sampling & community structure in densely connected networks - - PowerPoint PPT Presentation

SLIDE 1

sampling & community structure

in densely connected networks

Sune Lehmann, YY Ahn and JP Bagrow Technical University of Denmark

SLIDE 2

Yong-­‑Yeol ¡Ahn Jim ¡Bagrow

SLIDE 3

pervasive overlap metadata sampling

SLIDE 4

Illustra6on: ¡Newman ¡& ¡Girvan. ¡PRE ¡69, ¡026113 ¡(2004)

SLIDE 5

SLIDE 6

SLIDE 7 Annelise_Maren_Kromann Alice_Martucci Dorthe_Sandager_Bilde Kaspar_Bredahl_Rasmussen Rikke_Thorsteinsson Line_Friis_Frederiksen Stig_Aagaard Benny_Lautrup Nikolaj_Beuschel Dennis_Meyhoff_Brink Palle_Hyldga*rd_Poulsen Jonas_Jakobsen Rikke_Hadrup Peder_Holm-Pedersen Morten_Felding Marie_Louise_Scharff_Grandorf Kristian_Ingemann_Petersen Kristoffer_Gravgaard Gert_Nielsen Ida_Solhaug S*ren_Stenild_Juhl Christian_Nyhus_Andreasen Christian_Jensen Mogens_Skot-Hansen James_Kang Jacob_Berg Jesper_Levinsen Peter_Langen Karina_Kj*r Mikkjal_Gulklett Gert_Gadkj*r_Nielsen Ida_Marie_Heerfordt Kristian_West Signe_Torsbjerg_J*rgensen Aaron_Petersen_DiBona Ella_Caroline_DiBona Erik_West Karina_Louise_Petersen Niels_Arne_Dam Sofie_Falbe Helene_Christensen Joakim_Grundahl Erik_Christensen Mikkel_S*ndergaard Christina_Egelund_Pedersen Laura_Kirch_Kirkegaard Mikael_*rting_Kristiansen Michael_Nyhus_Andreasen Kasper_Vedel Christian_Christiansen Torbj*rn_Porsmose_Rokamp Anders_Iversen S*ren_Kj*rsgaard Steen_Thomsen Simon_Falbe-Hansen Esben_Sverdrup-Jensen Johan_Hjortsberg Thomas_R*nn Kristian_Krohn_Djurhuus Mads_Hofmeister Anders_Vedel_Holst Mikkel_Nissen S*ren_Iversen Andreas_Toft_S*rensen Frank_*stergaard Stine_Bang_Iversen Jessie_Hudecek_Kj*rsgaard Sebastian_Bernhardsson Brian_Larsen Lone_N*rgaard_Skoven Mikael_Caroc_Warner Jonas_Mengel-From Mads_Olesen Morten_Houmann_Jensen Maria_Glinvad Kirsten_Boelskifte Jens_Kj*rsgaard Kira_Clausen Andreas_Holst_Andersen Nina_Adelaide_Skeem_Geist Kalle_Hennie Rikke_Ankerstjerne_Schmidt Michael_Tronskov Lis_Agerb*k_J*rgensen Anne_Lynge_Agerb*k Mette_Agerb*k_Kj*ller Lisa_Agerb*k Carsten_Henriksen Karin_Agerb*k Karen_Agerb*k_J*rgensen Caroline_Buckee Nathan_Eagle Jure_Leskovec Anmol_Madan Aaron_Clauset Jose*_Fernando_Mendes Maximilian_Schich Julian_Candia Gourab_Ghoshal Sebastian_Ahnert Vinko_Zlatic* Dirk_Brockmann Pascal_Braun Skyler_Place Zehui_Qu Trevor_Gillaspy Annamaria_Talas Jim_Bagrow Cesar_A._Hidalgo Albert-Laszlo_Barabasi Ronaldo_Menezes Agi_Petroczky Suzanne_Aleva Yong-Yeol_Ahn Pu_WangMichele_Coscia Anne-Ruxandra_Carvunis Sabrina_Rabello Sang_Hoon_Lee Petter_Holme Sameet_Sreenivasan Peter_Csermely Janet_Kelley Ginestra_Bianconi Luigi_Cuccia Nitesh_Chawla Jozsef_Baranyi Joa*o_Gama_Oliveira Nicolle_Haley Marta_C._Gonzalez Mette_Miriam_Rakel_Bo*ll *****_****Dashun_Wang Chaoming_Song Martin_Schwartz Dan_Romescu Rut_Jesus Takashi_Iba J*rgen_Jensen Katrine_Bj*rnlund Torben_Jensen Ole_Jensen Louise_Fynbo_Jensen Torbj*rn_Jensen Sarah_Wasana Tim_SweeneyShannon_Walkley Carly_Wilcox James_SweeneyColeen_Filipinas Lauren_Knight_Lewicki Renee_Randall Hans_Laurberg Rasmus_Elsborg_Madsen Mikkel_N._Schmidt Morten_Hansen Finn_A*rup_Nielsen Morten_Morup Ling_Feng Kristoffer_Hougaard_Madsen Lek-Heng_Lim Anders_Petersen Lasse_M*lgaard Michael_Syskind_Pedersen Kaare_Brandt_Petersen Mads_Dyrholm Anders_Meng Tue_Lehn-Schi*ler Niels_Henrik_Bohl_Pontoppidan Jan_Larsen Peter_Lehmann_Syre_Fin S*ren_McLaks_Lehmann Lars_Lehmann_Hunnam B*rge_Lehmann Mikis_Theodorakis_Lehmann Signe_Lehmann Sanne_Lehmann_Nielsen Lene_Lehmann Helle_Lehmann_Becker Steen_Lehmann Winnie_R*dkj*r

