Predicting individual differences of fear and cognitive learning and extinction

Predicting individual differences of fear and cognitive learning and extinction | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Predicting individual differences of fear and cognitive learning and extinction View ORCID Profile C.A. Gomes , D. R. Bach , A. Razi , View ORCID Profile G. Batsikadze , S. Elsenbruch , H. Engler , T.M. Ernst , M.C. Fellner , C. Fraenz , E. Genç , A. Klass , F. Labrenz , S. Lissek , C.J. Merz , D. Metzen , A. Nostadt , R.J. Pawlik , J.E. Schneider , M. Tegenthoff , A. Thieme , O.T. Wolf , O. Güntürkün , H.H. Quick , R. Kumsta , D. Timmann , T. Spisak , N. Axmacher doi: https://doi.org/10.1101/2025.05.04.651880 C.A. Gomes 1 Department of Neuropsychology, Ruhr University Bochum , Germany 2 Department of Neurology, Essen University Hospital , Essen, Germany 3 Center Translational Neuro- and Behavioral Sciences (CTNBS), Essen University Hospital , Essen, Germany 9 Erwin L. Hahn Institute for Magnetic Resonance Imaging, University of Duisburg-Essen , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for C.A. Gomes For correspondence: carlos.assuncaodias{at}rub.de D. R. Bach 14 Department of Imaging Neuroscience, UCL Queen Square Institute of Neurology, University College London , London, UK 15 University of Bonn, Transdisciplinary Research Area “Life & Health”, Hertz Chair for Artificial Intelligence and Neuroscience , Bonn, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site A. Razi 16 Wellcome Centre for Human Neuroimaging, University College London , London, UK 17 Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University , Clayton, Australia 18 Monash Biomedical Imaging, Monash University , Clayton, Australia 19 CIFAR Azrieli Global Scholars Program, CIFAR , Toronto, Canada Find this author on Google Scholar Find this author on PubMed Search for this author on this site G. Batsikadze 2 Department of Neurology, Essen University Hospital , Essen, Germany 3 Center Translational Neuro- and Behavioral Sciences (CTNBS), Essen University Hospital , Essen, Germany 9 Erwin L. Hahn Institute for Magnetic Resonance Imaging, University of Duisburg-Essen , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for G. Batsikadze S. Elsenbruch 5 Department of Medical Psychology & Medical Sociology, Ruhr University Bochum , Bochum, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site H. Engler 3 Center Translational Neuro- and Behavioral Sciences (CTNBS), Essen University Hospital , Essen, Germany 11 Institute of Medical Psychology and Behavioral Immunobiology, University Hospital Essen , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site T.M. Ernst 2 Department of Neurology, Essen University Hospital , Essen, Germany 3 Center Translational Neuro- and Behavioral Sciences (CTNBS), Essen University Hospital , Essen, Germany 9 Erwin L. Hahn Institute for Magnetic Resonance Imaging, University of Duisburg-Essen , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site M.C. Fellner 1 Department of Neuropsychology, Ruhr University Bochum , Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site C. Fraenz 12 Department of Psychology and Neurosciences, Leibniz Research Centre for Working Environment and Human Factors at the Technical University of Dortmund (IfADo) , Dortmund, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site E. Genç 12 Department of Psychology and Neurosciences, Leibniz Research Centre for Working Environment and Human Factors at the Technical University of Dortmund (IfADo) , Dortmund, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site A. Klass 6 Department of Neurology, BG University Hospital Bergmannsheil, Ruhr University Bochum , Bochum, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site F. Labrenz 5 Department of Medical Psychology & Medical Sociology, Ruhr University Bochum , Bochum, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site S. Lissek 6 Department of Neurology, BG University Hospital Bergmannsheil, Ruhr University Bochum , Bochum, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site C.J. Merz 4 Department of Cognitive Psychology, Ruhr University Bochum , Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site D. Metzen 13 Institute of Psychology, Department of Educational Sciences and Psychology, TU Dortmund University , Dortmund, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site A. Nostadt 6 Department of Neurology, BG University Hospital Bergmannsheil, Ruhr University Bochum , Bochum, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site R.J. Pawlik 2 Department of Neurology, Essen University Hospital , Essen, Germany 3 Center Translational Neuro- and Behavioral Sciences (CTNBS), Essen University Hospital , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site J.E. Schneider 1 Department of Neuropsychology, Ruhr University Bochum , Germany 7 Genetic Psychology Lab, Ruhr University Bochum , Bochum, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site M. Tegenthoff 6 Department of Neurology, BG University Hospital Bergmannsheil, Ruhr University Bochum , Bochum, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site A. Thieme 2 Department of Neurology, Essen University Hospital , Essen, Germany 3 Center Translational Neuro- and Behavioral Sciences (CTNBS), Essen University Hospital , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site O.T. Wolf 4 Department of Cognitive Psychology, Ruhr University Bochum , Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site O. Güntürkün 8 Department of Biopsychology, Ruhr University Bochum , Germany 9 Erwin L. Hahn Institute for Magnetic Resonance Imaging, University of Duisburg-Essen , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site H.H. Quick 9 Erwin L. Hahn Institute for Magnetic Resonance Imaging, University of Duisburg-Essen , Essen, Germany 10 High Field and Hybrid MR Imaging, Essen University Hospital , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site R. Kumsta 7 Genetic Psychology Lab, Ruhr University Bochum , Bochum, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site D. Timmann 2 Department of Neurology, Essen University Hospital , Essen, Germany 3 Center Translational Neuro- and Behavioral Sciences (CTNBS), Essen University Hospital , Essen, Germany 9 Erwin L. Hahn Institute for Magnetic Resonance Imaging, University of Duisburg-Essen , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site T. Spisak 3 Center Translational Neuro- and Behavioral Sciences (CTNBS), Essen University Hospital , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site N. Axmacher 1 Department of Neuropsychology, Ruhr University Bochum , Germany 9 Erwin L. Hahn Institute for Magnetic Resonance Imaging, University of Duisburg-Essen , Essen, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Supplementary material Preview PDF Abstract The abilities to acquire new information and to modify previously learned knowledge are critical in an ever-changing world. However, the efficacy of learning is notably variable among individuals, with extinction learning being the epitome of such variability. Abundant studies have identified a core network of brain regions including amygdala, hippocampus, dorsal anterior cingulate cortex (ACC), ventromedial prefrontal cortex (PFC) and, more recently, the cerebellum, as key players in learning and extinction. Yet, the precise interactions within this network and their relationship to individual learning abilities and extinction have remained largely unexplored. In the present study, we examined how functional (FC), effective (EC), and structural (SC) connectivity patterns in the core learning network allow predicting individual differences in the efficacy of learning, extinction, and renewal. Analysing a large dataset of over 500 participants across a multitude of paradigms, our results revealed that FC predicted better acquisition, with a central role of ACC and hippocampus, whereas SC, involving ACC and amygdala, predicted higher levels of extinction learning. EC results suggested a predominantly inhibitory coupling among core learning network nodes, with paradigm-specific EC connectivity patterns predicting learning. Our predictions not only generalised between fear and cognitive predictive learning paradigms but were also successful in predicting learning from task-related FC and simulated data. Together, these results describe the multimodal neural determinants of learning, extinction, and renewal, and may inform individualised interventions for affective disorders based on neural connectivity patterns. Introduction The ability to learn from experience is a hallmark of every living system, from humans down to single-cell organisms. This ability differs strongly between individuals: While some are able to acquire new information quickly and display steep learning rates, others are much slower 1 . These differences concern abilities as widespread as the formation of new episodic memories, the development of novel practical skills, or the gradual learning about the putative outcome of actions. Since our world is constantly changing, it is equally important to cease responding to previously memorised information once it is no longer valid. This process is called extinction learning and involves the acquisition of two distinct memory traces. The first represents the initial association that is largely left intact, while the second is an association of inhibitory nature that suppresses the activation of the first trace 2 . These inhibited associated responses can return under diverse conditions and so turn into invasive components of psychopathology 3 . The societal and clinical relevance of extinction learning, and its associated problems can hardly be overestimated. According to Craske et al., more than 60 million European Union citizens suffer from anxiety disorders 4 . Importantly, extinction learning is itself strongly context-dependent, since presentation of a conditioned stimulus (CS) outside its extinction context tends to induce return of the conditioned response (CR), a phenomenon known as renewal 2 . The ability to extinguish previously acquired information and the propensity for renewal show pronounced individual differences, which may not only account for a person’s ability to flexibly update knowledge but also their vulnerability or resilience to psychopathology, specifically regarding anxiety disorders 5 . It is likely that these differences reflect a combination of both stable (trait) and variable (state) measures. For example, the ability to acquire novel and to update existing information, as well as the re-occurrence of extinguished memory traces, depend on age 6 , 7 , sex 8 , 9 , and personality traits such as trait anxiety and sensation seeking 10 , 11 , but are also modulated by acute psychosocial stress and/or state anxiety 12 , 13 . Understanding the neural determinants of individual differences in learning, extinction, and renewal, is thus not only a window into the mechanisms of extinction but may prove useful in our understanding of disorders that affect this ability and their potential treatments. The brain structures involved in fear conditioning are relatively well-known. Animal and human research converge towards the idea that the amygdala (AMY) stores the associations between the CS and the unconditioned stimulus (US), whereas the hippocampus (HIP) encodes context information 14 , 15 . The dorsal anterior cingulate cortex (ACC 1 ) and ventromedial prefrontal cortex (PFC) have prominent roles in fear appraisal and safety learning, respectively 16 . More recently, it has also been proposed that the cerebellum (CEB) provides predictions of upcoming sensory events during associative tasks 17 . The mechanisms of extinction learning putatively differ from those supporting initial learning and may be more complex since they require the formation of a second associative trace of an inhibitory nature. Moreover, despite the apparent ubiquity of learning and extinction in both fear conditioning and cognitive predictive learning contexts, it remains an open question whether these require the same or different neural determinants. For instance, whereas HIP and PFC support context-dependent extinction learning in both fear 18 – 20 and predictive learning 21 – 23 paradigms, the involvement and role of the AMY may be less universal than previously assumed 24 , 25 : Suppression of AMY activity by PFC enables extinction following aversive learning 14 , whereas AMY activity increases during extinction in both appetitive 26 and predictive learning 21 tasks, presumably related to salience or novelty processing. Thus, even though accumulated evidence points to the involvement of a similar set of brain regions in various conditioning-based learning paradigms, the macroscale network connectivity patterns that support these different kinds of learning remain unclear. Furthermore, there exists a paucity of studies devoted to investigating individual differences in learning efficacy from inter-areal connectivity patterns. Resting-state fMRI (rs-fMRI) has proved to be a reliable and convenient technique to measure intrinsic brain connectivity in large participant samples. Indeed, several studies have successfully predicted performance in various (non-)cognitive traits, as well as vulnerability or resilience to mental disorders 27 – 29 , from brain connectivity patterns. Still, the extant studies examining brain-behaviour relationships using rs-fMRI have mostly focused on functional connectivity (FC). While this method appears to reflect inter-regional interactions between local neural assemblies 30 , it does not convey information about the direction of these interactions. By contrast, recent advances in effective connectivity (EC) now allow the characterization of causal interactions among brain areas at rest 31 , which may provide important complementary information regarding the complex relationship between brain connectivity and individual differences in learning, extinction, and renewal. Even though correlation-based FC and EC are mathematically related, they differ fundamentally in that FC only accounts for linear, undirected statistical dependencies, whereas EC measures the directed causal influence that one brain region exerts over another 32 , 33 . In addition to functional and effective interactions, pronounced individual differences have been found in patterns of structural connectivity (SC) that reflect the integrity and effectivity of axonal information transfer. SC can be quantified using tractography, a technique that generates streamlines as a proxy for white matter fibre tracts across brain regions 34 . FC and SC are known to be related to some extent 35 , 36 , but this relationship is complex. FC-SC correlations have been shown to depend on the specific network connections being examined 37 , and the existence of strong FC in the absence of direct structural connections suggests that FC between two regions may rely on SC via a common third region 38 . The relationship between EC and SC is even less clear, although recent evidence suggests that constructing structurally-informed dynamic causal models (DCMs) of EC can outperform structurally-naïve DCMs by drastically improving group-level model evidence 39 . Nevertheless, it remains unclear how these two types of connectivity compare in terms of predicting cognitive variables. In summary, cognition depends on a complex interplay between FC, EC, and SC, which has prompted researchers’ calls for an integrative approach 40 , 41 . In the present study, we set