Content uploaded by Francesco Vaccarino

Author content

All content in this area was uploaded by Francesco Vaccarino on Oct 30, 2014

Content may be subject to copyright.

rsif.royalsocietypublishing.org

Research

Cite this article: Petri G, Expert P, Turkheimer

F, Carhart-Harris R, Nutt D, Hellyer PJ,

Vaccarino F. 2014 Homological scaffolds of

brain functional networks. J. R. Soc. Interface

11: 20140873.

http://dx.doi.org/10.1098/rsif.2014.0873

Received: 5 August 2014

Accepted: 3 October 2014

Subject Areas:

computational biology

Keywords:

brain functional networks, fMRI, persistent

homology, psilocybin

Author for correspondence:

P. Expert

e-mail: paul.expert@kcl.ac.uk

Electronic supplementary material is available

at http://dx.doi.org/10.1098/rsif.2014.0873 or

via http://rsif.royalsocietypublishing.org.

Homological scaffolds of brain functional

networks

G. Petri1, P. Expert2, F. Turkheimer2, R. Carhart-Harris3, D. Nutt3, P. J. Hellyer4

and F. Vaccarino1,5

1

ISI Foundation, Via Alassio 11/c, 10126 Torino, Italy

2

Centre for Neuroimaging Sciences, Institute of Psychiatry, Kings College London, De Crespigny Park,

London SE5 8AF, UK

3

Centre for Neuropsychopharmacology, Imperial College London, London W12 0NN, UK

4

Computational, Cognitive and Clinical Neuroimaging Laboratory, Division of Brain Sciences, Imperial College

London, London W12 0NN, UK

5

Dipartimento di Scienze Matematiche, Politecnico di Torino, C.so Duca degli Abruzzi no 24, Torino 10129, Italy

Networks, as efficient representations of complex systems, have appealed to

scientists for a long time and now permeate many areas of science, including

neuroimaging (Bullmore and Sporns 2009 Nat. Rev. Neurosci. 10, 186– 198.

(doi:10.1038/nrn2618)). Traditionally, the structure of complex networks has

been studied through their statistical properties and metrics concerned with

node and link properties, e.g. degree-distribution, node centralityand modular-

ity. Here, we study the characteristics of functional brain networks at the

mesoscopic level from a novel perspective that highlights the role of inhomo-

geneities in the fabric of functional connections. This can be done by focusing

on the features of a set of topological objects—homological cycles—associated

with the weighted functional network. We leverage the detected topological

information to define the homological scaffolds, a new set of objects designed to

represent compactly the homological features of the correlation network and

simultaneously maketheir homological properties amenable to networkstheor-

etical methods. As a proof of principle, we apply these tools to compare resting-

state functional brain activity in 15 healthy volunteers after intravenous infusion

of placebo and psilocybin—the main psychoactive component of magic mush-

rooms. The results show that the homological structure of the brain’s functional

patterns undergoes a dramatic change post-psilocybin, characterized by the

appearance of many transient structures of low stability and of a small

number of persistent ones that are not observed in the case of placebo.

1. Motivation

The understanding of global brain organization and its large-scale integration

remains a challenge for modern neurosciences. Network theory is an elegant frame-

work to approach these questions, thanks to its simplicity and versatility [1]. Indeed,

in recent years, networks have become a prominent tool to analyse and understand

neuroimaging datacoming from very diverse sources, such as functional magnetic

resonance imaging (fMRI), electroencephalography and magnetoencephalography

[2,3], also showing potential for clinical applications [4,5].

A naturalway of approaching these datasets is to devise a measure of dynami-

cal similarity between the microscopic constituents and interpret it as the strength

of the link between those elements. In the case of brain functional activity, this often

implies the use of similarity measures such as (partial) correlations or coherence

[6– 8], which generally yield fully connected, weighted and possibly signed adja-

cency matrices. Despite the fact that most network metrics can be extended to

the weighted case [9– 13], the combined effect of complete connectedness and

edge weights makes the interpretation of functional networks significantly

harder and motivates the widespread use of ad hoc thresholding methods

[7,14– 18]. However, neglecting weak links incurs the dangers of a trade-off

&2014 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution

License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original

author and source are credited.

between information completeness and clarity. In fact, it risks

overlooking the role that weak links might have, as shown for

example in the cases of resting-state dynamics [19,20], cognitive

control [21] and correlated network states [22].

In order to overcome these limits, Rubinov & Sporn

[13,23,24] recently introduced a set of generalized network and

community metrics for functional networks that among others

were used to uncover the contrasting dynamics underlying

recollection [25] and the physiology of functional hubs [26].

In this paper, we present an alternative route to the analy-

sis of brain functional networks. We focus on the combined

structure of connections and weights as captured by the hom-

ology of the network. A summary of all the keywords and

concepts introduced in this paper can be found in table 1.

2. From networks to topological spaces and

homology

Homology is a topological invariant that characterizes a topo-

logical space Xby counting its holes and their dimensions. By

hole, we mean a hollow region bounded by the parts of that

space. The dimension of a hole is directly related to the dimen-

sion of its boundary. The boundary of a two-dimensional

hole is a one-dimensional loop; the three-dimensional inner

part of a doughnut, where the filling goes, is bounded by

two-dimensional surface; for dimensions higher than 2, it

becomes difficult to have a mental representation of a hole,

but k-dimensional holes are still bounded by (k21) dimen-

sional faces. In our work, we start with a network and from

it construct a topological space. We now use figure 1 to

show how we proceed and make rigorous what we mean by

boundaries and holes.

In a network like that of figure 1a, we want the ring of nodes

(a,b,c,d) to bea good candidate fora one-dimensional boundary,

whereas the other rings of three nodes should not constitute

interesting holes. The reason for this choice comes from the for-

malization of the notion of hole. One way to formalize this is by

opposition that is we define what we mean by a dense subnet-

work in order to highlight regions of reduced connectivity, i.e.

holes. The most natural and conservative definition we can

adopt for a dense subnetwork is that of clique, a completely

jja

d

k

c

b

i

e

f

g

h

l

a

c

i

e

f

h

l

g

d

k

b

(a)(b)

Figure 1. Panels (a,b) display an unweighted network and its clique complex, obtained by promoting cliques to simplices. Simplices can be intuitively thought as

higher-dimensional interactions between vertices, e.g. as a simplex the clique (b,c,i) corresponds to a filled triangle and not just its sides. The same principle applies

to cliques—thus simplices—of higher order. (Online version in colour.)

