Information

How well does the NEF capture neuronal heterogeneity?

How well does the NEF capture neuronal heterogeneity?

From what I understand of the Neurogical Engineering Framework (NEF), groups of neurons are used to compute functions. However, I'm not clear if these calculations take into account neurons of different shapes/sizes and their attributes/uses in the brain. How well does the NEF capture heterogeneity and how is it used?


To review how neurons encode information, please check out these class notes review encoding.

In those notes, you'll notice the intercept $J_{bias}$ and the maximum firing rate $alpha$ are randomly selected when encoding functions in large populations of neurons. These variations can account for heterogeneity in attributes of neurons.


4 DISCUSSION

We used machine learning to identify the cognitive profiles within a large heterogeneous sample of children with learning-related problems. These profiles were represented as topographical maps. None of the known characteristics of the children (e.g., diagnosis or referral route) were predictive of the cognitive profiles identified by the machine learning. To highlight the cognitive profiles that exist within the dataset, we subsequently carved the topographical maps into four sections. The children that correspond to these four sections will necessarily have distinct cognitive profiles, but they could also be distinguished in terms of learning and behavioral scores, and patterns of brain organization. The four groups cut across any traditional diagnostic groups that existed within the data.

More than half of the sample fell into two extreme groups, one with age-appropriate cognitive abilities and the other with widespread cognitive deficits that were at least one standard deviation below age-typical levels across all tasks. There was no evidence that children with age-expected scores on the cognitive measures had learning difficulties. Their performance was in the age-typical range across all measures of learning and their structural communication skills were rated as normal for their age. But we should be very cautious in regarding these children as typically developing they have been referred by professionals in children's services, and as a group they have elevated behavioral difficulties. For this reason in our neuroimaging analysis we used an additional external comparison group.

The learning scores of the broad deficit group place them within the bottom 5% of the population on measures of spelling, reading, and maths, and they were rated as having difficulties in both structural and pragmatic aspects of communication. Generalized cognitive deficits therefore appear to constrain multiple aspects of learning. They also had behavioral problems related to executive function, although this was true for all four groups. Relative to both comparison groups, this group also had reduced structural connectivity in the left precentral gyrus, right inferior frontal gyrus, right lateral occipital cortex, and the left fusiform. These areas have been previously identified as playing a key role in multiple higher order cognitive skills. For example, the right inferior frontal gyrus is implicated in multiple different executive functions, most commonly measures of inhibitory control (Aron, Robbins, & Poldrack, 2014 ) the lateral occipital cortex has been found to be modulated by visual attention (Sprague & Serences, 2013 ) left premotor areas have been linked to language-related difficulties in both children and adults (Mayes, Reilly, & Morgan, 2015 Scott, McGettigan, & Eisner, 2009 ) and the fusiform gyrus has been suggested as a locus of immature processing of word forms in dyslexia (Tamboer, Vorst, Ghebreab, & Scholte, 2016 ). These general struggling learners are rarely studied, but our data suggest that they are common amongst those coming to the attention of children's specialist services. Their relative under-representation in studies of learning-related problems means that we have little understanding of the key underlying deficits, mechanisms or potential routes to effective intervention. It is also interesting to note that girls were disproportionately common in this group, relative to the sample as a whole or indeed relative to most studies of learning difficulties. Conversely very few girls appeared in the age-appropriate cognitive profile group. In short, the girls referred to the study tended to have more severe cognitive and learning difficulties. One possibility is that there is a gender bias in the reason for children coming to the attention of children's specialist services, with boys being identified more commonly for behavioral difficulties (which may be less closely tied to cognitive and learning profiles), whereas more severe cognitive or learning difficulties are needed for girls to come to the attention of specialists.

Two intermediate groups, both with fluid reasoning scores in the low-average range, were also identified. One intermediate group was characterized by problems on tasks requiring phonological processing, with performance around three quarters of a standard deviation below age-expected levels on measures of phonological awareness, and verbal short-term and working memory. These children had significant problems with structural aspects of communication, mirroring the well-documented link between phonological processing difficulties and specific difficulties with language (Bishop & Norbury, 2002 Bishop & Snowling, 2004 Ramus et al., 2013 ). However, the learning profile demonstrates equivalent and large deficits across measures of reading, spelling, and mathematics. Poor phonological processing is associated with both poor reading (Carroll & Snowling, 2004 Snowling, 1995 Wagner & Torgesen, 1987 ) and mathematical development (De Smedt, Taylor, Archibald, & Ansari, 2010 Hecht, Torgesen, Wagner, & Rashotte, 2001 Swanson & Sachse-Lee, 2001 ). A consistent finding within the field of learning difficulties is that phonological problems are linked selectively with reading. The majority of these findings come from studies that select poor readers, but this is not the same as demonstrating that phonological impairments will always result in selective reading difficulties. Our data suggest that children selected on the basis of phonological difficulties will actually have more widespread learning problems. Membership of the phonological deficit group was associated with reduced structural connectivity in the left precentral gyrus and rostral anterior cingulate, relative to both comparison groups. The precentral gyrus has been implicated in language processing and is thought to be involved in speech production and also decoding via articulatory simulation (Scott et al., 2009 ). This area has also been implicated in selective language impairment (Mayes et al., 2015 ). Furthermore, tracts of the perisylvian language network that connect temporal and frontal language areas deficits are passing the precentral gyrus and may be substantially contributing the connectomics differences. Differences in white matter properties of these tracts have been repeatedly implicated in language deficits (Rimrodt, Peterson, Denckla, Kaufmann, & Cutting, 2010 Roberts et al., 2014 ). This would also mirror the structural communication difficulties that these children demonstrate. Indeed, this is the only behavioral measure that aligns well with the cognitive profiles—children who perform poorly on phonological tasks are also rated as having significant structural language problems by their parents. Other behavioral measures of executive control do not align well with cognitive profiles.