out to investigate not only the neural connectivity patterns supporting learning and extinction but also whether these patterns generalise across different types of learning paradigms. As mentioned above, various forms of learning, such as fear learning and cognitive predictive learning, appear to rely on overlapping neural circuitry, as both involve acquiring and updating associative contingencies. While distinct paradigms may recruit additional regions based on task-specific demands (e.g., the piriform cortex in olfactory conditioning), a core learning network appears to be commonly engaged across different forms of associative learning. By comparing fear learning and cognitive predictive learning, our analysis aimed to identify this shared neural architecture and its role in learning and extinction. We analysed FC, EC and SC patterns within this network in a new large multi-center dataset of over 500 individuals from a collaborative project involving different types of learning and extinction (see Fig. S3 and Supplemental text: Experimental paradigms). Results We carefully optimised and homogenised data acquisition pipelines across centres, resulting in high test-retest reliability of functional and structural connectivity measures using either ROIs from the the whole brain or only the selected ROIs from our study (all Cronbach’s alpha > .80; Figure S1-2). We then acquired rs-fMRI and DWI data from a large group of participants (rs-fMRI: N=509; DWI: N=463) who conducted different fear and cognitive acquisition and extinction learning paradigms (Figure S3; Methods; Supplemental Methods: Experimental paradigms). Multimodal connectivity in the core learning network For the analysis of FC, we computed a composite score (concatenation of nine different metrics; Methods; Table S4) focusing on ipsilateral connections (e.g., left AMY - left HIP), with the exception of the CEB, given that (neo-)cerebellar regions are connected with the contralateral cerebral cortex. A mixed-effects model (participant nested within study) using FC as the outcome variable and ROI-pair as predictor revealed greater FC for HIP-AMY than any other connection (all ts > 36, ps FDR 26, ps FDR 12, ps FDR < .001; see Table S9 for remaining comparisons). This pattern was observed in both hemispheres ( Fig. 1C ). Download figure Open in new tab Figure 1. Number of datasets, regions of interest and connectivity estimates. (A) Number of resting-state fMRI (rs-fMRI) and diffusion weighted imaging (DWI) datasets acquired in each group of the consortium [ sfb1280.ruhr-uni-bochum.de ]. Blue and red columns indicate whether learning was assessed via skin conductance responses (SCR) or behavioural ratings, respectively. Striped and plain columns reflect fear conditioning or cognitive predictive learning paradigms, respectively. (B) ROIs used in the present study. All subject-specific ROIs were extracted from an automatic parcellation/segmentation using FreeSurfer (top left). For the cerebellum, the three cerebellar nuclei (fastigial, interposed and dentate nuclei) were extracted using the SUIT package (top middle; see Methods) and combined into one cerebellar ROI. For probabilistic tractography, surfaces of the dorsal anterior cingulate and ventral prefrontal cortices were used instead of their volumetric counterparts (top right). (C) Average functional connectivity between all pairs of ROIs. Functional connectivity was calculated across the entire sample based on a composite metric (see Methods). Greater FC was observed for the connections HIP - AMY and HIP – PFC. (D) Average structural connectivity between all pairs of ROIs. SC values are based on streamline counts across the entire sample. As expected, the connection HIP – AMY showed a disproportionately larger number of streamlines, followed by ACC – PFC and AMY – PFC connections. (E) Average effective connectivity based on spectral dynamic causal modelling (spDCM) estimates of directed connectivity among our ROIs (top: excitatory connections; bottom: inhibitory connections). (F) The winning model for effective connectivity showing the spDCM estimates of directed connectivity among our regions of interest computed using Parametric Empirical Bayes and Bayesian Model Averaging. rs-fMRI=resting-state functional magnetic resonance imaging; DWI=diffusion-weighted imaging; EDA=Electrodermal activity; ACC = Dorsal anterior cingulate cortex; AMY = Amygdala; CEB = Cerebellar nuclei; HIP = Hippocampus; PFC = Ventromedial prefrontal cortex. The general pattern of SC was similar to what we observed for FC ( Fig. 1D ). A corresponding mixed-effects model using streamlines as outcome variable revealed a disproportionate number of streamlines for the HIP-AMY, ACC-PFC, AMY-PFC and HIP-PFC connections (in this order) relative to all others (all zs > 12, p FDR < .001; see Table S10 for the remaining connections). The EC analyses showed that the core learning network was mostly characterized by inhibitory connections, with only a few excitatory connections, most notably, the bidirectional HIP-AMY connection ( Fig. 1E ; see also Fig. 1F for the group-level results using a Parametric Empirical Bayes model). Since the same modelling approach was used for FC and SC, we could also compare the relative strength of connectivity for each connection between these two modalities (see Methods). This analysis showed that relative FC between HIP-PFC and AMY-PFC was greater than relative SC (ts > 8.30, p FDR s 17.15, p FDR s < .001). Thus, despite their overall similarities, relative FC and SC values differed for some ROI pairs (Fig. S11). To examine whether the different types of connectivity were related to each other, we computed Pearson correlations between FC-EC, FC-SC and EC-SC on each individual connection (Fig. S12). Interestingly, functional connectivity between several connections was significantly correlated with the respective effective connectivity patterns as well as with SC between these regions. By contrast, we did not observe any correlation between SC and EC connection strengths even at uncorrected thresholds. Thus, while individual differences in FC were partially determined by (putatively more hard-wired, i.e. trait-like) differences in streamlines, these structural connectivity differences did not correspond to individual differences in EC. Learning measures Learning during acquisition, extinction and renewal was estimated separately for each study and experiment. For studies that collected skin conductance response (SCR) data, PsPM was used to estimate trial-by-trial SCR values, while participant ratings were used in behavioural-only studies (see Methods). For the group-level analysis, we included the initial eight trials (acquisition) and seven trials (extinction and renewal), which corresponded to the number of trials of the subject with the least number of trials. We observed a significant interaction of SCR data between time (i.e., trials) and condition (CS+ cs. CS-) for all experimental phases, indicating steeper increases in amplitude for CS+vs. CS-during acquisition (t = 3.85, p FDR < .001) and steeper decreases for extinction (t = −3.65, p FDR < .001) and renewal (t = −6.28, p FDR < .001) (see Fig. 2A ). Download figure Open in new tab Figure 2. Learning estimates. (A) Group-level averages of SCR amplitudes in individual CS+ (red) and CS-(blue) trials during acquisition (left), extinction (middle) and renewal (right). (B) Illustration of fixed-effect polynomial regression on the SCR data of an exemplar participant. After fitting the model, a unique learning score was computed comparing CS+ and CS-trials. (C) Average estimates of learning based on D and E for the entire sample (see Fig. S21 for the separate studies). (D) Multilevel generalised linear model using a logit link function to model behavioural ratings in predictive learning paradigms (exemplar participant). Individual parameter estimates were extracted from each participant and the expected rate of success after all 8 trials was computed. (E) Individual (blue: negative slopes; red: positive slopes) and group (black line) learning slopes estimated from the multilevel logistic regression in (D). CS+=conditioned stimulus (reinforced); CS-=conditioned stimulus (non-reinforced); SCR=skin conductance responses. For the extraction of subject-specific variables of learning, extinction and renewal, we used all trials available in each participant in either a subject-wise polynomial regressions for SCR data (studies S1, S2, S3, S5, S6; Fig. 2B ) or a generalised linear mixed-effects model using a logit link function for behavioural ratings (study S4; Fig. 2D ). Using these individual-level estimates of learning, we observed a much larger estimate for CS+ than CS-in the acquisition phase, indicating greater learning for CS+ trials, and the opposite pattern for extinction and renewal. Permutation testing on these learning scores confirmed that condition differences during all phases were significantly larger than would be expected by chance ( Fig. 2C ; acquisition: t = 5.69, p FDR < .001; extinction: t = −3.80, p FDR < .001; renewal: t = −4.99, p FDR < .001). Similarly, in the behavioural studies learning was highly significant at the group level during both acquisition (z = 19.75, p FDR s < .001) and extinction (z = 16.11, p FDR s 5.40, p FDR s .10), but the probability of making at least one renewal response (giving the same response as that given during acquisition) was still significant (t = 3.46, p FDR < .001). Further analyses showed that even though acquisition and extinction were correlated (r = .16, p FDR < .001; Fig S16), the amount of shared variance was very limited (≈ .03), suggesting that different factors may account for individual differences in these two phases. The correlation between acquisition/extinction and renewal was not significant, rs = -.06/-.08, ps FDR > .10. Prediction of individual differences We next investigated whether and how the three different types of connectivity (FC, SC, and EC) related to individual differences of learning, using a LASSO regression model (see Methods). Acquisition For acquisition, the relevant functional connections were lCEB-rPFC, lHIP-lPFC, rACC-rPFC, rAMY-rACC, rCEB-lHIP as well as bilateral HIP-ACC ( Fig. 3B ; using traditional p-values, these connections were all significantly above chance after correction for multiple comparisons, see Fig. S22). Results were similar for most combinations of individual studies (see also Fig. S27). In addition to these FCs that were relevant for learning across all studies, three FCs were only predictive in PL studies (lACC-lPFC, lAMY-lPFC and lCEB-rHIP). Note that it is possible for a connection to be identified as significant when data are pooled across all studies, even it was not significant in any individual study (see Fig. S23). Download figure Open in new tab Fig. 3. Predicting individual differences from brain connectivity. (A) Significant coefficient estimates from the LASSO model for the entire sample: FC only predicted acquisition, SC only predicted extinction learning, EC only predicted renewal. (B) Coefficient estimates were obtained using a ridge regression model for different groupings in the acquisition phase (see Fig. S23-S25 for lines of best fit for all connections). Ridge regression was chosen to provide estimates for all connections, as LASSO sets irrelevant coefficients to zero. Significant connections identified by LASSO are marked with stars (Ridge and non-zero LASSO coefficients were highly correlated, r = .87, p < .001). Only connections that were significant in at least one grouping for any connectivity type are displayed. (C) Same as (B) but for the extinction phase. (D) Same as (B) but for the renewal phase. (E) Bars represent the estimated marginal means for the number of times each ROI appeared in a significant connection across Monte-Carlo resampling iterations, where LASSO was run on random subsets of observations. This analysis assessed the robustness of each ROI as a “hub” within the fear and extinction network. Doughnut charts below each type of connectivity indicate how many times each ROI appeared in the original LASSO models (i.e., in B-D), providing a direct comparison between resampling-based estimates (bar plots) and the original results (doughnut charts). FC = Functional connectivity; SC = Structural connectivity; EC = Effective connectivity. All=All studies (S1,S2,S3,S4,S5,S6); FL=Fear Learning studies (S1,S2,S3,S5,S6); FLc=Fear Learning classical paradigm (S2,S3); PL=Predictive Learning studies (S4); FLr=Fear Learning renewal study (S2); PLr=Predictive learning renewal study (S4); Both=Both renewal studies (S2, S4). ACC=Dorsal anterior cingulate cortex; AMY=Amygdala; CEB=Cerebellar nuclei; HIP=Hippocampus; PFC=Ventro-medial prefrontal cortex. Neither structural nor effective connections significantly predicted learning for the entire sample. However, we did observe distinctive predictive connections for FL and PL paradigms when analysed separately. Regarding EC, disinhibition of the inhibitory connections lAMY→lPFC, rCEB→lAMY and bilateral HIP→AMY predicted acquisition in PL studies. In contrast, a more pronounced excitatory connection lCEB→rAMY predicted FL. For FLc, there was also an interesting association between acquisition and disinhibition of the inhibitory connection ACC→AMY, as well as increased PFC→AMY inhibition with greater acquisition. Thus, fear learning benefited from higher AMY inhibition by PFC and from AMY disinhibition by ACC. With regard to SC, only the connections lCEB-rPFC and rAMY-rPFC in PL were significant predictors. Even though our results indicate that certain connections are predictive of acquisition performance, they do not reveal which regions play a predominant role. To assess the extent to which each of the five ROIs could act as a “hub” governing individual differences in acquisition, we re-ran the LASSO model on one-hundred Monte-Carlo samples. In each iteration, we randomly selected 80% of the participants and recorded the frequency with which each ROI appeared in a significant connection. ROIs that appeared more frequently were considered more reliable and indicative of hub-like properties within the network. A Poisson mixed-linear model confirmed that, for functional connectivity (FC), all ROIs except the AMY were significantly different from zero (zs > 2, ps .10). Post-hoc tests revealed that ACC and HIP had larger numbers of appearances than the other ROIs (ps FDR 2, ps < .05; other ROIs: zs .10), with the AMY showing a larger number of counts than the other ROIs (zs > 4.5, ps FDR < .001). There was no significant ROI for the SC analysis (zs .10). In summary, the acquisition results indicated that FC was predictive of learning across the entire sample, with a strong focus on connections involving the ACC and HIP. Paradigm-specific EC- and SC-learning associations were also apparent, with differing dynamics for the PFC-AMY connections between PL and FL studies (disinhibition of AMY→PFC was beneficial for PL, whereas more pronounced inhibition PFC→AMY predicted greater fear acquisition). Extinction In stark contrast with our results during acquisition, we did not find any significant FC (or EC) connections that could predict individual differences in extinction learning across all paradigms. Interestingly, however, several structural connections consistently predicted extinction learning, most notably those involving the ACC, specifically, rHIP-rACC, lCEB-rACC, and bilateral AMY-ACC (see Fig. 3D ; using traditional p-values, these connections were all significantly