ARTICLE
A no ma i e modelling app oach e eals
age-a ypical co ical hickness in a subg oup
o males wi h au ism spec um diso de
Richa d A. I. Be hlehem 1,2✉, Jakob Seidli z1,3, Ra ael Rome o-Ga cia1, S a os T akoshis 4,5,
Guillaume Dumas 6,7,8 & Michael V. Lomba do 2,4
Unde s anding he e ogenei y is an impo an goal on he pa h o p ecision medicine o
au ism spec um diso de s (ASD). We examined how co ical hickness (CT) in ASD can be
pa ame e ized as an indi idualized me ic o a ypicali y ela i e o ypically-de eloping (TD)
age- ela ed no ms. Ac oss a la ge sample (n=870 pe g oup) and wide age ange (5–40
yea s), we applied no ma i e modelling esul ing in indi idualized whole-b ain maps o age-
ela ed CT a ypicali y in ASD and isola ing a small subg oup wi h highly age-a ypical CT. Age-
no med CT sco es also highligh s on-a e age di e en ia ion, and associa ions wi h beha-
iou al symp oma ology ha is sepa a e om insigh s gleaned om adi ional case-con ol
app oaches. This wo k showcases an indi idualized app oach o unde s anding ASD he -
e ogenei y ha could po en ially u he p io i ize wo k on a subse o indi iduals wi h co ical
pa hophysiology ep esen ed in age- ela ed CT a ypicali y. Only a small subse o ASD
indi iduals a e ac ually highly a ypical ela i e o age-no ms. d i ing small on-a e age case-
con ol di e ences.
h ps://doi.o g/10.1038/s42003-020-01212-9 OPEN
1B ain Mapping Uni , Depa men o Psychia y, Uni e si y o Camb idge, Camb idge CB2 0SZ, UK. 2Au ism Resea ch Cen e, Depa men o Psychia y,
Uni e si y o Camb idge, Camb idge CB2 8AH, UK. 3Depa men o Child and Adolescen Psychia y and Beha io al Science, Child en’s Hospi al o
Philadelphia, Philadelphia, PA 19104, USA. 4Labo a o y o Au ism and Neu ode elopmen al Diso de s, Cen e o Neu oscience and Cogni i e Sys ems
@UniTn, Is i u o I aliano di Tecnologia, Ro e e o, I aly. 5Depa men o Psychology, Uni e si y o Cyp us, Nicosia, Cyp us. 6Human Gene ics and Cogni i e
Func ions Uni , Ins i u Pas eu , Pa is, F ance. 7CNRS UMR3571 Genes, Synapses and Cogni ion, Ins i u Pas eu , Pa is, F ance. 8Human Gene ics and
Cogni i e Func ions Uni , Uni e si y Pa is Dide o , So bonne Pa is Ci é, Pa is, F ance. ✉email: [email p o ec ed]
COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio 1
1234567890():,;
Au ism spec um diso de (ASD) is a clinical beha iou al
consensus label we gi e o a di e se collec ion o pa ien s
wi h social-communica ion di ficul ies and p onounced
epe i i e, es ic ed, and s e eo yped beha iou s1.Beyond hesingle
label o ASD, pa ien s a e in ac widely he e ogeneous in pheno ype,
bu also wi h ega ds o he di e si y o di e en ae iologies2.E en
wi hin mesoscopic le els o analysis such as examining b ain
endopheno ypes, he e ogenei y is he ule a he han he excep ion3.
A he le el o s uc u al b ain a ia ion, neu oimaging s udies ha e
iden ified a ious neu oana omical ea u es ha migh help iden i y
indi iduals wi h au ism o e eal elemen s o a common unde lying
biology3. Howe e , he as neu oimaging li e a u e is also
inconsis en , wi h epo s o hypo- o hype -connec i i y, co ical
hinning e sus inc eased g ey o whi e ma e , b ain o e g ow h,
a es ed g ow h, o e en lack o mo phological di e ence al oge he ,
e c.4–13, lea ing s un ed p og ess owa ds unde s anding mechanisms
d i ing co ical pa hophysiology in ASD and ansla ing neu oima-
ging in o clinical u ili y.
Mul iple explana ions could be behind his inconsis ency
ac oss he li e a u e. Me hodology widely di e s ac oss s udies
(e.g., low s a is ical powe , di e en ways o es ima ing mo -
phology o olume) and is likely a e y impo an ac o 9,14.
Ini ia i es such as he au ism b ain imaging da a exchange
(ABIDE15); ha e made i possible o boos sample size by pooling
oge he da a om se e al di e en s udies. Howe e , wi hin-
g oup he e ogenei y in he au ism popula ion also immedia ely
s ands ou as ano he ac o obscu ing consis ency in he li e a-
u e, especially when he dominan app oach o case-con ol
models la gely igno es he e ogenei y wi hin he ASD popula ion.
In pa icula , some au ism- ela ed he e ogenei y epo ed in li -
e a u e migh be explained by ac o s such as age16,17. Indeed,
wi h ega ds o s uc u al b ain ea u es o in e es o s udy in
ASD (e.g., olume, co ical hickness (CT), su ace a ea), hese
ea u es change ma kedly o e de elopmen and may ollow
al oge he di e en ajec o ies in ASD18–20. Typical app oaches
owa ds dealing wi h age e ol e a ound g oup s a is ical mod-
elling o age as he a iable o in e es o emo ing age as a
co a ia e and hen pa ame ically modelling on-a e age di e -
ences be ween cases e sus con ols. While hese a e common
app oaches in he li e a u e, hey do no immedia ely p o ide
indi idualized es ima es o age- ela ed a ypicali y no do hey
accoun o indi idual a ia ion in de elopmen al ajec o ies. In
con as , no ma i e models o age- ela ed a ia ion may likely be
an impo an al e na i e o hese app oaches and may mesh
be e wi h some concep ual iews o a ypicali y in ASD as being
an ex eme o ypical popula ion no ms21. In con as o he
canonical case-con ol model, no ma i e age modelling allows o
compu a ion o indi idualized me ics ha can hone in on he
p ecision in o ma ion we a e in e es ed in— ha is, a ypicali y o
de elopmen exp essed in specific ASD indi iduals ela i e o
non-ASD no ms. Such an app oach may be a ui ul way o wa d
in isola ing indi iduals who a e ‘s a is ical ou lie s’. The easons
behind why hese indi iduals a e ou lie s ela i e o non-ASD
no ms may be o po en ial clinical and/o mechanis ic impo -
ance. Fu he mo e, con en ional case-con ol analyses may
obscu e mo e sub le indi idual di e ences as hey assume on-
a e age g oup di e ences. This is especially impo an in ligh o
p e iously epo ed null-findings14. Indeed, i we a e o mo e
o wa d owa ds s a ified psychia y and p ecision medicine o
ASD22, we mus go beyond case-con ol app oaches and employ
dimensional app oaches ha can ell us in o ma ion abou which
indi iduals a e a ypical and how o why hey exp ess such a y-
picali y. Thus, his app oach aims o p o ide mo e han a me e
s a is ical ad ance, i aims o be e concep ualize and cap u e
pe sonalized in e ences ha may ul ima ely esul in mo e
meaning ul and a ge ed clinical in e ence.
In he p esen s udy, we employ no ma i e modelling on age-
ela ed a iabili y as a means o indi idualize ou app oach o
isola e specific subse s o pa ien s wi h e y di e en neu al ea-
u es. He e we ocus specifically on a neu al ea u e o co ical
mo phology known as CT. CT is a well-s udied neu oana omical
ea u e hough o be di e en ially a ec ed in au ism and has
ecei ed inc easing a en ion in ecen yea s23–27. Recen wo k
om ou g oup also iden ified a gene ic co ela e o au ism-
specific CT a ia ion despi e conside able he e ogenei y in g oup-
specific CT in child en wi h au ism28. A s udy examining ABIDE
I coho da a disco e ed case-con ol di e ences in CT, albei
e y small in e ec size14. Simila ly, he mos ecen and la ges
s udy o da e, a mega-analysis combining da a om ABIDE and
he ENIGMA conso ium, also indica ed e y small on-a e age
case-con ol di e ences in CT es ic ed p edominan ly o a eas
o on al and empo al co ices, and indica e e y sub le age-
ela ed be ween-g oup di e ences and subs an ial wi hin-g oup
age- ela ed a iabili y27. Ano he ecen s udy also highligh ed
widesp ead di e ences in CT and se e al di e ences ha we e
sex-specific7. O e all, hese s udies emphasize h ee gene al
poin s o impo ance. Fi s , age o de elopmen al ajec o y is
ex emely impo an 16,29–31. Second, gi en he conside able
wi hin-g oup age- ela ed a iabili y and he p esence o a la ge
majo i y o null and/o e y small be ween-g oup e ec s, a he
han a emp ing o find on-a e age di e ences be ween all cases
e sus all con ols, we should shi ou ocus o capi alize on his
dimension o la ge age- ela ed a iabili y and isola e au ism cases
ha a e a he ex emes o his dimension o no ma i e a ia-
bili y. Thi d, biological sex is likely o be an impo an modula o
o ASD-specific mo phological di e ences7.
