Causal mapping of emotion networks in the human brain: Framework and initial findings
Introduction
How do networks of brain structures generate human emotions? Affective neuroscience has documented a wealth of data, primarily from activations observed in neuroimaging studies in response to emotional stimuli. This has provided us with an inventory of brain structures that participate in emotions, but little knowledge of their precise causal role. Studies in humans with direct electrical stimulation of structures such as the amygdala have shown causal links between brain regions and emotional responses, but these additional findings still leave us with scant knowledge of how emotions are implemented at the network level in the brain. The question is pressing for translational reasons as well. Deep-brain stimulation is being explored for a large number of neurological and psychiatric diseases, but with quite variable success. There are clear case studies of remarkable amelioration of depression, for instance—but only in some cases, limiting the generalizability of the results (Kennedy and al, 2011; Mayberg et al., 2005).
We think of emotions as functional, central brain states defined by their cause-and-effect relationships with other brain processes, and with stimuli and behaviors. Which stimuli reliably cause emotions? How do emotions in turn cause behavioral responses? And -- the topic of this paper -- how do different brain regions interact with one another during emotion processing? The basic problem can be sketched in relation to the amygdala as schematized in Fig. 1. The amygdala is activated by threat-related stimuli, lesions of the amygdala impair threat-related responses and (in humans) aspects of the experience of fear, and stimulation of the amygdala produces defensive behaviors (very roughly). Nobody nowadays would conclude that “fear is in the amygdala”. Instead, the amygdala helps to orchestrate the many different causal effects of a fear state. To understand these effects we need to map the causal relations between the amygdala and other brain regions, through which such effects are mediated. We know almost nothing about these causal relations in the human brain.
Studies in animals have begun to dissect the circuits responsible for processing emotion, and of course offer methodological tools that are unavailable in humans. For instance, experimental manipulation of brain activity in rodents and monkeys has provided insights into the causal roles of particular circuits, such as the extended amygdala (Amaral and Adolphs, 2016, Shackman and Fox, 2016) and the hypothalamus (Lin et al., 2011). Behavioral dependent measures, while they need to be interpreted carefully, have given us strong evidence for how specific neuronal populations can cause specific emotional behaviors related to fear and aggression. One main limitation with these animal studies has been achieving a whole-brain field-of-view. Although specific circuits can be manipulated, e.g. through optogenetic or chemogenetic activation, the downstream effects are typically measured in only a very small subset of brain regions. One exciting future combination of methods is concurrent optogenetic stimulation with whole-brain fMRI (Lee et al., 2010, Liang et al., 2015), or with ultrasound imaging. However, the homology to human emotions remains a major limitation (Adolphs and Anderson, 2018).
Elucidating the causal networks that underlie emotion processing is one of the most important but also most difficult challenges faced by affective neuroscience. It is important because only an account at the level of causal mechanisms can really explain brain processing, and because only such an account can yield insights that allow us to manipulate brain function (for instance, with interventions aimed to treat mood disorders). Yet it is difficult because most of the data from the human brain are correlational in nature, making it unclear how to infer causality from typical neuroimaging and electrophysiological studies. Here we demonstrate the promise of a new technique – concurrent electrical stimulation and fMRI – and a new method in causal discovery – the fast greedy equivalence search – to obtain large-scale causal models that describe how different brain regions interact. We begin by briefly reviewing some of the findings from affective neuroscience, with an emphasis on the amygdala, and then outline the logic of causal discovery, before presenting our approach and pilot data to support it.