above chance after correction for multiple comparisons, see Fig S22). In addition to these effects across both FL and PL experiments, the structural connections lHIP-lPFC, rAMY-rACC and rHIP-rACC predicted extinction in FL but not PL studies. However, the most interesting differences among the groupings were observed for EC, with a peculiar reversal in the direction of the AMY-HIP connectivity (benefit of more pronounced excitation of lHIP→lAMY for FL studies, benefit of higher disinhibition of lAMY→lHIP for PL studies), as well as a reversal in both directionality and valence of the HIP-ACC connection (less HIP inhibition by ACC being beneficial for PL extinction, and greater ACC inhibition by HIP being beneficial for FL extinction). In contrast, more pronounced PFC→ACC inhibitory connectivity was beneficial for extinction across the entire sample, as well as for FL separately. Post-hoc tests revealed that the ACC had the largest number of appearances relative to all other ROIs (zs > 8, ps FDR 2.14, p FDR s < .05; other ROIs: zs .10), with the PFC and CEB showing larger number of counts than the other ROIs (zs > 2.88, ps FDR < .01). There was no significant ROI for the FC analysis (zs .10). In summary, extinction was mostly predicted by higher density of structural connections involving the ACC (and to a lesser degree, AMY). Paradigm-specific learning was again predicted mostly by EC, specifically, a reversal in directionality and/or valence for the AMY-HIP and HIP-ACC connections. Renewal Out of the six studies analysed above, S2 (FLr) and S4 (PLr) contained data about renewal after extinction. When both studies were analysed together, there were no significant FC or SC connections that could predict individual differences in renewal. However, the analysis of EC indicated that the greater the disinhibition of the HIP by AMY and PFC the greater the renewal effect, whereas renewal benefited from a higher inhibition of the ACC by the PFC ( Fig 3D ; using traditional p-values, only rPFC→rHIP and rPFC→rACC were significantly above chance after correction for multiple comparisons, see Fig S22). Separate analysis for FLr renewal also revealed the relevance of disinhibition of the lHIP→lACC connectivity, whereas more pronounced inhibition of rCEB→lAMY, rCEB→lHIP, and rPFC→rACC connections, as well as disinhibition of rPFC→rHIP predicted renewal following PLr. The Poisson regression indicated that for EC, only the ACC, HIP and PFC were significantly different from zero (zs > 3.41, ps FDR < .001; other ROIs: zs .10). Pairwise comparisons revealed that the PFC had the largest number of appearances relative to all other ROIs (zs > 2.90, ps FDR 7.95, ps FDR < .001; see Fig. 3E , right panel). No ROIs were observed for FC or SC models. In summary, differences in EC were more sensitive than FC or SC in predicting renewal. Our results also indicated the importance of disinhibition of the HIP by both the PFC and AMY. Generalisability of learning predictors We found that the models showed significant generalisability. When the LASSO model was trained on one type of paradigm (e.g., FL) and tested on the other type of paradigm (e.g., PL), we observed a higher generalisability (i.e., lower mean squared errors) relative to a surrogate model in which the learning estimates had been shuffled across participants while keeping the structure of the remaining data matrix intact (all ps FDR < .001; Fig. 4A-C ). Selecting non-overlapping groups for FL and PL led to identical results (Fig. S32). Download figure Open in new tab Figure 4. Generalisability of learning and simulations. (A-C) Generalisability of acquisition (A), extinction (B) and renewal (C) was examined by running the LASSO model on one type of paradigm (e.g., fear learning) and testing on the other type of paradigm (e.g., cognitive predictive learning). Mean squared errors (MSEs) of each of the models were used as measure of fit. (D-F) Leave-on-group out (LOGO) cross-validation indicated that functional, structural and effective connectivity predictions were generalisable for acquisition (D), extinction (E) learning and renewal (F), respectively. (G) Strategy for computing task-based functional connectivity for study S2. (H) Significant functional connectivity was obtained for our LASSO model but not for other models using random connections. (I) Significant connections obtained from the significant model in (F). Red stripes indicate the estimates for those connections in the resting-state LASSO model. (J-L) Simulation analysis indicated that our model predictions were superior in predicting acquisition (J), extinction learning (K) and renewal (J) relative to either the same number of random connections from the model (random), or connectivity between the 4 th ventricle and lateral ventricles (ventricles). FC = Functional connectivity; SC = Structural connectivity; EC = Effective connectivity. FL=Fear Learning studies (S1,S2,S3,S5,S6); PL=Predictive Learning studies (S4); FLr=Fear Learning renewal study (S4); PLr=Predictive learning renewal study (S4). ACC=Dorsal anterior cingulate cortex; CEB=Cerebellar nuclei; HIP=Hippocampus. In addition, we performed a leave-one-group-out (LOGO) cross-validation analysis (group here referring to the six individual studies), using a multiple regression analysis on the selected predictors. As Fig. 4D-F shows, the acquisition predictors were highly generalisable only when the functional connections were used (EC: r = -.01, p FDR > .10; FC: r = .14, p FDR .10), whereas extinction predictors were only generalisable when the structural connections were used (EC: r = .03, p FDR > .10; FC: r = -.14, p FDR > .10, SC: r = .23, p FDR < .001). For renewal, generalisability was found only when the effective connections were included, although this effect only reached trend level significance after correction for multiple comparisons (EC: r = .11, p FDR = .06; FC: r = .04, p FDR > 0.10; SC: r = -.02, p FDR > .10). Follow-up analyses confirmed that for FC, prediction performance was indeed significantly greater for acquisition than for either extinction or renewal (ps FDR < .05), whereas SC prediction performance for extinction was significantly greater than for either acquisition or renewal (ps FDR < .01). EC prediction performance was greater for renewal than for acquisition (p .10). We also compared prediction performance during acquisition, extinction, and renewal between the three types of connectivity. During acquisition, the FC prediction performance was greater than the performances of either SC or EC (ps FDR < .05); during extinction, the SC prediction performance was greater than either FC or EC (ps FDR < .01); and during renewal, the EC prediction performance was significantly greater than SC (p .10). Thus, our LOGO results are consistent with our results from the LASSO models in that, at the whole-sample level, acquisition, extinction learning, and renewal are best predicted by FC, SC and EC, respectively. To the extent that the functional architecture of the individual’s brain is intrinsic, we should be able to observe a correspondence in the FC profile between task-based and resting-state connectivity 27 . Given that, in our study, resting-state FC in seven distinct connections were shown to predict acquisition across the entire group ( Fig. 3B ), we queried whether task-based FC would show a similar association. For that purpose, we selected two experiments from S2 for which task-fMRI data were available (N = 137). Task-based FC was computed for the time periods corresponding to the expected peaks of the BOLD signal for both CS+ and CS-trials during the acquisition phase. The difference in FC between CS+ and CS-was then calculated for each of the relevant ROIs and used as predictors in a subsequent multiple regression model (see Fig. 4G and Methods). The model including the LASSO predictors was significant (F(7,90) = 2.39, p < .05, R 2 = .16; Fig. 4H ). In addition, three out of six predictors (lHIP-lACC, rCEB-lHIP and rHIP-rACC) reached individual significance (see Fig. 4I ). This result indicates that task-dependent functional connectivity during acquisition using the predictors from the resting-state LASSO model was predictive of learning performances as well. This result was specific, because none of the models using pseudo-random functional connections within the core learning network was significant (all ps > .05, uncorrected), and the maximum R value that could be achieved (0.089) was almost half of that for the significant model with the LASSO predictors (0.157; see Fig. 4H ). Next, we enquired whether our combination of selected predictors would show a performance advantage over random connections between our ROIs that were not selected in the original LASSO models, by using simulated data. If so, this would show an additional layer of generalisability in the sense that our predictions could potentially be applicable to new datasets. For that purpose, we artificially generated 100 independent datasets (including all predictors, learning measure and covariates), each with 40 observations, based on purely simulated data drawn from a multivariate normal distribution that mimics the relationships between variables in our original dataset (see Methods). The performance of the selected LASSO model was tested against a pseudo-random model that consisted of an equal number of predictors as the LASSO model but chosen at random (excluding the LASSO predictors) within the same type of connectivity. Fig. 4J-L show the results of the simulation analysis (see also Fig. S35 for the results using bootstrapping). For both acquisition, extinction, and renewal, we observed significantly larger R values for the LASSO model relative to the pseudo-random model (all ps FDR < .001; black bars in Fig. 4J-L ). In addition, when the data were split into an FL and a PL group, better performances of the LASSO models were again observed for both groups for extinction and renewal (PL: p < .001, FL: p FDR = .01; coloured error bars in Fig. 4J-L ), and the PL group showed a significant effect for acquisition (PL: p FDR < .001, FL: p FDR = .16). In summary, generalisation of our predictions occurred between FL and PL paradigms, but only for the types of learning that could be predicted by the particular type of brain connectivity (i.e., transferable acquisition, extinction and renewal effects only for FC, SC, and EC, respectively). In addition, our (resting-state-based) model connections were more successful in predicting learning from task-based FC than alternative connections. Finally, our model predictions were also more successful in predicting learning in simulated datasets than pseudo-random connections within the core learning network. Discussion Here we show that individual abilities of learning, extinction, and renewal can be explained by distinct types of resting-state connectivity among a small set of regions within the core learning network. This is substantiated by the observation of a triple dissociation, with FC, SC and EC better predicting acquisition, extinction learning, and renewal, respectively. To our knowledge, this highlights for the first time the distinct functional roles of different types of brain connectivity for the different learning phases. Our findings reveal a core role of the ACC as a “hub” within this network. Surrounding this core, additional areas make specific contributions to learning, extinction, and renewal. Contrary to previous assumptions, we demonstrate that heterogeneity in learning measures, experimental setups, MR sequences, and statistical methods are no detriment for reliably detecting across-study effects, provided that careful harmonisation of sequence parameters and appropriate statistical modelling are employed. Indeed, the results of our generalisation analyses make it likely that our findings are applicable to a great variety of learning paradigms. Thus, we believe that the present study represents a major step forward in identifying a global neural architecture that determines individual abilities of learning and extinction, as well as the propensity for renewal. We will outline this in the following sections, one by one. Different types of connectivity within the core learning network Our results, shown in Fig. 1B-F , demonstrate that the three types of connectivity – expressed by structural, functional and effective connectivity – are clearly dissociable, suggesting that they reflect different neurophysiological processes. For example, HIP-AMY connectivity was noticeably higher than any other connection, which is not surprising given the myriad animal fear learning studies showing a tight coupling between these two regions (for reviews see 12 , 14 , 15 ). Importantly, however, the relative connectivity strengths were not always equivalent between FC, SC and EC – for instance, ACC and PFC showed stronger structural connections than HIP-PFC, even though FC was higher for the latter connection. Similarly, cerebellar effective connectivity to HIP, AMY, and PFC showed some of the strongest effects even though they shared the least number of streamlines (see Fig. 1C ). This dissociation between the three types of connectivity is the basis for their putatively different functional implications. The EC results revealed mostly inhibitory connections, with the notable exception of the HIP→AMY excitatory connection. Given that long-range connections are assumed to be predominantly excitatory 42 , our finding that most extrinsic (inter-areal) connections were inhibitory may appear puzzling. However, recent evidence suggests that a homeostatic brain makes abundant use of inhibitory connectivity, which allows for an effective control of the stability of memory patterns 43 , 44 . One mechanism contributing to this inhibitory network is through feedback-driven regulation. Neural predictors of acquisition Fear acquisition involves a conditioned stimulus (CS) that does not elicit any response on its own and an unconditioned stimulus (US) which elicits strong responding without need of prior training. Our analyses of FC patterns show that pre-existing functional interactions reliably predicted the amount of learning across the highly diverse fear conditioning and predictive learning paradigms that we employed. Since FC is strongly state-dependent 45 , our results point towards the relevance of pronounced intra-individual differences in learning. This is substantiated by the observation that acquisition could not be predicted by arguably more stable and trait-like patterns of structural connectivity. The gradually increasing learning curve observed across participants often results in the misleading interpretation that an association is learned slowly and with similar speed. However, individual learning curves often reveal a step-like rise of underlying associative strength that occurs with high differences 46 . Our results suggest that intra-individual differences captured by FC contribute to this variance. Fear conditioning and predictive learning are known to go along with ACC activation 23 , 24 . Accordingly, we found that ACC connectivity to HIP, PFC, and AMY were among the most reliable predictors of acquisition across the entire sample (see Fig. 3B and 5A ). The HIP has been commonly linked to the formation of new episodic memories 47 , whereas connectivity with the PFC may reflect recruitment of appraisal processes 16 . Furthermore, projections from ACC to AMY appear essential in mediating fear behaviour 48 , and our results suggest that this effect also applies to PL. Thus, the functional relevance