Gi en ou app oach o age- ela ed no ma i e CT modelling, we
fi s compa e he u ili y o age- ela ed no ma i e modelling
di ec ly agains mo e adi ional case-con ol models. We hen
desc ibe he p e alence o ASD cases ha show meaning ul age-
ela ed de iance in CT (i.e. >2 s anda d de ia ions om age-
ela ed no ms o ou side he 95% popula ion confidence bounds)
and show how a me ic o con inuous a iabili y in age- ela ed
a ypicali y in CT is exp essed ac oss he co ex in au ism. Finally,
we explo e age–a ypical CT–beha iou associa ions and assess
whe he such dimensional analyses associa ed wi h beha iou
iden i y simila o di e en egions han ypical case-con ol
analyses. To show applicabili y o his app oach we also applied
he same me hod o o he measu es o neu oana omy; gy ifica-
ion, olume and su ace a ea. Resul s and analyses o hese
me ics can be ound in he supplemen a y ma e ials and all code
and da a used a e a ailable on Gi Hub32.
Resul s
Age- ela ed no ma i e modelling. No ma i e modelling o age-
ela ed CT e ec s was done u ilizing male-only da a om he
ypically de eloping g oup (TD) (see “Me hods”sec ion o ull
sample desc ip ion, Fig. 1 o a schema ic o e iew and Supple-
men a y Figs. 1and 2 o mo e demog aphics in o ma ion). All
analyses we e done on CT a e aged wi hin 308 co ical egions33.
We used a local polynomial eg ession fi ing p ocedu e
(LOESS)34,35, whe e he local wid h o smoo hing ke nel o he
eg ession was de e mined by he model ha p o ided he o e all
smalles sum o squa ed e o s using hype pa ame e op imiza ion
ac oss 5–100% o he ull age ange using B en ’sme hod
36 as
implemen ed in he R op im unc ion om he s a s package.
We also assessed consis ency o ou ou pu using cen iles sco ing
and consis ency o he no ma i e model using ex ensi e boo -
s apping and sensi i i y analyses, bo h showed high ou come
consis ency (see “Me hods”sec ion and Supplemen a y Ma e ials;
Supplemen a y Figs. 3–5). To align he TD and ASD g oups,
ARTICLE COMMUNICATIONS BIOLOGY | h ps://doi.o g/10.1038/s42003-020-01212-9
2COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio
bo h we e binned in o one-yea age bins. Fo each age bin and e e y
b ain egion we compu ed a no ma i e mean and s anda d de ia-
ion om he TD g oup. This was done sepa a ely o each sex,
gi en known sex di e en ial de elopmen al ajec o ies. These s a-
is ical no ms we e hen used o compu e a w-sco e (analogous o a
z-sco e) o e e y indi idual wi h au ism and e e y b ain egion as
ollows:
W egion ¼CT egion μno m egion
σno m egion
The w-sco e o an indi idual hus eflec s how a away hei
CT is om TD no ms in uni s o s anda d de ia ion. Because w-
sco es a e compu ed o e e y b ain egion, we ge a w-sco e map
o each ASD pa icipan showing how each b ain egion o ha
indi idual is a ypical ela i e o TD no ms. Age bins ha
con ained ewe han fi e da a-poin s in he TD g oup we e
excluded om subsequen analysis as he s anda d de ia ions o
hese bins would essen ially be ze o (and hus he w-sco e could
no be compu ed). Wi h he inclusion o mo ion we also excluded
indi iduals o which no es ing-s a e MRI was a ailable. The
cha ac e is ics o he final au ism sample a e lis ed in Table 1.
1
2
3
4
–6
–3
0
3
10 20 30 40
Age
xAu ism -
σno m
µno m
W =
µno m
2σno m
=xAu ism Median µ and 2σ
pe age bin
2.0
2.5
3.0
3.5
4.0
20 40 60
µ
2σ
A
| Age dis ibu ion
2.5
5.0
7.5
10.0
12.5
15.0
17.5
20.0
22.5
25.0
27.5
30.0
32.5
35.0
37.5
40.0
42.5
45.0
47.5
50.0
52.5
55.0
57.5
60.0
62.5
65.0
67.5
0
20
40
60
80
100
120
140
160
180
200
220
240
260
280
300
320
0
20
40
60
80
100
120
140
160
180
200
220
240
260
280
300
320
Coun
20 40 60
B| Age no ma i e modelling o e iew
Male
Female
Age (bins)
Fig. 1 Demog aphics and desc ip i e s a is ics. a His og am o age dis ibu ion pe sex. Females we e excluded om u he analyses due o known sex
di e en ial e ec s in au ism and he lack o a ailable da a o es ima e popula ion no ms (see “Me hods”sec ion o de ails). bSchema ic o e iew o no ma i e
modelling. In he fi s ins ance LOESS eg ession is used o es ima e he de elopmen al ajec o y on CT o e e y indi idual b ain egion o ob ain an age-
specific mean and s anda d de ia ion. Then we compu ed median o each one-yea age-bin o hese mean and median neu o ypical es ima es o align hem
wi h he ASD g oup. Nex , o each indi idual wi h au ism and each b ain egion he no ma i e mean and s anda d de ia ion a e used o compu e a w-sco e
ela i e o hei neu o ypical age-bin. Con a y o con en ional boxplo s, he second panel shows mean, 1 sd and 2 sd o he neu o ypical g oup (in yellow) and
indi iduals wi h an au ism diagnosis in pu ple.
A| Case-con ol linea mixed e ec s model
B| Case-con ol linea mixed e ec s model a e ou lie emo al
00.3-0.3
00.3-0.3
Cohens d’
Cohens d’
Fig. 2 Case con ol di e ence analysis wi h linea mixed e ec model. Panel ashows e ec sizes o egions passing FDR co ec ion o linea mixed
e ec modelling o con en ional case con ol di e ence analysis. Cohen’sd alues ep esen ASD−con ol, hus blue deno es ASD<con ol and ed
deno es ASD>con ol. Panel bshows e ec sizes o egions passing FDR co ec ion a e ou lie emo al o he same linea mixed e ec modelling o
con en ional case con ol di e ence analysis.
COMMUNICATIONS BIOLOGY | h ps://doi.o g/10.1038/s42003-020-01212-9 ARTICLE
COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio 3
Case-con ol di e ences e sus age-no ma i e modelling. Ou
fi s analysis examined con en ional case-con ol di e ences
using linea mixed e ec modelling including si e, sex, age, in-
scanne head mo ion37 and Eule index38 as co a ia es. As
expec ed om p io pape s u ilizing la ge-scale da ase s o case-
con ol analysis14,27, a small subse o egions (8.7%, 27/308
egions) pass FDR co ec ion. O hese egions, mos a e o small
e ec size, wi h 26 o he de ec ed 27 egions showing an e ec
<0.2 s anda d de ia ions o di e ence (Fig. 2a). We suspec ed ha
such small e ec s could be la gely d i en by a ew ASD pa ien s39
wi h highly age-a ypical CT. Because we also had compu ed w-
sco es om ou no ma i e age-modelling app oach, we iden ified
specific‘s a is ical ou lie ’pa ien s o each indi idual egion wi h
w-sco es >2 s anda d de ia ions om ypical no ms and excluded
hem om he case-con ol analysis. This analysis gua ds agains
he influence o hese ex eme ou lie s, and i he e a e ue on-
a e age di e ences in ASD, he emo al o hese ou lie
pa ien * egions should ha e li le e ec on ou abili y o de ec
case-con ol di e ences. Howe e , emo al o ou lie pa ien s
now e ealed only 14 significan egions ins ead o 27 egions
wi h small case-con ol di e ences—a 1.9- old dec ease in he
numbe o egions de ec ed. Indeed, he majo i y o case-con ol
di e ences iden i ying small on-a e age e ec s we e p ima ily
d i en by his small subse o highly a ypical pa ien s (Fig. 2b).
