Gamma tACS over the temporal lobe increases the occurrence of Eureka! moments

The solution to a problem might manifest itself as a burst of unexpected, unpredictable clarity. Such Eureka! events, or Insight moments, are among the most fascinating mysteries of human cognition, whose neurophysiological substrate seems to include a role for oscillatory activity within the α and γ bands in the right parietal and temporal brain regions. We tested this hypothesis on thirty-one healthy participants using transcranial Alternating Current Stimulation (tACS) to externally amplify α (10 Hz) and γ (40 Hz) activity in the right parietal and temporal lobes, respectively. During γ-tACS over the right temporal lobe, we observed an increase in accuracy on a verbal insight task. Furthermore, electroencephalography (EEG) data revealed an increase in γ spectral power over bilateral temporal lobes after stimulation. Additionally, resting-state functional MRI data acquired before the stimulation session suggested a correlation between behavioral response to right temporal lobe tACS and functional connectivity of bilateral temporal lobes, in line with the bilateral increase in γ band revealed by EEG. Overall, results suggest the possibility of enhancing the probability of generating Eureka! moments in humans by means of frequency-specific noninvasive brain stimulation.


Linguistic -CRA problems
Participants first solved the Italian version of the recently validated CRA problems 1 (see Fig. 1C, for the original English version see 2 ). These types of problems have been consistently used to study insight problem-solving. Self-reports differentiating between insight and analytic problem-solving are reliable and their association with numerous behavioral and neuroimaging markers are well documented in literature 2,3,4,5 . CRA presents several advantages: they are composed of a large set of trials, allowing reliable data collection; they can be solved via insight or via analytical processes, allowing a direct comparison between their respective electrophysiological activations; they require short time responses and a small visual space to be presented. Each problem is formed by three words and the solution is represented by a forth word that can be associated with the others to form a compound word (i.e., manners, round, tennis; solution: table). We selected 105 items and divided them into 7 sets of 15 items each, balanced for difficulty as well as for the method preferentially used to solve them (i.e. we considered a given trial as "insight" or "analytic" based on the preferred strategy reported by participants included in the validation study by Salvi et. al., 2015 1 ). Three random blocks (= 15 trials each) were selected for each participant.

Visuo-linguistic Rebus Puzzles
Participants also solved the Rebus Puzzles, an insight task initially validated by MacGregor and Cunningham 6 and subsequently validated in Italian 1 . Unlike the CRA, this set of problems requires the integration of both visual and verbal information in order to find a common phrase fitting each item (i.e., "Cycle, Cycle, Cycle" solution: "Tricycle") (see Fig.   1C). Similarly to CRA problems, they can also be solved either via insight or via analytic processes. We divided all trials in 7 sets of 11 trials each, balanced for difficulty and for problem-solving strategy; 3 blocks (= 15 trials each) were selected for each participant.

EEG preprocessing and analysis
EEG preprocessing was performed separately for data collected before and after each stimulation block, with the examiner being blinded to the different stimulation conditions for all the preprocessing steps. First, the EEG file was imported to MATLAB (vR2014b, The Mathworks) using the EEGlab toolbox (v13.4.4b) 7 and channels were located on the scalp model using the DIPFIT plugin from eeglab. Data were bandpass filtered from 1 Hz Hz; γ = 30-50 Hz) was computed using the EEGlab function spectopo, then averaged across epochs respectively for data collected before and after the stimulation. The average power values for each subject were then included in four different multivariate ANOVA designs, respectively looking at changes in the theta (θ), α, beta (β) and γ frequency bands. Differences in spectral power before and after stimulation were investigated for each stimulation condition (factor "Stimulation" =Sham, tACS 10Hz and tACS 40Hz), including values for each EEG electrode (n=8) using SPSS software and a statistical significance threshold equal to p≤0.05.