of HIP-ACC, ACC-PFC and AMY-ACC connectivity may be related to the acquisition of novel memories and the evaluation of the motivational significance of the CS. Download figure Open in new tab Figure 5. Graphical depiction of all models in which significant connections between nodes were found. (A) Summary of the functional connectivity (FC) model using the combined sample of cognitive predictive learning and fear learning studies, showing significant connections in the acquisition phase. The presence of a connection in this figure reflects its selection by LASSO in Fig. 3 . For example, in Fig. 3B , LASSO identified a relevant ACC-HIP FC connection but not an ACC-CEB FC connection, which is depicted here as a line connecting ACC-HIP but not ACC-CEB. The size of each node represents its relative importance, as determined by Poisson regression analyses ( Fig. 3E ), with dashed outlines indicating non-significant nodes. (B) Same as (A) but showing the structural connectivity (SC) connections significant in the extinction phase. (C) Same as (A) but showing the SC connections significant in the renewal phase. (D-E) Graphical depiction of the PL (D) and FL (E) model, showing the effective connectivity (EC) connections significant in the acquisition phase. (F-G) Same as (D-E) but for the extinction phase. (H-I) Same as (D-E) but for the renewal phase. The bottom panels (D-I) integrate results from spectral DCM and LASSO analyses. For example, in Fig. 5E , the AMY→HIP connection is shown as a red arrow with a minus sign. The spectral DCM analysis ( Fig. 1E ) indicated that this connection was inhibitory (blue colour), which is indicated by a - sign in this figure, while the LASSO model for acquisition found a positive coefficient for this connection, which is indicated by the red arrow. In this way, the figure visually conveys which connections are inhibitory (-signs) and whether learning benefited from reduced inhibition (positive coefficients, red arrows) or increased inhibition (negative coefficients, blue arrows). Connections with alternative red and blue arrows indicate that their effects varied by hemisphere. Red and blue arrows indicate whether the association between learning and connectivity was positive or negative, respectively (see tile plots of Fig. 3B-D ). Plus and minus signs indicate excitatory and inhibitory connections (see Fig. 1E ). Please note that red arrows on inhibitory connections (-signs) indicate that disinhibition of these connections was beneficial in a given learning phase, while blue arrows on inhibitory connections (-signs) indicate a benefit of more pronounced inhibition. ACC = Dorsal anterior cingulate cortex; AMY = Amygdala; CEB = Cerebellar nuclei; HIP = Hippocampus; PFC = Ventro-medial prefrontal cortex. The connection to CEB might have a different character. During fear learning, large parts of the cerebellar cortex are activated by the prediction error 17 . Firing patterns of the cerebellar fastigial nucleus regulate fear-learning via thalamo-prefrontal dynamics, freezing behaviour through the periaqueductal grey, and anxiety behavior via thalamo-amygdala systems 49 . Given that during Pavlovian conditioning, the CS usually precedes the US with high contiguity, the CEB could serve as a fine-tuned predictor for the timepoint of the occurrence or the absence of the US. In summary, fear-related information from cortico-thalamic pathways may arrive in the AMY for initial processing (“quick and dirty” route 50 ). AMY connection to ACC and onwards to PFC could have important roles in appraisal and, for FL, suppress excess AMY activity, while HIP and CEB are implicated in the formation of new memory traces and indicating time points of prediction error information, respectively. These types of information may be relayed to the ACC, which, in turn, conveys this information back to the AMY for final integration and gating of fear responses. Neural predictors of extinction learning During extinction, the CS is no longer followed by the US. This ignites an expectancy-driven prediction error that does not erase the CS-US acquisition memory but establishes a second inhibitory trace that suppresses the occurrence of the conditioned response 2 . These learning events were only predicted by SC across the entire sample, but not by FC or EC. This suggests that the propensity of extinction is contingent on stable individual traits. Consequently, extinction learning is related to personality traits including tolerance of uncertainty 51 and trait anxiety 52 , and is particularly sensitive to rather stable microstructural white-matter measures and cortical thickness of selected emotional circuits 52 , 53 . As with acquisition, the ACC took a central role in extinction learning. Not only did its connectivity with HIP, AMY and CEB predict the speed of extinction, but also the connections involving the ACC were by far the most reliable. Indeed, two recent human neuroimaging meta-analyses reported activation of this region as the most consistent finding during extinction 25 , 54 , suggesting the ACC may function as a “hub” within the core learning network 55 . In addition to ACC, an intact HIP-PFC pathway seems crucial for the formation (and recall) of fear extinction and its contextual modulation in animal 15 as well as human neuroimaging studies 56 . Furthermore, PFC-AMY connectivity predicted extinction learning, corroborating early human neuroimaging connectivity findings 18 , 19 as well as results from rodents and non-human primates 57 , 58 . The integrity of the PFC-AMY pathway may thus be critical in allowing the PFC to regulate the formation and maintenance of extinction memories by active suppression of AMY output 14 . Finally, extinction could be predicted by HIP-AMY SC. Given their roles in the encoding of contextual representations and expression of CRs, respectively 14 , 15 , a stable HIP-AMY pathway would ensure that their integration results in a contextually-appropriate response 59 . Our results could imply that the extent of this ability is a trait factor. Regarding EC, analysis of FL indicated that inhibition of ACC by both PFC and HIP was related to extinction. Also, just as with acquisition, PFC suppression of AMY activity improved fear extinction in FL studies. We also found a modulation of AMY activity by HIP input (see Fig. 5G ). Since extinction learning is context-dependent, the excitatory HIP→AMY pathway would enable the HIP to relay back the crucial CS-context information necessary for the formation of context-dependent extinction memories in the AMY. In sum, our results indicate that structural connections among PFC-HIP-AMY may form part of a circuit in which the PFC suppresses excess AMY excitation (for FL only) under the HIP-driven representation of the appropriate context (for both FL and PL). The ACC may be a critical “hub” that synchronises these interactions and delivers the output to AMY for final integration. Neural predictors of renewal As outlined above, extinction is not an erasure of the old association, but the formation of a new memory trace of inhibitory nature. This can be demonstrated by several phenomena of which renewal is possibly the most interesting one. Renewal is the recovery of the extinguished response, induced by changing the context from that of the extinction phase back to that of acquisition. By this, it becomes obvious, how much context-dependent extinction learning is 2 , 60 . Across all studies, individual differences in renewal could only be explained by EC. Even though we cannot discard the possibility that absent SC (or FC) associations may partially relate to lower statistical power (only studies S2 and S4 were included), the EC results, nevertheless, indicate that the inhibitory network (see Fig. 5H-I ) convey more information about renewal than that afforded by either SC or FC. Also note that rs-fMRI was recorded shortly before acquisition/extinction, whereas renewal was tested the day after. If renewal is related to state variability, and if FC is more susceptible to decay over time than EC 61 , this could explain the prevalence of EC-related associations with renewal. The communication between HIP and PFC was particularly involved in renewal. Indeed, fear renewal increases connectivity between HIP and PFC 35 , inactivation of either of these two components reduces fear renewal 62 , 63 , and both PFC and HIP activity during extinction is positively correlated with renewal in predictive learning 21 – 23 . Renewal is also strongly context-dependent. A recent computational study made it likely that hippocampal replay is necessary and sufficient to generate context representations in the PFC 64 . Indeed, only participants with stronger hippocampal activation during extinction in a novel context relative to the acquisition context show renewal 65 . Accordingly, our PL studies showed that greater renewal was associated with HIP disinhibition by the PFC. Also, the unidirectional AMY→HIP connection was relevant for renewal. In rodents, AMY projections to ventral HIP modulate affective states 66 , 67 , such that this pathway may play a major role in the reemergence of fear memories 59 . Indeed, renewal probably represents a re-activation of the neuronal ensemble that was constituted during acquisition 60 . Simultaneous recordings from amygdala and hippocampal CA1 in fear-conditioned mice show synchronized activity that is related to the fear associated CS and results in freezing 68 . Thus, synchronized AMY→HIP connections could mediate fear memory retrieval that is associated with the re-exposure of the contextual cues that were present during acquisition but absent during extinction. This fits with the observation that a specific activation of a sparse ensemble of hippocampal cells elicits the recall of a context-bound memory engram of fear 69 . In summary, our results indicate that renewal is driven by the interaction of PFC and HIP, possibly by hippocampal replay that generates representations of the critical context in the PFC. Additionally, AMY-input to HIP possibly mediates the activation of fear-related memories in context-dependent ways. This could imply that prefrontal fear memories become contextually bound by hippocampal replay and can subsequently be activated during renewal by AMY activations that process the context that was present during acquisition. Generalisability and clinical relevance of our findings Across three different types of generalisability analysis, we were able to show that our model predictions are not only transferable between FL and PL studies, but also that the connections identified during resting states predict learning even when they are applied to task-based FC. A triple dissociation using across-studies predictions confirmed the higher predictive power of FC for acquisition, SC for extinction, and EC for renewal. The simulation results also indicate that our model could potentially be applied to new data – using our selected connections resulted in more explanatory variance than using pseudo-random connections from the same model. The generalisability of our findings may also have clinical implications for the psychotherapy practice. Anxiety disorders are typically treated within a particular therapeutic context, so one major challenge is to ascertain how to reduce fear over and beyond the therapeutic setting. To translate basic research to the treatment of fear-related disorders, it is crucial to understand how fears are both acquired and inhibited. We believe our study is a step closer to that aim. Across a large group of (healthy) participants and different paradigms, we showed that fear and extinction learning can be reasonably predicted by only a few connections between selected brain regions. The finding that the propensity for extinction may be more hardwired than renewal could imply that rather than optimising a person’s extinction ability, it may be more efficient to target (i.e., prevent) renewal. Fear learning and extinction are putative core mechanisms in the psychopathology and treatment of affective disorders. Thus, our findings may provide a foundation for future research into how individual differences in connectivity patterns related to learning and extinction contribute to differences in the risk for (or resilience against) affective disorders, and inform potential therapeutic interventions. Thus, our results complement current research in precision medicine on the role of different resting-state networks for distinct biotypes of affective disorders (including depression), which may help guiding decisions about specific therapeutic interventions 29 . Specifically, they indicate that structural connectivity is more directly related to extinction learning – and thus possibly, exposure therapy – than is functional connectivity, and suggest that the individual strengths of structural network connections should be considered in clinical populations in addition to currently applied measures of functional connectivity 70 . Finally, our data show that both acquisition and extinction involve extensive interactions with PFC and ACC. Indeed, research on the neural bases of cognitive-behavioural therapy (CBT; the most popular psychotherapy treatment for anxiety-related disorders) has suggested that therapy success depends on stable inhibitory control of AMY by PFC and ACC 71 . Thus, modulating the PFC/ACC during extinction (e.g. by altering schema representations in these regions) could potentially affect subsequent novel fear learning. This could be achieved by using non-invasive techniques such as repetitive transcranial magnetic stimulation (rTMS) and/or theta band transcranial alternating current stimulation (tACS), both of which have been shown to being able to specifically modulate PFC/ACC function 55 , 72 , 73 . Limitations of the present study One important limitation of the present study was the relatively small number of ROIs in our network. The inclusion of only a subset of regions was driven not only by theoretical reasons (see Supplemental Fig. S36 and Table S1), but also a statistical one. When computing EC, the number of possible ipsilateral connections is given by the formula 2*(n 2 – n). Thus, the addition of only two ROIs would have more than doubled the number of predictors in our models, surpassing the number of participants in some of the experiments. Future studies could combine our brain connectivity findings, while exploring additional connections which, together, may enhance predictive power. In addition, the lack of symmetry in explaining learning from the same connection in both hemispheres was surprising, given that the connectivity estimates were similar between hemispheres. This result may have partly been due to an inherent feature in LASSO’s variable selection procedure. Left and right hemispheric connections are often highly correlated. LASSO may, thus, favour the connection from one hemisphere which is a slightly better predictor than the corresponding connection from the other hemisphere, and penalise the other, leading to asymmetries in the identified contributing connections. Even though this feature does not directly affect the conclusions reached in our study, future work could explore the impact of different regularization strategies to better account for the relative contributions of both hemispheres in predictive modelling. Finally, all experiments analysed here fell under either the fear learning or the cognitive predictive learning paradigm. However, in order to establish the validity of the core fear and learning network, it will be