SLIDE 8

Sebastian_Bernhardsson Brian_Larsen Mikael_Caroc_Warner Jonas_Mengel-From Mads_Olesen Caroline_Buckee Nathan_Eagle Jure_Leskovec Anmol_Madan Aaron_Clauset Jose*_Fernando_Mendes Maximilian_Schich Julian_Candia Gourab_Ghoshal Sebastian_Ahnert Vinko_Zlatic* Dirk_Brockmann Pascal_Braun Skyler_Place Zehui_Qu Trevor_Gillaspy Annamaria_Talas Jim_Bagrow Cesar_A._Hidalgo Albert-Laszlo_Barabasi Ronaldo_Menezes Agi_Petroczky Suzanne_Aleva Yong-Yeol_Ahn Pu_WangMichele_Coscia Anne-Ruxandra_Carvunis Sabrina_Rabello Sang_Hoon_Lee Petter_Holme Sameet_Sreenivasan Peter_Csermely Janet_Kelley Ginestra_Bianconi Luigi_Cuccia Nitesh_Chawla Jozsef_Baranyi Joa*o_Gama_Oliveira Nicolle_Haley Marta_C._Gonzalez Mette_Miriam_Rakel_Bo*ll *****_****Dashun_Wang Chaoming_Song Martin_Schwartz Dan_Romescu Rut_Jesus Takashi_Iba J*rgen_Jensen Torben_Jensen Ole_Jensen Anders_Petersen Lasse_M*lgaard

SLIDE 9

Annelise_Maren_Kromann Alice_Martucci Dorthe_Sandager_Bilde Kaspar_Bredahl_Rasmussen Rikke_Thorsteinsson Line_Friis_Frederiksen Stig_Aagaard Benny_Lautrup Nikolaj_Beuschel Dennis_Meyhoff_Brink Palle_Hyldga*rd_Poulsen Jonas_Jakobsen Rikke_Hadrup Peder_Holm-Pedersen Morten_Felding Marie_Louise_Scharff_Grandorf Kristian_Ingemann_Petersen Kristoffer_Gravgaard Gert_Nielsen Ida_Solhaug Gert_Gadkj*r_Nielsen Ida_Marie_Heerfordt Kristian_West Signe_Torsbjerg_J*rgensen Aaron_Petersen_DiBona Ella_Caroline_DiBona Erik_West Karina_Louise_Petersen Niels_Arne_Dam Joakim_Grundahl Erik_Christensen Laura_Kirch_Kirkegaard Mikael_*rting_Kristiansen Steen_Thomsen Esben_Sverdrup-Jensen Kristian_Krohn_Djurhuus Mikkel_Nissen