The fourth group had a somewhat contrasting profile of cognitive deficit to the phonological deficit group. They were characterized by similar fluid IQ scores but had more pronounced difficulties in working memory. Their spatial short-term memory scores were over a standard deviation below age-expected levels, and half a standard deviation down on the verbal and spatial working memory measures. Their phonological abilities were less impaired, they were not rated as having the structural language difficulties reported for the phonological deficit group, and their neural profile was less homogenous. One possibility is that multiple different etiological routes can result in this profile of difficulties.

Despite contrasting cognitive and neural profiles, the learning profiles of the working memory and phonological deficit groups were nearly identical. This diverges strongly from a preceding literature that emphasizes a marked association between phonological difficulties and problems with literacy (Lyytinen et al., 2004 Snowling, Bishop, & Stothard, 2000 Tanaka et al., 2011 ), and an emerging literature that suggests strong associations between spatial short-term and working memory problems and numeracy difficulties (Bull et al., 2008 Raghubar, Barnes, & Hecht, 2010 Szucs et al., 2013 ). These previous studies all recruit on the basis of highly selective learning profiles (e.g., maths problems in the absence of reading difficulties) or diagnostic group, which will have overestimated the distinctiveness of these impairments within the general population of struggling learners.

Despite their utility, machine learning approaches to exploring cognitive profiles have limitations. The current combination of a multidimensional mapping method with a data-driven clustering algorithm suffers from the drawback that the number of groups within the data is underspecified. The mapping process is continuous, with no obvious boundaries, which makes it difficult to have a clear rationale about the formation of groups. Inevitably some children will sit close to a group boundary within the map. Our approach was to add clusters until the clusters did not differ on measures not included in the machine learning. This is how we arrived at four clusters. This is a relatively conservative approach, since different cognitive profiles could exist that genuinely have identical learning, behavioral, and neural correlates. Furthermore, we suspect that datasets with higher dimensionality, stemming from a more widespread battery of measures, could have greater success in identifying more widely differing cognitive profiles.

An alternative to machine learning is to use a network analysis with a community detection algorithm (e.g., Bathelt et al., 2018 Fair et al., 2012 ). An example of this approach applied to our data can be found in our Supplementary Materials section. This represents the children as nodes and the correlation between their profiles as edges. It is possible to use this approach to identify communities of clusters that maximize the correlation within cluster and the distinctiveness across clusters. This iterative process includes a quality of separation metric (Q) which the clustering algorithm is designed maximize. A major advantage of this approach is that no a priori assumptions about the number of clusters need to be made. However, there are also drawbacks to this alternative. The primary limitation is that a network analysis clusters children on the basis of a correlation matrix. As such it is blind to overall severity. The current sample contains a large number of children with relatively consistent poor scores across all cognitive measures and many children with stable age-appropriate scores. A network analysis would not be able to distinguish these two groups because the two profiles are highly correlated (this is indeed the case, see Supplementary Materials). The SOM uses Euclidean Distance as its primary means of locating children within the 2D topographical space, and as such is able represent both selective cognitive impairments and overall differences in severity. A further limitation is sample size. Whilst we included 530 children in the topographical mapping process, only 220 children were used in the structural neuroimaging comparison. This likely means that we only have sufficient power to detect the largest and most consistent group differences. More diffuse but equally important differences in whole brain connectome organization might exist, but a larger sample would be needed to identify them.

In summary, we used a machine learning approach that represents high-dimensional data as a 2D topography, to map the profiles of struggling learners. We combined this with a clustering algorithm to identify particular cognitive profiles represented within the map. Specifically, four profiles could be identified that comprise children with: (a) general and severe deficits, (b) age-appropriate performance, (c) working memory deficits, (d) phonological deficits. Furthermore, these data-driven groups are likely to align closely with underlying etiological mechanisms, as evidenced by differences in brain organization across two of the deficit groups, and provide the opportunity to devise interventions that more specifically target the cognitive difficulties faced by individuals with particular profiles.


ORIGINAL RESEARCH article

Cecilia U. D. Stenfors 1,2,3 * , Stephen C. Van Hedger 1 , Kathryn E. Schertz 1 , Francisco A. C. Meyer 1 , Karen E. L. Smith 1 , Greg J. Norman 1 , Stefan C. Bourrier 4 , James T. Enns 4 , Omid Kardan 1 , John Jonides 5 and Marc G. Berman 1 *
  • 1 Department of Psychology, University of Chicago, Chicago, IL, United States
  • 2 Department of Neurobiology, Care Science and Society, Aging Research Center, Karolinska Institute, Stockholm, Sweden
  • 3 Department of Psychology, Stockholm University, Stockholm, Sweden
  • 4 Department of Psychology, University of British Columbia, Vancouver, BC, Canada
  • 5 Department of Psychology, University of Michigan, Ann Arbor, MI, United States

Interactions with natural environments and nature-related stimuli have been found to be beneficial to cognitive performance, in particular on executive cognitive tasks with high demands on directed attention processes. However, results vary across different studies. The aim of the present paper was to evaluate the effects of nature vs. urban environments on cognitive performance across all of our published and new/unpublished studies testing the effects of different interactions with nature vs. urban/built control environments, on an executive-functioning test with high demands on directed attention—the backwards digit span (BDS) task. Specific aims in this study were to: (1) evaluate the effect of nature vs. urban environment interactions on BDS across different exposure types (e.g., real-world vs. artificial environments/stimuli) (2) disentangle the effects of testing order (i.e., effects caused by the order in which experimental conditions are administered) from the effects of the environment interactions, and (3) test the (mediating) role of affective changes on BDS performance. To this end, data from 13 experiments are presented, and pooled data-analyses are performed. Results from the pooled data-analyses (N = 528 participants) showed significant time-by-environment interactions with beneficial effects of nature compared to urban environments on BDS performance. There were also clear interactions with the order in which environment conditions were tested. Specifically, there were practice effects across environment conditions in first sessions. Importantly, after parceling out initial practice effects, the positive effects of nature compared to urban interactions on BDS performance were magnified. Changes in positive or negative affect did not mediate the beneficial effects of nature on BDS performance. These results are discussed in relation to the findings of other studies identified in the literature. Uncontrolled and confounding order effects (i.e., effects due to the order of experimental conditions, rather than the treatment conditions) may explain some of the inconsistent findings across studies in the literature on nature effects on cognitive performance. In all, these results highlight the robustness of the effects of natural environments on cognition, particularly when confounding order effects have been considered, and provide a more nuanced account of when a nature intervention will be most effective.