Table 1. List of notations.

name symbol deﬁnition

graph Ga graph G¼(V,E) is a representation of a set Vof nodes iinterconnected by edges or links eij [E; this

interaction can be weighted, directional and signed

clique ca completely connected subgraph C¼(V’,E’) contained in an undirected and unweighted graph

G¼(V,E)(V0,V,eij [E08i,j[E0)

k-simplex

s

k

formally, a convex hull of kþ1 nodes [ p

0

,p

1

,...,p

k

], it is used here as a generalization to higher

dimensions of the concept of link, e.g. a 2-simplex is a triangle, a 3-simplex a tetrahedron. The faces f

of

s

k

are obtained as subsets of [ p

0

,p

1

,...,pk]

simplicial complex Ka topological space composed by attaching simplices

s

, with two conditions: (i) if

s

[Kthen all its

faces f[Kand (ii) the intersection of any two

s

i,

s

j[Kis empty or a face of both

s

i

and

s

j

.

clique complex l(G) a simplicial complex built from a graph Gby promoting every k-clique c,Gto a (k21)-simplex

deﬁned by the nodes of c, e.g. a 3-clique becomes a 2-simplex (a full triangle)

kth homology group Hk(K) a group describing the holes of a simplicial complex Kbounded by k-dimensional boundaries, e.g. H

1

describes two-dimensional cycles bounded by 1-simplices, H

2

describes three-dimensional voids bounded

by 2-simplices, etc.

H

k

generator gan element of the generating set of H

k

, a subset of H

k

such that all elements can be expressed as

combination of generators

homological scaffold H(K) a weighted graph constructed from the persistent homology generators of H

1

of a simplicial complex K

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

2

connected subgraph [27]. Moreover, cliques have the crucial

property, which will be important later, of being nested, i.e. a

clique of dimension k(k-clique) contains all the m-cliques

defined by its nodes with m,k. Using this definition and filling

in all the maximal cliques, the network in figure 1acan be rep-

resented as in figure 1b: 3-cliques are filled in, becoming tiles,

and the only interesting structure left is the square (a,b,c,d). It

is important at this point to note that a k-clique can be seen as

ak21 simplex, i.e. as the convex hull of k-points. Our represen-

tation of a network can thus be seen as a topological space

formed by a finite set of simplices that by construction satisfy

the condition that defines the type of topological spaces called

abstract simplicial complexes [28]: each element of the space is

a simplex, and each of its faces (or subset in the case of

cliques) is also a simplex.

This condition is satisfied, because each clique is a sim-

plex, and subsets of cliques are cliques themselves, and the

intersection of two cliques is still a clique.

The situation with weighted networks becomes more

complicated. In the context of a weighted network, the

holes can be thought of as representing regions of reduced

connectivity with respect to the surrounding structure.

Consider, for example, the case depicted in figure 2a:the

network is almost the same as figure 1 with the two excep-

tions that it now has weighted edges and has an additional

very weak edge between nodes aand c. The edges in the

cycle [a,b,c,d] are all much stronger than the link (a,c)that

closes the hole by making (a,b,d) and (b,c,d) cliques and there-

fore fills them. The loop (e,f,g,h,i) has a similar situation, but

the difference in edge weights between the links along the

cycle and those crossing, is not as large as in the previous

case. It would be therefore useful to be able to generalize

the approach exposed earlier for binary networks to the case

of weighted networks in such a way as to be able to measure

the difference between the two cases (a,b,c,d)and(e,f,g,h,i). As

shown by figure 2b, this problem can be intuitively thought of

as a stratigraphy in the link-weight fabric of the network,

where the aim is to detect the holes, measure their depth

and when they appear as we scan across the weights’ range.

From figure 2b, it becomes clear that the added value of

this method over conventional network techniques lies in its

capability to describe mesoscopic patterns that coexist over

different intensity scales, and hence to complement the infor-

mation about the community structure of brain functional

networks. A way to quantify the relevance of holes is given

by persistent homology. We describe it and its application to the

case of weighted networks in full detail in §3.

3. A persistent homology of weighted networks

The method that we adopt was introduced in references

[29,30] and relies on an extension of the metrical persistent

homology theory originally introduced by references [31,32].

Technical details about the theory of persistent homology

and how the computation is performed can be found

in the works of Carlsson, Zomorodian and Edelsbrunner

[28,31– 35]. Persistent homology is a recent technique in com-

putational topology developed for shape recognition and

the analysis of high dimensional datasets [36,37]. It has been

used in very diverse fields, ranging from biology [38,39]

and sensor network coverage [40] to cosmology [41]. Similar

approaches to brain data [42,43], collaboration data [44] and

network structure [45] also exist. The central idea is the con-

struction of a sequence of successive approximations of the

original dataset seen as a topological space X. This sequence

of topological spaces X

0

,X

1

,...,X

N

¼Xis such that

Xi#Xjwhenever i,jand is called the filtration. Choosing

how to construct a filtration from the data is equivalent to

choosing the type of goggles one wears to analyse the data.

In our case, we sort the edge weights in descending

order and use the ranks as indices for the subspaces. More

specifically, denote by

V

¼(V,E,

v

) the functional network

with vertices V, edges Eand weights

v

:E!R. We then con-

sider the family of binary graphs G

v

¼(V,E

v

), where an edge

e[Eis also included in G

v

if its weight

v

e

is larger than

v

(e[E

v

,

v

(e)

v

).