These emaining 14 egions wi h small on-a e age e ec s we e
es ic ed o a eas nea he pos e io cingula e co ex, empo o-
pa ie al co ex and a eas o isual co ex.
In con as o a canonical case-con ol model, we compu ed
no ma i e models o age which esul ed in indi idualized w-
sco es ha indica e how a ypical CT is o an indi idual
compa ed o ypical no ms o ha age. This modelling app oach
allows o compu a ion o w-sco es o e e y egion and in e e y
pa ien , hus esul ing in a w-sco e map ha can hen i sel be
es ed o di e ences om a null hypo hesis o w-sco e =0,
indica ing no significan on-a e age ASD a ypicali y in age-
no med CT. These hypo hesis es s on no ma i e w-sco e maps
e ealed no egions su i ing FDR co ec ion.
Isola ing ASD indi iduals wi h age- ela ed CT a ypicali y.
While he no ma i e modelling app oach can be sensi i e o
di e en pa hology han adi ional case-con ol models, ano he
s eng h o he app oach is he abili y o isola e indi iduals
exp essing high CT-a ypicali y. We ope a ionalized ‘significan ’
a ypicali y in s a is ical e ms as w-sco es >2 SD away om TD
no ms. By applying his cu -o , we can hen desc ibe wha p o-
po ion o he ASD popula ion alls in o his CT subg oup ca e-
go y o each indi idual b ain egion. O e all b ain egions he
median p e alence o hese pa ien s is a ound 7.6% (Fig. 3).
Meaning ha in each b ain egion he e a e ~7.6% o indi iduals
ha would be conside ed an ou lie . This di e ence om an
expec ed p opo ion o 5% in he p esen sample co esponds o a
X2o 3.85 (wi h Ya es con inui y co ec ion40) ha is significan
a p=0.049 (wi hou con inui y co ec ion: X2=4.32, p=0.038).
The dis ibu ion o p e alence ac oss b ain egions also has a
posi i e ail indica ing ha o a small numbe o b ain egions
he p e alence can jump up o mo e han 10%. Exp essed back
in o sample size numbe s, i 10% o he ASD popula ion had
significan CT abno mali ies, wi h a sample size o n=699, his
means ha n=70 pa ien s possess such a ypicali y. Unde sco ing
he p e alence o hese cases is impo an since as shown ea lie ,
i is likely ha p ima ily hese ‘s a is ical ou lie ’pa ien s d i e
mos o he iny case-con ol di e ences obse ed.
The e a e o he in e es ing a ibu es abou his subse o b ain
egions. Wi h ega d o age, hese pa ien s we e almos always in
he age ange o 6–20, and we e much less p e alen beyond age
20 (Supplemen a y Figs. 6 and 7). The median age o ou lie s
ac oss b ain egions anged om [10.6–20.2] yea s old, wi h an
o e all skewed dis ibu ion owa ds he younge end o he
spec um (Supplemen a y Fig. 7), showing ha CT a ypicali y
po en ially no malizes wi h inc easing age in ASD, hough i
should be no ed ha his may pa ially be explained by he o e all
skewed age dis ibu ion in he o e all da ase .
Pa ien s wi h CT a ypicali y we e also la gely hose ha
exp essed such a ypicali y wi hin specific b ain egions and we e
no p ima ily subjec s wi h globally a ypical CT. To show his we
compu ed a w-sco e a io ac oss b ain egions (Supplemen a y
Table 1 Sample age cha ac e is ics a e no ma i e
modelling selec ion.
Dx Mean SD NMedian Min Max
Au ism 14.93 5.97 699 13.40 5.53 39.2
TD 15.35 6.37 624 13.34 5.89 39.4
0
5
10
15
20
0.04 0.08 0.12 0.16
Pe cen age o subjec s wi h a ypical w–sco e
numbe o
b ain egions
A| Region speci ic p e alence
B| Subjec pe cen age pe b ain egion
0.0 0.2
P e alence
Fig. 3 Region specific p e alence o a ypical w-sco es. Panel ashows he by egion p e alence o indi iduals wi h a w-sco e o g ea e han ±2SD. Fo
isualiza ion pu poses hese images a e h esholded a he median p e alence o 0.076. Panel bshows he o e all dis ibu ion o p e alence ac oss all
b ain egions.
ARTICLE COMMUNICATIONS BIOLOGY | h ps://doi.o g/10.1038/s42003-020-01212-9
4COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio
Figs. 6 and 7) ha helps us isola e pa ien s ha show globally
a ypical CT ac oss mos b ain egions. The small numbe o
pa ien s wi h a a io indica ing a global di e ence ( a io > 0.5,
n=14) we e hose ha had globally hinne co ices. This small
subse o indi iduals was much smalle han he numbe o
egion-wise ou lie s as shown in Fig. 3. Upon isual inspec ion o
he aw da a o hese pa icipan s, i is clea ha he global
hinning e ec is no likely a ue biological di e ence bu a he
one likely d i en by he quali y o he aw images, e en hough
he Eule index did no indica e ailu e in econs uc ion.
Un o una ely, we did no ha e enough comple e pheno ypic
da a on hese subjec s o wa an u he in-dep h pheno ypic
analysis.
Explo a o y analysis o b ain–beha iou ela ionships.An
addi ional ad an age o he use o no ma i e modelling o e he
adi ional case-con ol modelling is ha we can use he indi i-
dualized a ypicali y as a no el me ic o finding associa ions wi h
pheno ypic ea u es. He e we used w-sco es o compu e Spea -
man co ela ions o he mos commonly sha ed pheno ypic
ea u es in he ABIDE da ase : ADOS, SRS, SCQ, AQ, FIQ and
Age. A e co ec ing o mul iple compa isons ac oss pheno ype
and egion (6 pheno ypic measu es*308 egions =1848 es s) we
iden ified a numbe o b ain egions ha su i e mul iple com-
pa ison co ec ions o he SRS and ADOS sco es (Fig. 4, Sup-
plemen a y Fig. 8). SRS is associa ed wi h w-sco es p ima ily in
a eas o la e al on al and pa ie al co ex, while ADOS is asso-
cia ed wi h w-sco es p ima ily in la e al and in e io empo al
co ex. No ably, hese egions a e la gely di e en om egions
ha appea o show on-a e age di e en ia ion in case-con ol
and w-sco e analyses.
Sensi i i y analysis. Sensi i i y analyses on he e ec s o econ-
s uc ion quali y using Eule index as well as esidual e ec s o in-
scanne head mo ion om he es ing-s a e acquisi ion did no
e eal a significan impac on h esholded case-con ol di e ences
o w-sco e. Specifically, sys ema ic exclusion o op mo ion and
Eule subjec s esul ed in highly spa ially consis en e ec size
maps (all < 0.7). Indi iduals iden ified as s a is ical ou lie s did
no ha e disp opo ionally high mo ion o high Eule indices. Fo
mo e de ails see “Me hods”sec ion and supplemen a y ma e ials
(Supplemen a y Figs. 2 and 3).
Discussion
In he p esen s udy, we find ha wi h a highly powe ed da ase ,
con en ional case-con ol analyses e eal small di e ences in CT
in au ism and a e es ic ed o a small subse o egions. In
gene al, his idea abou sub le e ec sizes o case-con ol com-
pa isons is compa ible wi h o he ecen pape s u ilizing pa ially
o e lapping da a—Haa and colleagues u ilized only ABIDE I
da a14, while an Rooji and colleagues27 u ilized bo h ABIDE I
and II da ase combined wi h u he da a om he ENIGMA
conso ium. While hese s a emen s abou small e ec sizes a e
no no el, ou findings sugges ha e en hese small e ec sizes
may be misleading and o e -op imis ic. U ilizing no ma i e
modelling as a way o iden i ying and emo ing CT-a ypical
ou lie pa ien s, we find he e ha mos small case-con ol di -
e ences a e d i en by a small subg oup o pa ien s wi h high CT-
a ypicali y o hei age, which indeed begs he ques ion o he
exis ence o on-a e age a ypical co ical mo phology in au ism14.
