Supplementary Material

Hyett MP, Parker GB, Guo CC, Zalesky A, Nguyen VT, Yuen T, Breakspear MJ. Scene unseen: disrupted neuronal adaptation in melancholia during emotional film viewing.

sTable 1. Symptoms and Signs Expressed by Melancholic (Mel1-Mel16) and Non-melancholic (NMel1-NMel16) Participants.

sTable 2. Demographic and Clinical Characteristics Across Melancholic, Non-melancholic, and Control Groups.

sTable 3. Prediction of Presence or Absence of Differing Drug Classes, Controlling for Clinical Group, from Interaction of Rest and Negative Film viewing Sub-Network Edge Weights.

sFigure1. a) Overall and b) Continuous Ratings of Emotional Valence for the Two Film Clips, “Bill Cosby” and “The Power of One”, Averaged across 18 Healthy Participants. Error Bars Signify Standard Error of the Mean.

sFigure2. Average Inter-Subject Correlations of Hidden Neural States (Observed = Red) and Corresponding Null Distributions. Left = Strong Effect. Middle = Marginally Significant. Right = Non-significant Examples.

sFigure 3. Intra-Class Correlations of Hidden Neural States (Observed = Red) and Corresponding Null Distributions. Left, Middle and Right Panels Derived from same Condition/Group/Mode as sFigure 2.

sFigure 4. Average Correlations of BOLD (Observed = Red) and Corresponding Null Distributions. Left = Strong Effect. Left, Middle and Right Panels Derived from same Condition/Group/Mode as sFigure 2.

sFigure 5. Group comparisons of rank-ordered distributions of all 64 edge weights across resting state and neutral film viewing. Left column shows melancholic versus healthy controls: Right column shows melancholia versus non-melancholic MDD.

Appendix I. Methods.

1

Essential Symptoms (Both Required) Specifiers (Five of Nine Required)

Patient

Psychomotor Disturbance

Distinct Anhedonia

Concentration Impairment

Mood Non-reactivity

Anergia

Diurnal Mood Variation

Appetite /Weight Loss

Early Morning Wakening

Mel1 + + + + + - - -Mel2 + + + + + + - +Mel3 + + + + + + - +Mel4 + + + + + + - -Mel5 + + + + + + - -Mel6 + + + + + + - -Mel7 + + + + + - + +Mel8 + + + + + - + +Mel9 + + - - + + - -Mel10 + + + + + - + +Mel11 + + + + + + - +Mel12 + + + - + + - -Mel13 + + + - + + - +Mel14 + + + + + + - -Mel15 + + + - + - + -Mel16 + + + + + + + +NMel1 - - - - - - - -NMel2 - - - - - - - -NMel3 - - - - - - - -NMel4 - - - - - - - -NMel5 - + - - - - - -NMel6 - - - - - - - -NMel7 - + - - - - - -NMel8 - - - - - - - -NMel9 - - - - - - - -NMel10 + + + - + - + -NMel11 - - - - - - - -NMel12 + - - - - - - -NMel13 - - - - - - - -NMel14 - + - - - - - -NMel15 + + + - + + - -NMel16 - - - - - - - -

sTable 1. Symptoms and Signs Expressed by Melancholic (Mel1-Mel16) and Non-melancholic (NMel1-NMel16) Participants‘+’ Indicates presence of symptom/sign; ‘-’ Indicates absence of symptom/sign.

sTable 2. Demographic and Clinical Characteristics Across Melancholic, Non-melancholic, and Control GroupsGroup Group Comparisona

Test VariableMelancholic

Non-melancholic Control

Melancholic vsNon-melancholic

P value Melancholic vs

Control

P value

Non-melancholic vs

ControlAge, mean (SD) 38.00 (9.94) 40.44 (10.73) 43.75 (14.10) t = -0.68 .51 t = -1.33 .19 t = -0.75Female sex, No (%) 8 (50.0) 10 (62.5) 9 (56.3) χ2 = 0.51 .48 χ2 = 0.13 .72 χ2 = 0.13Years of education, mean (SD) 14.81 (3.31) 15.88 (2.44) 17.44 (3.58) t = -1.03 .31 t = -2.15 .04 t = -1.44