To each of the G

v

, we associate its clique,orflag complex K

v

,

that is the simplicial complex that contains the k-simplex [n

0

,

n

1

,n

2

,...n

k21

] whenever the nodes n

0

,n

1

,n

2

,...n

k21

define

acliqueinG

v

[27]. As subsets of cliques and intersections of

cliques are cliques themselves, as we pointed out in §2, our

clique complex is thus a particular case of a simplicial

complex.

The family of complexes fK

v

gdefines a filtration, because

we have K

v

#K0

v

for

v

.

v

0. At each step, the simplices in K

v

inherit their configuration from the underlying network

structure and, because the filtration swipes across all weight

scales in descending order, the holes among these units con-

stitute mesoscopic regions of reduced functional connectivity.

Moreover, this approach also highlights how network

properties evolve along the filtration, providing insights

about where and when lower connectivity regions emerge.

This information is available, because it is possible to keep

track of each k-dimensional cycle in the homology group H

k

.

A generator uniquely identifies a hole by its constituting

elements at each step of the filtration process. The importance

of a hole is encoded in the form of ‘time-stamps’ recording its

birth

b

g

and death

d

g

along the filtration fK

v

g[31]. These two

time-stamps can be combined to define the persistence

p

g

¼

d

g

2

b

g

of a hole, which gives a notion of its importance in

terms of it lifespan. Continuing the analogy with stratigraphy,

b

g

and

d

g

correspond, respectively, to the top and the bottom

of a hole and

p

g

would be its depth. As we said above, a genera-

tor gk

i, or hole, of the kth homology group H

k

is identified by its

birth and death along the filtration. Therefore, gk

iis described by

the point (

b

g,

d

g)[R2. A standardway to summarize the infor-

mation about the whole kth persistent homology group is then

to consider the diagram obtained plotting the points corre-

sponding to the set of generators. The (multi)set f(

b

g

,

d

g

gis

called the persistence diagram of H

k

. In figure 2c, we show the

persistence diagram for the network shown in figure 2afor

H

1

. Axes are labelled by weights in decreasing order. It is easy

to check that the coordinates correspond exactly to the appear-

ance and disappearance of generators. The green vertical bars

highlight the persistence of a generator along the filtration.

The further a point is from the diagonal (vertically), the more

persistent the generator is. In §4, we introduce two objects, the

persistence and the frequency homological scaffolds, designed

to summarize the topological information about the system.

4. Homological scaffolds

Once one has calculated the generators {gk

i}iof the kth persist-

ent homology group H

k

, the corresponding persistence

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

3

diagram contains a wealth of information that can be used,

for example, to highlight differences between two datasets.

It would be instructive to obtain a synthetic description of

the uncovered topological features in order to interpret the

observed differences in terms of the microscopic components,

at least for low dimensions k. Here, we present a scheme to

obtain such a description by using the information associated

with the generators during the filtration process. As

each generator, gk

iis associated with a whole equivalence

class, rather than to a single chain of simplices, we need to

j10

appearance of

(a,b,c,d) hole

10

(a)

(b)

(c)

8

6

4

weights

2

0

0

1

2

3

4

5

death weight

6

7

8

9

98765

birth wei

g

ht

43210

closure of

(a,b,c,d) hole

appearance of

(e,f,g,h,i) hole

closure of

(e,f,g,h,i) hole

8

8

8

1

10

10

10

6

5

6

6

4

5

4

8

10

a

d

k

c

b

i

e

f

g

h

l

(a,b,c,d)

(e,f,g,h,i)

Figure 2. Panels (a–c) display a weighted network (a), its intuitive representation in terms of a stratigraphy in the weight structure according the weight filtration

described in the main text (b) and the persistence diagram for H

1

associated with the network shown (c). By promoting cliques to simplices, we identify network

connectivity with relations between the vertices defining the simplicial complex. By producing a sequence of networks through the filtration, we can study the

emergence and relative significance of specific features along the filtration. In this example, the hole defined by (a,b,c,d) has a longer persistence (vertical solid

green bars) implying that the boundary of the cycle are much heavier than the internal links that eventually close it. The other hole instead has a much shorter

persistence, surviving only for one step and is therefore considered less important in the description of the network homological properties. Note that the births and

deaths are defined along the sequence of descending edge weights in the network, not in time. (Online version in colour.)

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

4

choose a representative for each class, we use the representa-

tive that is returned by the javaplex implementation [46] of the

persistent homology algorithm [47]. For the sake of simplicity

in the following, we use the same symbol gk

ito refer to a

generator and its representative cycle.

We exploit thisto definetwo new objects, the persistence and

the frequency homological scaffolds Hp

Gand Hf

Gof a graph G.The

persistence homological scaffold is the network composed of all

the cycle paths corresponding to generators weighted by their

persistence. If an edge ebelongs to multiple cycles g

0

,g

1

,...,g

s

,

its weight is defined as the sum of the generators’ persistence:

vp

e¼X

gije[gi

p

gi:(4:1)

Similarly, we define the frequency homological scaffold Hf

Gas

the network composed of all the cycle paths corresponding

to generators, where this time, an edge eis weighted by the

number of different cycles it belongs to

v

f

e¼X

gi

1e[gi,(4:2)

where 1e[giis the indicator function for the set of edges com-

posing g

i

. By definition, the two scaffolds have the same edge

set, although differently weighted.

The construction of these two scaffolds therefore high-

lights the role of links which are part of many and/or long

persistence cycles, isolating the different roles of edges

within the functional connectivity network. The persistence

scaffolds encodes the overall persistence of a link through

the filtration process: the weight in the persistence scaffold

of a link belonging to a certain set of generators is equal to

the sum of the persistence of those cycles. The frequency scaf-

fold instead highlights the number of cycles to which a link

belongs, thus giving another measure of the importance of

that edge during the filtration. The combined information

given by the two scaffolds then enables us to decipher the

nature of the role different links have regarding the homolo-

gical properties of the system. A large total persistence for a

link in the persistence scaffold implies that the local structure

around that link is very weak when compared with the

weight of the link, highlighting the link as a locally strong

bridge. We remark that the definition of scaffolds we gave

depends on the choice of a specific basis of the homology

group, and the choice of a consistent basis is an open problem

in itself, therefore the scaffolds are not topological invariants.