Functional Connectivity: Probing the Brain’s Astounding Complexity

One of the greatest scientific challenges for 21st century medicine is to illuminate the relationship between the brain and what we call “mind.” Psychiatrists want to know how we get from neurons and synapses to mental suffering. How do learning, development, and cognitive flexibility arise? What accounts for the uniqueness of every human self? What goes wrong in psychiatric disease? Within the psychiatric community, brain science has generally been regarded as being still too elementary to explain such complex phenomena.

In this article I describe how analysis of the brain’s intrinsic functional connectivity has become an important approach for expanding our understanding of the astounding complexity of the human brain. Utilizing this new paradigm, it is possible to explore questions that earlier seemed virtually unfathomable many of these are relevant, even pivotal, to psychiatry.

In news parlance, functional connectivity is a “rapidly developing scientific story.” And for psychiatrists, it is a story worth following.

What is functional connectivity?

Functional “imaging” measures physiological factors that are considered to be a gauge of neuronal functioning (such as changes in regional oxygen utilization) the accumulated data are then transformed into “human readable” images. Functional “connectivity” utilizes functional imaging data and analyzes the statistical associations between measurements of neurophysiological activity in 2 or more spatially remote areas of the brain. Functional connectivity studies are a mathematical, non-theoretical look at activity over the whole brain, in an attempt to discern in which areas the activity is either correlated or anti-correlated. For example, do areas A and X display increased metabolic activity when area D exhibits decreased metabolic activity?

Traditional functional connectivity methods do not tell us about the direction of connectivity-which region is influencing which. Functional connectivity also does not tell us whether 2 regions are simultaneously being influenced by a third. Nor does functional connectivity say anything about the way in which various brain areas might be structurally connected. Functional connectivity takes a purely statistical look at larger patterns of neurophysiologic activity that emerge from the brain’s hundreds of billions of neurons, reciprocally interacting over both short and long distances at microsecond speed.

Early functional connectivity research was designed to study the brain while an individual was performing a task or interacting with the environment-termed “psychophysiological interactions.” More recent work has shifted to a focus on the brain’s intrinsic neural activity while the subject is at “rest” or engaged in undirected thought.

The historical context

It was not until the early 1800s that modern scientific notions about the brain began to take precedence. 2 Franz Joseph Gall promoted the idea that various mental capacities were localized to different brain regions he also believed that the strength of these qualities could be measured by looking at protrusions in the skull overlying these areas. This notion was at the heart of phrenology, extremely popular in the 19th century.

Gall assumed that the brain was symmetrically organized. His categorization of what constituted fundamental, functional capacities of the brain drew on cultural values of his day, including vanity, guile, kindness, and pride. Although scientific advances have discredited many of these claims, Gall’s fundamental insight about localization was profound.

Localization was a powerful new idea at the time it exhibited the 2 fundamental attributes of a paradigm as first defined by the science historian Thomas S. Kuhn as “. . . sufficiently unprecedented to attract an enduring group of adherents away from competing modes of scientific activity. Simultaneously, it was sufficiently open-ended to leave all sorts of problems for the redefined group of practitioners to resolve.” 3

The localization paradigm opened new questions for legitimate scientific inquiry neurological investigators became the Lewis and Clark of neuro-anatomic territories. More than 200 years of explorations in neurology have been spent mapping the localization of various functional elements within the brain and simultaneously parsing brain activity into its most fundamental capabilities.

For example, until the mid-1950s, memory was considered to be a widely distributed, unitary function of the brain. Then, studies of patient H. M. elucidated a central role for the hippocampus in memory formation. In addition, episodic and procedural memory processes were differentiated.

While the concept of localization drove much of the work of scientific exploration during this time, there also were global, overarching theories of brain function. The eminent British neurologist John Hughlings Jackson posited that the more evolutionarily developed regions of the brain exerted control over primitive brain areas he articulated how disturbances in this organization were evident in disease states. Indeed, many scientists who are well known for their work on localization (Wernicke, Penfield) also were aware that the brain was extraordinarily complex and that the parts had to work together, like an astoundingly accomplished orchestra.

Structural connectivity

An important step in expanding the localization concept and focusing on how various parts of the brain communicated was the recognition of disconnection syndromes in the 1960s. 4 Disconnection syndromes did not result from lesions in the cortical gray matter itself but rather from interruptions in the white matter fibers and tracts running to and from cell bodies in gray matter-the lines of communication within the brain. If a lesion in white matter isolates or “disconnects” a crucial cortical region, this could lead to a clinical picture similar to one seen with a lesion in the cortical area itself. In other words, now neurologists were turning their attention to structural connectivity, the so-called wiring diagram of the brain.

Astoundingly, all of this work was accomplished by utilizing only clinical observation, scientific reasoning, and post-mortem findings. The only imaging available was pneumoencephalography, a painful technique that allowed clinicians to visualize the shape of the ventricles in the brain of a living patient by using x-rays after injecting air into the spinal column.