important that future studies include a greater variety of paradigms, such as blink conditioning, appetitive and olfactory learning. Conclusion Individual differences in learning and extinction are core determinants of cognitive flexibility and are believed to be crucial for explaining treatment success of fear-related disorders. Despite its obvious fundamental relevance and clinical importance, the lack of consensus regarding the neural mechanisms of underlying individual abilities in learning and extinction has hindered progress in the translation of neurobiological models of extinction to clinical applications. We believe our study takes a step forward in bridging that gap. By a careful process of homogenisation, using novel approaches to identify subject-specific learning across different paradigms, applying complementary types of brain connectivity, and conducting state-of-the-art statistical modelling, we were able to show both similar and distinct neural mechanisms of learning and extinction across a multitude of paradigms. These results have profound implications for understanding why the abilities of learning and extinction, as well as the propensity to show renewal, are highly variable in both healthy and clinical populations. Methods Participants. The study described here was conducted as part of a large-scale collaborative research project, SFB1280 “Extinction Learning” ( sfb1280.ruhr-uni-bochum.de ). Participants were recruited for different studies within this project. In addition to task-based fMRI, participants in these studies also took part in resting-state and/or diffusion-weighted imaging scanning sessions. Only participants with neuroimaging data from at least one of these modalities were included in the present analysis (see Fig. 1A ). This study was approved by the respective local ethical committees, and all participants gave written informed consent and were monetarily reimbursed or received course credits (see Table S2 for demographical information). For resting-state fMRI, 513 participants in total were included in 6 different studies: S1, N=28 [age = 24.4 (3.51 SD), 19 women]; S2, N=152 [age = 22.0 (2.20 SD), 95 women]; S3, N=44 [age = 23.5 (3.56 SD), 22 women]; S4, N=177 [age = 25.8 (4.19 SD), 89 women]; S5, N=56 [age = 24.1 (3.74 SD), 26 women]; S6, N=56 [age = 26.3 (4.66 SD), 38 women]. For diffusion-weighted imaging, 467 participants in total were included in 5 studies: S2, N=166 took part in S2 [age = 21.9 (2.17 SD), 103 women]; S3, N=44 [age = 23.5 (3.56 SD), 22 women]; S4, N=175 [age = 25.7 (4.04 SD), 85 women]; S5, N=56 [age = 24.1 (3.55 SD), 25 women]; S6=56 [age = 26.5 (4.85 SD), 27 women]. Scanning sequences All MRI images were acquired using a 3T MRI scanner (S1, S2, S4, S5, Philips Achieva; S3, Siemens MAGNETOM Vida; S6, Siemens Skyra). Because participants were scanned at three different locations using 3 different 3T MRI systems from two different vendors, distinct sequence and imaging parameters had to be used. To ensure that our connectivity estimates were stable across sites and time, we scanned two travelling heads over the course of three years on the three different scanners in which the imaging data for the actual studies were acquired. Both resting-state and diffusion weighted imaging scans were acquired using the exact same parameters that were used in the actual individual studies (see below). Fifty-six ROIs selected from FreeSurfer’s automatic parcellation/segmentation (Fig. S1A) were used to compute FC between all pairs of ROIs, as well as fractional anisotropy (FA) within each ROI. Test-retest reliability, estimated using Cronbach’s alpha, was very high across all sessions, for each site, and for each travel head (αs > 0.84, ps .60, p FDR s < .001; Fig. S1B, middle). In addition, FC and FA estimates for each ROI remained relatively stable across the different sessions (Fig. S1B, right). Similar results were obtained when using only the ROIs selected for the main study (Fig. S2). For the high-resolution T1-weighted (MP-RAGE) structural images the following parameters were used: TR = 8 ms, TE = 4 ms, flip angle = 8°, voxel size = 1 x 1 x 1 mm 3 , FOV = 24 x 24 cm (studies S1,S2,S4,S5); TR = 2.53 ms, TE = 2 ms, flip angle = 7°, voxel size = 1 x 1 x 1 mm 3 , FOV = 19.2 x 25.6 cm (study S3); TR = 1.77 ms, TE = 3 ms, flip angle = 8°, voxel size = 1 x 1 x 1 mm 3 , FOV = 19.2 x 25.6 cm (study S6). Whole-brain T2*-weighted images during rs-fMRI were acquired using a gradient echo, echo-planar imaging (EPI) sequence with the following parameters: TR = 2.5 s, TE = 30 ms, flip angle = 90°, voxel size = 3 x 3 x 3 mm 3 , FOV = 24 x 24 cm, 80 x 80 voxels, number of slices = 47, number of volumes = 190 (studies S1,S2,S4); TR = 1.43 s, TE = 30 ms, acceleration factor = 2, flip angle = 69°, voxel size = 3 x 3 x 3 mm 3 , FOV = 24 x 24 cm, 80 x 80 voxels, number of slices = 48, number of volumes = 190 (study S3); TR = 2.5 s, TE = 30 ms, flip angle = 90°, voxel size = 3 x 3 x 3 mm 3 , FOV = 28 x 31 cm, 92 x 92 voxels, number of slices = 46, number of volumes = 192 (study S6). Diffusion-weighted images were acquired using the following parameters: TR = 9.5 s, TE = 88 ms, flip angle = 90°, voxel size = 2 x 2 x 2 mm 3 , FOV = 24 x 24 cm, 112 x 112 voxels, number of slices = 60, number of directions = 60 (b = 1000 s/mm 2 ) (studies S1,S2,S4); TR = 5.5 s, TE = 114 ms, flip angle = 90°, voxel size = 1.6 x 1.6 x 1.6 mm 3 , FOV = 24 x 24 cm, 132 x 128 voxels, number of slices = 60, number of directions = 60 (b = 1000 s/mm 2 ) (study S3); TR = 10.2 s, TE = 87 ms, flip angle = 90°, voxel size = 2 x 2 x 2 mm 3 , FOV = 60 x 60 cm, 120 x 120 voxels, number of slices = 70, number of directions = 60 (b = 1000 s/mm 2 ) (study S6). Experimental paradigms S1: The paradigm for this study consisted of four phases over two days: fear acquisition and fear reversal (day 1), followed by fear extinction (day 2). Participants viewed images of household appliances (16 CS in total), some paired with an electric shock (US). CS were presented for 1s, embedded within 2s video contexts that preceded them. The US, when delivered, lasted 0.75s and followed the CS presentation. During fear acquisition, half of the CS were reinforced (CS+; 50% probability) whereas the other half were not (CS-). In fear reversal, contingencies were switched for half of the CS. US expectancy ratings were collected on each trial using a 4-point scale (2.5s duration), and trials were separated by a fixation cross (7–9s). S2: In this study, participants viewed office scenes where a desk lamp’s colour indicated CS type: one colour (CS+) was paired with a shock (US; 62.5% probability), while another (CS-) was not. Each trial included a fixation cross (6.8–9.5s), a context image (1s), and the CS presentation (6s). Fear acquisition involved 16 trials per CS type, followed by extinction and renewal phases without reinforcement. S3: In this study, two geometric shapes served as CS, with one (CS+) paired with a shock (US; 62.5% probability) during acquisition, while the other (CS-) remained unpaired. The experiment spanned two days: habituation (6 trials), acquisition (16 CS-, 16 CS+, 10 CS+US trials), and extinction (16 trials per CS type) occurred on day 1, while recall (not analysed) was on day 2. Trials lasted 8s, with shocks (when present) delivered at 7.9s. Inter-trial intervals varied between 14.3s and 17.9s. S4: Participants learned to predict whether specific foods would cause a stomach-ache based on context (restaurant). During acquisition (80 trials; 8 stimuli × 10 repetitions), each food was shown in one of two contexts for 3s, followed by a question screen (max 4s) and 2s feedback. In extinction (80 trials), half of the stimuli were shown in the same context (AAA), half in a different one (ABA); extinction and distractor stimuli were included. The renewal phase (24 trials; 3 repetitions per stimulus) was conducted in the original context without feedback. Inter-trial intervals varied between 5–9s. S5: In this study, participants viewed three geometric shapes (CS+G, CS+N, CS–) matched in luminescence and surface area. During acquisition (day 1), CS+G and CS+N were followed by an electrical shock (US; 62.5% probability), while CS– was never reinforced (8 trials). Each trial lasted 20Zs: a jittered 0–2.5Zs black screen, 8Zs CS presentation, and 9.5–12Zs inter-trial interval. During extinction (day 2), each CS was shown 8 times; CS+N and CS– appeared in original size, while CS+G was presented across four sizes (100%, 75%, 50%, 25%, each size shown twice). S6: This paradigm included two randomized-controlled studies examining the effects of systemic inflammation on fear learning. On day 1, participants underwent acquisition training, where three visual CS were paired with either visceral pain (CS+US+vis), an aversive tone (CS+US+aud), or no US (CS–). A total of 36 CS trials were presented (12 per CS type), with 75% reinforcement for CS+ (9 US+vis, 9 US+aud). CS were shown 6–10Zs before US onset, and US lasted 14Zs, with CS and US co-terminating. On day 2, extinction included the same CS sequence presented without reinforcement. Inter-stimulus intervals consisted of a fixation cross (8Zs). Participants received either intravenous LPS or placebo 2Zh before acquisition (study 1) or extinction (study 2). Only responses to CS+US+vis were analysed. For more details regarding the description of these studies see the section “Experimental paradigms (extended text)” and Fig. S3 in the supplementary information. Preprocessing of neuroimaging data For the preprocessing of the resting-state fMRI data, fmriprep (version 20.1.1) was used, which included, removal of the first two volumes, motion correction, slice timing correction and co-registration to the T1w image. The BOLD time-series were resampled onto native space. For denoising, we extracted the expanded motion regressors from the fmriprep output (6 standard motion parameters, their quadratic terms and corresponding temporal derivatives; total of 24 regressors), in addition to the global signals and the mean signals within WM and CSF masks. In total, 36 regressors were used for denoising, as recommended by Satterthwaite and colleagues 14 . Next, we fitted a GLM using these sources of noise, and extracted the residuals of the resulting demeaned time series, which we then used for the functional/effective connectivity analysis described below. FreeSurfer was run within fmriprep, thus, additionally providing the segmentation and parcellation maps (in native space) which were needed for the ROI extraction (see below). For the preprocessing of diffusion-weighted imaging data, we initially ran the function dwidenoise from MRTrix3 ( https://www.mrtrix.org/ ), which implements dMRI noise level estimation and denoising based on random matrix theory, followed by mrde gibbs, which additionally removes Gibbs ringing artifacts. Topup was then applied in order to estimate and correct susceptibility-induced distortions, followed by eddy-current correction, in order to correct eddy currents and movements in the diffusion data 75 . Finally, we ran FSL’s tool eddy_quad for quality assessment of individual datasets. ROI extraction Two different parcellation maps from FreeSurfer (Desikan-Killiany and Destrieux; version 6) were used for extracting regions-of-interest (ROIs). A total of ten ROIs were extracted: amygdala (AMY), hippocampus (HIP), ventro-medial prefrontal cortex (PFC), dorsal anterior cingulate cortex (ACC) and cerebellar nuclei (CEB), for both the left and right hemispheres. We decided to include cerebellar nuclei as ROI and not the cerebellar cortex, because the cerebellar nuclei are the sole output structure of the cerebellum (see 76 for a recent review). The choice of ROIs was determined by the SFB1280 consortium prior to any data acquisition (see also Fig. S36 and Table S1, for a literature review implicating these regions in learning and extinction). AMY and HIP were taken from the automatic volumetric segmentation of the subcortical regions (aseg). We extracted the labels from the Desikan-Killiany atlas “medial-orbito frontal” and “caudal anterior cingulate”, which correspond to the PFC and ACC, respectively (see Fig. 1B and Fig. S4). Given that the automatic FreeSurfer’s parcellation does not output the cerebellar nuclei (only the cerebellum as a whole), additional processing was required to construct the CEB ROIs. We employed the SUIT pipeline ( https://www.diedrichsenlab.org/imaging/suit.htm ). After aligning the T1w image of each individual to the ACPC, the cerebellum was cropped from the rest of the brain and normalised using DARTEL for the purpose of matching it to the SUIT atlas (in MNI space). We then applied an inverse normalisation to reslice the SUIT atlas into the functional space of each participant. Finally, the cerebellar nuclei (interposed, dentate and fastigial nuclei) were extracted and merged into one single ROI ( Fig. 1B ). All ROIs were resampled into functional (rs-fMRI) and FreeSurfer space. For both functional and effective connectivity, the ROIs in functional space were used. For structural connectivity, all ROIs were kept in FreeSurfer’s native space, in order to take advantage of surface files (e.g., pial surface) that is known to improve the accuracy of tractography 77 . In addition, the PFC and ACC volumetric ROIs were also converted into surfaces (Fig. S4). These surface ROIs were only used for computing streamlines during probabilistic tractography. Tracking from surfaces from cortical brain regions has advantages relative to their volumetric counterparts 77 . We also extracted the bilateral thalamus using FreeSurfer’s automatic segmentation. The thalamus of the contralateral hemisphere with respect to the seed/target CEB was used as a waypoint to more accurately guide tractography that included CEB ROIs (see Structural Connectivity section below for more details). Functional connectivity (FC) Recently Mohanty et al. 78 tested the accuracy of several FC metrics by evaluating a support vector machine classifier using a neighbourhood component analysis feature selection. They found that FC was better characterized by a combination of nine different but complementary metrics (a composite metric) than any metric alone. The authors reasoned that Pearson correlations - the most common measure of FC - only look for statistical linear time-dependencies, whereas there could still be underlying statistical dependencies between BOLD signals that are poorly captured by Pearson correlations (e.g., non-linear dependencies). Therefore, for the present study, we followed Mohanty’s recommendation and computed a total of nine functional connectivity metrics: Pearson correlation, cross-correlation, dynamic time warping, Euclidean distance, Manhattan distance, Wasserstein distance, mutual information, coherence and wavelet coherence (see Supplemental Methods, Table S4 for the mathematical implementation, Fig. S7 for correlations among these metrics, and Fig. S8 for comparisons among the different FC metrics within our network). For the functional connectivity analysis within the core learning network, a multilevel model was fit with the learning estimates as the outcome variable and the connectivity values for the 20 ROI pairs as the predictors of interest. Age and sex were included as covariates, and participant nested within study was treated as a random effect. Correction for multiple comparisons between the different connections was performed using the false discovery rate (FDR) method. Effective connectivity (EC) EC was estimated using spectral dynamic causal