In con as , we u he showed ha analysis o CT-no med sco es
(i.e. w-sco es) hemsel es e eals a comple ely di e en se o
egions ha a e on-a e age a ypical in ASD. The di ec ionali y o
such di e ences also e e ses in some cases. Fo ins ance, Haa
and colleagues disco e ed ha a eas o he isual co ex a e
hicke in ASD compa ed o TD in ABIDE I14. Ou case-con ol
analyses he e la gely mi o ha finding. Howe e , e-analysis
a e w-sco e ou lie emo al o ally emo es he e ec s p e-
iously epo ed in he isual co ex. Thus, he e is a clea case
whe eby ou no ma i e age modelling app oach iden ifies e ec s
ha a e likely d i en by only a small subse o indi iduals. New
insigh s ia no ma i e age modelling, alongside cleaning up
in e p e a ions behind case-con ol models, bo h highligh he
meaning ul u ili y o his app oach. The p esence o small egion-
dependen ou lie e ec s in ASD misleadingly d i es on-a e age
in e ences om case-con ol models. Thus, i is impo an o he
field o be e unde s and how p e alen his a ypicali y is o a
gi en b ain egion (i.e. ou analyses did no e eal a consis en
b ain egion o g oup o indi iduals wi h spa ially o e lapping
pa e ns o ex eme w-sco es).
We also no ed ha his egion-specific small subg oup showing
highly age-a ypical CT was p edominan ly es ic ed o he
childhood o ea ly adul age ange. In la e adul ages, he p e-
alence o his subg oup d ops o . This could be a po en ial
indica o ha highly a ypical CT is mo e p e alen and de ec able
a ea lie ages. I will be impo an o assess e en ea lie age
A
| ADOS ~ W Co ela ion
B| SRS ~ W Co ela ion
Fig. 4 Pheno ype–w-sco e co ela ions. Spea man co ela ions be ween ADOS and w-sco e in he op panel. The lowe panel shows he same o he SRS.
COMMUNICATIONS BIOLOGY | h ps://doi.o g/10.1038/s42003-020-01212-9 ARTICLE
COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio 5

anges such as he fi s yea s o li e30, as well as la e adul yea s
when aging p ocesses begin o ake e ec 41. Again, i should be
no ed ha his may pa ially be explained by he o e all skewed
age dis ibu ion in he o e all da ase . Fu u e s udies wi h ei he
mo e balanced de elopmen al samples o samples ha co e he
en i e li espan will be be e posi ioned o confi m his age- ela ed
skewed p ofile in ‘a ypical’b ain egions.
In addi ion, we also iden i y a e y small g oup o indi iduals
ha ha e a ypical pa e ns in o e 50% o b ain egions. Un o -
una ely, no much beha iou al o pheno ypic in o ma ion was
a ailable o his sub-g oup. We hope ha u u e s udies will
ob ain mo e de ailed pheno ypic in o ma ion in o de o
delinea e mo e p ecisely wha he clinical and o mo e b oad
beha iou al implica ions migh be o his a ypicali y. Fu he -
mo e, i is clea om he p esen wo k ha his subg oup only
co e s a e y small subse in he au ism popula ion and hus
u u e s udies will equi e la ge sample sizes o be able o iden i y
his subg oup. Howe e , mi o ing wo k in au ism gene ics,
whe eby disco e ies a e con inually being made ega ding e y
small p opo ions o he ASD popula ion being explained by
highly pene an gene ic mechanisms42, i also may be he case
ha such indi iduals wi h highly age-a ypical CT a e indi iduals
wi h specific highly pene an biological mechanisms unde lying
hem, and possibly ela ed o neu ogenesis and o he ac o s ha
a e implica ed in CT changes28. Wi h animal models o highly
pene an gene ic mechanisms linked o au ism, i is no able ha
such mechanisms ha e he e ogeneous e ec s on b ain olume43.
Thus, he ac ha his is only a small subse need no be an
obs acle o he disco e y o co e biological mechanisms. Ideally,
u u e s udies will also collec de ailed gene ic and/o o he bio-
logical in o ma ion in o de o p obe he co e biological ae iology
unde lying he pa e n o b oad a ypical CT.
We also conduc ed explo a o y analyses o ela e he w-sco es
back o pheno ypic in o ma ion mo e b oadly, inso a as his was
a ailable in ABIDE. He e, we find collec ions o a eas ha a e
la gely di e en om egions no mally de ec ed wi h on-a e age
case-con ol o on-a e age non-ze o w-sco e di e ences. In e -
es ingly, he associa ions wi h ADOS and SRS show somewha
di e en ial spa ial opog aphy which may sugges ha he o e all
sco es a e ela ed o di e en unde lying neu obiological
mechanisms. O e all, hese esul s could sugges ha he no -
ma i e model is sensi i e o signal ela ed o beha iou al a ia-
bili y. Howe e , i should be emphasized ha ADOS and SRS
sco es a e no a ailable o he ull da ase and he epo ed e ec s
we e small and should hus be conside ed explo a o y. In addi-
ion, ADOS and SRS a e o en only collec ed on indi iduals wi h a
diagnosis al eady, which makes he gene al in e ence o his
b ain–beha iou ela ionship po en ially biased. Based on p esen
esul s howe e we expec u u e s udies, wi h mo e comp e-
hensi e pheno ypic in o ma ion such as EU-AIMS2- ials
(h ps://www.aims-2- ials.eu), o be able o confi m his
b ain–beha iou ela ionship. I will be in e es ing o see whe he
in a la ge mo e comp ehensi e sample he same opological
dissocia ion becomes appa en as well.
The cu en esul s can be con as ed wi h a ecen s udy on
he EU-AIMS LEAP coho 44. This s udy di e s om he cu en
wo k in being based on a comple ely independen da ase (EU-
AIMS LEAP s. ABIDE). The s udies also di e in how no ma i e
models a e es ima ed—LOESS and cen iles s. Gaussian p ocess
eg ession. This s udy applied no ma i e modelling only o males
o educe sex- ela ed he e ogenei y whe eas Zahibi e al. u ilized
bo h males and emales and used sex as a ac o in he model. The
cu en s udy also u ilizes a la ge sample size (au ism n=699,
TD n=624; au ism n=321, TD n=206 in e . 44). Despi e
hese di e ences, some impo an consis encies eme ge. In pa -
icula , ou map o p e alence o he CT ou lie g oup (Fig. 3)is
somewha consis en wi h he spa ial opology Zahibi and col-
leagues epo o nega i e de ia ions om he no ma i e model
(e.g., Fig. 4 o e . 44). Fu he mo e, while ou analyses o
b ain–beha iou al ela ionships is limi ed, he e is some con-
sis ency ac oss his s udy and Zahibi e al. wi h he co ela ion
be ween ADOS o al sco es and le in e io on al gy us. Thus,
despi e he me hodological di e ences, he o e all consis ency
sugges s ha many o he in e ences om hese wo ks gene alize
o he au ism popula ion.
The e a e a numbe o ca ea s o conside in he p esen s udy.
Fi s and o emos , he p esen da a a e c oss-sec ional and he
no ma i e age modelling app oach canno make claims abou
ajec o ies a an indi idual le el. Wi h longi udinal da a, his
no ma i e modelling app oach could be ex ended. Howe e , a
he momen he classifica ions o highly age-a ypical CT indi i-
duals a e limi ed o s a ic no ma i e s a is ics wi hin disc e e age-
bins a he han based on s a is ics om obus no ma i e a-
jec o ies. The da ase also ep esen s ASD wi hin an age ange
ha misses e y ea ly de elopmen al and also e y la e adul hood
pe iods. Second, he da ase also p esen s a pos -hoc collec ion o
si es accumula ed h ough he ABIDE ini ia i e, whe eby scan-
ne s, imaging acquisi ion sequences and pa ame e s, sample
asce ainmen , e c., a e highly he e ogeneous. As a esul , we
obse ed ha si e had a la ge e ec on explaining a iance in CT
and his is compa ible wi h obse a ions made by o he s udies14.
Fu he mo e, i is likely ha he e may be sys ema ic in e ac ions
be ween scanne si e and some a iables o in e es such as age
(e.g. di e en scanning si es will likely ha e ec ui ed specific age
coho s). Thi d, he e a e a numbe o di e en app oaches o
no ma i e modelling ha all ha e p os and cons (see e . 45 o an
excellen e iew). We chose o use LOESS es ima ion as i is
compu a ionally e ficien and he esul ing w-sco es a e easily
in e p e able. Howe e , since i is based on es ima ion o s anda d
de ia ion om a no ma i e sample i is po en ially sensi i e o
small samples in a gi en age-bin (e.g. i he e a e only ou da a-
poin s o a gi en age-bin he e is likely o be a less eliable sd).