Moreover, it is possible for an edge to be added to a cycle

shortly after the cycle’s birth in such a way that it creates a

triangle with the two edges composing the cycle. In this

way, the new edge would be part of the shortest cycle, but

the scaffold persistence value would be misattributed to the

two other edges. This can be checked, for example, by moni-

toring the clustering coefficient of the cycle’s subgraph as

edges are added to it. We checked for this effect and found

that in over 80% of the cases the edges do not create triangles

that would imply the error, but instead new cycles are cre-

ated, whose contribution to the scaffold is then accounted

for by the new cycle. Finally, we note also that, when a

new triangle inside the cycle is created, the two choices of

generator differ for a path through a third strongly connected

node, owing to the properties of boundary operators. Despite

this ambiguity, we show in the following that they can be

useful to gain an understanding of what the topological

differences detected by the persistent homology actually

mean in terms of the system under study.

5. Results from fMRI networks

We start from the processed fMRI time series (see Methods for

details). The linear correlations between regional time series

were calculated after covarying out the variance owing to all

other regions and the residual motion variance represented

by the 24 rigid motion parameters obtained from the pre-

processing, yielding a partial-correlation matrix

x

a

for each

subject. The matrices

x

a

were then analysed with the algorithm

described in the previous sections. We calculated the generators

g1

iof the first homological group H

1

along the filtration. As

mentioned before, each of these generators identifies a lack of

mesoscopic connectivity in the form of a one-dimensional

cycle andcan be represented in a persistence diagram. Weaggre-

gate together the persistence diagrams of subjects belonging to

each group and compute an associated persistence probability

density (figure 3). These probability density functions constitute

the statistical signature of the groups’ H

1

features.

We find that, although the number of cycles in the groups

are comparable, the two probability densities strongly differ

(Kolmogorov– Smirnov statistics: 0.22, p-value less than 10

210

).

The placebo group displays generators appearing and per-

sisting over a limited interval of the filtration. On the contrary,

most of the generators for the psilocybin group are situated in

a well-defined peak at small birth indices, indicating a shorter

average cycle persistence. However, the psilocybin distribution

is also endowed with a longer tail implying the existence of

a few cycles that are longer-lived compared with the placebo

condition and that influences the weight distribution of the psi-

locybin persistence scaffold. The difference in behaviour of the

two groups is made explicit when looking at the probability dis-

tribution functions for the persistence and the birth of

generators (figure 4), which are both found to be significantly

different (Kolmogorov–Smirnov statistics: 0.13, p-value ,

10

230

for persistence and Kolmogorov–Smirnov statistics:

0.14, p-value ,10

235

for births). In order to better interpret

and understand the differences between the two groups,

we use the two secondary networks described in §4, Hf

Pla and

Hp

Pla for the placebo group and Hf

Psi and Hp

Psi for the psilocybin

group. The weight of the edges in these secondary networks is

proportional to the total number of cycles an edge is part of, and

the total persistence of those cycles, respectively. They comp-

lement the information given by the persistence density

distribution, where the focus is on the entire cycle’s behaviour,

with information on single links. In fact, individual edges

belonging to many and long persistence cycles represent func-

tionally stable ‘hub’ links. As with the persistence density

distribution, the scaffolds are obtained at a group level by

aggregating the information about all subjects in each group.

These networks are slightly sparser than the original complete

x

a

networks

r

(Hf,p

pla )¼2m(Hp

pla)

n(n1) ¼0:92 (5:1)

and

r

(Hf,p

Psi )¼2m(Hp

Psi)

n(n1) ¼0:91 (5:2)

and have comparable densities. A first difference between the

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

5

two groups becomes evident when we look at the distributions

for the edge weights (figure 5a). In particular, the weights of

Hp

pla display a cut-off for large weights, whereas the weights

of Hp

psi have a broader tail (Kolmogorov– Smirnov statistics:

0.06, p-value ,10

220

;figure5a). Interestingly, the frequency

scaffold weights probability density functions cannot be distin-

guished from each other figure 5a(inset) (Kolmogorov–

Smirnov statistics: 0.008, p-value ¼0.72). Taken together,

these two results imply that while edges statistically belong to

the same number of cycles, in the psilocybin scaffold, there

exist very strong, persistent links.

The difference between the two sets of homological scaf-

folds for the two groups becomes even more evident when

one compares the weights between the frequency and

persistence scaffolds of the same group. Figure 5bis a scatter

plot of between the weights of edges from both scaffolds for

the two groups. The placebo group has a linear relationship

between the two quantities meaning that edges that are per-

sistent also belongs to many cycles (R

2

¼0.95, slope ¼0.23).

Although the linear relationship is still a reasonable fit for

the psilocybin group (R

2

¼0.9, slope ¼0.3), the data in this

case display a larger dispersion. In particular, it shows that

edges in Hf,p

psi can be much more persistent/longer-lived

than in Hf,p

pla but still appear in the same number of cycles,

i.e. the frequency of a link is not predictive of its persistence

or simply put: some connections are much more persistent in

the psychedelic state. Moreover, the slopes of linear fits of the

two clouds are statistically different ( p-value ,10

220

,n

pla

¼

1.0

0.9

0.8

0.7

0.6

0.5

death

death

0.4

0.3

0.2

birth birth

(a)(b)

0.1

0

–10.0 –9.5 –9.0 –8.5 –8.0 –7.5 –7.0 –6.5 –10.0 –9.6 –9.2 –8.8 –8.4 –7.6–8.0 –7.2 –6.8

1.0

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

1.0

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0

Figure 3. Probability densities for the H

1

generators. Panel (a) reports the (log-)probability density for the placebo group, whereas panel (b) refers to the psilocybin

group. The placebo displays a uniform broad distribution of values for the births– deaths of H

1

generators, whereas the plot for the psilocybin condition is very

peaked at small values with a fatter tail. These heterogeneities are evident also in the persistence distribution and find explanation in the different functional

integration schemes in placebo and drugged brains. (Online version in colour.)