modelling (spDCM), a toolbox for SPM12 31 . spDCM is a variant of standard DCM that it especially developed for modelling resting-state data. It is based on the cross (power) spectral density of the observed BOLD time-series (see the Supplemental Methods for a comprehensive mathematical description of spDCM; see also Fig. S10 for the correlation between EC estimates and those from the FC metrics coherence and cross-correlation). From the preprocessed and denoised functional images, a single representative timeseries was computed for each ROI by performing a principal component analysis across voxels and retaining the principal eigenvariate. This procedure has the advantage that it is more robust to outliers when computing EC. The DCM model was then specified for each participant. Because we were interested in ipsilateral connections (with the exception of the contralateral connections involving the cerebellum), we set a prior with zero mean and zero variance for those connections that we wanted to exclude from our models. Estimates of parameter uncertainty were computed by extracting the diagonal of the covariance matrices for each individual connection (these estimates were used in the EC LASSO models described below). After spDCM was computed for each subject individually, a structure was returned with the values in the matrix containing the mean value of the Gaussian distribution for the connections of interest (commonly known as A-matrix). These mean values comprise the connectivity parameters that indicate the individual-level effective connectivity from one brain region to another brain region. Subsequently, we computed a group-level DCM by assembling all subject-level DCMs and providing them as an input to the group DCM analysis using the framework of Parametric Empirical Bayes (PEB). PEB takes into account the estimated covariance between parameters and subject in a random effect analysis (see Supplemental Methods for technical details about PEB). Only extrinsic connections with posterior probabilities above 95% (which is equivalent to a log Bayes factor of 3) were considered, which corresponds to “strong evidence” within a Bayesian framework. Structural connectivity (SC) After preprocessing the diffusion data, we proceeded with the tractography analysis using FSL. First, we fitted a diffusion tensor model at each voxel, which returned the first three eigenvectors and corresponding eigenvalues, as well as a fractional anisotropy (FA) map. Because all subsequent analyses were done in diffusion space, we used the FA map to create all necessary transformation matrices from diffusion (dwi) to anatomical (T1w) and FreeSurfer spaces. The reason for using the FA image was because it had a more similar contrast to the anatomical image than non-weighted diffusion (B0) volumes, so it provided a slightly more accurate registration. All transformation matrices (diffusion -> FreeSurfer, FreeSurfer -> diffusion, diffusion -> T1w, T1w -> diffusion) were computed using boundary-based registration (BBR). Next, we ran an analysis using “Bayesian Estimation of Diffusion Parameters Obtained using Sampling Techniques” (BEDPOSTX), which models crossing fibres within each voxel of the raw diffusion data, and creates the distributions on diffusion parameters at each voxel. Probabilistic tractography was then estimated using the samples from the distributions above. Note that tractography was run in diffusion space but the transformation matrix diffusion -> FreeSurfer was provided, since all ROIs were in FreeSurfer space and the pial surface was used as a stopping mask. For the tracking of fibres from cortical regions (i.e., ACC and PFC) we used surfaces instead of volumetric ROIs, as it has been suggested that seeding from surfaces is superior for the cortex 77 . However, for subcortical regions and cerebellar nuclei, we used the volumetric ROIs, since these regions tend to contain a high degree of anisotropy. As a means to guide tractography, we included the pial mask as a stopping mask, which effectively prevented tracts from crossing the grey/white-matter interface (which would be biologically implausible). Furthermore, for any seed regions, we further restricted tractography to exclude streamlines that did not stop by the target region. In other words, tracts were terminated and included in the total streamline count if and only if they reached the target region from the seed region. Finally, we also constrained tracts to bypass any regions containing non-white matter voxels (e.g., CSF, skull). One limitation with the approach described above is that it can only be meaningfully used for ipsilateral connections, as including the grey/white-matter interface will inevitably discard any streamlines that attempt to cross hemispheres. This is particularly problematic for connections involving the cerebellum, since anatomical connections of the cerebellar hemispheres involve to a large extent the contralateral cerebral hemisphere with cerebellar output crossing at the level of the brainstem 79 . In order to provide a similar degree of tractography accuracy for the cerebellar connections, we restricted the movement of the tracts to/from the cerebellar nuclei by (1) defining two stop masks - the pial surface ipsilateral to the cerebellar seed, and the cerebellum hemisphere, and (2) using the thalamus as a first waypoint, such that only tracts that initially run via the thalamus en route to the target ROI were considered valid streamlines. The rationale for point (2) is that output of the cerebellar nuclei is connected with various cortical areas via the thalamus 79 . Learning measures SCR measures For analysis of the SCR data we used the Matlab toolbox Psychophysiological Modelling (PsPM, version 6.0.0) 80 . Similar to analysis of fMRI data, PsPM creates an explicit mathematical “forward” model of the data-generating process with unknown parameters. This model is then inverted to yield the most likely parameter values, given the data. In the present case, we estimated the amplitudes of centrally generated CS- and US-related sudomotor responses on each trial, which serve as learning index. The raw data for each subject were initially trimmed to the task duration and, subsequently, filtered using an adaptive filtering algorithm with varying amount of time points, such that an optimal number could be determined that maximally reduced gradient artefacts. Subsequently, residual artifacts in the filtered data were detected using an automated quality assessment procedure 81 . Finally, each dataset was visually checked for residual artifacts that may have been missed by the automated artifact detection tool, which were marked by one of three researchers with extensive training in preprocessing SCR data. Any detected artefact periods were logged and later ignored during data analysis. Participants with an excessive amount of artifacts were excluded from further statistical analysis, which was agreed by the three researchers, who were blinded to the results of the actual analysis. Note that most of these exclusions were the result of failed recordings due to technical difficulties with the equipment or sudden abortion of the scanning session. In total, 56 acquisition and 79 extinction datasets were excluded due to flat or extremely coarse SCR data (see Fig. S13 for examples of typical excluded datasets). After preprocessing, data for all but one study (S6) were analysed with a standard non-linear model 82 using a canonical skin conductance response function 83 (see also Supplemental Methods for a mathematical description of this model). This approach is appropriate for the CS—US duration used in these studies (around 7s). For all phases (acquisition, extinction and renewal), we modelled the following events: Context (fixed latency), CS onset (fixed latency), CS interval (flexible latency, sudomotor burst dispersion fixed at 0.3s) and US (fixed latency). In order to avoid bias, PsPM assumes the same event sequence for all trials, so in the case of non-reinforced trials (no US), we modelled US omissions by adding an event of the same duration as the actual US in the place where the shock would have occurred during reinforced trials. To exclude the possibility that the estimated response to the CS could have been confounded by the response to the US (due to the overlap the elicited SCR), we applied two additional steps. First, we gradually decreased the modelled time interval between CS onset and US onset, to the point that there would be no differences between CS+US+ and CS+US-. Second, we ensured post hoc that the estimated sudomotor impulse did not overlap in time with the onset of the US. Both these procedures ensured that the response to the CS during reinforced trials was unlikely to have been contaminated by the response to the US. In consequence, there was no evidence for a difference between CS+US+ and CS+US-(in all projects, ps > .10, uncorrected), suggesting that our method was not biased by the US presence. For the acquisition phase of S6, in which the US duration was very long (14 seconds), we used a different approach. Crucially, the standard non-linear model requires assumptions on the number and distribution of anticipatory responses during the CS—US interval, and these have not been thoroughly tested and validated in long-interval paradigms. Hence, we opted for a method that dispenses with these assumptions and has been successfully used to estimate spontaneous fluctuations which can occur any time 84 . Specifically, we modelled one response per two seconds and estimated its amplitude and onset, across the entire acquisition. We then assigned the estimated sudomotor response to the experimental events that occurred at the same time. Next, we summed the estimated amplitude of all response occurring during each CS, and divided them by CS duration, thus providing a score of the anticipatory sudomotor activity per CS. Once the modelling was completed for all studies, the results for each trial were manually inspected to ensure that a proper model fit was attained. Finally, for each subject, we extracted the amplitude for each trial and for each phase of the experiment. These data were subsequently analysed using R (version 4.2.2; https://www.r-project.org/ ). Using each study’s parameter estimates, we then selected the conditions of interest that we wished to model (see Table S3). Some studies (e.g., S5) contained fewer than four non-reinforced CSs in their experimental design, which did not allow for an accurate assessment of learning across time. Therefore, we combined reinforced trials (CS+US+) and non-reinforced trials (CS+US-) in our modelling. As noted above, we ensured the amplitude differences between CS+US+ and CS+US-were indistinguishable and were not significant in each single study. Further tests confirmed that the response amplitudes for CS+US+ were not contaminated by subsequent US responses (see Fig. S14). For generating a single participant-specific learning score that could characterise individual learning performance, we employed the following sequential steps (see Fig. 2B for a visual example): A theory-free polynomial regression of the 2nd order was conducted on the amplitude scores extracted from PsPM as the dependent variable; trial number, CS type and their interaction were used as predictors of interest. After the model fit, the model predictions were extracted for each participant. The difference in the predictions between each trial and the previous trial was calculated for each CS type separately and summed into a single score. Thus, each participant’s learning was characterised by two scores (one for CS+ and one for CS-). The difference between the CS scores (CS+ minus CS-) gave us an indication of the learning rate relative to the baseline. Behavioural responses The behavioural data in S4 consisted of binary decisions (whether a food item gave stomach-ache) as the dependent variable (see above for the description of the paradigm used in S4) and trial number as a predictor of interest. Because we were interested in individual rates of learning, we ran a multi-level logistic regression analysis using a random intercept for each participant and a random slope for trial number. The extracted coefficients included, for each participant, an intercept, representing the average learning, as well as a slope for trial number, representing the learning rate. However, the slopes derived from logistic regression models cannot be interpreted on their own, since we require at least two parameters (i.e., β 0 and β time ) to determine the shape of the logistic curve. Specifically, β 0 indicates the general ability of the subject, whereas β time indicates the subject’s ability to learn through time without considering the overall skill. Therefore, we computed the probability of success for each participant and trial, and, subsequently, calculated the expected number of correct trials after 8 attempts (since there were always 8 unique trial types in each experiment) based on these probabilities. Scores closer to 8 would mean that participants learned successfully, whereas scores closer to 4 would mean that participants were at chance levels. These scores were thus used as a proxy for how well each individual learned through time (see Fig. S15 for a distribution of these scores). This procedure was used for the analysis of acquisition, extinction and renewal. Predicting individual differences The main goal of the present study was to predict the efficacy of learning and extinction by multimodal brain connectivity patterns. This purpose required three learning parameters per participant (one for acquisition, one for extinction and one for renewal) in each and every study included in the present research. Different studies utilised different experimental paradigms, different number of trials/conditions, and different dependent variables. To reduce this inhomogeneity, we used a modelling strategy such that the different types of models that we implemented for the SCR and behavioural studies resulted in similar learning parameters (i.e., a single score per subject that reflected learning across time during acquisition, extinction and renewal). Furthermore, we standardised the learning estimates on a study-by-experiment level, such that each study/experiment had an average estimate of 0 and variance of 1. This step was important because it brought all study-experiment combination into a common unit of measurement, while preserving the relative distances between individuals within that study-experiment combination. An additional advantage of this procedure is that it effectively helps to reduce differences across studies due to unwanted site effects (e.g., scanner, sample size, etc.). To explore how much non-trivial variation was present after standardisation, we ran a simple nested multilevel model using participant, study and experiment as separate random effects and the standardised learning variable as the dependent variable. The variance estimates of the random effects “study” and “experiment” were zero (see Table S3), indicating a degenerate model (i.e., the between-study and between-experiment variability is insufficient to justify their inclusion as random effects, over and above the participant random effect). Because all models above were fitted using maximum likelihood, we could compare models with and without the random effects of study and/or experiment. The results of these model comparisons indicated equivalent AICs to using participant as the sole random factor (see Table S4). LASSO regression