Estimated IQ, mean 107.93 (12.40) 108.19 (9.96) 117.94 (7.55) t = -0.06 .95 t = -2.69 .01 t = -3.12

QIDS-SR, mean 16.69 (4.22) 15.06 (4.10) 1.19 (1.51) t = 1.10 .28 t = 13.82 <.001 t = 12.68

STAI-State, mean 49.73 (16.18) 46.25 (12.37) 25.44 (6.52) t = 0.68 .50 t = 5.42 <.001 t = 5.95

STAI-Trait, mean 55.00 (13.33) 62.88 (8.43) 31.19 (6.45) t = -1.98 .58 t = 6.27 <.001 t = 11.94

2

GAF, mean (SD) 58.13 (7.04) 69.38 (6.29) 95.00 (0.00) t = -4.77 <.001 t = -20.95 <.001 t = -16.29CORE, mean (SD)

Noninteractiveness 3.44 (3.10) 0.75 (1.69) NA t = 3.05 .005 NA NARetardation 4.88 (3.40) 1.06 (2.41) NA t = 3.66 <.001 NA NAAgitation 0.69 (0.00) 0.00 (0.00) NA t = 2.55 .02 NA NATotal 9.00 (6.12) 1.81 (4.02) NA t = 3.93 <.001 NA NA

Current medications, No. (%)

No medication 1 (6.3) 5 (31.3) NA χ2 = 3.28 .07 NA NASSRI 2 (12.5) 8 (50.0) NA χ2 = 5.24 .02 NA NAAny medication other than SSRI 13 (81.3) 5 (31.3) NA χ2 = 8.13 .004 NA NA

Dual-action antidepressantb 8 (50.0) 5 (31.3) NA χ2 = 1.17 .28 NA NA

Tricyclic or monoamine oxidase inhibitor

4 (25.0) 2 (12.5) NA χ2 = 0.82.36

NA NA

Mood stabiliserc 1 (6.3) 2 (12.5) NA χ2 = 2.13 .34 NA NAAntipsychotic 4 (25.0) 0 NA χ2 = 4.57 .03 NA NA

Abbreviations: GAF, Global Assessment of Functioning; NA, not applicable; QIDS-SR, Quick Inventory of Depressive Symptomatology; SSRI, selective serotonin reuptake inhibitor; STAI State-Trait Anxiety Inventory.Uncorrected P values for between-group comparisons; differences significant at P < .05. b Serotonin noradrenaline reuptake inhibitor. c Lithium or valproate.

3

Demographic and Clinical Characteristics

The melancholic, non-melancholic and control groups did not differ significantly by age or

gender (see sTable 2). Age ranges for the groups were 20-52 (melancholic), 21-56 (non-

melancholic), 22-75 (controls) – with two controls aged over 60. Two non-melancholic

participants and one healthy control participant were left-handed. The control group reported

more years of education compared to the melancholic group (t = -2.15, p = 0.04), and their

estimated IQ was higher than both melancholic (t = -2.69, p = 0.01) and non-melancholic

groups (t = -3.12, p = 0.004). The depressed groups did not differ by years of education,

estimated IQ, depression severity or state and trait anxiety scores. Consistent with the

diagnostic primacy of psychomotor disturbance, the melancholic group had higher scores

compared to the non-melancholic group on all CORE sub-scales (noninteractiveness: t =

3.05, p = 0.005; retardation: t = 3.66, p < 0.001; agitation: t = 2.55, p = 0.02) and higher total

CORE scores (t = 3.93, p < 0.001). All groups differed on the GAF with the melancholic

group having the most severe functional impairment, followed by the non-melancholic and

then the control group participants. More non-melancholic participants reported receiving a

selective serotonin reuptake inhibitor (SSRI) antidepressant drug compared with the

melancholic participants (χ2 = 5.24, p = .02); a higher proportion of melancholic participants

were receiving non-SSRI medications (χ2 = 8.13, p = .004). More melancholic participants

reported being on antipsychotic medication (χ2 = 4.57, p = .03).