Structural brain imaging

Until computed tomography (CT) came into clinical use in 1973, it was not possible to visualize brain parenchyma during life. Not until 1994 was there a technique that spared radiation exposure-magnetic resonance imaging (MRI). CT, MRI, and other brain imaging approaches have allowed us to make remarkable advances in mapping human brain structures and in diagnosing human disease, but these modalities also have inherent limitations. An MRI scan is not a “picture” in the usual sense but rather a computer-generated readout of summary information into a digital image. One voxel (essentially a 3-dimensional pixel) in brain imaging contains information from approximately 1 cubic millimeter. Within 1 cubic millimeter of gray matter are perhaps a million neurons plus glia cells, blood vessels, and extracellular space. 5

In other words, none of these structural imaging techniques approaches the level of the individual neuron. Also, a structure-alone perspective of the human brain has other inherent limitations. For example, even if we take into account ongoing remodeling of neuronal connections (neuroplasticity), how can we account for the brain’s astounding moment-to-moment flexibility?

Functional brain imaging

Functional imaging techniques measure fluctuations over time in factors that gauge neuronal activity such as regional glucose utilization, blood flow, or oxygen consumption. The most widely used functional imaging techniques are positron emission tomography (PET), which utilizes radioactively tagged molecules, and functional MRI (fMRI), which differentiates molecules by their behavioral responses within a magnetic field.

One important functional imaging approach is based on the observation that, in a magnetic field, oxygenated blood behaves differently than de-oxygenated blood. This allows researchers to track fluctuations in blood-oxygen levels in human subjects in the scanner without the use of radiation. This blood-oxygen level dependent (BOLD) signal has become a major investigative tool.

Functional imaging techniques allow clinicians to evaluate patients who may have deficits in neuronal functioning even when structural deficits are not apparent. For example, early in the disease course of frontal dementias, patients may have decreased functioning in frontal regions before any changes are seen on structural imaging.

Functional imaging also has made it possible for investigators to pursue a myriad of fundamental questions. One line of such inquiry examines which parts of the brain are active as someone performs a carefully designed task. For instance, how is viewing a familiar face different from viewing an unfamiliar one? Which parts of the brain are involved in moral decision-making? Also, for a given task, are normal individuals functionally different from people who have specific psychiatric conditions?

Limitations of functional imaging

A serious obstacle for functional imaging research has been distinguishing a “task signal” of significance from background “noise.” The brain is metabolically very active, utilizing approximately 20% of body energy resources even though it represents only 2% of body mass. 5 Most of this energy utilization is from ongoing neuronal metabolism. When performing a demanding cognitive task, the brain’s energy utilization increases by less than 5%. Furthermore, differences that might exist between the normal and the study populations are even smaller.6 These factors make the task signal difficult to detect.

Also, individuals vary in their level of effort, degree of anxiety associations related to the task that come to consciousness, movement during the study, and so on. Although these brain-based activities are not the focus of the study, the metabolic activity they produce shows up in the scanner. Therefore, to amplify the task signal and also draw broad conclusions, a widely used approach to studying task-related questions has been to pool the findings from numerous individuals and average the results onto a standard anatomical brain atlas.

Along with these technical challenges, functional imaging studies also have a theoretical limitation built into their fundamental design. Task-based functional connectivity studies focus on the brain correlates of the task and assume the brain’s ever-present background neuronal metabolic activity is simply “noise.”

Resting state connectivity and its importance

In 2001, an important observation changed the field of functional imaging. Marcus Raichle compared PET and, later, fMRI BOLD signal findings from research subjects who had been engaged in task-based studies with those who were in control groups. Dr. Raichle’s laboratory routinely used the “rest condition” as a control rather than using, for example, a neutral task as a control for an emotional one.

At some point in our work, and I do not recall the motivation, I began to look at the resting state scans minus the task scans. What immediately caught my attention was the fact that regardless of the task under investigation, activity decreases were clearly present and almost always included the posterior cingulate and the adjacent precuneus. . . . Initially puzzled by the meaning of this observation, I began collecting examples from our work and placed them in a folder which I labelled [sic] MMPA for mystery medial parietal area. 7

Further analysis of data by Marcus Raichle and his colleagues led to the identification of a network of specific brain regions in which activity was anti-correlated with task-based activity no matter what the task was. This network was named the Default Mode Network (DMN) by Michael D. Greicius. 8 The work of Raichle and others was consistent with the first published report of intrinsic resting state functional connectivity by Biswal and colleagues in 1995 that “functionally related brain regions exhibited correlation of low frequency fluctuations in the resting state.” 9

The importance of these discoveries has been far-reaching

Consider that previously, in task-based functional imaging studies, the challenge had been to find the signal within the experimental background “noise.” Now it had become clear that this “noise” was data. The fluctuating BOLD signal could be mathematically mined as a source of information about the intrinsic functional organization of the brain. Moreover, the data could be obtained relatively easily by placing a person in a scanner and instructing him or her to “rest” or to visually fixate on 1 spot: this made it possible to study some patients who had difficulty cooperating with other protocols.

Moreover, using functional connectivity did not require patient averaging. Subjects could be studied individually, making it possible to compare different individuals or 1 individual at different times. These advantages made the prospect of utilizing functional connectivity as a clinical tool more viable.

For researchers, “. . . functional brain connectivity . . . [had] become one of the most influential concepts in modern cognitive neuroscience, especially given the current shift in emphasis from studies of functional segregation to studies of functional integration.” 10 We had long appreciated that detailing the synapse-to-synapse, structural organization of the brain would not capture the brain’s vast neuronal networks at work, operating as a dynamic system at speeds that would support the myriad manifestations of complex human behavior. Now “. . . task-free analysis of intrinsic connectivity networks may help elucidate the neural architectures that support fundamental aspects of human behavior.” 11

What neuroscience has learned from studies of functional connectivity

The DMN is only one of numerous large-scale, intrinsically synchronized, dynamic and interacting, functionally organized networks in the brain. These intrinsic functional networks and important nodes or hubs in those networks are consistent with synaptic maps of the brain. In other words, these networks do not violate our previous understanding of the anatomical organization of the brain into systems for motor behavior, perception, cognition, and so on.