0

0.01

0.02

P(p)

0.03

0.04

0.05

10–5

0.0001

P(b)

0.001

0.01

placebo

psilo

0.1

(a)(b)

0.2 0.4 0.6

persistence p

0.8 1.0 0 0.2 0.4 0.6

birth

b

0.8 1.0

Figure 4. Comparison of persistence

p

and birth

b

distributions. Panel (a) reports the H

1

generators’ persistence distributions for the placebo group (blue line) and

psilocybin group (red line). Panel (b) reports the distributions of births with the same colour scheme. It is very easy to see that the generators in the psilocybin

condition have persistence peaked at shorter values and a wider range of birth times when compared with the placebo condition. (Online version in colour.)

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

6

13 200 and n

psi

¼13 275 [48]) pointing to a starkly different

local functional structure in the two conditions.

The results from the persistent homology analysis and the

insights provided by the homological scaffolds imply that

although the mesoscopic structures, i.e. cycles, in the psilocy-

bin condition are less stable than in the placebo group, their

constituent edges are more stable.

6. Discussion

In this paper, we first described a variation of persistent hom-

ology that allows us to deal with weighted and signed

networks. We then introduced two new objects, the homolo-

gical scaffolds, to go beyond the picture given by persistent

homology to represent and summarize information about

individual links. The homological scaffolds represent a new

measure of topological importance of edges in the original

system in terms of how frequently they are part of the genera-

tors of the persistent homology groups and how persistent

are the generators to which they belong to. We applied this

method to an fMRI dataset comprising a group of subjects

injected with a placebo and another injected with psilocybin.

By focusing on the second homology group H

1

,wefound

that the stability of mesoscopic association cycles is reduced by

the action of psilocybin, as shown by the difference in the

probability density function of the generators of H

1

(figure 3).

It is here that the importance of the insight given by the

homological scaffolds in the persistent homology procedure

becomes apparent. A simple reading of this result would be

that the effect of psilocybin is to relax the constraints on brain

function, ascribing cognition a more flexible quality, but when

looking at the edge level, the picture becomes more complex.

The analysis of the homological scaffolds reveals the existence

of a set of edges that are predominant in terms of their persist-

ence although they are statistically part of the same number of

cycles in the two conditions (figure 5). In other words, these

functional connections support cycles that are especially stable

and are only present in the psychedelic state. This further

implies that the brain does not simply become a random

system afterpsilocybin injection, but instead retains some organ-

izational features, albeit different from the normal state, as

suggested by the first part of the analysis. Further work is

required to identify the exact functional significance of these

edges. Nonetheless, it is interesting to look at the community

structure of the persistence homological scaffolds in figure 6.

The two pictures are simplified cartoons of the placebo (figure

6a) and psilocybin (figure 6b) scaffolds. In figure 6a,b,the

nodes are organized and coloured accordingto their community

membership in the placebo scaffold (obtained with the Louvain

algorithm for maximal modularity and resolution 1 [50]). This is

done in order to highlight the striking difference in connectivity

structure in the two cases. When considering the edges in the tail

of the distribution, weight greater than or equal to 80, in figure

5a, only 29 of the 374 edges present in the truncated psilocybin

scaffold are shared with the truncated placebo scaffold (165

edges). Of these 374 edges, 217 are between placebo commu-

nities and are observed to mostly connect cortical regions. This

supports our idea that psilocybin disrupts the normal organiz-

ation of the brain with the emergence of strong, topologically

long-range functional connections that are not present in a

normal state.

The two key results of the analysis of the homological scaf-

folds can therefore be summarized as follows (i) there is an

increased integration between cortical regions in the psilocybin

state and (ii) this integration is supported by a persistent scaffold

of a set of edges that support cross modular connectivity prob-

ably as a result of the stimulation of the 5HT2A receptors in

the cortex [51].

We can speculate on the implications of such an organiz-

ation. One possible by-product of this greater communication

across the whole brain is the phenomenon of synaesthesia

which is often reported in conjunction with the psychedelic

state. Synaesthesia is described as an inducer-concurrent

pairing, where the inducer could be a grapheme or a visual

stimulus that generates a secondary sensory output—like a

colour for example. Drug-induced synaesthesia often leads

to chain of associations, pointing to dynamic causes rather

1

10

–1

10

–2

P(w

f

)

10

–3

10

–4

10 10

2

link frequency w

f

10.0 100.0

1.0

0.1

0.01

0.001

0.0001 1000.0

10

3

0

50

100

150

200

250

300

350

(a)(b)

200 400 600

psi

pla

800 1000 1200

pla

p

psi

p

e

e

link persistence w

p

e

link frequency w f

e

P(wp)

e

link persistence wp

e

Figure 5. Statistical features of group homological scaffolds. Panel (a) reports the (log-binned) probability distributions for the edge weights in the persistence

homological scaffolds (main plot) and the frequency homological scaffolds (inset). While the weights in the frequency scaffold are not significantly different, the

weight distributions for the persistence scaffold display clearly a broader tail. Panel (b) shows instead the scatter plot of the edge frequency versus total persistence.

In both cases, there is a clear linear relationship between the two, with a large slope in the psilocybin case. Moreover, the psilocybin scaffold has a larger spread in

the frequency and total persistence of individual edges, hinting to a different local functional structure within the functional network of the drugged brains. (Online

version in colour.)

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

7

than fixed structural ones as may be the case for acquired

synaesthesia [52]. Broadly consistent with this, it has been

reported that subjects under the influence of psilocybin

have objectively worse colour perception performance

despite subjectively intensified colour experience [53].

To summarize, we presented a new method to analyse

fully connected, weighted and signed networks and applied

it to a unique fMRI dataset of subjects under the influence

of mushrooms. We find that the psychedelic state is associ-

ated with a less constrained and more intercommunicative

mode of brain function, which is consistent with descriptions

of the nature of consciousness in the psychedelic state.