Section snippets
Emotion and the amygdala
Data from lesion studies and fMRI in humans, and from a range of approaches in animals, consistently implicate the amygdala (Fig. 2), the medial prefrontal cortex, the insula, the hypothalamus, and the periaqueductal gray in emotions. These structures function as components of considerably more distributed systems, and attempts to localize particular emotions (fear, sadness, etc.) to any one of these structures have been largely unsuccessful (Lindquist et al., 2012), even though specific
Causal discovery
The range of methods for analyzing brain function at the network level vary from essentially descriptive (such as looking at correlations between regions, i.e., standard functional connectivity approaches) to methods for inferring parameters of pre-specified models (such as dynamic causal modeling (DCM) (Friston, 2011; Friston et al., 2013)). All of these have tradeoffs: standard “functional connectivity” from correlations does not provide a causal model; DCM is limited by our knowledge of the
Methods
Our approach is roughly hierarchical in nature (Fig. 4) and consists of three main steps. These three steps are based on the analysis of existing resting-state data from large databases, comparison to resting-state data from individual neurosurgical patients, and comparison with the patient's graph using direct electrical stimulation-fMRI data from that patient. Future extensions beyond the scope of the present paper include (in red): formally integrating information from the large datasets as
Parameter setting for causal discovery and reproducibility across datasets
We first determined the reproducibility of our causal discovery analysis by deriving causal graphs from high-quality, large-sample size, statistically independent datasets. We began by using the HCPs dataset, which maximizes sample size, cross-subject generalizability, and statistical independence of the datasets. Comparing across 11 independent datasets from the HCPs, we obtained graphs at 10 different sparsity settings (Fig. 6). As expected, increasing the sparsity parameter resulted in
Summary of findings
We outlined a workflow for discovering causal connections in the human brain, and provided initial validation, measures of reliability, and comparisons across datasets. We then demonstrated the application of this workflow to the connectivity of the amygdala, as a case study for the investigation of causal networks that subserve emotion processing. However, our approach is quite general and we intend it to be applicable to the investigation of any brain structure, not just the amygdala (and
Acknowledgments
Supported by NIMH grant 2P50MH094258, NINDS grant 1U01NS103780-01, and grant 542941 from the Simons Foundation Collaboration on the Global Brain to R.A.; and by NSF grant 1564330 to F.E.
Data sharing
The HCP and MCP datasets are publicly available. All other data and analyses are available from the authors upon request.
Author contributions
J. Dubois conducted most of the analyses and generated all data figures in the main text. H. Oya and M. Howard conducted the electrical stimulation-fMRI experiments, and H. Oya processed the resulting data and conducted GLM analyses. J.M. Tyszka provided help with analysis of fMRI data. F. Eberhardt conducted all causal discovery analyses. All authors discussed and planned
References (82)
- et al.
Case report: stimulation of the right amygdala induces transient changes in affective bias
Brain Stimul.
(2014) AFNI: what a long strange trip it's been
Neuroimage
(2012)- et al.
Building a science of individual differences from fMRI
Trends Cogn. Sci.
(2016) - et al.
The human amygdala and the induction and experience of fear
Curr. Biol.
(2011) Test-retest reliability of dynamic causal modeling for fMRI
Neuroimage
(2015)Regression DCM for fMRI
Neuroimage
(2017)- et al.
Analysing connectivity with Granger causality and dynamic causal modelling
Curr. Opin. Neurobiol.
(2013) - et al.
The minimal preprocessing pipelines for the Human Connectome Project
Neuroimage
(2013) - et al.
The Rhesus monkey connectome predicts disrupted functional networks resulting from pharmacogenetic inactivation of the amygdala
Neuron
(2016) Functional System and Areal Organization of a Highly Sampled Individual Human Brain
Neuron
(2015)
Mapping the functional network of medial prefrontal cortex by combining optogenetics and fMRI in awake rats
Neuroimage
Probing and regulating dysfunctional circuits using deep brain stimulation
Neuron
Deep brain stimulation for treatment-resistant depression
Neuron
Ventromedial prefrontal cortex is critical for the regulation of amygdala activity in humans
Biol. Psychiatry
Efficacy and safety of transcranial magnetic stimulation in the acute treatment of major depression: a multisite randomized controlled trial
Biol. Psychiatry
Mapping effective connectivity in the human brain with concurrent intracranial electrical stimulation and BOLD-fMRI
J. Neurosci. Methods
MSM: a new flexible framework for multimodal surface matching
Neuroimage
The future of fMRI connectivity
Neuroimage
Network modelling methods for fMRI
Neuroimage
Functional connectomics from resting-state fMRI
Trends Cogn. Sci.