models were then built using learning score as the dependent variable, and either FC, EC, or SC values as predictors of interest (i.e., in three separate models). Age and sex were used as covariates in all models. Regularised regression methods such as LASSO can be difficult to interpret in the presence of multicollinearity since they will discard almost arbitrarily one of the collinear predictors. Thus, before each LASSO model, we ran a multiple regression analysis in order to assess several multicollinearity diagnostics, which included the variance inflation factor (VIF), tolerance, eigenvalues, condition indexes and variance proportions. None of the models showed a sufficiently high degree of multi-collinearity that would warrant further investigation 85 (see Fig. S17). Each model was built either using all studies or a subset of studies (see Table S8). LASSO was run with a k-fold cross-validation procedure (using the cv.glmnet function from the glmnet library), separately for acquisition, extinction, and renewal. The function initially randomly splits the data into 10 equal folds. For each fold (i.e., test data), the model is trained on the remaining nine folds and tested on the held-out fold. This process is repeated for all 10 folds, ensuring that each fold serves as a test set once. We then extract the average mean squared error (MSE) across these 10 different splits as a measure of how well the model generated from the training data can predict the test data. In addition to this inner cross-validation loop, we also included an outer loop of cross-validation (100 iterations) to ensure the stability of our LASSO results and determine the optimal regularization parameter λ. Specifically, for each of the 100 outer-loop cross-validations, cv.glmnet was run with a custom lambda path, and the MSE was computed across 10 folds. Each of the 100 cross-validations used different random splits of the data. To maximise the chances of finding a lambda with the minimum cross-validation error, we constructed a manual sequence of lambda values (i.e., the lambda path). This was done by supplying the function with a decreasing sequence of 1000 values from a maximal lambda to a minimum lambda based on the following formulas: In other words, the entire cross-validation was repeated 1,000 times with different regularization parameters λ, given by λ path computed above. Predictors were standardised within each fold prior to model fitting to prevent leakage 86 . In addition, the L1-regularisation penalty was applied to all predictors of interest except the covariates. The model was selected based on the lambda that minimised the average MSE across these 100 outer-loop cross-validations. The calculation of significance in regularised regression is controversial and an area of much debate, as p-values and confidence intervals in LASSO models are biased due to the regularisation process which is conducted to reduce variance in the parameter estimates 87 . Alternative methods to compute p-values have been suggested, but none have been found robust. In the present study, we defined the non-zero coefficients from LASSO, obtained via the cross-validation procedure described above, as the “significant” variables under the penalised optimisation framework L1. Nevertheless, we also provide standard p-values for each of the relevant connections from the LASSO models as comparison. LASSO was run on one-hundred random samples of observations, counting the non-zero connections, and testing each connection using binomial tests while correcting for multiple comparisons (see Fig. S22). Overall model performance was assessed for each experimental phase separately (acquisition, extinction and renewal). First, the data was split in half, such that one half was used as the training set and the other half as the testing set. Surrogate datasets were constructed by shuffling the learning-connectivity correspondences of the testing sets across subjects. The mean-squared errors (MSEs) were computed for the original and surrogate models, and the difference taken as a measure of performance. Permutation tests were performed by comparing the original difference against the null distribution, which was constructed by randomly multiplying the MSE differences by either −1 or 1 ten thousand times (see Table S11). For EC models, we also used the previously-computed estimates of variability for each connection (see above) as weights during model fit, so as to mimic the group-level PEB that also uses this information. Finally, in order to ensure our results were not due to the specific statistical LASSO method we used, we re-ran all of our models using standard multiple regression, ridge regression and elastic net. The LASSO coefficients were well within the range of the coefficients provided by these alternate methods, and there were very robust positive correlations among the coefficients from the different methods (Fig. S18). Generalisability analysis After fitting our LASSO models, we examined the generalisability of these models by running four different types of analyses. First, we grouped studies into two larger groups: a fear learning (FL; S1,S2,S3,S5,S6) and a predictive learning (PL; S4) group. We applied Monte-Carlo cross-validation to extract mean-squared errors (MSEs) of different combinations of the testing set participants. Next, we trained a LASSO model on all data from one paradigm (e.g., fear learning; using the same lambda path to compute the optimal lambda, as in the main LASSO model described above) and applied it to randomly selected 50% of data from the other paradigm (e.g., predictive learning). This was repeated 100 times, and each time we selected a different random subsample of 50% of the test data. The training data were always the same across the 100 repetitions. As a result of the 100 repetitions, we obtained a distribution of MSE values (blue error bars in Figures 4A-C ). Subsequently, we created surrogate datasets by randomly shuffling the relationship between learning and connectivity estimates in the test (unseen) data. Again, we selected 50% of these surrogate test and computed how well the model (based on the actual data in the training set) could predict these test data, resulting in a surrogate MSE value. This was repeated 100 times with different shuffles of the test data (black error bars in Figures 4A-C ). Then, the difference between the original and surrogate MSEs was used as a performance measure. Statistical significance was then assessed via using non-parametric permutation tests. Specifically, the null distribution was constructed by randomly multiplying each of the 100 MSE differences by either −1 or 1 with equal probability and again averaging these differences. Next, we recalculated the mean difference across 10,000 permutations, representing the expected variation if the observed differences were due to chance. Finally, we compared the observed mean difference to this null distribution and calculated a two-tailed p-value, which reflects the proportion of permuted means that were as extreme as or more extreme than the observed difference. In the second generalisability analysis, we employed a leave-one-group-out (LOGO) cross-validation by splitting the data such that each training set comprised of all studies except one (e.g., S2-S6, but not S1), and the left-out study (e.g., S1) used as the test data. A multiple regression was fit on the training data and the model tested on the testing data. The Pearson correlation coefficient for the predicted and observed values served as an index of generalisation. In the third generalisability analysis, we enquired whether our model predictions could be useful to ascertain task-based FC, since the model for the acquisition phase revealed some significant functional connections with respect to the entire sample. For that purpose, we selected one fear conditioning study for which task fMRI data were available (S2, two experiments, 152 participants) and computed FC during both CS+ and CS-trials using all FC metrics described above and for all combinations of the relevant ROIs. A multiple linear regression model was constructed using the significant connections from the overall LASSO model, and the model p-value was contrasted with one from a model that contained randomly-selected connections. As the vast majority of FC studies computed Pearson correlation coefficients as a proxy for connectivity, our FC predictors were also based on Pearson correlations. Finally, to extrapolate our findings to potential new data, we also ran two types of analyses: simulations and bootstrapping. For the simulations, we drew samples from a multivariate normal distribution since the outcome variable of learning was normally distributed (Fig. S19). The covariance matrices for each group (PL and FL) were computed for all numerical predictors and covariates, to account for correlations among these variables. To accommodate potential variations around the sample mean and covariance (as we would not expect the sample mean and covariance matrix of subsequent studies to equal exactly the empirical estimates), we allowed some degree of sampling error around our estimates by unscaling the variables. For the bootstrapping, we performed the same analysis as above but using bootstrap samples once with and once without replacement. Two multiple regression models were computed (one for FL and another for PL studies), using learning as the outcome variable, the connectivity pairs from the selected LASSO models as predictors and sex and age as covariates. One-hundred independent datasets were generated, each contained 40 observations (we tested various other sample sizes [90, 300, 1000], but the results were identical; see Fig. S20). For each dataset, r-squared from a multiple regression was computed for the actual LASSO predictors, as well as for pseudo-random predictors that were not present in the actual LASSO model but consisted of regions from the core learning network. Importantly, we selected the exact same number of pseudo-random connections as the LASSO model within each specific type of connectivity (e.g., if the LASSO model for effective connectivity contained 3 connections, the random model for the effective connectivity would also contain 3 connections not present in the LASSO model). We also computed connectivity within the lateral and fourth ventricle, as we did not expect these regions to predict learning, and, thus, functioned as an additional baseline. Acknowledgements This study was supported by the DFG SFB 1280 “Extinction Learning” (Project Nr. 316803389). We thank Bianca Hagedorn for her help in data collection, Cosima Clotten, Julia Stengel, Esther Yadgarova, Fabian Wissing and Lina Schmidt for their help with the PsPM analysis and figures, and Aarti Swaminathan for her help with some sections of the discussion. Funder Information Declared Deutsche Forschungsgemeinschaft, https://ror.org/018mejw64 , 316803389 Footnotes 1 For the dorsal anterior cingulate cortex and ventromedial prefrontal cortex we chose the acronyms ACC and PFC respectively, to be consistent with other three-letter labels (AMY, HIP, CEB) and avoid excessive lettering in ROI pair labels in figures and tables. References 1. ↵ Byrom , N. C . Accounting for individual differences in human associative learning . Front. Psychol . 4 , ( 2013 ). 2. ↵ Bouton , M. E . Context, ambiguity, and unlearning: sources of relapse after behavioral extinction . Biol. Psychiatry 52 , 976 – 986 ( 2002 ). OpenUrl CrossRef PubMed Web of Science 3. ↵ Bouton , M. E. , Maren , S. & McNally , G. P . Behavioral and neurobiological mechanisms of pavlovian and instrumental extinction learning . Physiol. Rev . 101 , 611 – 681 ( 2021 ). OpenUrl CrossRef PubMed 4. ↵ Craske , M. G. , et al. Anxiety disorders . Nat. Rev. Dis. Primer 3 , 17024 ( 2017 ). OpenUrl CrossRef 5. ↵ Rauch , S. L. , Shin , L. M. & Phelps , E. A . Neurocircuitry Models of Posttraumatic Stress Disorder and Extinction: Human Neuroimaging Research—Past, Present, and Future . Biol. Psychiatry 60 , 376 – 382 ( 2006 ). OpenUrl CrossRef PubMed Web of Science 6. ↵ Adolph , D. et al. Measuring extinction learning across the lifespan – Adaptation of an optimized paradigm to closely match exposure treatment procedures . Biol. Psychol . 170 , 108311 ( 2022 ). 7. ↵ Pattwell , S. S. et al. Altered fear learning across development in both mouse and human . Proc. Natl. Acad. Sci . 109 , 16318 – 16323 ( 2012 ). OpenUrl Abstract / FREE Full Text 8. ↵ Rattel , J. A. et al. Peritraumatic unconditioned and conditioned responding explains sex differences in intrusions after analogue trauma . Behav. Res. Ther . 116 , 19 – 29 ( 2019 ). OpenUrl CrossRef PubMed 9. ↵ Milad , M. R. et al. Fear conditioning and extinction: Influence of sex and menstrual cycle in healthy humans . Behav. Neurosci . 120 , 1196 – 1203 ( 2006 ). OpenUrl CrossRef PubMed Web of Science 10. ↵ Rattel , J. A. et al. Sensation seeking and neuroticism in fear conditioning and extinction: The role of avoidance behaviour . Behav. Res. Ther . 135 , 103761 ( 2020 ). 11. ↵ Sjouwerman , R. , Scharfenort , R. & Lonsdorf , T. B . Individual differences in fear acquisition: multivariate analyses of different emotional negativity scales, physiological responding, subjective measures, and neural activation . Sci. Rep . 10 , 15283 ( 2020 ). 12. ↵ Maren , S. & Holmes , A. Stress and Fear Extinction . Neuropsychopharmacology 41 , 58 – 79 ( 2016 ). OpenUrl CrossRef PubMed 13. ↵ Kuhn , M. , Mertens , G. & Lonsdorf , T. B . State anxiety modulates the return of fear . Int. J. Psychophysiol . 110 , 194 – 199 ( 2016 ). OpenUrl CrossRef PubMed 14. ↵ Maren , S. & Quirk , G. J . Neuronal signalling of fear memory . Nat. Rev. Neurosci . 5 , 844 – 852 ( 2004 ). OpenUrl CrossRef PubMed Web of Science 15. ↵ Maren , S. , Phan , K. L. & Liberzon , I . The contextual brain: implications for fear conditioning, extinction and psychopathology . Nat. Rev. Neurosci . 14 , 417 – 428 ( 2013 ). OpenUrl CrossRef PubMed 16. ↵ Etkin , A. , Egner , T. & Kalisch , R . Emotional processing in anterior cingulate and medial prefrontal cortex . Trends Cogn. Sci . 15 , 85 – 93 ( 2011 ). OpenUrl CrossRef PubMed Web of Science 17. ↵ Doubliez , A. et al. The cerebellum and fear extinction: evidence from rodent and human studies . Front. Syst. Neurosci . 17 , 1166166 ( 2023 ). 18. ↵ Phelps , E. A. , Delgado , M. R. , Nearing , K. I. & LeDoux , J. E . Extinction Learning in Humans . Neuron 43 , 897 – 905 ( 2004 ). OpenUrl CrossRef PubMed Web of Science 19. ↵ Milad , M. R. et al. Recall of Fear Extinction in Humans Activates the Ventromedial Prefrontal Cortex and Hippocampus in Concert . Biol. Psychiatry 62 , 446 – 454 ( 2007 ). OpenUrl CrossRef PubMed Web of Science 20. ↵ Kalisch , R. et al. Context-Dependent Human Extinction Memory Is Mediated by a Ventromedial Prefrontal and Hippocampal Network . J. Neurosci . 26 , 9503 – 9511 ( 2006 ). OpenUrl Abstract / FREE Full Text 21. ↵ Lissek , S. , Glaubitz , B. , Uengoer , M. & Tegenthoff , M . Hippocampal activation during extinction learning predicts occurrence of the renewal effect in extinction recall . NeuroImage 81 , 131 – 143 ( 2013 ). OpenUrl CrossRef PubMed 22. Lissek , S. , Glaubitz , B. , Schmidt-Wilcke , T. & Tegenthoff , M . Hippocampal Context Processing during Acquisition of a Predictive Learning Task Is Associated with Renewal in Extinction Recall . J. Cogn. Neurosci . 28 , 747 – 762 ( 2016 ). OpenUrl CrossRef PubMed 23. ↵ Lissek , S. , Golisch , A. , Glaubitz , B. & Tegenthoff , M . The GABAergic system in prefrontal cortex and hippocampus modulates context-related extinction learning and renewal in humans . Brain Imaging Behav . 