7. Methods

7.1. Dataset

A pharmacological MRI dataset of 15 healthy controls was used

for a proof-of-principle test of the methodology [54]. Each subject

was scanned on two separate occasions, 14 days apart. Each scan

consisted of a structural MRI image (T1-weighted), followed by a

12 min eyes-close resting-state blood oxygen-level-dependent

(BOLD) fMRI scan which lasted for 12 min. Placebo (10 ml

saline, intravenous injection) was given on one occasion and psi-

locybin (2 mg dissolved in 10 ml saline) on the other. Injections

were given manually by a study doctor situated within the scan-

ning suite. Injections began exactly 6 min after the start of the

12-min scans, and continued for 60 s.

7.1.1. Scanning parameters

The BOLD fMRI data were acquired using standard gradient-echo

EPI sequences, reported in detail in reference [54]. The volume

repetition time was 3000 ms, resulting in a total of 240 volumes

acquired during each 12 min resting-state scan (120 pre- and 120

post-injection of placebo/psilocybin).

7.1.2. Image pre-processing

fMRI images were corrected for subject motion within individual

resting-state acquisitions, by registering all volumes of the

functional data to the middle volume of the acquisition using

the FMRIB linear registration motion correction tool, generating

a six-dimension parameter time course [55]. Recent work demon-

strates that the six parameter motion model is insufficient to

correct for motion-induced artefact within functional data,

instead a Volterra expansion of these parameters to form a 24

parameter model is favoured as a trade-off between artefact cor-

rection and lost degrees of freedom as a result of regressing

motion away from functional time courses [56]. fMRI data

were pre-processed according to standard protocols using a

high-pass filter with a cut-off of 300 s.

Structural MRI images were segmented into n¼194 cortical

and subcortical regions, including white matter cerebrospinal

fluid (CSF) compartments, using FREESURFER (http://surfer.nmr.

mgh.harvard.edu/), according to the Destrieux anatomical atlas

[57]. In order to extract mean-functional time courses from

the BOLD fMRI, segmented T1 images were registered to the

middle volume of the motion-corrected fMRI data, using bound-

ary-based registration [58], once in functional space mean

time-courses were extracted for each of the n¼194 regions in

native fMRI space.

7.1.3. Functional connectivity

For each of the 194 regions, alongside the 24 parameter motion

model time courses, partial correlations were calculated between

all couples of time courses (i,j), non-neural time courses (CSF,

white matter and motion) were discarded from the resulting

functional connectivity matrices, resulting in a 169 region corti-

cal/subcortical functional connectivity corrected for motion

and additional non-neural signals (white matter/CSF).

7.2. Persistent homology computation

For each subject in the two groups, we have a set of persistence

diagrams relative to the persistent homology groups H

n

. In this

paper, we use the H

1

persistence diagrams of each group to

construct the corresponding persistence probability densities

for H

1

cycles.

Filtrations were obtained from the raw partial-correlation

matrices through the PYTHON package Holes and fed to javaplex

[46] via a Jython subroutine in order to extract the persistence

(a)(b)

Figure 6. Simplified visualization of the persistence homological scaffolds. The persistence homological scaffolds Hp

pla (a) and Hp

psi (b) are shown for comparison.

For ease of visualization, only the links heavier than 80 (the weight at which the distributions in figure 5abifurcate) are shown. This value is slightly smaller than

the bifurcation point of the weights distributions in figure 5a. In both networks, colours represent communities obtained by modularity [49] optimization on the

placebo persistence scaffold using the Louvain method [50] and are used to show the departure of the psilocybin connectivity structure from the placebo baseline.

The width of the links is proportional to their weight and the size of the nodes is proportional to their strength. Note that the proportion of heavy links between

communities is much higher (and very different) in the psilocybin group, suggesting greater integration. A labelled version of the two scaffolds is available as GEXF

graph files as the electronic supplementary material. (Online version in colour.)

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

8

intervals and the representative cycles. The details of the

implementation can be found in reference [30], and the software

is available at Holes [59].

Funding statement. G.P. and F.V. are supported by the TOPDRIM project

supported by the Future and Emerging Technologies programme of

the European Commission under Contract IST-318121. I.D. P.E. and

F.T. are supported by a PET methodology programme grant from

the Medical Research Council UK (ref no. G1100809/1). The authors

acknowledge support of Amanda Feilding and the Beckley

Foundation and the anonymous referees for their critical and

constructive contribution to this paper.

References

1. Baronchelli A, Ferrer-i-Cancho R, Pastor-Satorras R,

Chater N, Christiansen MH. 2013 Networks in

cognitive science. Trends Cogn. Sci. 17, 348– 360.

(doi:10.1016/j.tics.2013.04.010)

2. Bullmore E, Sporns O. 2009 Complex brain

networks: graph theoretical analysis of structural

and functional systems. Nat. Rev. Neurosci. 10,

186–198. (doi:10.1038/nrn2618)

3. Meunier D, Lambiotte R, Bullmore E. 2010 Modular

and hierarchically modular organization of brain

networks. Front. Neurosci. 4, 200. (doi:10.3389/

fnins.2010.00200)

4. Pandit AS, Expert P, Lambiotte R, Bonnelle V, Leech

R, Turkheimer FE, Sharp DJ. 2013 Traumatic brain

injury impairs small-world topology. Neurology 80,

1826–1833. (doi:10.1212/WNL.0b013e3182929f38)

5. Lord LD et al. 2012 Functional brain networks

before the onset of psychosis: a prospective fMRI

study with graph theoretical analysis. Neuroimage:

Clin. 1, 91–98. (doi:10.1016/j.nicl.2012.09.008)

6. Lord LD, Allen P, Expert P, Howes O, Lambiotte R,

McGuire P, Bose SK, Hyde S, Turkheimer FE. 2011

Characterization of the anterior cingulate’s role in

the at-risk mental state using graph theory.

Neuroimage 56, 1531– 1539. (doi:10.1016/j.

neuroimage.2011.02.012)

7. Zalesky A, Fornito A, Bullmore E. 2012 On the use

of correlation as a measure of network connectivity.