Hence in si ua ions whe e da a is spa e he LOESS app oach may
allow o less eliable no ma i e sco es. In o de o assess he
sensi i i y o ou app oach in he p esen da a we implemen ed
he a o emen ioned boo s apping p ocedu e o iden i y obus -
ness o ou lie de ec ion. In addi ion, we also conduc ed a cen iles
es ima ion ha is ela i ely s anda d in o example epidemiol-
ogy46, simila o quan ile ank maps47 and a guably less sensi i e
o small sample unce ain y. Bo h app oaches showed highly
significan co ela ion in de e mining whole-b ain w-sco e a ios
( =0.87, p=4e−119 and =0.66, p=5.7e−39 o ABIDE I and
ABIDE II, espec i ely; see supplemen a y ma e ials, Supple-
men a y Fig. 5). Fou h, ou cu en sample was ma ched on IQ
and as a esul excluded indi iduals wi h low IQ sco es (<70).
While highe IQ does no au oma ically imply highe o e all
unc ioning48 i does limi he gene alisabili y o ou findings o
indi iduals wi h no mal o high IQ. Finally, al hough in-scanne
head mo ion is a well-known con ounde in es ing-s a e con-
nec i i y s udies49,50 i has ecen ly been shown ha he same
mo ion may also a ec s uc u al image quali y and su ace
econs uc ion, especially in clinical coho s37. To add ess his
issue in he p esen analysis we included mean amewise dis-
placemen in ou models. In his line, Sa alia e al. 51 ecen ly
showed ha amewise displacemen is a sensi i e p oxy o
mo ion- ela ed bias in s uc u al images. We find ha , while his
se e ely impac ed he con en ional case con ol analysis (e.g.
educing he numbe o significan ROI’s om 38 o 27), i did
no impac he ou lie h esholded analysis o he same ex en . To
u he assess he sensi i i y o mo ion on he p esen app oach
we include sensi i i y analyses based on sys ema ic emo al o
high mo ion subjec s and find ha he spa ial opology o e ec s
ARTICLE COMMUNICATIONS BIOLOGY | h ps://doi.o g/10.1038/s42003-020-01212-9
6COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio
was s ongly conse ed. Gi en he impac on he con en ional
analysis app oach we s ongly encou age u u e s udies o con-
side mo ion as an impo an con ounde .
In conclusion, he p esen s udy shows how no ma i e age
modelling app oach in ASD ques ions ou in e p e a ion o
con en ional case-con ol modelling while shedding new insigh
in o he e ogenei y in ASD. We show ha esul s om case-
con ol analyses, e en wi hin la ge da ase s, can be highly sus-
cep ible o he influence o ‘ou lie ’subjec s. Remo ing hese
ou lie subjec s om analyses can conside ably clean up he
in e ences being made abou on-a e age di e ences ha apply o
a majo i y o he ASD popula ion. Ra he han only being nui-
sances o s anda d g oup-le el analyses, hese ou lie pa ien s a e
meaning ul in hei own ligh , and can be iden ified wi h ou
no ma i e age modelling app oach. No ma i e models may
p o ide an al e na i e o case-con ol models ha es hypo heses
a a g oup-le el, by allowing addi ional insigh o be made a
mo e indi idualized le els, and hus help u he p og ess
owa ds pe sonalized medicine o ASD. Fu he mo e, he cu -
en app oach is in line wi h he o iginal no ma i e modelling
app oach ad oca ed by Ma quand and colleagues21 which sug-
ges ed he de elopmen o me hods o mo e away om he a-
di ional case- s.-con ol analyses. No ma i e modelling was
o iginally p oposed as one solu ion among o he s like s a ifica-
ion. He e, a clea pa h o wa d would be o combine bo h, o
ins ance by using ou pu o no ma i e models as ea u es used in
he pa icipan s a ifica ion, hus a oiding i ial clus e ing
caused by con ounding ac o s. In he p esen wo k we show ha
no ma i e modelling is mo e han a pu ely s a is ical ad ance-
men o imp o e obus ness. I allows us o iden i y a small
subg oup ha we expec o ha e s ong ele ance o he dis-
co e y o co e biological o pheno ypic clinical a ge s. I allowed
o explo a ion o b ain–beha iou ela ionships ha e eal di -
e en ial spa ial opology o ADOS and SRS sco es. Mo e
impo an ly howe e , i mo es us concep ually close o making
p ecise dimensional in e ences a he han pu ely elying on
diagnos ic ca ego ies.
Me hods
Pa icipan s. In his s udy, we fi s sough o le e age la ge neu oimaging da ase s
o yield g ea e s a is ical powe o iden i ying sub le e ec s. To achie e his, we
u ilized he ABIDE da ase s (ABIDE I and II; 15) (see Supplemen a y Fig. 1).
In o med consen was gi en a each si e included in he ABIDE s udies, see he
websi e o mo e de ails: h p:// con_1000.p ojec s.ni c.o g/indi/abide/. Gi en ha
he no malized modelling app oach gi es us indi idual le el measu es we chose o
also include si es wi h limi ed numbe s o subjec s. G oups we e subsequen ly
ma ched on age using he non-pa ame ic nea es neighbou ma ching p ocedu e
implemen ed in he Ma chi package in R (h ps://c an. -p ojec .o g/web/packages/
Ma chI /index.h ml)52. A e ma ching case and con ol g oups and excluding
scans o poo e quali y (see supplemen a y ma e ials) we we e le wi h a sample
size N=870 pe g oup (Tables 2and 3).
Imaging p ocessing and quan ifica ion. Co ical su ace econs uc ion was
pe o med using he MPRAGE (T1) image o each pa icipan wi h F eeSu e
(h p://su e .nm .mgh.ha a d.edu/) e sion ( 5.3.0, o ensu e compa abili y wi h
p e ious ABIDE publica ions). The econs uc ion pipeline pe o med by F ee-
Su e “ econ-all”in ol ed in ensi y no maliza ion, egis a ion o Talai ach space,
skull s ipping, WM segmen a ion, essella ion o he WM bounda y, and au o-
ma ic co ec ion o opological de ec s. B iefly, non-uni o mi y in ensi y co ec ion
algo i hms we e applied be o e skull s ipping53, esul ing in esampled iso opic
images o 1 mm. An ini ial segmen a ion o he whi e ma e issue was pe o med
o gene a e a essella ed ep esen a ion o he WM/GM bounda y. The esul ing
su ace was de o med ou wa ds o he olume ha maximizes he in ensi y con as
be ween GM and ce eb ospinal fluid, gene a ing he pial su ace54. Resul ing su -
aces we e cons ained o a sphe ical opology and co ec ed o geome ical and
opological abno mali ies. CT o each e ex was defined as he sho es dis ance
be ween e ices o he GM/WM bounda y and he pial su ace55. We chose o no
conduc manual segmen a ions and excluded ailed subjec s om any subsequen
analysis (and hese subjec s we e emo ed p io o he ma ching and QC p oce-
du es). To assess he quali y o F eesu e econs uc ions we compu ed he Eule
index38. The Eule numbe is a quan i a i e p oxy index o segmen a ion quali y
and has shown high o e lap wi h manual quali y con ol labelling38. The index
coun s he numbe o imes he eesu e has had o in e pola e su ace gaps
du ing he econs uc ion o ensu e a con inuous ou come su ace. As such he
index is e ec i ely a measu e o he eliabili y o he su ace econs uc ion and
he esul ing CT es ima es. In he ull sample we ound a small bu significan
di e ence in bo h hemisphe es (Supplemen a y Fig. 2) wi h he au ism g oup
ha ing o e all sligh ly wo se scan quali y (d=0.176 and d=0.187 o le and igh
hemisphe es, espec i ely). The e o e, we chose o exclude he op 10% o subjec s
wi h an ex eme Eule index (co esponding o a Eule index o ~300) and e an
he Ma chi gene ic ma ching algo i hm o check o ma ched samples. To u he
ensu e adequa e con ol o scan quali y we included he index i sel as a con ound
a iable in all models.
Ac oss bo h ABIDE I and ABIDE II CT was ex ac ed o each subjec using
wo di e en pa cella ions schemes: an app oxima ely equally sized pa cella ion o
308 egions (~500 mm2each pa cel)33,56,57 and a pa cella ion o 360 egions
de i ed om mul i-modal ea u es ex ac ed om he Human Connec ome P ojec
(HCP) da ase 58. The 308- egion pa cella ion was cons uc ed in he F eeSu e
sa e age empla e by subdi iding he 68 egions defined in he Desikan–Killiany
a las59. Thus, each o he 68 egions was sequen ially sub-pa cella ed by a
back acking algo i hm in o egions o ~500 mm2, esul ing in a high- esolu ion
pa cella ion ha p ese ed he o iginal ana omical bounda ies defined in he
o iginal a las33. Su ace econs uc ions o each indi idual we e co- egis e ed o he
sa e age subjec . The in e se ans o ma ion was used o map bo h pa cella ion
schemes in o he na i e space o each pa icipan .
