摘要
The study of human cognition through neuroscience has benefitted from a constant development and transformation of associated technology and analysis. Taking working memory (WM) as an example, and also as the area in which Mark Stokes has demonstrated this, it is a field that has often found itself right at the forefront of cognitive neuroscience, with a fascinating progression of theoretical insight and development as a consequence (see D'Esposito and Postle [2014] and Postle [2006] for thorough reviews of the field). Often, this innovation and discovery has produced theoretical upheaval, such as understanding the relationship/dependence/separation between attention and WM (Nobre & Stokes, 2011), challenging the limited capacity of WM (Bays & Husain, 2008), and the role of pFC in WM (Stokes, 2015) to name a few. (At this stage, the reader may have noted the involvement of Mark Stokes in several of these upheavals. It is not likely to be coincidence.)Throughout this journey over these last decades, TMS has been present, providing causal results to complement correlational neuroimaging, and behavioral studies. Despite its theoretical strengths as a class of research tools, the practical challenges and theoretical unknowns regarding physiological effects have meant that TMS has not always lived up to its promise as a “causal silver bullet” that can confirm correlational findings (Parkin, Ekhtiari, & Walsh, 2015; Siebner, Hartwigsen, Kassuba, & Rothwell, 2009).1 The future of TMS is, however, looking decidedly up (again). Advances in concurrent invasive recordings to visualize neuronal responses to TMS have provided invaluable in vivo evidence on the acute effects of stimulation (Allen, Pasley, Duong, & Freeman, 2007; Moliadze, Zhao, Eysel, & Funke, 2003). In humans, combining TMS with neuroimaging has also given the field a boost, showing effects across the brain (Bergmann et al., 2021), and on electrophysiological markers such as ERPs (Taylor, Walsh, & Eimer, 2008) and oscillations (Thut & Miniussi, 2009). Thus, as this commentary will argue, a crucial way forward for investigating cognition through neuroscience will be through innovative multimodal experimental approaches that can produce convergent evidence in one shot; examples will follow here as to how this approach can drive the field forward, with the focus being on TMS, given its superior spatial and temporal resolution compared with transcranial electrical stimulation.A game-changer for TMS has been in its simultaneous combination with fMRI and EEG. Although these innovations were demonstrated decades ago—for TMS-fMRI: Bohning et al. (1998); TMS-EEG: Ilmoniemi et al. (1997)—their progression from niche technical feats to essential contributors to cognitive neuroscience is coming to fruition. Given that one can now acquire reliable, well-supported systems for TMS-fMRI and TMS-EEG, the next step is underway in which advanced experimental and analysis approaches will answer questions that would otherwise be difficult to address.The first growth area of cognitive TMS research is in the combination of TMS-fMRI with functional and effective connectivity analysis approaches; TMS can be applied to intervene with brain networks necessary for behavior and causally establish the necessity of associated network nodes and communication pathways. Although there exist many sophisticated methods for establishing connectivity to uncover brain networks through advanced mathematical and statistical algorithms, they all provide indirect “best current estimates” (Johansen-Berg, 2013); what remains to be established is whether such connections exist beyond the statistical level and TMS is arguably the best placed to accomplish this. Nee and D'Esposito (2017) provided a valuable demonstration of what a TMS-fMRI-effective connectivity—here, dynamic causal modeling—approach can contribute theoretically. The use of TMS to causally test dynamic causal modelings has been suggested previously (Hartwigsen et al., 2015; Bestmann & Feredoes, 2013), and perturbation of a network with TMS has also been proposed in the context of network control theory (Medaglia, Pasqualetti, Hamilton, Thompson-Schill, & Bassett, 2017). A recent study by Sydnor et al. (2022) is a clear indication of the strengths of TMS-fMRI-perturbation, in which single pulses of TMS to ventrolateral pFC produced fMRI signal changes in the amygdala, the magnitude of which were predicted by the density of the white matter fiber