Neuroimage 60, 2096– 2106. (doi:10.1016/j.

neuroimage.2012.02.001)

8. Achard S, Salvador R, Whitcher B, Suckling J,

Bullmore E. 2006 A resilient, low-frequency, small-

world human brain functional network with highly

connected association cortical hubs. J. Neurosci. 21,

63–72. (doi:10.1523/JNEUROSCI.3874-05.2006)

9. Newman MEJ. 2001 Scientific collaboration

networks. II. Shortest paths, weighted networks,

and centrality. Phys. Rev. E 64, 016132. (doi:10.

1103/PhysRevE.64.016132)

10. Opsahl T, Agneessens F, Skvoretz J. 2010 Node

centrality in weighted networks: generalizing

degree and shortest paths. Soc. Netw. 32, 245–

251. (doi:10.1016/j.socnet.2010.03.006)

11. Opsahl T, Colizza V, Panzarasa P, Ramasco JJ. 2008

Prominence and control: the weighted rich-club

effect. Phys. Rev. Lett. 101, 168702. (doi:10.1103/

PhysRevLett.101.168702)

12. Colizza V, Flammini A, Serrano MA, Vespignani A.

2006 Detecting rich-club ordering in complex

networks. Nat. Phys. 2, 110– 115. (doi:10.1038/

nphys209)

13. Rubinov M, Sporns O. 2010 Complex network

measures of brain connectivity: uses and

interpretations. Neuroimage 52, 1059– 1069.

(doi:10.1016/j.neuroimage.2009.10.003)

14. Zuo XN, Ehmke R, Mennes M, Imperati D. 2012

Network centrality in the human functional

connectome. Cereb. Cortex 22, 1862– 1875.

(doi:10.1093/cercor/bhr269)

15. Qian S, Sun G, Jiang Q, Liu K, Li B, Li M, Yang X,

Yang Z, Zhao L. 2013 Altered topological patterns of

large-scale brain functional networks during passive

hyperthermia. Brain Cogn. 83, 121– 131. (doi:10.

1016/j.bandc.2013.07.013)

16. Sepulcre J, Liu H, Talukdar T, Martincorena I. 2010

The organization of local and distant functional

connectivity in the human brain. PLoS Comput. Biol.

6, e1000808. (doi:10.1371/journal.pcbi.1000808)

17. Ginestet CE, Nichols TE, Bullmore ET, Simmons A.

2011 Brain network analysis: separating cost from

topology using cost-integration. PLoS ONE 6,

e21570. (doi:10.1371/journal.pone.0021570)

18. Power JD, Cohen AL, Nelson SM, Wig GS, Barnes KA.

2011 Functional network organization of the human

brain. Neuron 72, 665– 678. (doi:10.1016/j.neuron.

2011.09.006)

19. Schwarz AJ, McGonigle J. 2011 Negative edges and

soft thresholding in complex network analysis of

resting state functional connectivity data.

J. Neuroimage 55, 1132–1146. (doi:10.1016/j.

neuroimage.2010.12.047)

20. Bassett DS, Nelson BG, Mueller BA, Camchong J,

Lim KO. 2012 Altered resting state complexity in

schizophrenia. Neuroimage 59, 2196– 2207.

(doi:10.1016/j.neuroimage.2011.10.002)

21. Cole MW, Yarkoni T, Repovs G, Anticevic A, Braver TS.

2012 Global connectivity of prefrontal cortex

predicts cognitive control and intelligence. J. Neurosci.

32, 8988– 8999. (doi:10.1523/JNEUROSCI.0536-

12.2012)

22. Schneidman E, Berry MJ, Segev R, Bialek W. 2006

Weak pairwise correlations imply strongly correlated

network states in a neural population. Nature 440,

1007–1012. (doi:10.1038/nature04701)

23. Rubinov M, Sporns O. 2011 Weight-conserving

characterization of complex functional brain

networks. Neuroimage 56, 2068–2079. (doi:10.

1016/j.neuroimage.2011.03.069)

24. Sporns O. 2013 Structure and function of complex

brain networks. Dialogues Clin. Neurosci. 15,

247–262.

25. Fornito A, Harrison BJ, Zalesky A, Simons JS. 2012

Competitive and cooperative dynamics of large-scale

brain functional networks supporting recollection.

Proc. Natl Acad. Sci. USA 109, 12 788 –12 793.

(doi:10.1073/pnas.1204185109)

26. Liang X, Zou Q, He Y, Yang Y. 2013 Coupling of

functional connectivity and regional cerebral blood

flow reveals a physiological basis for network hubs

of the human brain. Proc. Natl Acad. Sci. USA 110,

1929–1934. (doi:10.1073/pnas.1214900110)

27. Evans TS. 2010 Clique graphs and overlapping

communities. JSTAT 2010, P12037. (doi:10.1088/

1742-5468/2010/12/P12037)

28. Edelsbrunner H, Harer J. 2010 Computational

topology: an introduction. Providence, RI: American

Mathematical Society.

29. Petri G, Scolamiero M, Donato I, Vaccarino F. 2013

Networks and cycles: a persistent homology approach

to complex networks. In Proc. the European Conf. on

Complex Systems 2012 (eds T Gilbert, M Kirkilionis,

G Nicolis), pp. 93– 99. Springer Proceedings in

Complexity. Berlin, Germany: Springer. (doi:10.1007/

978-3-319-00395-5)

30. Petri G, Scolamiero M, Donato I, Vaccarino F. 2013

Topological strata of weighted complex networks.

PLoS ONE 8, e66506. (doi:10.1371/journal.pone.