S a is ics and ep oducibili y. Because o powe limi a ions in pas wo k wi h
small samples, we conduc ed an a p io i s a is ical powe analysis indica ing ha a
minimum case-con ol e ec size o d=0.1752 could be de ec ed a his sample
size wi h 80% powe a a conse a i e alpha se o 0.00560. Fo co ela ional
analyses looking a b ain–beha iou associa ions, we examined a subse o pa ien s
wi h he da a om he SRS (Nau ism_male =421) and ADOS o al sco es (Nau-
ism_male =505). Wi h he same powe and alpha le els, he minimum e ec o
SRS is =0.1765 and =0.1651 o he ADOS.
The e a e likely many a iables ha con ibu e o a iabili y in CT be ween
indi iduals and ac oss he b ain. In o de o isually assess he con ibu ion o some
p ominen sou ces o a iance we adop ed a isualiza ion amewo k de i ed om
gene exp ession analysis (h p://bioconduc o .o g/packages/ a iancePa i ion)61 and
included he mos commonly a ailable co a ia es in he ABIDE da ase : age, sex,
diagnosis, scanne si e, ull-scale IQ, e bal IQ, handedness and SRS. Gi en ha
ABIDE was no designed as an in eg a ed da ase om he ou se , i seems plausible
ha he scanne si e migh be ela ed o au ism o au ism- ela ed a iables (e.g., some
si es migh ha e di e en case-con ol a ios o only ec ui ed specific subg oups).
Figu e 5shows he anked con ibu ion o hose co a ia es. Pe haps unsu p isingly,
scanne si e and age p o ed o be he mos dominan sou ces o a iance (each
explaining on a e age a ound 15% o he o al a iance). Ou ini ial con en ional
analysis was aimed o delinea e po en ial b oad case-con ol di e ences, as has been
done in p e ious s udies14,27. We used a linea mixed e ec s model wi h scanne si e
Table 2 Sample cha ac e is ics o age.
Dx Sex Mean SD NMedian Min Max
Au ism Male 16.32 9.09 754 13.75 5.13 64
Au ism Female 15.06 8.43 116 12.57 5.22 54
TD Male 16.64 8.98 660 13.69 5.89 64
TD Female 13.25 5.33 210 11.09 5.91 32
Table 3 Sample cha ac e is ics.
Measu e Dx Sex Mean SD NMedian
IQ Au ism Male 106.13 16.51 754 107
Au ism Female 105.88 16.21 116 106.5
TD Male 111.28 12.13 660 111
TD Female 112.07 13.21 210 112
ADOS Au ism Male 11.15 3.86 505 11
Au ism Female 11.41 3.9 63 11
Con ol Male 1.55 1.58 38 1
Con ol Female 3 1.05 10 3
SRS Au ism Male 80.42 21.41 421 77
Au ism Female 85.95 22.07 61 88
Con ol Male 38.43 15.25 337 41
Con ol Female 39.93 12.13 120 42
COMMUNICATIONS BIOLOGY | h ps://doi.o g/10.1038/s42003-020-01212-9 ARTICLE
COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio 7
as a andom e ec . Gi en he po en ially s ong con ibu ion o age we chose o
include his as fixed e ec s co a ia es in he model. Mul iple compa ison co ec ion
was implemen ed wi h Benjamini–Hochbe g FDR a q< 0.0562.Allmodelsalso
included Eule indices38 and mean amewise displacemen 37 as con ound eg esso s
(see also Supplemen a y Fig. 3 o sensi i i y analyses on hese con ound eg esso s).
No ma i e modelling eliabili y. To assess he eliabili y o he no ma i e w-sco e
we pe mu ed he no ma i e sample (1000 boo s aps, wi h eplacemen ) and
compu ed 1000 pe mu ed w-sco es o each indi idual and each b ain egion. To
subsequen ly quan i y he eliabili y o he w-sco e we compu ed an FDR co ec ed
analogous p- alue o each subjec by compu ing he absolu e posi ion o he eal
w-sco e in he dis ibu ion o pe mu ed w-sco es. The a ionale being ha i a eal
w-sco e would be in he op 5% o he boo s apped dis ibu ion i would likely no
be a eliable sco e (e.g. he sco e would be influenced by only a small subse o he
no ma i e da a). The median numbe o b ain egions pe subjec wi h a significan
p- alue in he no ma i e sample was 1 (ou o 308), indica ing ha he no ma i e
sample is opologically obus and ha he w-sco e is a obus eflec ion o a y-
picali y. Mo e de ails on he boo s apping p ocedu e a e p o ided in he sup-
plemen a y ma e ial (Supplemen a y Fig. 4). To u he assess he dis ibu ion in
he no ma i e g oup we also conduc ed one-sample linea mixed e ec s modelling
in he no ma i e g oup only o de e mine i any o all b ain egions would show
ou lie consis ency. The e we e no b ain egions o which he w-sco e showed a
de ia ion significan om ze o in he no ma i e g oup (e en wi hou co ec ing o
mul iple compa isons ac oss all b ain egions).
Because w-sco e maps a e compu ed o each indi idual, we an hypo hesis es s
a each b ain egion o iden i y egions ha show on-a e age non-ze o w-sco es
s a ified by sex (FDR co ec ed a q< 0.05). To assess he e ec o age- ela ed
indi idual ou lie s on he global case-con ol di e ences we e- an he hypo heses
es s on w-sco es a e emo ing egion-wise indi idual ou lie s (based on a 2 SD
cu -o ). Al hough o cla i y he p esen manusc ip s p esen only esul s on CT,
o comple eness esul s om he same analysis on co ical olume, su ace a ea
and gy ifica ion a e shown in Supplemen a y Figs. 9–12.
Un o una ely, despi e a significan emale sub-g oup, he age-wise binning
g ea ly educed he numbe o bins wi h enough da a-poin s in he emale g oup.
Gi en he educed sample size in he emale g oup and he known in e ac ion
be ween au ism and biological sex63,64, as well as he known sex di e ences in
de elopmen al ajec o ies65, we conduc ed no ma i e modelling on he male
g oup only (Fig. 2a).
To explo e isola ed subse s o indi iduals wi h significan age- ela ed CT
a ypicali y, we used a cu -o sco e o 2 s anda d de ia ions (i.e. w ≥2o w≤2).
This cu -o allows us o isola e specific ASD pa ien s wi h abno mal CT ela i e o
age-no ms o each indi idual b ain egion. We hen calcula ed sample p e alence
(pe cen age o all ASD pa ien s wi h a ypical w-sco es), in o de o desc ibe how
equen such indi iduals a e in he ASD popula ion and o each b ain egion
indi idually. A sample p e alence map can hen be compu ed o show he
equency o hese pa ien s ac oss each b ain egion. We also wan ed o assess how
many pa ien s ha e ma kedly a ypical w-sco es (beyond 2 SD) ac oss a majo i y o
b ain egions. This was achie ed by compu ing an indi idual global w-sco e a io
as ollows:
gW ¼Σwjj>2
Σwjj<2
We also compu ed global w-sco e a ios o posi i e and nega i e w egions
sepa a ely.
Explo a o y analyses. In addi ion o assessing he e ec o no ma i e ou lie on
con en ional case-con ol analyses we also conduc ed some explo a o y analysis on
he no ma i e w-sco es. Fi s , o explo e whe he he w-sco es eflec a po en ially
meaning ul pheno ypic ea u e we also compu ed Spea man co ela ions o each
b ain egion be ween he mos commonly sha ed pheno ypic ea u es in ABIDE:
ADOS, SRS, SCQ, AQ, FIQ and Age. Resul ing p- alues ma ices we e co ec ed o
mul iple compa isons using Benjamini–Hochbe g FDR co ec ion and only egions
su i ing and FDR-co ec ed p- alue o < 0.05 a e epo ed.