pathway between the TMS target and amygdala. One could envisage a similar approach in which multiple lines of evidence including structure and function are combined to reveal patterns of connectivity for different types of behavioral tasks and/or different task stages. In the context of WM, such an approach may contribute to the ongoing debate on the nature of communication between sensory, parietal, and prefrontal cortex during WM maintenance (Gayet, Paffen, & Van der Stigchel, 2018; Scimeca, Kiyonaga, & D'Esposito, 2018; Xu, 2017), or the involvement of, and interactions between, more established networks such as the default mode and frontoparietal networks during WM (Murphy, Bertolero, Papadopoulos, Lydon-Staley, & Bassett, 2020).In addition to brain networks, the cognitive neuroscience field is interested in what information is being coded by brain regions engaged for a behavior. Here, decoding techniques such as multivoxel pattern analysis (MVPA) have proven to be useful (e.g., Woolgar, Jackson, & Duncan, 2016; Postle, 2015), and recently, they have been combined with fMRI (Jackson, Feredoes, Rich, Lindner, & Woolgar, 2021) and EEG (Rose et al., 2016) to provide answers to otherwise difficult-to-resolve questions, with the bonus of causal inference. Jackson et al. (2021) addressed a long-standing debate on relevant information representation across the brain in the face of irrelevant information. In this study, concurrent TMS-fMRI (i.e., TMS applied online during task performance in the MRI scanner) data were analyzed with MVPA, to show causally that the multiple demand network (Duncan, 2010) enhances relevant information rather than suppresses irrelevant information. Specifically, right dorsolateral pFC (DLPFC) TMS disrupted task performance, resulting in a decrease in the decoding of relevant information across the multiple demand network, whereas irrelevant information decoding was not significantly modulated by disruptive TMS. MVPA provided a unique way in which to view the effects of TMS, that is, on information coding, rather than the magnitude of the BOLD response and which in this case was critical for testing the impact of DLPFC on the same information being represented under different behavioral contexts.Rose et al. (2016) combined TMS with EEG (with fMRI informing the location of the cortical TMS target) to test a cognitive account of WM in which information can be held in different states depending on task demands (Oberauer, 2009). A latent, undecodable WM item (rendered thus because it was not required during a specific epoch of the task) could be “reactivated” with a single TMS “ping.”2 Specifically, the inactive item could be decoded for a brief period post-TMS, indicating that it had been brought into a different physiological state by TMS such that it could be detected with MVPA. This result marries very well with multiple-states theoretical accounts of WM (Stokes, 2015; Oberauer, 2009; Cowan, 2005), and demonstrates the unique ability of a TMS ping to access an otherwise inaccessible brain state.Combining TMS with neuroimaging has taken a further, dramatic step with simultaneous TMS-fMRI-EEG. While a technically extremely challenging feat, this approach is arguably a “holy grail” for some in the brain stimulation field, providing a bridge between the spatial and temporal, to capture oscillatory activity changes across the whole brain in response to perturbation with TMS. Proof of concept was demonstrated by Peters et al. (2013), with a cognitive application published more recently (Peters et al., 2020) in which right dorsal premotor cortex was stimulated with a short burst of below motor-threshold TMS and the state of pre-TMS alpha and low beta oscillations were examined to determine how trial-by-trial fluctuations influenced the effect of TMS throughout a functional motor network. TMS-induced signal propagation, as indexed by the fMRI response throughout the cortical–subcortical network of interest, was impeded by stronger pre-TMS alpha power, in line with alpha serving a suppressive function in terms of information transmission. The promise of this technique is in its ability to probe dynamic brain states, including those which exist during attention, memory, or cognitive control (Sack, 2022), and it is certainly in line with views such as those put forward by, for example, Mark Stokes, in which multiple brain areas combine while maintaining a highly dynamic state, to achieve flexible behavior (Stokes & Duncan, 2014). TMS-fMRI-EEG would very much be in the “watch this exciting space closely” category.Computational modeling of behavior can now also be added as an experimental approach benefitting from the addition of TMS. Again, relying on the perturbation logic, the spread of the consequences of TMS throughout hidden layers can provide a causal missing link, particularly if applied to existing TMS data (e.g., TMS-behavioral or TMS-neuroimaging data). Very recent work (Whyte, 2022) has taken this exact approach, applying the WM model of Manohar, Pertzov, and Husain (2017) to explain the behavioral and TMS effects on fMRI decoding of stimuli of Jackson et al. (2021). According to the model, stimuli and associated response rules are coded through rapid synaptic changes. Modeling a short TMS train to DLPFC as a brief excitatory input, the effect on the model was to produce a disruption of conjunction units coding task rules, assumed to be in the multiple demand network (Woolgar, Hampshire, Thompson, & Duncan, 2011). The results of the modeling could explain the fMRI decoding and behavioral results reported by Jackson et al. (2021). This plastic attractor model provides an account of DLPFC that permits adaptive, flexible behavior to select task-relevant stimulus information and to produce appropriate motor actions in response. The added value of TMS here was to provide a controlled perturbation of the network with behavioral and neural outcomes, with the computational model revealing the otherwise hidden underlying mechanisms. It is a fruitful area of future research, given that at some stage, computational models of brain and behavior will need causal verification in the same way as for mathematically derived functional brain networks, as described above.A final area that will be discussed here, for future cognitive TMS studies to pursue more ardently, is combining of patterned stimulation protocols with neuroimaging. There are a range of patterns in which TMS can be applied, producing complex physiological effects such as spike timing-dependent plasticity, to investigate cortico-cortico connectivity, for example, cortico-cortical paired associative stimulation (ccPAS; Romei, Thut, & Silvanto, 2016; Rizzo et al., 2009). Casula, Pellicciari, Picazio, Caltagirone, and Koch (2016) applied a ccPAS protocol to investigate the direction of communication between key nodes of the frontoparietal network, DLPFC, and posterior parietal cortex. Concurrent EEG (in the absence of behavior) showed an increase in high frontal gamma activity following DLPFC-parietal ccPAS, with the reverse when parietal cortex was stimulated first. The timing of these changes also fit with a GABA-dependent mechanism, suggesting a change in the amount of GABA available to DLPFC pyramidal neurons that generated the EEG signal (Casula et al., 2016). The interesting confluence of findings from this study can certainly motivate similar studies combining neuroimaging, and perhaps also neuropharmacology to generate unique research questions regarding the neural pathways and mechanisms underlying cognition (Pitcher, Parkin, & Walsh, 2021; Romei et al., 2016).The clear improvements in the field of combining TMS with neuroimaging and analysis, such as developments in hardware (Navarro de Lara et al., 2015), understanding the effects of TMS at the level of brain regions (Tik et al., 2019) and neuronally (Romero, Davare, Armendariz, & Janssen, 2019), and modeling of TMS effects (Shirinpour et al., 2021), all contribute to increasing the accuracy of TMS for causally perturbing brain and behavior. There are more sophisticated questions being asked by this field, producing more sophisticated evidence, compared with the “early days” of cognitive TMS. It is indeed an exciting time to be in the noninvasive neurostimulation field, with its potential to have a central role in informing cognitive neuroscientific theory growing with each publication.The author is supported by a Royal Society Leverhulme Trust Senior Research Fellowship.Reprint requests should be sent to Eva Feredoes, School of Psychology and Clinical Language Sciences, University of Reading, Earley Gate, Reading, RG2 7AG, UK, or via e-mail: e.a.feredoes@reading.ac.uk.Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article’s gender citation balance. The authors of this article report its proportions of citations by gender category to be as follows: M/M = .615; W/M = .256; M/W = .077; W/W = .051.