11 , 1885 – 1900 ( 2017 ). OpenUrl CrossRef PubMed 24. ↵ Fullana , M. A. et al. Neural signatures of human fear conditioning: an updated and extended meta-analysis of fMRI studies . Mol. Psychiatry 21 , 500 – 508 ( 2016 ). OpenUrl CrossRef PubMed 25. ↵ Fullana , M. A. et al. Fear extinction in the human brain: A meta-analysis of fMRI studies in healthy participants . Neurosci. Biobehav. Rev . 88 , 16 – 25 ( 2018 ). OpenUrl CrossRef PubMed 26. ↵ Kruse , O. , Klein , S. , Tapia León , I. , Stark , R. & Klucken , T . Amygdala and nucleus accumbens involvement in appetitive extinction . Hum. Brain Mapp . 41 , 1833 – 1841 ( 2020 ). OpenUrl CrossRef PubMed 27. ↵ Finn , E. S. et al. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity . Nat. Neurosci . 18 , 1664 – 1671 ( 2015 ). OpenUrl CrossRef PubMed 28. Dubois , J. , Galdi , P. , Han , Y. , Paul , L. K. & Adolphs , R . Resting-State Functional Brain Connectivity Best Predicts the Personality Dimension of Openness to Experience . Personal. Neurosci . 1 , e6 ( 2018 ). OpenUrl CrossRef 29. ↵ Williams , L. M . Precision psychiatry: a neural circuit taxonomy for depression and anxiety . Lancet Psychiatry 3 , 472 – 480 ( 2016 ). OpenUrl CrossRef PubMed 30. ↵ Siegel , M. , Donner , T. H. & Engel , A. K . Spectral fingerprints of large-scale neuronal interactions . Nat. Rev. Neurosci . 13 , 121 – 134 ( 2012 ). OpenUrl CrossRef PubMed 31. ↵ Razi , A. et al. Large-scale DCMs for resting-state fMRI . Netw. Neurosci . 1 , 222 – 241 ( 2017 ). OpenUrl CrossRef PubMed 32. ↵ Friston , K. J . Functional and effective connectivity in neuroimaging: A synthesis . Hum. Brain Mapp . 2 , 56 – 78 ( 1994 ). OpenUrl CrossRef 33. ↵ Novelli , L. , Friston , K. & Razi , A . Spectral dynamic causal modeling: A didactic introduction and its relationship with functional connectivity . Netw. Neurosci . 8 , 178 – 202 ( 2024 ). OpenUrl CrossRef PubMed 34. ↵ Dell’Acqua , F. & Catani , M . Structural human brain networks: hot topics in diffusion tractography . Curr. Opin. Neurol . 1 ( 2012 ) doi: 10.1097/WCO.0b013e328355d544 . OpenUrl CrossRef PubMed 35. ↵ Greicius , M. D. , Supekar , K. , Menon , V. & Dougherty , R. F . Resting-State Functional Connectivity Reflects Structural Connectivity in the Default Mode Network . Cereb. Cortex 19 , 72 – 78 ( 2009 ). OpenUrl CrossRef PubMed Web of Science 36. ↵ Liégeois , R. , Santos , A. , Matta , V. , Van De Ville , D. & Sayed , A. H . Revisiting correlation-based functional connectivity and its relationship with structural connectivity . Netw. Neurosci . 4 , 1235 – 1251 ( 2020 ). OpenUrl CrossRef PubMed 37. ↵ De Pasquale , F. , Della Penna , S. , Sabatini , U. , Caravasso Falletta , C. & Peran , P . The anatomical scaffold underlying the functional centrality of known cortical hubs . Hum. Brain Mapp . 38 , 5141 – 5160 ( 2017 ). OpenUrl CrossRef PubMed 38. ↵ Honey , C. J. et al. Predicting human resting-state functional connectivity from structural connectivity . Proc. Natl. Acad. Sci . 106 , 2035 – 2040 ( 2009 ). OpenUrl Abstract / FREE Full Text 39. ↵ Greaves , M. D. , Novelli , L. & Razi , A. Structurally informed resting-state effective connectivity recapitulates cortical hierarchy . Preprint at doi: 10.1101/2024.04.03.587831 ( 2024 ). OpenUrl Abstract / FREE Full Text 40. ↵ Litwińczuk , M. C. , Muhlert , N. , Cloutman , L. , Trujillo-Barreto , N. & Woollams , A . Combination of structural and functional connectivity explains unique variation in specific domains of cognitive function . NeuroImage 262 , 119531 ( 2022 ). 41. ↵ Rasero , J. , Sentis , A. I. , Yeh , F.-C. & Verstynen , T . Integrating across neuroimaging modalities boosts prediction accuracy of cognitive ability . PLOS Comput. Biol . 17 , e1008347 ( 2021 ). OpenUrl CrossRef PubMed 42. ↵ Gulyás , A. I. , Megıás , M. , Emri , Z. & Freund , T. F. Total Number and Ratio of Excitatory and Inhibitory Synapses Converging onto Single Interneurons of Different Types in the CA1 Area of the Rat Hippocampus . J. Neurosci . 19 , 10082 – 10097 ( 1999 ). OpenUrl Abstract / FREE Full Text 43. ↵ Mongillo , G. , Rumpel , S. & Loewenstein , Y . Inhibitory connectivity defines the realm of excitatory plasticity . Nat. Neurosci . 21 , 1463 – 1470 ( 2018 ). OpenUrl CrossRef PubMed 44. ↵ Barron , H. C. , Vogels , T. P. , Behrens , T. E. & Ramaswami , M . Inhibitory engrams in perception and memory . Proc. Natl. Acad. Sci . 114 , 6666 – 6674 ( 2017 ). OpenUrl Abstract / FREE Full Text 45. ↵ Shirer , W. R. , Ryali , S. , Rykhlevskaia , E. , Menon , V. & Greicius , M. D . Decoding Subject-Driven Cognitive States with Whole-Brain Connectivity Patterns . Cereb. Cortex 22 , 158 – 165 ( 2012 ). OpenUrl CrossRef PubMed Web of Science 46. ↵ Gallistel , C. R. , Fairhurst , S. & Balsam , P . The learning curve: Implications of a quantitative analysis . Proc. Natl. Acad. Sci . 101 , 13124 – 13131 ( 2004 ). OpenUrl Abstract / FREE Full Text 47. ↵ Alvarez , R. P. , Biggs , A. , Chen , G. , Pine , D. S. & Grillon , C . Contextual Fear Conditioning in Humans: Cortical-Hippocampal and Amygdala Contributions . J. Neurosci . 28 , 6211 – 6219 ( 2008 ). OpenUrl Abstract / FREE Full Text 48. ↵ Jhang , J. et al. Anterior cingulate cortex and its input to the basolateral amygdala control innate fear response . Nat. Commun . 9 , 2744 ( 2018 ). OpenUrl CrossRef PubMed 49. ↵ Frontera , J. L. et al. The cerebellum regulates fear extinction through thalamo-prefrontal cortex interactions in male mice . Nat. Commun . 14 , 1508 ( 2023 ). OpenUrl CrossRef PubMed 50. ↵ LeDoux , J. E. & Pine , D. S . Using Neuroscience to Help Understand Fear and Anxiety: A Two-System Framework . Am. J. Psychiatry 173 , 1083 – 1093 ( 2016 ). OpenUrl CrossRef PubMed 51. ↵ Wake , S. , Morriss , J. , Johnstone , T. , Van Reekum , C. M. & Dodd , H . Intolerance of uncertainty, and not social anxiety, is associated with compromised extinction of social threat . Behav. Res. Ther . 139 , 103818 ( 2021 ). 52. ↵ Nees , F. et al. White matter correlates of contextual pavlovian fear extinction and the role of anxiety in healthy humans . Cortex 121 , 179 – 188 ( 2019 ). OpenUrl CrossRef PubMed 53. ↵ Fani , N. et al. Fear-potentiated startle during extinction is associated with white matter microstructure and functional connectivity . Cortex 64 , 249 – 259 ( 2015 ). OpenUrl CrossRef PubMed 54. ↵ Picó-Pérez , M. et al. Common and distinct neural correlates of fear extinction and cognitive reappraisal: A meta-analysis of fMRI studies . Neurosci. Biobehav. Rev . 104 , 102 – 115 ( 2019 ). OpenUrl CrossRef PubMed 55. ↵ Onoda , K. , Kawagoe , T. , Zheng , H. & Yamaguchi , S . Theta band transcranial alternating current stimulations modulates network behavior of dorsal anterior cingulate cortex . Sci. Rep . 7 , 3607 ( 2017 ). OpenUrl CrossRef PubMed 56. ↵ Hennings , A. C. , McClay , M. , Drew , M. R. , Lewis-Peacock , J. A. & Dunsmoor , J. E . Neural reinstatement reveals divided organization of fear and extinction memories in the human brain . Curr. Biol . 32 , 304 – 314.e5 ( 2022 ). OpenUrl CrossRef PubMed 57. ↵ McDonald , A. J. , Mascagni , F. & Guo , L . Projections of the medial and lateral prefrontal cortices to the amygdala: a Phaseolus vulgaris leucoagglutinin study in the rat . Neuroscience 71 , 55 – 75 ( 1996 ). OpenUrl CrossRef PubMed Web of Science 58. ↵ Freedman , L. J. , Insel , T. R. & Smith , Y . Subcortical projections of area 25 (subgenual cortex) of the macaque monkey . J. Comp. Neurol . 421 , 172 – 188 ( 2000 ). OpenUrl CrossRef PubMed Web of Science 59. ↵ Nguyen , R. , Koukoutselos , K. , Forro , T. & Ciocchi , S . Fear extinction relies on ventral hippocampal safety codes shaped by the amygdala . Sci. Adv . 9 , eadg4881 ( 2023 ). 60. ↵ Packheiser , J. , Donoso , J. R. , Cheng , S. , Güntürkün , O. & Pusch , R . Trial-by-trial dynamics of reward prediction error-associated signals during extinction learning and renewal . Prog. Neurobiol . 197 , 101901 ( 2021 ). 61. ↵ Frässle , S. & Stephan , K. E . Test-retest reliability of regression dynamic causal modeling . Netw. Neurosci . 6 , 135 – 160 ( 2022 ). OpenUrl CrossRef PubMed 62. ↵ Corcoran , K. A. & Maren , S . Hippocampal Inactivation Disrupts Contextual Retrieval of Fear Memory after Extinction . J. Neurosci . 21 , 1720 – 1726 ( 2001 ). OpenUrl Abstract / FREE Full Text 63. ↵ Maren , S. & Hobin , J. A . Hippocampal regulation of context-dependent neuronal activity in the lateral amygdala . Learn. Mem . 14 , 318 – 324 ( 2007 ). OpenUrl Abstract / FREE Full Text 64. ↵ Walther , T. et al. Context-dependent extinction learning emerging from raw sensory inputs: a reinforcement learning approach . Sci. Rep . 11 , 2713 ( 2021 ). OpenUrl CrossRef PubMed 65. ↵ Lissek , S. , Glaubitz , B. , Güntürkün , O. & Tegenthoff , M . Noradrenergic stimulation modulates activation of extinction-related brain regions and enhances contextual extinction learning without affecting renewal . Front. Behav. Neurosci . 9 , ( 2015 ). 66. ↵ Pi , G. et al. Posterior basolateral amygdala to ventral hippocampal CA1 drives approach behaviour to exert an anxiolytic effect . Nat. Commun . 11 , 183 ( 2020 ). 67. ↵ Felix-Ortiz , A. C. et al. BLA to vHPC Inputs Modulate Anxiety-Related Behaviors . Neuron 79 , 658 – 664 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 68. ↵ Seidenbecher , T. , Laxmi , T. R. , Stork , O. & Pape , H.-C . Amygdalar and Hippocampal Theta Rhythm Synchronization During Fear Memory Retrieval . Science 301 , 846 – 850 ( 2003 ). OpenUrl Abstract / FREE Full Text 69. ↵ Liu , X. et al. Optogenetic stimulation of a hippocampal engram activates fear memory recall . Nature 484 , 381 – 385 ( 2012 ). OpenUrl CrossRef PubMed Web of Science 70. ↵ Tozzi , L. et al. Personalized brain circuit scores identify clinically distinct biotypes in depression and anxiety . Nat. Med . 30 , 2076 – 2087 ( 2024 ). OpenUrl CrossRef PubMed 71. ↵ Brooks , S. J. & Stein , D. J . A systematic review of the neural bases of psychotherapy for anxiety and related disorders . Dialogues Clin. Neurosci . 17 , 261 – 279 ( 2015 ). OpenUrl CrossRef PubMed 72. ↵ Tzabazis , A. et al. Shaped Magnetic Field Pulses by Multi-Coil Repetitive Transcranial Magnetic Stimulation (rTMS) Differentially Modulate Anterior Cingulate Cortex Responses and Pain in Volunteers and Fibromyalgia Patients . Mol. Pain 9 , 1744–8069 - 9–33 ( 2013 ). OpenUrl 73. ↵ Modirrousta , M. et al. The efficacy of deep repetitive transcranial magnetic stimulation over the medial prefrontal cortex in obsessive compulsive disorder: results from an open-label study . Depress. Anxiety 32 , 445 – 450 ( 2015 ). OpenUrl CrossRef PubMed 74. Satterthwaite , T. D. et al. An improved framework for confound regression and filtering for control of motion artifact in the preprocessing of resting-state functional connectivity data . NeuroImage 64 , 240 – 256 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 75. ↵ Andersson , J. L. R. , Skare , S. & Ashburner , J . How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging . NeuroImage 20 , 870 – 888 ( 2003 ). OpenUrl CrossRef PubMed Web of Science 76. ↵ Kebschull , J. M. et al. Cerebellum Lecture: the Cerebellar Nuclei—Core of the Cerebellum . The Cerebellum 23 , 620 – 677 ( 2023 ). OpenUrl CrossRef PubMed 77. ↵ Glasser , M. F. et al. The Human Connectome Project’s neuroimaging approach . Nat. Neurosci . 19 , 1175 – 1187 ( 2016 ). OpenUrl CrossRef PubMed 78. ↵ Mohanty , R. , Sethares , W. A. , Nair , V. A. & Prabhakaran , V . Rethinking Measures of Functional Connectivity via Feature Extraction . Sci. Rep . 10 , 1298 ( 2020 ). OpenUrl CrossRef PubMed 79. ↵ Middleton , F. A. & Strick , P. L . Anatomical Evidence for Cerebellar and Basal Ganglia Involvement in Higher Cognitive Function . Science 266 , 458 – 461 ( 1994 ). OpenUrl Abstract / FREE Full Text 80. ↵ Bach , D. R. & Friston , K. J . Model-based analysis of skin conductance responses: Towards causal models in psychophysiology . Psychophysiology 50 , 15 – 22 ( 2013 ). OpenUrl CrossRef PubMed 81. ↵ Kleckner , I. R. , et al. Simple, Transparent, and Flexible Automated Quality Assessment Procedures for Ambulatory Electrodermal Activity Data . IEEE Trans. Biomed. Eng . 65 , 1460 – 1467 ( 2018 ). OpenUrl CrossRef PubMed 82. ↵ Bach , D. R. , Daunizeau , J. , Friston , K. J. & Dolan , R. J . Dynamic causal modelling of anticipatory skin conductance responses . Biol. Psychol . 85 , 163 – 170 ( 2010 ). OpenUrl CrossRef PubMed Web of Science 83. ↵ Bach , D. R. , Flandin , G. , Friston , K. J. & Dolan , R. J . Modelling event-related skin conductance responses . Int. J. Psychophysiol . 75 , 349 – 356 ( 2010 ). OpenUrl CrossRef PubMed Web of Science 84. ↵ Bach , D. R. , Daunizeau , J. , Kuelzow , N. , Friston , K. J. & Dolan , R. J . Dynamic causal modeling of spontaneous fluctuations in skin conductance . Psychophysiology 48 , 252 – 257 ( 2011 ). OpenUrl CrossRef PubMed 85. ↵ Belsley , D. A. , Kuh , E. & Welsch , R. E . Regression Diagnostics: Identifying Influential Data and Sources of Collinearity . ( Wiley , 1980 ). doi: 10.1002/0471725153 . OpenUrl CrossRef 86. ↵ Rosenblatt , M. , Tejavibulya , L. , Jiang , R. , Noble , S. & Scheinost , D . Data leakage inflates prediction performance in connectome-based machine learning models . Nat. Commun . 15 , 1829 ( 2024 ). OpenUrl CrossRef PubMed 87. ↵ Williams , D. R . The Confidence Interval that Wasn’t: Bootstrapped “Confidence Intervals” in L1-Regularized Partial Correlation Networks . Preprint at doi: 10.31234/osf.io/kjh2f ( 2021 ). OpenUrl CrossRef View the discussion thread. Back to top Previous Next Posted May 05, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Predicting individual differences of fear and cognitive learning and extinction Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Predicting individual differences of fear and cognitive learning and extinction C.A. Gomes , D. R. Bach , A. Razi , G. Batsikadze , S. Elsenbruch , H. Engler , T.M. Ernst , M.C. Fellner , C. Fraenz , E. Genç , A. Klass , F. Labrenz , S. Lissek , C.J. Merz , D. Metzen , A. Nostadt , R.J. Pawlik , J.E. Schneider , M. Tegenthoff , A. Thieme , O.T. Wolf , O. Güntürkün , H.H. Quick , R. Kumsta , D. Timmann , T. Spisak , N. Axmacher bioRxiv 2025.05.04.651880; doi: https://doi.org/10.1101/2025.05.04.651880 Share This Article: Copy Citation Tools Predicting individual differences of fear and cognitive learning and extinction C.A. Gomes , D. R. Bach , A. Razi , G. Batsikadze , S. Elsenbruch , H. Engler , T.M. Ernst , M.C. Fellner , C. Fraenz , E. Genç , A. Klass , F. Labrenz , S. Lissek , C.J. Merz , D. Metzen , A. Nostadt , R.J. Pawlik , J.E. Schneider , M. Tegenthoff , A. Thieme , O.T. Wolf , O. Güntürkün , H.H. Quick , R. Kumsta , D. Timmann , T. Spisak , N. Axmacher bioRxiv 2025.05.04.651880; doi: https://doi.org/10.1101/2025.05.04.651880 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Neuroscience Subject Areas All Articles Animal Behavior and Cognition (7635) Biochemistry (17691) Bioengineering (13892) Bioinformatics (41937) Biophysics (21452) Cancer Biology (18588) Cell Biology (25504) Clinical Trials (138) Developmental Biology (13378) Ecology (19899) Epidemiology (2067) Evolutionary Biology (24320) Genetics (15609) Genomics (22506) Immunology (17736) Microbiology (40394) Molecular Biology (17181) Neuroscience (88605) Paleontology (666) Pathology (2832) Pharmacology and Toxicology (4824) Physiology (7641) Plant Biology (15156) Scientific Communication and Education (2045) Synthetic Biology (4294) Systems Biology (9825) Zoology (2271)