The intrinsic functional networks can be found during cognitive tasks and in the rest condition, even when the systems for motor behavior, perception of various kinds, cognition, etc, are not being consciously engaged. Indeed, these networks persist during sedation, sleep, and under anesthesia.

While the intrinsic functional networks agree with earlier understanding of anatomical organization, the networks are not restricted to neuroanatomical regions with single synapse connections. The astounding degree of structural neural-network complexity in the brain likely explains how regions of the brain might be “functionally connected” even when their “structural connections” are not clear.

There is rapid coordination and interaction among the intrinsic brain networks and their hub regions. The brain is constantly switching connectivity patterns and reorganizing according to demands of the moment. Although the brain is a massively complex dynamic system, it can be studied by utilizing advanced imaging techniques and innovative mathematical and computational approaches. The importance of collaboration between experts in different fields cannot be overstated.

The most studied networks that relate to cognition are the Central Executive Network (CEN), a Salience Network (SN), and the DMN. The rapid interplay of these and perhaps other networks underpins behavioral changes that are based on the individual’s homeostatic needs, given that conditions (both internal and external) shift rapidly. The CEN is most active during cognitive tasks. The SN is activated in response to salient stimuli and plays a role in emotional processing and in switching from the DMN to the CEN. These networks are found in everyone however, there is individual variation in features such as the strength of connectivity within each network.

The most far-reaching question we posed was whether regional functional connectivity within the salience and executive-control networks in the task-free setting would correlate with subject attributes measured outside the scanner. In other words, do individual differences in intrinsic connectivity strength correlate with how one feels and thinks in daily life? 11

This idea has been used to look at a large variety of traits and conditions. Indeed, individual variations in the coherence of these intrinsic networks appear to correlate with patient capabilities (eg, fluid intelligence, stressor-associated anticipatory anxiety, executive task performance). 11,12 Should studies confirm that variations in functional connectivity are reliable indicators of specific behavioral variations outside the scanner, the potential clinical uses of this technology are staggering.

Theoretically, the mechanisms by which genetic variation, maturation, and experience affect an individual may be understood within this paradigm of intrinsic functional network organization in the brain. Early variations in neural development as well as experiences over time have an impact on synaptic strength and network functioning these individual variations, in turn, have far-reaching effects on information processing within these networks over a lifetime.

Functional networks are differentially affected in various disease states. This has been intensively studied in the neurodegenerative diseases. There is mounting evidence, for example, that the pattern of disruption found in the coherence of various intrinsic brain networks is different for Alzheimer disease than for behavioral variant frontotemporal dementia.

Many psychiatric conditions have also been studied using functional connectivity approaches, including autism spectrum disease, PTSD, ADHD, bipolar disorder, and others. 7 However, the heterogeneity of these disorders has made them difficult to study.

In MDD, there appear to be changes in connectivity within nodes of the anterior and posterior sub-networks of the DMN and altered connectivity between these areas and the SN as well as the CEN. The anterior sub-network of the DMN is believed to be involved in emotional and self-referential processing while the posterior sub-network of the DMN is more involved in memory and consciousness. Findings suggest that alterations in functional connectivity associated with MDD “. . . reflect a state of increased interaction between self-referential and emotional networks, and the dominance of negative self-referential over cognitive processing which corresponds to the clinical symptoms of depression.” 13

Some of the potential applications for functional connectivity in clinical situations include as an aid in diagnosing, measuring disease severity, providing prognostic information, or monitoring disease progress and treatment effectiveness. For psychiatry, functional connectivity studies may help to differentiate subtypes within heterogeneous populations for disorders such as depression or schizophrenia.

Some neuropsychiatric “mysteries” are being solved through functional connectivity studies. For example, how do single lesions (as from a stroke) in different parts of the brain produce similar, complex neuropsychiatric syndromes in different individuals? Darby and colleagues 14 were interested in this question as applied to the emergence of delusional misidentification syndromes after single lesions. Capgras syndrome is perhaps the best known of the delusional misidentification syndromes, exemplified by: “You look like my wife, but I know you are not my real wife you are an imposter my real wife is somewhere else.”

Darby found that the disparate lesions that had led to delusional misidentification were functionally connected to brain regions involved in familiarity assessment and in belief evaluation. This finding supports earlier theories about what goes wrong in delusional misidentification: first, the patient recognizes the individual or place but fails to experience that person or place as familiar and second, there is a failure in “belief evaluation,” namely a failure to realize that it defies logic to believe that this is anything but the real person or place.

Extremely promising work by Emily S. Finn has shown that functional connectivity profiles are specific enough to distinguish individuals, including across different sessions in the scanner and across task and resting states. 12 This identification of individuals is referred to as functional connectome fingerprinting. The ability to capture an individual’s uniqueness from data in the scanner is a truly remarkable achievement.

We have come a long way. It took centuries for the scientific world to understand the fundamentals of brain anatomy and neuro-cellular architecture. Even as progress was being made in mapping the brain’s neural circuitry, the goal of truly probing the complexity of human behavior felt like a distant star we would likely never reach. Yet, in the mere 25 years since MRI was first introduced into medicine, we have made astounding progress in probing the brain’s complexity. As psychiatrists, we feel a pressing need for new light to be shed on what goes wrong in mental disease. Powerful new scientific paradigms, advancing technology, and cross-discipline collaborations give us reason to be hopeful.

References:

1. Pawela C, Biswal B. Brain connectivity: a new journal emerges. Brain Connect. 20111:1-2.

2. Trimble MR. The Intentional Brain: Motion, Emotion, and the Development of Modern Neuropsychiatry. Baltimore, MD: Johns Hopkins University Press 2016.