SLIDE 10

Sebastian_Bernhardsson Brian_Larsen Lone_N*rgaard_Skoven Mikael_Caroc_Warner Jonas_Mengel-From Morten_Houmann_Jensen Lis_Agerb*k_J*rgensen Anne_Lynge_Agerb*k Mette_Agerb*k_Kj*ller Lisa_Agerb*k Carsten_Henriksen Karin_Agerb*k Karen_Agerb*k_J*rgensen Martin_Schwartz Dan_Romescu J*rgen_Jensen Katrine_Bj*rnlund Torben_Jensen Ole_Jensen Louise_Fynbo_Jensen Torbj*rn_Jensen Sarah_Wasana Tim_SweeneyShannon_Walkley James_SweeneyColeen_Filipinas Lauren_Knight_Lewicki Peter_Lehmann_Syre_Fin S*ren_McLaks_Lehmann Lars_Lehmann_Hunnam B*rge_Lehmann Mikis_Theodorakis_Lehmann Signe_Lehmann Sanne_Lehmann_Nielsen Lene_Lehmann Helle_Lehmann_Becker Steen_Lehmann Winnie_R*dkj*r

SLIDE 11

1

102 103

100 101 102 103 50 100 150 200 number of metabolites number of communities per metabolite

H2O, H+ ATP ADP Pi

SLIDE 12

A B

SLIDE 13

A B

SLIDE 14

SLIDE 15

Clauset, Moore, Newman. Nature 453, 98 (2008)

Hierarchy and community structure were thought to be two sides of the same story

SLIDE 16

Hierarchy and community structure were thought to be two sides of the same story

X

Clauset, Moore, Newman. Nature 453, 98 (2008)

SLIDE 17

Even in the case when nodes belong to multiple communities, their links can be well categorized.

SLIDE 18

eik ejk i j k

A B

a b c a b c

C

S(eac, ebc) = 1 3 S(eac, ebc) = 1

n+(i) ≡ {x | d(i, x) ≤ 1}

S(eik, ejk) = |n+(i) ∩ n+(j)| |n+(i) ∪ n+(j)|

SLIDE 19

LinkComm R Package by Alex T. Kalinka

SLIDE 20

Dc = mc − (nc − 1)

nc(nc−1) 2

− (nc − 1)

− −

D =

• c

mc M Dc = 2 M

• c

mc mc − (nc − 1) (nc − 1)(nc − 1)

2

SLIDE 21

SLIDE 22

BRUSH HAIR GROOM COMB HAIRSPRAY TOOTHPASTE TOOTHBRUSH PAINTER PAINTING PAINT BROOM SWEEP

• SUNSET, SUNRISE, ORANGE
• SUNSET, SUNRISE, RED
• SUNSET, SUNRISE, PRETTY,

BEAUTIFUL

• SUNSET, SUNRISE, MOON
• SUNSET, SUNRISE, BEACH
• SUNSET, SUNRISE, SUN, DAWN, DUSK,

SUNSHINE

• SUNSET, SUNRISE, DAWN, DUSK,

AFTERNOON, EVENING

SLIDE 23

LinkComm R Package by Alex T. Kalinka (Pavel Tomancak's group)

SLIDE 24

LinkComm R Package by Alex T. Kalinka (Pavel Tomancak's group)

SLIDE 25

LinkComm R Package by Alex T. Kalinka (Pavel Tomancak's group)

SLIDE 26

LinkComm R Package by Alex T. Kalinka (Pavel Tomancak's group)

SLIDE 27

LinkComm R Package by Alex T. Kalinka (Pavel Tomancak's group)

SLIDE 28

pervasive overlap metadata sampling

SLIDE 29

Building a (synthetic) benchmark graph assumes a model. A specific view of the community structure.

SLIDE 30

SLIDE 31

SLIDE 32

A synthetic graph cannot be used to compare methods with different models of community structure.

SLIDE 33

A synthetic graph cannot be used to compare methods with different models of community structure. Problem: Use metadata to test the detected structure. Solution:

SLIDE 34

Community coverage Overlap quality Overlap coverage

community memberships no membership high coverage low coverage high overlap coverage low overlap coverage high overlap low overlap

Community quality

Subjects

HIV / AIDS Medical Nonfiction / General Infectious Diseases

Subjects

Africa - General Africa History

Amazon.com

Subjects

HIV / AIDS Medical Africa

Acetyl-CoA

• 1. Glycolysis / Gluconeogenesis
• 2. TCA cycle
• 3. Fatty acid biosynthesis
• 4. ...

