Brain stimulation boosts perceptual learning by altering sensory GABAergic plasticity and functional connectivity

Interpreting cluttered scenes —a key skill for successfully interacting with our environment— relies on our ability to select relevant sensory signals while filtering out noise. Training is known to improve our ability to make these perceptual judgements by altering local processing in sensory brain areas. Yet, the brain-wide network mechanisms that mediate our ability for perceptual learning remain largely unknown. Here, we combine transcranial direct current stimulation (tDCS) with multi-modal brain measures to modulate cortical excitability during training on a signal-in-noise task (i.e. detection of visual patterns in noise) and test directly the link between processing in visual cortex and its interactions with decision-related areas (i.e. posterior parietal cortex). We test whether brain stimulation alters inhibitory processing in visual cortex, as measured by magnetic resonance spectroscopy (MRS) of GABA and functional connectivity between visual and posterior parietal cortex, as measured by resting state functional magnetic resonance imaging (rs-fMRI). We show that anodal tDCS during training results in faster learning and decreased GABA+ during training, before these changes occur for training without stimulation (i.e. sham). Further, anodal tDCS decreases occipito-parietal interactions and time-varying connectivity across the visual cortex. Our findings demonstrate that tDCS boosts learning by accelerating visual GABAergic plasticity and altering interactions between visual and decision-related areas, suggesting that training optimises gain control mechanisms (i.e. GABAergic inhibition) and functional inter-areal interactions to support perceptual learning.

It is unlikely that these changes in GABA+ for the intervention groups were due to differences 158 in MRS data quality (i.e. linewidth, Signal-to-noise ratio: SNR) between groups (Table S1). In our results are specific to GABA+. 168 Next, we tested whether changes in OCT GABA+ relate to behavioural performance. 169 We computed percent GABA+ change during tDCS (During) compared to GABA+ before 170 stimulation (Pre) to control for variability in baseline GABA+ measures (i.e. Pre). We  Figure 4b). These correlations were significantly different between groups (Fisher's 176 z test: z=-2.41, p=0.016). Further, this relationship remained significant when controlling for tissue composition within the MRS voxel, controlling for MRS data quality (i.e. linewidth, 178 SNR), and using GABA+ referenced to NAA rather than water (Table S2). There was no 179 significant correlation for OCT Glutamate (Glu) change and learning rate, suggesting that this 180 result is specific to GABA (Table S2). We found no significant correlation between learning 181 rate on the contrast-detection task and change in GABA+ for the Control group (r(20)=0.06, 182 p=0.790). Finally, there was no significant correlation between learning rate and OCT GABA+   Anodal tDCS alters functional connectivity 193 We next tested whether anodal tDCS during training on the SN task alters extrinsic (i.e. 194 between OCT and intra-parietal sulcus [IPS]) or intrinsic (i.e. within OCT) connectivity as 195 measured by rs-fMRI. First, we tested for changes in extrinsic OCT-IPS connectivity after vs.   the Anodal group (t(21)=-2.16, p=0.042), but no significant change for the Sham group 201 (t(14)=1.96, p=0.071). Next, we asked whether changes in extrinsic connectivity relate to 202 behaviour (i.e. learning rate) and OCT GABA+ change during stimulation, as our analysis 203 showed GABA+ changes during stimulation that relate to behaviour for the Anodal rather than 204 the Sham group. We found a significant positive correlation of OCT-IPS connectivity change 205 with learning rate for the Sham (r(11)=0.74, p=0.003; Figure 5b), but not for the Anodal group 206 (r(20)=-0.24, p=0.286; Figure 5b). Comparing these correlations showed a significant 207 difference between groups (Fisher's z test: z=-3.06, p=0.002). Further, we found a significant within the MRS voxel, controlling for MRS data quality, and using GABA+ referenced to NAA 212 rather than water (Table S3). There was no significant correlation for OCT Glu change and 213 learning rate, suggesting that this result is specific to GABA (Table S3). 214 These results demonstrate that anodal OCT stimulation results in decreased occipito-215 parietal connectivity after training that relates to decreased OCT GABA+ during stimulation, 216 suggesting that enhanced GABAergic plasticity due to tDCS in the OCT may relate to local 217 visual processing rather than occipito-parietal interactions. In contrast, for task training without 218 stimulation (i.e. sham stimulation), learning-dependent changes in occipito-parietal 219 connectivity relate to faster learning but not changes in GABA+.    Figure 6b). However, we observed a significant negative correlation for 230 change in intrinsic OCT connectivity with change in OCT GABA+ for the Anodal group 231 (r(14)=-0.52, p=0.039; Figure 6c), but not for the Sham group (r(13)=0.21, p=0.453; Figure   232 6c). This relationship remained significant when controlling for tissue composition within the 233 MRS voxel, controlling for MRS data quality, and using GABA+ referenced to NAA rather 234 than water (Table S3). There was no significant correlation for OCT Glu change and learning 235 rate, suggesting that this result is specific to GABA (Table S3). Taken together, our results

236
show that increased local OCT connectivity relates to decreases in OCT GABA+ during anodal 237 but not sham stimulation, providing converging evidence that enhanced GABAergic plasticity 238 due to anodal tDCS in the OCT relates to local visual processing.  Our functional connectivity analysis shows that our intervention (anodal tDCS during task 243 training) alters occipito-parietal interactions that relate to GABAergic plasticity. However, 244 static connectivity offers a summary measure of the synchrony between two brain regions 245 across long timescales (i.e. 8mins for our rs-fMRI scans) that does not capture short-lived 246 changes in inter-regional synchrony and how they propagate across different brain regions.   a time-varying connectivity analysis (i.e. HMM) to detect brain states that capture recurring patterns of activity and connectivity over time and test whether our intervention alters these 253 brain states. 254 We conducted this analysis using time courses from early and higher visual areas and      We show that anodal stimulation alters inter-regional synchrony at both coarse (i.e. static