The human connectome: origins and challenges
Neuroimage
The WU-Minn Human Connectome Project: an overview
Neuroimage
How should neuroscience study emotions? By distinguishing emotion states, concepts, and experiences
Soc. Cogn. Affect. Neurosci.
Reply to Barrett: affective neuroscience needs objective criteria for emotions
Soc. Cogn. Affect. Neurosci.
The Neurobiology of Emotion: A New Synthesis
Impaired recognition of emotion in facial expressions following bilateral damage to the human amygdala
Nature
Functionalism cannot save the classical view of emotion
Soc. Cogn. Affect. Neurosci.
The theory of constructed emotion: an active inference account of interoception and categorization
Soc. Cogn. Affect. Neurosci.
Failure to respond autonomically to anticipated future outcomes following damage to prefrontal cortex
Cereb. Cortex
Optimal structure identification with greedy search
J. Mach. Learn. Res.
Feeling emotions: subcortical and cortical brain activity during the experience of self-generated emotions
Nat. Neurosci.
The role of the amygdala in fear and anxiety
Annu. Rev. Neurosci.
Noninvasive brain stimulation: from physiology to network dynamics and back
Nat. Neurosci.
Breathing inhibited when seizures spread to the amygdala and upon amygdala stimulation
J. Neurosci.
Introduction to the foundations of causal discovery
Int. J. Data Sci. Anal.
On the number of experiments sufficient and in the worst case necessary to identify all causal relations among N variables
Fear and panic in humans with bilateral amygdala damage
Nat. Neurosci.
Statistical Methods: Experimental Design and Scientific Inference
The central extended amygdala in fear and anxiety: closing the gap between mechanistic and neuroimaging research
Neurosci. Lett.
Resting-state networks link invasive and noninvasive brain stimulation across diverse psychiatric and neurological diseases
PNAS
Cited by (27)
Alpha and theta oscillations during the cognitive reappraisal of aversive pictures: A spatio-temporal qEEG investigation
2023, International Journal of PsychophysiologySensing fear: fast and precise threat evaluation in human sensory cortex
2023, Trends in Cognitive SciencesThe resting-state causal human connectome is characterized by hub connectivity of executive and attentional networks
2022, NeuroImageCitation Excerpt :FGES proceeds in two stages, first adding edges until the BIC stops improving, then removing edges until the BIC stops improving (Ramsey et al., 2017). While FGES has not commonly been used for analysis of empirical neuroimaging data, this method was applied (in combination with direct stimulation) to test causal connectivity patterns of the amygdala (Dubois et al., 2020) with promising initial findings for mapping human emotion networks. We computed FGES with causal-cmd v1.2.0 (https://bd2kccd.github.io/docs/causal-cmd/) using default parameters (BIC penalized likelihood score, penalty discount = 1 corresponding to the classic BIC score).
Neuropsychologia special issue editorial: The neural basis of emotion
2020, NeuropsychologiaAssessing the Primacy of Human Amygdala-Inferotemporal Emotional Scene Discrimination with Rapid Whole-Brain fMRI
2019, NeuroscienceCitation Excerpt :Interestingly, however, the amplitude of the BOLD signal response to neutral scenes is quite consistent across FG and PHG (Fig. 2), while the response to emotional scenes falls either above (for FG) or below that level (for PHG). This pattern is not inconsistent with the results of a recent study in which intracranial stimulation of the human amygdala (interleaved with fMRI scanning) was shown to strongly inhibit BOLD signal in this medial inferotemporal region of the visual cortex (Dubois et al., 2017). While this will of course require explicit study, it may be that the increasing activity of the amygdala during emotional scene perception serves to selectively reduce activity in PHG.
Non-human contributions to personality neuroscience: From fish through primates - a concluding editorial overview
2024, Personality Neuroscience