG3FIUcyiwW3IuUAgQ87W5eRmvVAY9YJh44iFzfKcsY239Rq4GdKO5Xg PV4BpmsODoz7FPX2uVgNuxPX0BgNbWilQJ9pijSvfcScCMVEA8Yq8CpFR6sCjdY9G0kTzeBy4BiX Tok2SVMGHxm0vvGwlq2HfLxU/xxh9idws1Atnu8EjeLIevFuSD2DHswkwV9KG8oNj94WaidiTyss mv9a6eIRm1JU11JUm2Ef+MQWOuQ5/HEtZqFXkzeA6nFI0CFl2XYdetFUSUEfXNEJ3BQYxAQ2y0Qi L+NLwDQQvyQQsjSCzRQgNsIivbRNM7JMFBJOsyQ61YBNb1NLu+RMyRKCyqSDs5QdKWhO3ySAQwV1 itcWjOc4Q3VpWQEIh0R13ZMQmrNUTLKBMxMVdyJtPHU40VETBVVRWxFME8V0alURB5VTKMFgDMVB +/U1HxUgGyU34gZTC/VQNSVRFmVSbHFTC7VTSQJTxkcRh9dO7UQKpQAIk/cMgP/gdCdRT4x4QY+Y iMO0iM1FhlGhiPdidUi1T/s0EEpRVwqnOLI0h9hFiqr4VjAREnQ1PVaVKbX0alV1VZtUVwylRYHh VoYxWg1GhgBlFcjli5nIcL84FoWIGE4RPymmjEiHFNAYEyvjjNRIVRdUjR0RjdrYLsmIjfLBTN4I E9s4jlbWjeEoFCt3jkJEjuwoPur4jvDYjvLIcOYIj/bIFAqVj/q4j/zYj/74jwBJQPV4jwT5WfM4 jw5xDAW5kIiBFw75kBAZkRI5kRRZkRZ5kRiZkRq5kRzZkR75kRB5hSI5kiRZkiZ5kiiZkiq5kizZ ki75kjAZkzLpkmogkjU5kyj/OQY5gJNksAZroDlqEJRCSZI+eZP6ogZrYJQliZQ/CZQ+KZJF6ZRJ 2ZJCGZQ8GZNWaZNaKZMyAABXmAMAsJNXSZJdKZYx2ZUA8JNqcADqYZb6gpYA8AP6MgZsCQBKKZJg WRYyoC9rgJY4UJNqgAN6+ZN5CQAHcJcmqQalcgBUOZZo2ZRkoAZduZdHCQCUCZOPqS9g6ZZjqTll iZgreQArwJZNKQOmiZYiWZZ5GZQAIJiHuZSWSQZsKQN9GZafCZYygJtqkAM5sAaCeZlXiJhqUBZk kJeQeZJrmZY8iZZyqTl66ZllcZwtiZYrQJhhGZlksFfYiQNmuZc1uVc3mQNj/0AGMdCbY1CeNdmV ZCADMRCcMrACmiOe6zmSa0CawbkGlgma9XmdSGmYoAmUNRmdhumTZrEGmOGT9hmZYxCbIvkD+Rmc pLmfapmd8GmUK0CZZcGb2Hmh4dmbpnmSaPmakvmcPQkAMXCdmBmWlrkGm7mbZoEZvRmdycmilmkA domfSZgDI2qZZdGcAGAA6omjxGmTCcqX0VmSONqU+PmayDmZw+mdpOmf6+mVZDAGvymdZGBcRpmc UzqexfmiYRmgB3AAuamXtDmbxCmkamGSZUml+ImbmgkAjyClL9mVj8CWVRqWS4qgaRmkK9qajzCn tCmj0bmnk0mgacmWgTqgMv86kjN6hQsqA/oJo5qzpP95hdEJpzMqpZ8ZoiN5omoBmcmpFpK6pxJ6 or3Zkwm6onuKpzQKnFfYlfhJm+oZm2t5ALWJpSopqziapTrJoJ9pmWBJo3ExqIdpqYYal5EZpeqx pyT5qEaqnD+wAjFAmS4qqZVKp0h6pJG6rGnJqQDgpT95pJhqPFsaoV35A3A6pTG6oo+art4qoWTg oLAKnfs5pzv5oIYpA7OZok2pFnLZmvoimKpqFq+pp0+JrP6ZrHwZpbhalJb6rEUKljigL9NarZEJ ltV5hRG7rXZZkwqLqxFapPvJmfSqlNAamwJLBgTbk5hxpsrZleMpofa5riP/KasOeqKvWpunSa7T qZwuqqeFeqSrWaJpqaQLm7S42pUPO6Cu+pMdS6TKabS6CpaHqaFGyqTcGZwZWpw6Spq6ma7D+ZpJ iaPC6aiKup9Um6S9OaeJerRZSpuzKq+6+QOwyqt6ea1zW7D+epRlqS9xka0GYJRx0a79eaPRCZdH OpxqgatlAZqM25SY4aVQKVRiabaAC5zfh7hlgasFe6SKy5mJ2ZY3GbiKuaZBm5aCqRasqZy4iZ+D G6vKKZjWWaayy6QtKZ2QyaKpWqlKuQbqqqQ3OZU9ObyEubvFaZXE+6yiepI++bzHSbyyCpXQO7xy uaXzKqq8qauJ+by/u71HRlmcx2mWLNqTR7m7AYqYyMuX7Muxl9qZ8Bu/M4mfzSm/9suTT/q+97u/ /Nu//vu/LKm/ADzABFzABnzACJzACrzADCyTAQEA ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/global_images/scionline_home_top1_off.gif R0lGODlhVQATAJH/AP///0tLSwAAAMDAwCH5BAEAAAMALAAAAABVABMAQAJtjI+py+0PIwKJGoqn zTfYemCbN3njJ6XqyrbuC8fyTNf2jed6w5E9aQL6Fh8O4Gf87ZbMpvMJjUqnVNnxeMH6rsKN9ooy jTqonkir4grJ32FYHYpnkG4xCxxq44tZL7rfwYZlJhZWdYiYqFhTAAA7 ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/global_images/scionline_home_top2_on.gif R0lGODlheQATAKL/AP////+ZAEtLSwAAAMDAwAAAAAAAAAAAACH5BAEAAAQALAAAAAB5ABMAQAOY KLrc/jDKSau9LzSttGcB93UC54TlWK7gYrKeyG7uHKczpu987//AoHBILBqPyKRyyWw6nyvT5yVS qWgkqW2bG6FetWxNC4aaz+i0es1uu9/weDKESsno1k4dj4XV7VWBX1t8YX4yWXQWijlRe1GNjC1T fFWAWmKUYHh5U3aLkowxgySfgH15p2SVhJJcomOucrO0tba3UAkAOw== ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/global_images/scionline_home_top3_off.gif R0lGODlhRwATAJH/AP///0tLSwAAAMDAwCH5BAEAAAMALAAAAABHABMAQAJfjI+py+3/QJJGWgTo rYFO3R0bCJXmiabqyrbuC8fyTNecd3lht4W6WOmNOLai8YhMKpfM5iuTCYKgQqlUh7PuSBAoEej1 iTFRjDhL+jGoorLFm+Vpwt857u3M6/f8RwEAOw== ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/global_images/scionline_home_top4_off.gif R0lGODlhPwATAJH/AP///0tLSwAAAMDAwCH5BAEAAAMALAAAAAA/ABMAQAJUjI+py+1/ADBzonqV 3cH2T4HQSJbmiabqyrbuC8fnhmWeGIWfx3HyDwwKh8Si8ZjgdSS1HfPmo/mckhGmUs1lkjosU+Ql KZ9gbRKrK1O+yLb7fSwAADs= ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/KE_off.gif R0lGODlhTAATAIAAAEtLS////yH5BAAAAAAALAAAAABMABMAAAJshI+py+0Po5y02ouz3rz7D4bi SC7BGQDoqR7p6hqorM6y3das7s4wrUgBhULEK3akFY25mDKx7C2LuOrN1yTusrtVkoWDOnNEJrCV bD6D2XZUNx2/sOfj7y4tD33hkv8PGCg4SFhoeIiYqFEAADs= ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/global_images/scionline_home_top6_off.gif R0lGODlhXAATAJH/AP///0tLSwAAAMDAwCH5BAEAAAMALAAAAABcABMAQAJ0jI+py+0PYwOJGooR sPkGWx3dR3pg6ZnSyrbuC8fyTNf2jef6zvfQmDlxgCiNCEUUGlNHn/MJjUqn1Kr1it1sLluSlvjp fivd8FCsNWPS3PJPHGIHi+aQZjyu6/P6CF+uhtTWdrJ3BLK2hXfIhuX4CBnpWAAAOw== ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/global_images/scionline_home_top7_off.gif R0lGODlhWQATAJH/AP///0tLSwAAAMDAwCH5BAEAAAMALAAAAABZABMAQAJvjI+py+0PYwIMUBuo uYfnrWgfKI6f6EnqyrbuC8fyTNf2jef6bpcIakp5NL7g5NcxFnnMpvMJjUqn1CrVgkGCMlksMenl YsXfX1a8GbPUY84lHEanSyi2uT05r+JJEvhPomZ0JzQ3t2SVqLjIOFUAADs= ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/global_images/shim.gif R0lGODlhAQABAIAAAP///////yH5BAUUAAAALAAAAAABAAEAAAICRAEAOw== ------=_NextPart_000_0000_01C395B3.D0924F80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://www.sciencemag.org/icons/science_logo_sm.gif R0lGODlhbgAiAMQAAP/////x7//j3//Vz//Iv/+6sP+tof+soP+Zmf+Rgf+Dcf91Yf9mUP9ZQf9L Mf8+Iv89If8wEv8iAv4BAgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACH5BAUU ABMALAAAAABuACIAAAX/oCSOZGmeaKqubOu+cCzPdG3feK7vfO//wKBwSBw5EIRkMqFoRGYKAEBR LEYI0kICIZBKG7KHF/CsBiMDQMBBWgS+MscYYg4epNQSBC5DSBN1QW8ADCddYDMRZYE+DV4HJ3eI JRENDEwNTpQNTSaWTQ8nDpcLIg8LTGyiDagMDaolnw2hLo5epbEEsCIOBVIBBGlSAg4LWF6AIwpd wMOFEg5dyA7HXgOLEgnSAwSDAAjKzMcCzyt7Ywa0KQvWZRDCIrZ/IlfDT/KQEvIAXQMH0vMkoLml TAq4evyesJOSb0U1L+lQMBiDTc6AEWOSVcMlQRgARGMCTAIoQES1hiLu/5TaOMLjpBQQvKFTNwJg ARMAGmaUsFAKrARebkoYQ4CEHy8