382
Participants 383 We tested forty-five healthy volunteers (27 female; mean age 22.9 ± 3.3 years) in two 384 intervention groups, twenty-four in the stimulation group (Anodal) and twenty-one in the no-385 stimulation group (Sham). We tested an additional no-intervention group of twenty-two healthy 386 volunteers who did not receive training nor stimulation (Control: 17 female; mean age 25.8 ± 387 4.2 years). All participants were right-handed, had normal or corrected-to-normal vision, did 388 not receive any prescription medication, were naïve to the aim of the study, gave written Stimuli 392 We presented participants with Glass patterns (Glass, 1969) generated using previously   Radial (spiral angle: 0°) and concentric stimuli (spiral angle: ± 90°) were presented at 23% or 409 25% signal level counterbalanced across trials; noise dipoles were presented at random position 410 and orientation. To control for potential local adaptation due to stimulus repetition and ensure 411 that learning related to global shape rather than local stimulus features, we jittered (± 1-3°) the 412 spiral angle across stimuli.

415
All participants in the intervention groups took part in a single brain imaging session during 416 which they were randomly assigned to the Anodal or Sham group. Participants in the Anodal 417 group received anodal tDCS on the right OCT, whereas participants in the Sham group did not 418 receive stimulation. We recorded three MRS measurements from the right OCT during 419 training: before, during and after stimulation. In addition, we recorded whole-brain rs-fMRI 420 data before and after training while participants fixated on a cross at the centre of the screen 421 ( Figure 1b). Participants in the no-intervention Control group took part in a single brain 422 imaging session without stimulation or training; we recorded three MRS measurements from 423 right OCT at the same timings of the MRS measurements as for the intervention groups. We 424 did not record rs-fMRI data for this group due to time constraints.

425
During training, participants in the intervention groups were presented with Glass 426 patterns and were asked to judge and indicate by button press whether the presented stimulus 427 in each trial was radial or concentric. Two stimulus conditions (radial vs. concentric Glass 428 patterns; 100 trials per condition), were presented for each training block. For each trial, a 429 stimulus was presented for 300ms and was followed by fixation (i.e., blank screen with a central 430 fixation dot) while waiting for the participant's response (self-paced training paradigm). Trial-431 by-trial feedback was provided by means of a visual cue (green tick for correct, red 'x' for 432 incorrect) followed by a fixation dot for 500ms before the onset of the next trial.

433
In the no-intervention control group, participants were tested in a contrast change 434 detection task. In particular, participants were presented with Glass patterns where 100% of the 435 dipoles were randomly oriented (0% signal patterns). In each trial, participants were asked to   Behavioural data analysis 489 We measured behavioural performance per training block as the mean accuracy per 200 trials. 490 To quantify learning-dependent changes in behaviour, we computed the behavioural 491 performance before, during and after stimulation as the average performance of blocks 1-2 492 (Pre), 3-5 (During) and 6-9 (Post), respectively. Further, we quantified learning rate by fitting  500 We pre-processed the MRS data using MRspa v1.5c (www.cmrr.umn.edu/downloads/mrspa/). 501 We applied Eddy current, frequency and phase correction before subtracting the average ON   (Table S1). To control for potential 517 differences in data quality across participants and blocks, we performed control analyses that 518 accounted for changes in linewidth and SNR (Table S2, Table S3). We did not include control  (Table S2, Table S3 530 We pre-processed the structural and the rs-fMRI data in SPM12.4 (v7219; 531 www.fil.ion.ucl.ac.uk/spm/software/spm12/) following the Human Connectome Project 532 pipeline for multi-band data (Smith et al., 2013). In particular, we first coregistered (non-533 linearly) the T1w structural images (after brain extraction) to MNI space to ensure that all 534 participant data were in the same stereotactic space for statistical analysis. We then (a) 535 corrected the EPI data for susceptibility distortions (fieldmap correction) and any spatial 536 misalignments between EPI volumes due to head movement (i.e. aligned each run to its single 537 band reference image), (b) coregistered the second EPI run to the first (rigid body) to correct any spatial misalignments between runs, (c) coregistered the first EPI run to the structural 539 image (rigid body) and (d) normalised them to MNI space for subsequent statistical analyses 540 (applying the deformation field of the structural images). Data were only resliced after MNI 541 normalisation to minimise the number of interpolation steps. Following MNI normalisation, 542 (e) data were skull-stripped, (f) spatially smoothed with a 4mm Gaussian kernel to improve the 543 signal-to-noise ratio and the alignment between participant data (two times the voxel size;  To clean the fMRI signals from signals related to motion and the noise components, we 563 followed a soft clean-up ICA denoise approach (Griffanti et al., 2014). That is, we first                  We fitted model spectra of γ-amino-butyric acid (GABA), Glutamate (Glu), Glutamine (Gln) and N acetylaspartate (NAA) to the edited spectra. The model spectra of were generated based on previously reported chemical shifts and coupling constants using the GAMMA/PyGAMMA simulation library of VESPA for carrying out the density matrix formalism. A 20 x 20 spatial matrix was used to simulate the spatial variations inside and outside the nominal PRESS dimensions. Simulations were performed with the same RF pulses and sequence timings as that on the 3T system in use.

Data Quality a. Reported variables
See Table S1 b. Data exclusion criteria Water peak linewidth > 10 Hz CRLB > 15% GABA+ concentration outside three standard deviations from the mean across all groups and blocks.
c. Quality measures of postprocessing Model fitting See Table S1 d. Sample Spectrum See Figure 3