4

sTable 3. Prediction of Presence or Absence of Differing Drug Classes, Controlling for

Clinical Group, from Interaction of Rest and Negative Film viewing Sub-Network Edge

Weights

Dependent Variables

Predictor Variables SSRI (Yes/No) Non-SSRI Drug (Yes/No)

Rest * Negative Film Viewing†

Exp (β) Wald Sig. Exp (β) Wald Sig.

0.000 1.331 0.249 0.000 1.261 0.262

† Controlling for clinical group (melancholic/non-melancholic)

We used logistic regression to test whether differing medication classes confounded our

observed sub-network interaction differences. Clinical participants were divided into two

subsets: those in receipt of an SSRI medication, versus those not on any such medication; and

those on any other non-SSRI medication (eg, antipsychotics, mood stabilisers, all broad-

action antidepressants) compared to those who were not taking non-SSRI medication. We

controlled for clinical group and showed that sub-network scores, averaged across rest and

negative film viewing, were not predictive of differing medication classes.

Emotion Ratings During Emotional Film viewing

An independent cohort of 18 healthy participants was recruited to provide emotion ratings of

the stand up comedy (“Bill Cosby – Himself”) and The Power of One films. While viewing

each film, participants provided continuous ratings of their emotion using rating software

custom-built in LabView. They were instructed to continuously report their emotion by

moving a computer mouse as they viewed the film. Participants were required to move the

mouse all the way to the left if they felt completely sad, depressed, disgusted or unpleasant;

and move the mouse all the way to the right if they felt completely happy, joyful and pleased.

A vertical bar, indicating their current rating (between -1 and 1), provided visual feedback.

5

Negative ratings corresponded to values towards -1, whilst positive ratings corresponded to

values towards +1. In addition, participants provided an overall rating of the film

immediately after viewing. The order of film presentation was counterbalanced between

participants. Overall, valence ratings of the films were consistent with the choice of valence

labels in the current study (eg, ‘positive’/ Bill Cosby and ‘negative’/The Power of One).

sFigure 1. a) Overall and b) Continuous Ratings of Emotional Valence for the Two Film

Clips, “Bill Cosby” and “The Power of One”, Averaged across 18 Healthy Participants. Error

Bars Signify Standard Error of the Mean

Inter-subject correlations of BOLD fluctuations, and comparison of inferred (hidden)

neural states and BOLD.

Consistent neuronal responses to film stimuli between participants is a hallmark finding in

the use of dynamic stimuli during functional imaging experiments. Previous research has

imputed these responses from direct observation of the blood-oxygen-level-dependent

(BOLD) signal. To validate the inversion of dynamic causal models (DCMs) during film

6

viewing, we examined inter-subject correlations (ISCs) of the inferred neuronal states

following model inversion, by calculating the average pair-wise correlations between all

participants (giving ISCs). The statistical significance of these neuronal ISCs was tested

using a permutation approach. In particular, for any choice of mode (auditory, visual etc.), the

time series from each subject was cycled forward in time by independent random integer

increment (modulus N=181). This ensures that correlations within each time series are

preserved, while those between subjects are destroyed. Average pair-wise correlations from

an ensemble of 1000 of these surrogate data were used to represent the null distribution that

the measured values represent trivial effects arising from serial auto-correlations and finite

sample length. Family-wise control for multiple comparisons (across all possible pairs) was

achieved using false discovery rate (FDR) correction ( = 0.0055). Examples of strong,

marginal and null observations are provided in sFigure 2.

sFigure2. Average Inter-Subject Correlations of Hidden Neural States (Observed = Red) and

Corresponding Null Distributions. Left = Strong Effect. Middle = Marginally Significant.