3. Kuhn TS. The Structure of Scientific Revolutions 3rd ed. Chicago, IL: University of Chicago Press 1996.

4. Geschwind N. Disconnexion syndromes in animals and man. Selected Papers on Language and the Brain. Houten, the Netherlands: Springer 1974.

5. Weinberger DR, Radulescu E. Finding the elusive psychiatric “lesion” with 21st-century neuroanatomy: a note of caution. Am J Psychiatry. 2015173:27-33.

6. Fox MD, Greicius M. Clinical applications of resting state functional connectivity. Front Syst Neurosci. 20104:19.

7. Raichle ME. The restless brain: how intrinsic activity organizes brain function. Phil Trans R Soc Lond B Biol Sci. 2015370:pii:20140172.

8. Greicius MD, Krasnow B, Reiss AL, Menon V. Functional connectivity in the resting brain: a network analysis of the default more hypothesis. Proc Natl Acad Sci USA. 2003100:253-258.

9. Biswal B, Yetkin FZ, Haughton VM, Hyde JS. Functional connectivity in the motor cortex of resting human brain using echo-planar MRI. Magn Reson Med. 199534:537-541.

10. Fingelkurts AA, Fingelkurts AA, Kähkönen S. Functional connectivity in the brain-is it an elusive concept? Neurosci Biobehav Rev. 200528:827-836.

11. Seeley WW, Menon V, Schatzberg AF, et al. Dissociable intrinsic connectivity networks for salience processing and executive control. J Neurosci. 2007 27:2349-2356.

12. Finn ES, Shen X, Scheinost D, et al. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nat Neurosci. 201518:1664-1671.

13. Mulders PC, van Eijndhoven PF, Schene AH, et al. Resting-state functional connectivity in major depressive disorder: a review. Neurosci Biobehav Rev. 201556:330-344.

14. Darby RR, Laganiere S, Pascual-Leone A, et al. Finding the imposter: brain connectivity of lesions causing delusional misidentifications. Brain. 2016140:497-507.


Acknowledgements

The authors thank P. Tobler, K. Kondrakiewicz, T. Hensler, A. Walsh, and the reviewers for comments on an earlier version of the manuscript. A.O. was supported by the Knut and Alice Wallenberg Foundation (KAW 2014.0237), a European Research Council Starting Grant (284366 Emotional Learning in Social Interaction project) and a Consolidator Grant (2018-00877) from the Swedish Research Foundation (VR). B.L. was partially supported by Forte (COFAS2: 2014-2785 FOIP) and E.K. by a European Research Council Starting Grant (H 415148) and a National Science Centre grant (2015/19/B/HS6/02209).


Discussion

Social impairments are a hallmark of ASD, yet phenotypic and genetic heterogeneity is thought to contribute to discrepant evidence in the literature. We explore a neural mechanism associated with ASD by considering patterns of social discrimination as measured by mu attenuation over time for children with different genetic etiologies. From the diagnostic comparisons, we show aberrant patterns of mu attenuation in ASD are specific to the upper mu band, while the lower mu band reflects less atypical patterns, consistent with prior work [34]. The dynamic patterns indicate that children with ASD show an increasing lower mu difference between social and nonsocial motions, which may help resolve diagnostic inconsistencies within the literature. For instance, prior evidence of atypical mu attenuation in ASD between observed motion conditions (i.e., social relative to nonsocial motion, as in the current study) [12, 28,29,30] has relied on individual averages. Our findings suggest that the discrimination pattern may not be evident if there are too few trials (i.e., before children with ASD habituated to nonsocial motion observations). Although it is a concern that children with ASD did contribute fewer trials than the typical controls, the eventual condition differentiation in ASD (i.e., noted by approximately trial 15 in Fig. 2) indicates that our study had a sufficient number of trials.

Mu attenuation has been proposed to reflect a human corollary to the mirror neuron system [24, 61], which describes activation recorded over the sensorimotor cortex during both action execution and observation of human actions. Although it is possible that mu attenuation reflects the conductance of occipital or posterior alpha rhythm more broadly (i.e., responsivity to general motion information) [62], our results indicate differentiation of social and nonsocial motions. One working hypothesis of ASD suggests mirror neuron system deficits that disrupt neural correlates supporting the action/observation system, subsequently eliciting atypical mu attenuation [63]. Aligned with this theory, other evidence suggest that atypical functioning of the mirror neuron system may lead to a downstream effect of poor imitative abilities [33] or disrupted higher order social cognitive abilities (i.e., theory of mind) [64]. However, it is important to note that similar to prior work by Dumas and colleagues, we found mu attenuation diagnostic differences within the upper mu band (10–12 Hz), but no group difference within the lower mu band (8–10 Hz). This is consistent with prior work suggesting that this lower frequency may reflect primary sensory processing [36, 37] that habituates over the course of the exposure. Yet, sensory processing of biological motion occurred more rapidly in the TYP group compared to longer processing in the ASD group, perhaps indicative of functional connectivity reductions related to social cognition [65]. Our results offer further evidence of atypical mu attenuation patterns in ASD, although unique neural mechanisms underlying atypical social discrimination may be derived from specific genetic etiologies. In other words, a mirror neuron hypothesis may indeed describe a subset of children with ASD, while a more general, distributed network of neural correlates may be impacted in other ASD subgroups.