MRI data acquisition and functional connectivity analysis
In order to characterize the response to tACS in different individuals by means of functional connectivity analysis, resting-state fMRI (rs-fMRI) was acquired for all participants. fMRI connectivity analyses provide a non-invasive way to examine spontaneous brain functioning, which utilizes the temporal correlation between spontaneous blood oxygen level-dependent (BOLD) signal fluctuations of different brain regions to identify broadly connected networks 8  http://www.fil.ion.ucl.ac.uk/spm/) and MATLAB 7.5 (MathWorks, MA, USA). The first five volumes of functional images of each subject were discarded to allow for steady-state magnetization. EPI images were slice-time corrected using the interleaved descending acquisition criteria, and realigned and re-sliced to correct for head motion using a mean functional volume derived from the overall fMRI scans. Subjects whose head motion exceeded 2.0 mm or rotation exceeded 1.0 • during scanning were excluded. In order to obtain a better estimation of brain tissues maps, we implemented an optimized segmentation and normalization process using the DARTEL (Diffeomorphic Anatomical Registration using Exponentiated Lie Algebra) 9 module for SPM8. Briefly, this approach is based on the creation of a customized anatomical template built directly from participants' T1-weighted images instead of the canonical one provided with SPM (MNI template, ICBM 152, Montreal Neurological Institute). DARTEL allows for a finer normalization into standard space and consequently avoids under or over estimation of brain regions' volume potentially induced by the adoption of an external template. A Hidden Markov Random Field model was applied in all segmentation processes in order to remove isolated voxels. Customized tissue prior images and a T1-weighted template were smoothed using an 8 mm full-width at halfmaximum (FWHM) isotropic Gaussian kernel. Functional images were consequently nonlinearly normalized to standard space and a voxel resampling to (isotropic) 3 x 3 x 3 mm were applied. Linear trends were removed to reduce the influence of the rising temperature of the MRI scanner and all functional volumes were band pass filtered at (0.01 Hz < f < 0.08 Hz) to reduce low-frequency drift. Finally, a CompCor algorithm was applied in order to control physiological high-frequency respiratory and cardiac noise 10 .
In order to analyze fMRI and behavioral scores individual delta scores were computed (i.e. accuracy during Sham minus accuracy during Active stimulation, e.g. [Sham -tACS 10Hz]), and correlated with seed-brain connectivity patterns by using two predefined regions of interest (ROI) placed in correspondence of the regions being targeted with tACS. Two 5mm diameter spheres were created at (46, -66, 35) and (51, -4, -25), respectively corresponding to the right inferior parietal (representing P4 EEG electrode) and right anterior temporal lobe (T8) in MNI space. The average correlation between the BOLD signal from each ROI and those of each voxel in the brain were then correlated with the aforementioned delta scores, to identify patterns of connectivity co-varying with the propensity to respond to electrical stimulation during insight problem-solving. Two subjects were discarded due to excessive head motion in the scanner, while one subject did not complete the MRI exam due to an unpleasant feeling during the acquisition. Therefore, the final analysis was conducted on 28 subjects.

tACS computational model
To check for the degree of focality of tACS solution, two models had been created separately for tACS 10Hz over P4 and tACS 40HZ over T8. Distribution of current and normal components of generated electrical fields are reported for each montage in Fig. S1. A realistic head model based on T1-weighted and Proton Density (PD) weighted phantom MRI images of the single-subject template Colin27 was used to simulate the electric field distribution using the Stimweaver software (Neuroelectrics, Barcelona), as previously

Cognitive Tasks
Cognitive assessment was based on a battery of PC-based tasks, encompassing global cognitive functioning scores such as Intelligence Quotient (i.e. Fullscale IQ, Performance IQ and Verbal IQ) obtained at the TIB (Premorbid Intelligence Battery) test 12,13 and fluid intelligence (Raven Advance Progressive Matrices -RAPM 14 ). Additionally, domain specific tasks were administered, including: Go-NoGo task for inhibition 15 , Pop-Task for Switching abilities 16 , Change detection task for visuo-spatial working memory 17 , Digit span forward/backward for verbal working and short-term memory 18 , Visual search task for sustained attention 19 , Global and Local features task for filtering abilities 20 . All the tasks were delivered using a Windows based PC, using E-prime software 2.0. Participants completed the cognitive evaluation the day of the MRI acquisition. Accuracy and Reaction Times (for correct responses) were calculated and included in the correlation analysis. Figure S1. Surface models of the tACS montages adopted for right temporal (a) and right parietal (b) stimulation conditions. Color code refers to the amount of current delivered at the brain cortical level (Volts). Models are built using a standard head model as described in the Supplementary Information.