Right = Non-significant Examples.

We used the mean correlation coefficient, averaged over all pairs of subjects because

this measure has been used widely in the analysis of film viewing fMRI data. However, as

subjects are included more than once in all possible pair combinations, there is some

7

redundancy in this approach. We thus also used an arguably more principled measure,

namely the intra-class correlation. Corresponding results for the same three modes are shown

in sFigure 3. The results are highly consistent with the averaged correlation coefficients.

sFigure 3. Intra-Class Correlations of Hidden Neural States (Observed = Red) and

Corresponding Null Distributions. Left, Middle and Right Panels Derived from same

Condition/Group/Mode as sFigure 2.

To validate the novel approach of inverting DCMs from film viewing data, we also

compared the ISCs of the inferred neuronal states to those of the observed BOLD signals,

within and between each of the three participant groups. Corresponding results for the same

three modes are shown in sFigure 4. In general, the ISCs of the raw BOLD signals are

stronger than those of the hidden neural states. However, the statistical significance of the

two data sets are strongly consistent.

8

sFigure 4. Average Correlations of BOLD (Observed = Red) and Corresponding Null

Distributions. Left = Strong Effect. Left, Middle and Right Panels Derived from same

Condition/Group/Mode as sFigure 2.

The Influence of Neutral Film Viewing on Network Parameters

We examined effective connectivity strengths across all 64 edges during neutral film

viewing. sFigure5 compares the distribution of edge weights between melancholic and

control, and melancholic and non-melancholic groups, using the resting state condition as a

reference. In keeping with the other movie viewing conditions, edge strengths were skewed

towards the stronger tails (generally in the positive direction) in the melancholic group. The

distribution of edge strengths in the non-melancholic group during neutral film viewing was

between those of the melancholic and control groups.

Using the same approach as for the positive and negative film viewing conditions (see “DCM

Analysis of Naturalistic Film Viewing” of main text), we assessed for group differences in

sub-networks of directed edge weights between resting state and neutral film viewing using

the NBS. No significant sub-networks were identified between rest and neutral conditions for

any of the group contrasts using the same threshold (4.25) that was used to identify the

effects shown in Figure 4.

9

We next analyzed whether there was an interaction between melancholic and control groups,

and condition (rest and neutral film viewing) using the same sub-network of edges as in

Figure 4. Univariate analysis of variance revealed a significant group x condition interaction

(F = 4.91, p = 0.031). Sub-network strengths increased from rest to neutral film viewing in

those with melancholia (means of 0.033 and 0.075 respectively), and decreased in controls

(means of 0.084 and 0.048 for rest and neutral film viewing respectively). The mean

differences between rest and neutral film viewing conditions were thus not as large as the

difference between rest and negative film viewing and hence do not survive family wise error

correction. Nonetheless the direction of change, whilst muted in size, was consistent with that

in the emotionally salient movies. In other words, the presence of emotionally salient content

in the films was a more effective probe in eliciting between group effects.

10

sFigure 5. Group comparisons of rank-ordered distributions of all 64 edge weights across resting state and neutral film viewing. Left column

shows melancholic versus healthy controls: Right column shows melancholia versus non-melancholic MDD.