As part of a preliminary analysis, we examined the neural social indices associated with different functional genetic roles of LGDMs as a first step to explore a possible shared neural social phenotype. We implemented a post hoc clustering strategy in order to examine potential convergent pathways between LGDMs that are and are not functionally expressed during embryonic development [18, 66]. The choice to cluster LGDMs around functional expression during embryonic development is based on early genetic regulatory control supporting regional differentiation within the embryonic brain [67, 68], including key social neural structures (e.g., amygdala). We had predicted that the LGDM within the embryonic development group would have a more severely disrupted neural social index due to evidence from animal and human models indicating significant impairments related to social behavior [19, 20] and/or information encoding [21, 22]. The results indicated that children with an LGDM primarily expressed during embryonic development exhibit sensitization of lower mu attenuation to social motion. In other words, these children initially exhibited more mu attenuation for nonsocial motion, but eventually demonstrate more for social motion. This pattern was distinct from children with an LGDM not primarily expressed during embryonic development that exhibited greater lower mu attenuation discrimination throughout the entire experiment (i.e., greater mu attenuation to social than nonsocial motion beginning at the first few trials).

Our results suggest that social motion perception may be conserved despite early genetic disruption, though the delayed processing supports the notion of potentially delayed information processing. It is important to note that this delay was specific to the social motion condition (increasing neural response over time) but not the nonsocial motion condition (i.e., no change over time), which may help clarify the mechanism by which prior models [19, 20] derive impaired social behavior. An interpretation of the results may be that children with an LGDM primarily expressed during embryonic development are increasing their attention to, or interest in, social stimuli after an initial period, which may reflect a delayed social engagement (e.g., motivation or salience). One explanation may be that the impact of embryonic genes on social perception is greater [69], suggesting that functional timing of genetic expression may differentially affect the neural social phenotype. Importantly, these findings align with genetics research indicating that ASD genes converge on several select pathways [70, 71], which may help to further explain the underlying neural social heterogeneity.

An important limitation of the current study is the continued genetic heterogeneity despite functionally classifying the expression of LGDM within early development. Within our LGDM groups, there are only several children with a shared LGDM (i.e., SETD2, n = 2 DYRK1A, n = 2 CHD8, n = 2). Thus, the discoveries of this work are not to be taken as firm conclusions, but rather considered in order to motivate and guide continued use of a genetics-first approach to elucidate potential etiological mechanisms of ASD. For instance, most of the children within the early embryonic LGDM group exhibit the social sensitization pattern described here (six out of eight cases see Additional file 3: Figure S1 for individual patterns), except for one child with MED13L and one child with DYRK1A. In part, this qualitative finding is consistent with the overall group clustering approach indicating delayed social processing, suggesting a potential neural index associated with this particular genetic etiology. However, the specificity for specific LGDMs may be poor, considering that only two out of three children with a DYRK1A LGDM exhibited this pattern. Similar to prior work linking core social symptoms to biomarkers of ASD [11, 72,73,74,75,76], we encourage the use of this data as a way to bridge the gap between genetic and phenotypic characterization as a means to facilitate the discovery of ASD etiological mechanisms and accelerate progress for ASD therapeutic interventions.

It may be surprising that our task elicited mu attenuation during nonsocial motion observation (i.e., ball bouncing, tubes swinging) that is not biological and subsequently should not be simulated within the action/observation system. However, to a large extent, the majority of studies implementing mu attenuation as an outcome utilized comparisons between self-executed, social observed, and nonsocial observed motion. It may be the case that by engaging the motor execution system during these tasks, the threshold for the action/observation system is elevated, reducing the amount of mu attenuation for nonsocial comparisons. In fact, neural regions implicated in mu suppression during execution vs. observation [77] involve regions that also play a role in general motion perception, including the occipital, premotor, and somatosensory cortices. Moreover, this study replicated prior work with this same task that indicated a modest degree of mu attenuation to nonsocial motion, in addition to social motion [17]. We posit that our task measured more globally distributed neural differences between social and nonsocial motions compared to other tasks that have used self-initiated actions to target the premotor cortex. Of note, this passive viewing task is more conducive for children with reduced capacity for following behavioral instructions (i.e., to make self-initiated motions), while still providing a robust neural index, which specifies individual patterns.

The neural social indices were correlated with features of social cognition (i.e., social responsiveness), particularly with the lower mu band. This finding is compelling evidence that these indices accurately capture subtle levels of social impairments in vivo, as opposed to relying on parental reports (e.g., SRS-2). Additionally, average patterns of mu attenuation were unaffected by general cognition, despite drastic cognitive differences for children with a LGDM. Although this may not negate a contributory role of cognitive ability for higher-order operations related to social motion (e.g., action prediction), this evidence from this study suggests that motion perception is intact for children with lower cognitive abilities (i.e., cognitive scores under 50). Much of the existing research investigating neural social indices is restricted to children and adults with moderate to average cognitive capabilities. The majority of ASD-LGDM cases with low verbal IQ show typical mu attenuation patterns (i.e., greater for social motion in five out of eight cases with verbal IQ < 50). Taken together, these neural social indices can provide a robust characterization of the underlying neural mechanisms supporting social cognition, regardless of level of cognitive function, thereby improving our understanding of the social phenotype.

This study is the first to use a genetics-first approach to explore the genetic etiologies of autism associated with severe LGDMs in the context of neural social indices. Our use of a unique statistical method to measure ongoing dynamic changes associated with social motion perception demonstrates the utility of this method to better understand underlying processes relevant to ASD and LGDMs. Although this study is limited by a small sample size and thus should be considered exploratory, the analysis of neural social phenotypes based on functional clustering offers a promising approach for narrowing in on convergent pathways that may reflect shared phenotypes and provide insight for targeted treatment [5, 18]. Future research will need to take into account the variety and combination of genetic functional roles. Ongoing efforts to recruit a larger, more genetically homogenous group will help target specific functional outcomes during early childhood and adolescence. However, due to the rarity of this population, these preliminary results are informative and can help guide future research by better describing the functional processes during social motion perception and similar processes that are impaired in ASD.


Impaired executive function exacerbates neural markers of posttraumatic stress disorder

A major obstacle in understanding and treating posttraumatic stress disorder (PTSD) is its clinical and neurobiological heterogeneity. To address this barrier, the field has become increasingly interested in identifying subtypes of PTSD based on dysfunction in neural networks alongside cognitive impairments that may underlie the development and maintenance of symptoms. The current study aimed to determine if subtypes of PTSD, based on normative-based cognitive dysfunction across multiple domains, have unique neural network signatures.