Finally, we explo ed whe he he aw CT alues could be used in a mul i a ia e
ashion o sepa a e g oups by diagnosis o illumina e s a ifica ion wi hin ASD in o
subg oups. He e we used k-medoid clus e ing on -dis ibu ed s ochas ic neighbou
embedding ( SNE)66. Ba nes–Hu SNE was used o cons uc a wo-dimensional
embedding o all pa cels in o de o be able o un k-medoid clus e ing in a 2D
ep esen a ion and in o de o isually assess he mos likely scena io wi hin he
amewo k sugges ed by Ma quand and colleagues21. Nex , we pe o med
pa i ioning a ound medoids (PAM), es ima ing he op imum numbe o clus e s
using he op imum a e age silhoue e wid h67. De ails o his explo a o y analysis
a e epo ed in he supplemen a y ma e ials (Supplemen a y Fig. 13).
Repo ing summa y. Fu he in o ma ion on esea ch design is a ailable in he Na u e
Resea ch Repo ing Summa y linked o his a icle.
Da a a ailabili y
All da a is openly a ailable on Gi Hub32, his includes all measu es ex ac ed om he
aw imaging da a alongside he ele an pheno ypic and quali y con ol measu es.
O iginal unp ocessed neu oimaging da a is openly a ailable h ough he ABIDE
conso ium: h p:// con_1000.p ojec s.ni c.o g/indi/abide/abide_I.h ml.
Code a ailabili y
All code is openly a ailable on Gi Hub32, Cohen’sdwe e compu ed using: h ps://gi hub.
com/m lomba do/u ils/blob/mas e /cohens_d.R and he cen iles c oss- alida ion code
can be ound in h ps://gi hub.com/deep-in ospec ion/PyNM.
Recei ed: 16 Ap il 2020; Accep ed: 5 Augus 2020;
Re e ences
1. Lai, M.-C., Lomba do, M. V. & Ba on-Cohen, S. Au ism. Lance 383, 896–910
(2014).
2. Lomba do, M. V., Lai, M. -C. & Ba on-Cohen, S. Big da a app oaches o
decomposing he e ogenei y ac oss he au ism spec um. Mol. Psychia y
h ps://doi.o g/10.1038/s41380-018-0321-0 (2019).
3. Ecke , C. The neu oana omy o au ism spec um diso de : an o e iew o
s uc u al neu oimaging findings and hei ansla abili y o he clinical
se ing. Au ism 21,18–28 (2017).
4. Hong, S.-J., Be nha d , B. C., Gill, R. S., Be nasconi, N. & Be nasconi, A. The
spec um o s uc u al and unc ional ne wo k al e a ions in mal o ma ions o
co ical de elopmen . B ain 140, 2133–2143 (2017).
5. Hong, S.-J. e al. A ypical unc ional connec ome hie a chy in au ism. Na .
Commun. 10, 1022 (2019).
6. Hong, S.-J., Valk, L., Ma ino, A. D., Milham, M. P. & Be nha d , B. C.
Mul idimensional neu oana omical sub yping o au ism spec um diso de .
Ce eb al Co ex 10, 3578–3588 (2018).
7. Bed o d, S. A. e al. La ge-scale analyses o he ela ionship be ween sex, age
and in elligence quo ien he e ogenei y and co ical mo phome y in au ism
spec um diso de . Mol. Psychia y h ps://doi.o g/10.1038/s41380-019-0420-6
(2019).
8. Schue ze, M. e al. Mo phological al e a ions in he halamus, s ia um, and
pallidum in au ism spec um diso de . Neu opsychopha macology 41,
2627–2637 (2016).
9. Visse s, M. E., Cohen, M. X. & Geu s, H. M. B ain connec i i y and high
unc ioning au ism: a p omising pa h o esea ch ha needs efined models,
me hodological con e gence, and s onge beha io al links. Neu osci.
Biobeha . Re . 36, 604–625 (2012).












































































































































0
25
50
75
100
Si e
Age
FIQ
VIQ
SRS
Handedness
Diagnosis Sex
Residuals
Va iance explained (%)
Fig. 5 Explained a iance in co ical hickness o each co a ia e. Age age
a he ime o scanning, FIQ unc ional in elligen quo ien , VIQ e bal
in elligence quo ien , SRS o al sco e o he social esponsi e scale,
Diagnosis diagnos ic g oup, i.e. ASD o TD.
ARTICLE COMMUNICATIONS BIOLOGY | h ps://doi.o g/10.1038/s42003-020-01212-9
8COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio
10. Ecke , C. e al. B ain su ace ana omy in adul s wi h au ism: he ela ionship
be ween su ace a ea, co ical hickness, and au is ic symp oms. JAMA
Psychia y 70,59–70 (2013).
11. Lai, M.-C. e al. Biological sex a ec s he neu obiology o au ism. B ain 136,
2799–2815 (2013).
12. Yang, D. Y.-J., Beam, D., Pelph ey, K. A., Abdullahi, S. & Jou, R. J. Co ical
mo phological ma ke s in child en wi h au ism: a s uc u al magne ic
esonance imaging s udy o hickness, a ea, olume, and gy ifica ion. Mol.
Au ism 7, 11 (2016).
13. Mensen, V. T. e al. De elopmen o co ical hickness and su ace a ea in
au ism spec um diso de . Neu oImage: Clin. 13, 215–222 (2016).
14. Haa , S., Be man, S., Beh mann, M. & Dins ein, I. Ana omical abno mali ies
in au ism? Ce eb. Co ex 26, 1440–1452 (2016).
15. Di Ma ino, A. e al. The au ism b ain imaging da a exchange: owa ds a la ge-
scale e alua ion o he in insic b ain a chi ec u e in au ism. Mol. Psychia y
19, 659–667 (2014).
16. Geo giades, S., Bishop, S. L. & F azie , T. Edi o ial pe spec i e: longi udinal
esea ch in au ism—in oducing he concep o ‘ch onogenei y’.J. Child
Psychol. Psychia y 58, 634–636 (2017).
17. Lo d, C., Bishop, S. & Ande son, D. De elopmen al ajec o ies as au ism
pheno ypes. Am. J. Med. Gene . C 169, 198–208 (2015).
18. Mills, K. L. e al. S uc u al b ain de elopmen be ween childhood and
adul hood: con e gence ac oss ou longi udinal samples. Neu oimage 141,
273–281 (2016).
19. Raznahan, A. e al. Pa e ns o coo dina ed ana omical change in human
co ical de elopmen : a longi udinal neu oimaging s udy o ma u a ional
coupling. Neu on 72, 873–884 (2011).
20. Raznahan, A. e al. How does you co ex g ow? J. Neu osci. 31, 7174–7177
(2011).
21. Ma quand, A. F., Rezek, I., Bui elaa , J. K. & Beckmann, C. F. Unde s anding
he e ogenei y in clinical coho s using no ma i e models: beyond case-con ol
s udies. Biol. Psychia y 80, 552–561 (2016).
22. Kapu , S., Phillips, A. G. & Insel, T. R. Why has i aken so long o biological
psychia y o de elop clinical es s and wha o do abou i ? Mol. Psychia y 17,
1174–1179 (2012).
23. Jiao, Y. e al. P edic i e models o au ism spec um diso de based on b ain
egional co ical hickness. Neu oimage 50, 589–599 (2011).
24. Khund akpam, B. S., Lewis, J. D., Kos opoulos, P., Ca bonell, F. & E ans, A. C.
Co ical hickness abno mali ies in au ism spec um diso de s h ough la e
childhood, adolescence, and adul hood: a la ge-scale MRI s udy. Ce eb. Co ex
27, 1721–1731 (2017).
25. Mo adi, E., Khund akpam, B., Lewis, J. D., E ans, A. C. & Tohka, J.
Neu oImage P edic ing symp om se e i y in au ism spec um diso de based
on co ical hickness measu es in agglome a i e da a. Neu oimage 144,
128–141 (2017).
26. Zielinski, B. A. e al. Longi udinal changes in co ical hickness in au ism and
ypical de elopmen . B ain 137, 1799–1812 (2014).
27. an Rooij, D. e al. Co ical and subco ical b ain mo phome y di e ences
be ween pa ien s wi h au ism spec um diso de and heal hy indi iduals
ac oss he li espan: esul s om he ENIGMA ASD Wo king G oup. Am. J.
Psychia y 175, 359–369 (2018).