Many pathway Memberships

IDP (Inosine diphosphate)

• 1. Purine metaboilsm

Few pathway Memberships

Metabolic network

SLIDE 35

Community coverage Overlap quality Overlap coverage Community quality

Subjects

HIV / AIDS Medical Nonfiction / General Infectious Diseases

Subjects

Africa - General Africa History

Amazon.com

Subjects

HIV / AIDS Medical Africa

SLIDE 36

Community coverage Overlap coverage

community no membership high coverage low coverage ects

S Medical al Infectious Diseases

rk

SLIDE 37

Overlap quality Overlap coverage

high overlap low overlap

Nonfiction / General Infectious Diseases

Acetyl-CoA

• 1. Glycolysis / Gluconeogenesis
• 2. TCA cycle
• 3. Fatty acid biosynthesis
• 4. ...

Many pathway Memberships

IDP (Inosine diphosphate)

• 1. Purine metaboilsm

Few pathway Memberships

Metabolic network

SLIDE 38

Overlap coverage

community memberships high coverage low coverage high overlap coverage low overlap coverage

Infectious Diseases

lic network

SLIDE 39

metadata network description N k⇥ community

• verlap

PPI (Y2H) PPI network of S. cerevisiae

• btained by yeast two-hybrid

(Y2H) experiment [41] 1647 3.06 Set of each protein’s known functions (GO terms)a The number of GO terms PPI (AP/MS) Affinity purification mass spectrometry (AP/MS) experiment 1004 16.57 GO terms GO terms PPI (LC) Literature curated (LC) 1213 4.21 GO terms GO terms PPI (all) Union of Y2H, AP/MS, and LC PPI networks 2729 8.92 GO terms GO-terms Metabolic Metabolic network (metabolites connected by reactions) of E. coli 1042 16.81 Set of each metabolite’s pathway annotations (KEGG)b The number of KEGG pathway annotations Phone Social contacts between mobile phone users [44, 45, 46] 885989 6.34 Each user’s most likely geographic location Call activity (number of phone calls) Actor Film actors that appear in the same movies during 2000–2009 [47] 67411 8.90 Set of plot keywords for all of the actor’s films Length of career (year of first role) US Congress Congressmen who co-sponsor bills during the 108th US Congress [48, 49] 390 38.95 Political ideology, from the common space score [50, 51] Seniority (number

• f congresses

served) Philosopher Philosophers and their philosophical influences, from the English Wikipediac 1219 9.80 Set of (wikipedia) hyperlinks exiting in the philosopher’s page Number of wikipedia subject categories Word Assoc. English words that are often mentally associated [52] 5018 22.02 Set of each word’s senses, as documented by WordNetd Number of senses Amazon.com Products that users frequently buy together 18142 5.09e Set of each product’s user tags (annotations) Number of product categories

SLIDE 40

• 25. Clauset, A., Newman, M. E. J. & Moore, C.

Finding community structure in very large

• networks. Phys. Rev. E 70, 066111 (2004).
• 26. Rosvall, M. & Bergstrom, C. T. Maps of random walks on complex networks reveal commu-

nity structure. Proceedings of the National Academy of Sciences 105, 1118–1123 (2008).

• 9. Palla, G., Der´

eny, I., Farkas, I. & Vicsek, T. Uncovering the overlapping community structure

• f complex networks in nature and society. Nature 435, 814 (2005).

SLIDE 41

Other networks Social networks Biological networks 885989 6.34 k N 1042 16.81 1647 3.06 1004 16.57 1213 4.21 2729 8.92 67411 8.90 390 38.95 1219 9.80 5018 22.02 18142 5.09 1 2 3 4 L C G I L C G I L C G I L C G I L C G I L C G I L C G I L C G I L C G I L C G I L C G I

Composite performance

Amazon.com Word Assoc. Philosopher US Congress Actor Phone PPI (all) PPI (LC) PPI (AP/MS) PPI (Y2H) Metabolic