0066506)

31. Edelsbrunner H, Letscher D, Zomorodian A. 2002

Topological persistence and simplification. Discrete

Comput. Geom. 28, 511– 533. (doi:10.1007/s00454-

002-2885-2)

32. Carlsson G, Zomorodian A. 2005 Computing

persistent homology. Discrete Comput. Geom. 33,

249–274. (doi:10.1007/s00454-004-1146-y)

33. Steiner DC, Edelsbrunner H, Harer J. 2007 Stability

of persistence diagrams. Discrete Comput. Geom. 37,

103–120. (doi:10.1007/s00454-006-1276-5)

34. Carlsson G. 2009 Topology and data. Bull. Am. Math.

Soc. 46 255– 308. (doi:10.1090/S0273-0979-09-

01249-X)

35. Ghrist R. 2008 Barcodes: the persistent topology of

data. Bull. Am. Math. Soc. 45, 61– 75. (doi:10.1090/

S0273-0979-07-01191-3)

36. Carlsson G, Ishkhanov T, Silva V, Zomorodian A.

2008 On the local behaviour of spaces of natural

images. Int. J. Comput. Vision 76, 1–12. (doi:10.

1007/s11263-007-0056-x)

37. Lum PY et al. 2012 Extracting insights from the

shape of complex data using topology. Sci. Rep. 3,

1236. (doi:10.1038/srep01236)

38. Chan JM, Carlsson G, Rabadan R. 2013 Topology of

viral evolution. Proc. Natl Acad. Sci. USA 110,

18 566–18 571. (doi:10.1073/pnas.1313480110)

39. Nicolau M, Levine A, Carlsson G. 2011 Topology

based data analysis identifies a subgroup of breast

cancers with a unique mutational profile and

excellent survival. Proc. Natl Acad. Sci. USA 108,

7265–7270. (doi:10.1073/pnas.1102826108)

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

9

40. De Silva V, Ghrist R. 2007 Coverage in sensor

networks via persistent homology. Algebr. Geom.

Topol. 7, 339– 358. (doi:10.2140/agt.2007.7.339)

41. van de Weygaert R et al. 2011 Alpha, betti and the

megaparsec universe: on the topology of the cosmic

web. Trans. Comput. Sci. XIV, 60 – 101. Special issue

on Voronoi diagrams and Delaunay triangulation.

(http://arxiv.org/abs/1306.3640)

42. Lee H, Chung MK, Kang H, Kim B-N, Lee DS. 2011

Discriminative persistent homology of brain

networks. In IEEE Int. Symp. on Biomedical Imaging:

From Nano to Macro,30 March 2011– 2 April 2011,

pp. 841–844. Piscataway, NJ: IEEE. (doi:10.1109/

ISBI.2011.5872535)

43. LeeH,KangH,ChungMK,KimBN,LeeDS.2012

Persistent brain network homology from the

perspective of dendrogram. IEEE Trans. Med. Imaging

31, 2267 –2277. (doi:10.1109/TMI.2012.2219590)

44. Carstens CJ, Horadam KJ. 2013 Persistent homology

of collaboration networks. Math. Probl. Eng. 2013

1–7. (doi:10.1155/2013/815035)

45. Danijela H, Maletic

´S, Rajkovic

´M. 2009 Persistent

homology of complex networks. J. Stat. Mech.

2009, P03034. (doi:10.1088/1742-5468/2009/03/

P03034)

46. Computational Topology workgroup, Stanford

University. 2014 JavaPlex: Java library for

computing persistent homology and other

topological invariants. See http://git.

appliedtopology.org/javaplex/.

47. Skraba P, Vejdemo-Johansson M. 2013 Persistence

modules: algebra and algorithms. (http://arxiv.org/

abs/1302.2015)

48. Cohen J, Cohen P. 1983 Applied multiple regression/

correlation analysis for the behavioral sciences.

Hillsdale, NJ: Lawrence Erlbaum Associates.

49. Newman MEJ. 2006 Modularity and

community structure in networks. Proc. Natl Acad.

Sci. USA 103, 8577–8582. (doi:10.1073/pnas.

0601602103)

50. Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E.

2008 Fast unfolding of communities in large

networks. J. Stat. Mech. Theory Exp. 2008, P10008.

(doi:10.1088/1742-5468/2008/10/P10008)

51. Eastwood SL, Burnet PW, Gittins R, Baker K, Harrison PJ.

2001 Expression of serotonin 5-HT(2A) receptors in the

human cerebellum and alterations in schizophrenia.

Synapse 42, 104–114. (doi:10.1002/syn.1106)

52. Sinke C, Halpern JH, Zedler M, Neufeld J, Emrich HM,

Passie T. 2012 Genuine and drug-induced synesthesia:

a comparison. Conscious. Cogn. 21, 1419 –1434.

(doi:10.1016/j.concog.2012.03.009)

53. Hartman AM, Holister LE. 1963 Effect of mescaline,

lysergic acid diethylamide and psylocibin on colour

perception. Psychomarmacolgia 4, 441– 451.

(doi:10.1007/BF00403349)

54. Carhart-Harris RL et al. 2012 Neural correlates of

the psychedelic state as determined by fMRI

studies with psilocybin. Proc. Natl Acad. Sci.

USA 109, 2138–2143. (doi:10.1073/pnas.

1119598109)

55. Smith SM et al. 2004 Advances in functional and

structural MR image analysis and implementation

as FSL. Neuroimage 23, 208–219. (doi:10.1016/j.

neuroimage.2004.07.051)

56. Satterthwaite TD et al. 2013 An improved

framework for confound regression and filtering for

control of motion artifact in the preprocessing of

resting-state functional connectivity data.

Neuroimage 64, 240– 256. (doi:10.1016/j.

neuroimage.2012.08.052)

57. Destrieux C, Fischl B, Dale A, Halgren E. 2010

Automatic parcellation of human cortical gyri and

sulci using standard anatomical nomenclature.

Neuroimage 15, 1– 15. (doi:10.1016/j.neuroimage.

2010.06.010)

58. Greve DN, Fischl B. 2009 Accurate and robust brain

image alignment using boundary-based registration.

Neuroimage 48, 63– 72. (doi:10.1016/j.neuroimage.

2009.06.060)

59. Computational Topology workgroup, Stanford

University. 2013 Holes: Java library for computing

persistent homology and other topological invariants.

See https://code.google.com/p/javaplex/.

rsif.royalsocietypublishing.org J. R. Soc. Interface 11: 20140873

10