11

Appendix I. Methods

Detailed Diagnostic Approach to Depression Sub-typing

Clinical diagnoses of melancholic or non-melancholic depression were made by psychiatrists

weighting previously detailed criteria.1,2 For a diagnosis of melancholia, two compulsory

criteria were required (see sTable 1): (A) psychomotor disturbance (expressed as motor

slowing and/or agitation); and (B) an anhedonic mood state. In addition, five of the following

nine clinical features were required (and met) in all assigned melancholic patients: (1)

concentration and/or decision making impairment; (2) nonreactive affect; (3) distinct anergia;

(4) diurnal mood variation – being worse in the morning; (5) appetite and/or weight loss; (6)

early morning wakening; (7) no preceding stressors accounting for the depth of the

depressive episode; (8) previous good response to adequate antidepressant therapy; and (9)

normal personality functioning. Whilst respecting the DSM diagnostic approach to

melancholia, these have been customised to take into account criteria that have historically

characterised melancholia.1,2 sTable 1 shows specific criteria for each patient, as

crosschecked against their clinical assessment material (ie, clinical notes, assessment letters

and referral material to the Black Dog Institute Depression Clinic).

fMRI Image Acquisition

All participants underwent a 6 ¼-minute resting state fMRI scan (186 volumes) and were

instructed to keep their eyes shut for the duration of the scan. Resting state fMRI was

acquired prior to three separate fMRI sequences during which participant’s viewed positive

(“Bill Cosy – Himself”), negative (“The Power of One”) and neutral films (landscape

footage). The presentation order of the films was pseudo-randomly counterbalanced across

participants. All participants explicitly reported remaining awake across the scanning

sequences. Scanning was conducted using a Philips 3.0-T scanner (Philips Medical Systems;

12

Best, Netherlands). Functional data were acquired using T2*-weighted gradient echo-planar

sequences (33 axial slices; repetition time/echo time: 2000/30 msec; 76° flip angle;

reconstruction matrix size: 128 × 128; field of view (anterior-posterior): 240 mm; voxel size:

3.0 × 3.0 × 3.0 mm; no gap).

Data Preprocessing

Resting state and film viewing fMRI images were preprocessed using statistical parametric

mapping (SPM8) software (http://www.fil.ion.ucl.ac.uk/spm/).3 For each participant, each

image was realigned to the first acquired session-specific image, normalised (unwarped) to

standard Montreal Neurological Institute (MNI) space and smoothed with a full-width half-

maximum (FWHM) kernel of 4 mm. All preprocessed functional data (resting state and film

viewing) were then used as inputs for probabilistic concatenated independent component

analysis (ICA) using the MELODIC (Multivariate Exploratory Linear Decomposition into

Independent Components) toolbox in the FMRIB Software Library (FSL)

(http://www.fmrib.ox.ac.uk/fsl/).4 For the ICA, non-brain voxels were masked with voxel-

wise demeaning of the data and normalisation of the voxel-wise variance. Preprocessed data

were next whitened and projected into a 70-dimensional subspace using Principle

Components Analysis, providing for a reasonably fine-grained decomposition of functionally

relevant brain regions.5 These whitened observations were decomposed into sets of vectors

that describe signal variation across the temporal domain (giving time courses), the

session/subject domain, and across the spatial domain (giving spatial maps). This was

implemented through a non-Gaussian spatial source distribution using a fixed-point iteration

technique.6 Estimated component maps were divided by the standard deviation of the residual

noise, with a threshold of 0.5 set (the probability that needed to be exceeded by a voxel to be

13

considered ‘active’ in the component of interest) by fitting a mixture model to the histogram

of intensity values.4

Node Selection

We selected eight ‘modes’ from the group-level spatial maps as generated by ICA, with the

aim of capturing emotional, cognitive and perceptual systems relating to attention and

interoception. The components selected were: (A) Auditory (AUD); (B) Default mode

network (DMN); (C) Executive control (EXC); (D) Left insula (L-INS); (E) Right insula (R-

INS); (F) Left frontoparietal attention (LFP); (G) Medial visual pole (MVP); and (H) Right

frontoparietal attention (RFP; see Figure 1). All components were checked for

correspondence with previously identified cognitive and sensory networks using spatial

cross-correlation,5 except for the L-INS and R-INS modes. These components were identified

from the ICA maps by, i) first obtaining the centre coordinates of the bilateral anterior insula

using PickAtlas, and then, ii) using these coordinates to identify the spatial maps with the

most specificity from the 70 components of the ICA. These maps were then used in

specifying and estimating stochastic dynamic causal models (sDCMs) across all conditions

for all study participants.