Measures

• verlap coverage
• verlap quality

community coverage community quality

Methods

L C G I – – – – Links Clique Percolation Greedy Modularity Infomap

SLIDE 42

Other networks Social networks Biological networks 885989 6.34 k N 1042 16.81 1647 3.06 1004 16.57 1213 4.21 2729 8.92 67411 8.90 390 38.95 1219 9.80 5018 22.02 18142 5.09 1 2 3 4 L C G I L C G I L C G I L C G I L C G I L C G I L C G I L C G I L C G I L C G I L C G I Amazon.com Word Assoc. Philosopher US Congress Actor Phone PPI (all) PPI (LC) PPI (AP/MS) PPI (Y2H) Metabolic

SLIDE 43

pervasive overlap metadata sampling

SLIDE 44

But Sune, we often find ‘good’ non-

• verlapping communities in networks

that should possess pervasive overlap according to your argument.

SLIDE 45

SLIDE 46

• hmm. could sampling cause networks

with pervasive overlap to appear non-

• verlapping?

SLIDE 47

a simple model for pervasive overlap

SLIDE 48

is µ ≡ P

m mrm

is ν ≡ P

n nsn.

a simple model for pervasive overlap

pervasively overlapping network characterized by two degree distributions rm and sn these determine the fraction of elements that belong to m modules and fraction of modules that contain n elements with averages projection provides module and element networks respectively

SLIDE 49

failure model

SLIDE 50

failure model

failures occurs on the element network. before projection, elements fail with probability (1 - p) and are removed from the network we say that modules fail when fewer some critical fc of the nodes in the module remain failed modules are removed from the module network, but their elements remain in the element network

SLIDE 51

quantity of interest

We wish to determine S(p), the fraction of remaining nodes within the giant component as a function of p, for both the element and module networks The giant component in the element network disappears when the network loses global connectivity. In the module network the giant connected component vanishes when the modules become uncoupled (non-

• verlapping)

SLIDE 52

with these things in place, we can sharpen

• ur question

SLIDE 53

1 - p

could we end up in a situation where the element network remains globally connected, but module network has under gone a percolation transition?

SLIDE 54

f0(z) =

∞

X

m=0

rmzm, f1(z) = 1 µ

∞

X

m=0

mrmzm−1, g0(z) =

∞

X

n=0

snzn, g1(z) = 1 ν

∞

X

n=0

nsnzn−1. probabilityofarandomlychosen elementtobelongtommodules probabilitythatarandom elementwithinarandomly chosenmodulebelongstom

• thermodules

probabilityofarandom moduletocontainn elements probthatarandommodule

• farandomlychosen

elementtocontainnother elements

SLIDE 55

Likewise, the total number of elements that a randomly chosen neighbor of A is connected to is generated by G1(z) = f1(h(z)). (6) Before determining S, we first identify the critical point pc where the giant component

• emerges. This happens when the expected number of elements two steps away from a random

element exceeds the number one step away, or @zG0(G1(z))

• z=1 @zG0(z)
• z=1 > 0.

(7) Substituting Eqs. (5) and (6) gives f 0

0(1)h0(1)[f 0 1(1)h0(1) 1] > 0 or f 0 1(1)h0(1) > 1. Finally,

the condition for a giant component to exist, since h0(1) = pg0

1(1), is

pf 0

1(1)g0 1(1) > 1.

(8) For the uniform case, rm = (m, µ) and sn = (n, ⌫), this gives p(µ 1)(⌫ 1) > 1. If µ = 3 and ⌫ = 3, then the transition occurs at pc = 1/4. To find S, consider the probability u for element A to not belong to the giant component. A is not a member of the giant component only if all of A’s neighbors are also not members, so u satisfies the self-consistency condition u = G1(u). The size of the giant component is then S = 1 G0(u).

1 Element network

Consider a randomly chosen element A that belongs to a group of size n. Let P(k|n) be the probability that A still belongs to a connected cluster of k nodes (including itself) in this group after failures occur: P(k|n) = ✓n − 1 k − 1 ◆ pk−1(1 − p)n−k. (2) The generating function for the number of other elements connected to A within this group is hn(z) =

n

X

k=1

P(k|n)zk−1 = (zp + 1 − p)n−1 . (3) Averaging over module size: h(z) = 1 ν

∞

X

n=0

nsnhn(z) = g1(zp + 1 − p). (4) The total number of elements that A is connected to, from all modules it belongs to, is then generated by G0(z) = f0(h(z)). (5)