In a sample of 271 veterans (90% male) that completed both neuropsychological testing and resting-state fMRI, two complementary, whole-brain functional connectivity analyses explored the link between brain functioning, PTSD symptoms, and cognition.

At the network level, PTSD symptom severity was associated with reduced negative coupling between the limbic network (LN) and frontal-parietal control network (FPCN), driven specifically by the dorsolateral prefrontal cortex and amygdala Hubs of Dysfunction. Further, this relationship was uniquely moderated by executive function (EF). Specifically, those with PTSD and impaired EF had the strongest marker of LN-FPCN dysregulation, while those with above-average EF did not exhibit PTSD-related dysregulation of these networks.

These results suggest that poor executive functioning, alongside LN-FPCN dysregulation, may represent a neurocognitive subtype of PTSD.


Frontiers in Neuroscience

The editor and reviewers' affiliations are the latest provided on their Loop research profiles and may not reflect their situation at the time of review.



SHARE ON

V. Results

An adult male rhesus macaque (monkey L) was trained to perform variants of a point-to-point arm movement task in a 3D experimental apparatus for juice reward [1]. 1 A 96-electrode silicon array (Blackrock Microsystems) was then implanted in premotor/motor cortex. Array recordings (𢄤.5 RMS threshold crossing applied to each electrode’s signal) yielded tuned activity for the direction and speed of arm movements. As detailed in [1], a standard Kalman filter model was fit by correlating the observed hand kinematics with the simultaneously measured neural signals, while the monkey was performing the point-to-point reaching task ( Fig. 3 ). The resulting model was used online to control an on-screen cursor in real time. This model and 500 of these trials (2010-03-08) serves as the standard against which the SNN implementation’s performance is compared.

Neural and kinematic measurements for one trial. a. The ninety-six cortical recordings that were fed as input to the Kalman filter and the spiking neural network (spike counts in 50ms bins). b. Arm x- and y-velocity measurements that were correlated with the neural data to obtain the Kalman filter’s matrices, which were also used to engineer the neural network.

Starting with the matrices obtained by correlating the observed hand kinematics with the simultaneously measured neural signals, we built a SNN using the NEF methodology and simulated it in Nengo using the parameter values listed in Table I . We ensured that the time constants τ i RC , τ i ref , and τ i PSC were smaller than the implementation’s time step (50ms).

TABLE I

SymbolRangeDescription
max G(Jj (x))200� HzMaximum firing rate
G(Jj (x)) = 0𢄡 to 1Normalized x-axis intercept
J j bias Satisfies first twoBias current
αjSatisfies first twoGain factor
φ ∼ j x ‖ φ ∼ j x ‖ = 1 Preferred-direction vector
τ j RC 20 msRC time constant
τ j ref 1 msRefractory period
τ j PSC 20 msPSC time constant

We had the choice of two network architectures for the aj(t) units: a single 3D integrator or two 1D integrators ( Fig. 4 ). The latter were more stable, as reported previously [14], and yielded better results given the available computer resources. We also had the choice of representing the 96 neural measurements with the bk(t) units (see Fig. 2b ) or simply replacing these units’ spike rates with the measurements (spike counts in 50ms bins). The latter was more straight forward, avoided error in estimating the measurements, and conserved computer resources. Replacing bk(t) with y(t)’s k th component is equivalent to choosing φ k y from a standard basis (i.e., a unit vector with 1 at the k th position and 0 everywhere else), which is what we did.

Spiking neural network architectures. a. 3D integrator: A single population represents three scalar quantities—x and y-velocity and a constant. b. 1D integrators: A separate population represents each scalar quantity—x or y-velocity in this case.

The SNN performed better as we increased the number of neurons ( Fig. 5a,b ). For 20,000 neurons, the x and y-velocity decoded from its two 10,000-neuron populations matched the standard decoder’s prediction to within 0.03% (RMS error normalized by maximum velocity). 2 As reported in [10], the RMS error was roughly inversely proportional to the square-root of the number of neurons ( Fig. 5c,d ). There is a tradeoff between accuracy and computational time. For real-time operation—on a 3GHz PC with a 1ms simulation time-step—the network size is limited to 1,600 neurons. Encouragingly, this small network’s error was only 0.27%.

Comparing the x and y-velocity estimates decoded from 96 recorded cortical spike trains (10s of data) by the standard Kalman filter (blue) and the SNN (red). a,b. Networks with 2,000 and 20,000 spiking neurons. c. Dependence of RMS error (between SNN and Kalman filter) on network size (note log scale). d. Product of RMS error and neuron count’s (NC) square root is roughly constant (for NC > 200), implying that they are inversely proportional.


Associated Data

Neurobiological and behavioral findings suggest that the development of delinquent behavior is associated with atypical social-affective processing. However, to date, no study has examined neural processes associated with social interactions in severely antisocial adolescents. In this study we investigated the behavioral and neural processes underlying social interactions of juvenile delinquents and a matched control group. Participants played the mini-Ultimatum Game as a responder while in the MRI scanner. Participants rejected unfair offers significantly less when the other player had ‘no alternative’ compared with a �ir’ alternative, suggesting that they took the intentions of the other player into account. However, this effect was reduced in the juvenile delinquents. The neuroimaging results revealed that juvenile delinquents showed less activation in the temporal parietal junction (TPJ) and inferior frontal gyrus (IFG). However, the groups showed similar activation levels in the dorsal anterior cingulate cortex (dACC) and the right anterior insula (AI) when norms were violated. These results indicate that juvenile delinquents with severe antisocial behavior process norm violations adequately, but may have difficulties with attending spontaneously to relevant features of the social context during interactions.