DCM Specification

Dynamic causal modelling infers effective connectivity amongst neuronal populations by

combining dynamic models of neuronal states and detailed biophysical models of

haemodynamics. Traditionally, DCM has been employed to provide generative models of

task-related data, where stimulus or task manipulations are introduced as known inputs to

regions identified through use of the general linear model.3 With stochastic DCM (sDCM),

the system perturbations are modelled as unknown system fluctuations arising

14

endogenously.7,8 Broadly speaking, sDCM otherwise proceeds in a similar vein to classic

DCM, namely: 1. The user specifies a model (or models) through the choice of nodes and

inputs; 2. The empirical data are introduced as time series of each node; 3. The model

evidence (the probability of observing the data given the model) is maximised using a

variational scheme to minimise an objective function (the free energy). This also yields

posterior model parameter estimates as well as estimates of the unknown state fluctuations;8,9

and 4. If more than one model is specified, model comparison is performed using the

evidence for each model. This Bayesian model evidence (aka marginal likelihood) penalises

the accuracy of each model by a measure of its complexity.10 In the present setting, we

implemented a single, fully connected bilinear DCM with unknown fluctuations at every

node. No external inputs were specified. This model was estimated in all participants.

For each of the components, for each condition, the peak activation voxel was

identified, with MNI coordinates of these peaks used to define a regionally specific voxel of

interest to allow initial estimation of the DCMs. For visualisation purposes (eg, see Figure 1),

a 6 mm sphere was used to represent the spatial location of the peak weight of the

corresponding ICA mode. Dual regression was used to extract participant-specific time series

from condition-specific, group-level spatial maps (representing an average of the voxels

within each map, weighted by their relative expression in that map), with the corresponding

component time series used as inputs for the sDCMs.8,9 In specifying the DCMs, no inputs

were selected for the first and second levels, and a fully connected model was chosen for the

search space.

15

References

1. Parker G, Fink M, Shorter E, et al. Issues for DSM-5: whither melancholia? The case

for its classification as a distinct mood disorder. Am J Psychiatry. 2010;167(7):745-

747.

2. Taylor MA, Fink M. Melancholia: The Diagnosis, Pathophysiology and Treatment of

Depressive Illness. New York, NY: Cambridge University Press; 2006.

3. Friston KJ, Holmes AP, Worsley KJ, Poline J-P, Frith CD, Frackowiak RSJ.

Statistical parametric maps in functional imaging: A general linear approach. Hum

Brain Mapp. 1995;2(4):189-210.

4. Beckmann CF, Smith SM. Probabilistic independent component analysis for

functional magnetic resonance imaging. IEEE Trans Med Imaging. 2004;23(2):137-

152.

5. Smith SM, Fox PT, Miller KL, et al. Correspondence of the brain's functional

architecture during activation and rest. Proc Natl Acad Sci U S A.

2009;106(31):13040-13045.

6. Hyvarinen A. Fast and robust fixed-point algorithms for independent component

analysis. IEEE Trans Neural Netw. 1999;10(3):626-634.

7. Friston KJ, Li B, Daunizeau J, Stephan KE. Network discovery with DCM.

NeuroImage. 2011;56(3):1202-1221.

8. Li B, Daunizeau J, Stephan KE, Penny W, Hu D, Friston K. Generalised filtering and

stochastic DCM for fMRI. NeuroImage. 2011;58(2):442-457.

9. Daunizeau J, David O, Stephan KE. Dynamic causal modelling: a critical review of

the biophysical and statistical foundations. NeuroImage. 2011;58(2):312-322.

10. Marreiros AC, Stephan KE, Friston KJ. Dynamic Causal Modelling. Scholarpedia.

2010;5(7):9568.

16