2 Module network

Consider a random module C and then a random member element A. Let Q(`|m) be the probability that C is connected to ` modules, including itself, through element A, who was

• riginally connected to m modules including C:

Q(`|m) = ✓m 1 ` 1 ◆ q`1

1

(1 q1)m` , (9) where q1 = 1 ⌫

1

X

n=0

nsn

n

X

i=x

✓n 1 i 1 ◆ pi1(1 p)ni. (10) (Notice that q1 = 1 when x(n) ⌘ dnfce = 1 for all n.) The generating function jm for the number of modules that C is connected to, including itself, through A is jm(z) =

m

X

`=1

Q(`|m)z`1 = (zq1 + 1 q1)m1 . (11) Once again, averaging jm over memberships gives j(z) = 1 µ

1

X

m=0

mrmjm(z) = f1(zq1 + 1 q1). (12) The total number of modules that C is connected to is not generated by g0(j(z)) but by ˜ g0(j(z)), where the ˜ gi are the generating functions for module size after elements fail: ˜ g0(z) =

1

X

n=0

˜ snzn, ˜ g1(z) = P1

n=0 n˜

snzn1 P1

n=0 n˜

sn . (13) The probability ˜ sk to have k member elements remaining in a module after percolation is given by ˜ sk = P

n

n

k

• pk(1 p)nksn

P

n

Pn

k0=x

n

k0

• pk0(1 p)nk0sn

(14) The denominator is necessary for normalization since we cannot observe modules with fewer than dnfce members. Notice that ˜ sn = sn when sn = δ(n, ν) and dnfce = n = ν. Finally, the total number of modules connected to C through any member elements is generated by F0(z) = ˜ g0(j(z)) and the total number of modules connected to a random neighbor of C is generated by F1(z) = ˜ g1(j(z)). As before, the module network has a giant component when ∂zF0(F1(z))|z=1 ∂zF0(z)|z=1 > 0 and S = 1F0(u) = 1 ˜ g0(j(u)), where u satisfies u = F1(u) = ˜ g1(j(u)). For the uniform case with µ = 3, ν = 3, and fc > 2/3, the critical point for the module network is pc = 1/2, a considerably higher threshold than for the element network (pc = 1/4). In Fig. 2 we show S for µ = 3 and ν = 6. The “robustness gap” between the element and module networks widens as the module failure cutoff increases, covering a significant range

• f p for the larger values of fc.

... a couple of pages of math

SLIDE 56

Simulations 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

S p

fc = 12 fc = 23 fc = 56 fc = 1 Modules Theory

Elements

• for rm = δ(m, µ), s
• ), sn = δ(n, ν),
• ith µ = 3
• ν = 6.

SLIDE 57

scale free networks

It is known that scale-free networks are robust to random failures when 2 < λ < 3 (meaning that pc → 0). (This result requires max value

• f distribution, K,

to be large.)

SLIDE 58

scale free networks

0.2 0.4 0.6 0.8 0.05 0.1 0.15 0.2 0.25 0.3

p

λ = 2.5

0.2 0.4 0.6 0.8

S

λ = 3.0

0.2 0.4 0.6 0.8 1

λ = 3.5

N = 100 N = 500 N = 5000

Here we take rm = δ(m, μ) as before, but now sn ∼ n−λ, with λ ≥ 2 As we lower λ (increasing K), the elements become more robust (as expected), but the module network becomes less robust. For modular networks, it may not be feasible to build extremely large

• modules. Interestingly, enforcing on sn a maximum module size cutoff

N = max{n | sn > 0} only improves element robustness.

Figure 3: Robustness of scale-free networks. Here rm = δ(m, 3), sn ∼ n−λ, fc = 1/2, and N ≡ max{n | sn > 0}. Increasing N and decreasing λ, measures known to improve

SLIDE 59

0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 Brain (Functional) 0.2 0.4 0.6 0.8 Metabolic 0.2 0.4 0.6 0.8 1 Protein-Protein Interaction 0.2 0.4 0.6 0.8 1 Word Association

fc = 0.7 fc = 0.6 fc = 0.5 elements

Collaborations Web Links

p S'

real world networks

S’(p) the fraction of original nodes in the giant connected component Shaded regions provide a guide to the eye for the robustness gap (fc = 0.7).

SLIDE 60