28. Rome o-Ga cia, R., Wa ie , V., Bullmo e, E. T., Ba on-Cohen, S. &
Be hlehem, R. A. I. Synap ic and ansc ip ionally down egula ed genes a e
associa ed wi h co ical hickness di e ences in au ism. Mol. Psychia y
h ps://doi.o g/10.1038/s41380-018-0023-7 (2018).
29. Cou chesne, E. e al. Mapping ea ly b ain de elopmen in au ism. Neu on 56,
399–413 (2007).
30. Cou chesne, E., Campbell, K. & Solso, S. B ain g ow h ac oss he li e span in
au ism: age-specific changes in ana omical pa hology. B ain Res. 1380,
138–145 (2011).
31. Schumann, C. M. e al. Longi udinal magne ic esonance imaging s udy o
co ical de elopmen h ough ea ly childhood in au ism. J. Neu osci. 30,
4419–4427 (2010).
32. Be hlehem, R. A. I., Seidli z, J., Rome o-Ga cia, R., Dumas, G. & Lomba do,
M. V. No ma i e age modelling o co ical hickness in au is ic males.
Zenodo h ps://doi.o g/10.5281/ZENODO.1325171 (2018).
33. Rome o-ga cia, R., A ienza, M., Clemmensen, L. H. & Can e o, J. L. E ec s o
ne wo k esolu ion on opological p ope ies o human neoco ex.
Neu oimage 59, 3522–3532 (2012).
34. Cle eland, W. S., De lin, S. J. & G osse, E. Reg ession by local fi ing:
me hods, p ope ies, and compu a ional algo i hms. J. Econom. 37,87–114
(1988).
35. Le eb e, A. e al. Alpha wa es as a neu oma ke o au ism spec um diso de :
he challenge o ep oducibili y and he e ogenei y. F on . Neu osci. 12, 662
(2018).
36. B en , R. P. An algo i hm wi h gua an eed con e gence o finding a ze o o a
unc ion. Compu . J. 14, 422–425 (1971).
37. Pa doe, H. R., Kucha sky Hiess, R. & Kuzniecky, R. Mo ion and mo phome y
in clinical and nonclinical popula ions. Neu oimage 135, 177–185 (2016).
38. Rosen, A. F. G. e al. Quan i a i e assessmen o s uc u al image quali y.
Neu oimage 169, 407–418 (2018).
39. By ge, L., Dubois, J., Tyszka, J. M., Adolphs, R. & Kennedy, D. P. Idiosync a ic
b ain ac i a ion pa e ns a e associa ed wi h poo social comp ehension in
au ism. J. Neu osci. 35, 5837–5850 (2015).
40. Ya es, F. Con ingency ables in ol ing small numbe s and he χ2 es . Suppl. J.
R. S a . Soc. 1, 217–235 (1934).
41. Happé, F. & Cha l on, R. A. Aging in au ism spec um diso de s: a mini-
e iew. Ge on ology 58,70–78 (2012).
42. Geschwind, D. H. & S a e, M. W. Gene hun ing in au ism spec um diso de :
on he pa h o p ecision medicine. Lance Neu ol. 14, 1109–1120 (2015).
43. Ellegood, J. e al. A highly specific pa e n o olume ic b ain changes
due o 22q11.2 dele ions in bo h mice and humans. Mol. Psychia y 19,6–6
(2014).
44. Zabihi, M. e al. Dissec ing he he e ogeneous co ical ana omy o au ism
spec um diso de using no ma i e models. Biol. Psychia y Cogn. Neu osci.
Neu oimaging 4, 567–578 (2019).
45. Ma quand, A. F. e al. Concep ualizing men al diso de s as de ia ions om
no ma i e unc ioning. Mol. Psychia y h ps://doi.o g/10.1038/s41380-019-
0441-1 (2019).
46. Visse , G. H. A., Eile s, P. H. C., El e ink-S inkens, P. M., Me kus, H. M. W.
M. & Wi , J. M. New Du ch e e ence cu es o bi hweigh by ges a ional
age. Ea ly Hum. De . 85, 737–744 (2009).
47. Chen, H. e al. Quan ile ank maps: a new ool o unde s anding indi idual
b ain de elopmen . Neu oimage 111, 454–463 (2015).
48. Tillmann, J. e al. In es iga ing he ac o s unde lying adap i e unc ioning in
au ism in he EU-AIMS Longi udinal Eu opean Au ism P ojec . Au ism Res.
12, 645–657 (2019).
49. Van Dijk, K. R. A., Sabuncu, M. R. & Buckne , R. L. The influence o head
mo ion on in insic unc ional connec i i y MRI. Neu oimage 59, 431–438
(2012).
50. Powe , J. D., Ba nes, K. A., Snyde , A. Z., Schlagga , B. L. & Pe e sen, S. E.
Spu ious bu sys ema ic co ela ions in unc ional connec i i y MRI ne wo ks
a ise om subjec mo ion. Neu oimage 59, 2142–2154 (2012).
51. Sa alia, N. K. e al. Mo ion- ela ed a i ac s in s uc u al b ain images e ealed
wi h independen es ima es o in-scanne head mo ion. Hum. B ain Mapp. 38,
472–492 (2017).
52. Ho, D. E., Imai, K., King, G. & S ua , E. A. Nonpa ame ic p ep ocessing o
pa ame ic causal in e ence. J. S a . So w. 42,1–28 (2011).
53. Ségonne, F. e al. A hyb id app oach o he skull s ipping p oblem in MRI.
Neu oimage 22, 1060–1075 (2004).
54. Dale, A. M., Fischl, B. & Se eno, M. I. Co ical su ace-based analysis. I.
Segmen a ion and su ace econs uc ion. Neu oimage 9, 179–194 (1999).
55. Fischl, B. & Dale, A. M. Measu ing he hickness o he human ce eb al co ex
om magne ic esonance images. P oc. Na l Acad. Sci. USA 97, 11050–11055
(2000).
56. Be hlehem, R. A. I., Rome o-Ga cia, R., Mak, E., Bullmo e, E. T. & Ba on-
Cohen, S. S uc u al co a iance ne wo ks in child en wi h au ism o ADHD.
Ce eb. Co ex 27, 4267–4276 (2017).
57. Whi ake , K. J. e al. Adolescence is associa ed wi h genomically pa e ned
consolida ion o he hubs o he human b ain connec ome. P oc. Na l Acad.
Sci. USA 113, 9105–9110 (2016).
58. Glasse , M. F. e al. A mul i-modal pa cella ion o human ce eb al co ex.
Na u e 536, 171–178 (2016).
59. Desikan, R. S. e al. An au oma ed labeling sys em o subdi iding he human
ce eb al co ex on MRI scans in o gy al based egions o in e es . Neu oimage
31, 968–980 (2006).
60. Benjamin, D. J. e al. Redefine s a is ical significance. Na . Hum. Beha . 2,
6–10 (2018).
61. Ho man, G. E. & Schad , E. E. a iancePa i ion: in e p e ing d i e s o
a ia ion in complex gene exp ession s udies. BMC Bioin o ma. 17, 483
(2016).
62. Benjamini, Y. & Hochbe g, Y. Con olling he alse disco e y a e: a p ac ical
and powe ul app oach o mul iple es ing. J. R. S a . Soc. Se . B 57, 289–300
(1995).
63. Lai, M.-C., Lomba do, M. V., Auyeung, B., Chak aba i, B. & Ba on-Cohen, S.
Sex/gende di e ences and au ism: se ing he scene o u u e esea ch. J. Am.
Acad. Child Adolesc. Psychia y 54, 283–297 (2015).
64. Lai, M.-C., Ba on-Cohen, S. & Buxbaum, J. D. Unde s anding au ism in he
ligh o sex/gende . Mol. Au ism 6, 24 (2015).
65. Kaczku kin, A. N., Raznahan, A. & Sa e hwai e, T. D. Sex di e ences in he
de eloping b ain: insigh s om mul imodal neu oimaging.
Neu opsychopha macology 44,71–85 (2019).
66. an de Maa en, L. Accele a ing -SNE using ee-based algo i hms. J. Mach.
Lea n. Res. 15, 3221–3245 (2014).
COMMUNICATIONS BIOLOGY | h ps://doi.o g/10.1038/s42003-020-01212-9 ARTICLE
COMMUNICATIONS BIOLOGY | (2020) 3:486 | h ps://doi.o g/10.1038/s42003-020-01212-9 | www.na u e.com/commsbio 9