Learning to optimize perceptual decisions through suppressive interactions in the human brain

Translating noisy sensory signals to perceptual decisions is critical for successful interactions in complex environments. Learning is known to improve perceptual judgments by filtering external noise and task-irrelevant information. Yet, little is known about the brain mechanisms that mediate learning-dependent suppression. Here, we employ ultra-high field magnetic resonance spectroscopy of GABA to test whether suppressive processing in decision-related and visual areas facilitates perceptual judgments during training. We demonstrate that parietal GABA relates to suppression of task-irrelevant information, while learning-dependent changes in visual GABA relate to enhanced performance in target detection and feature discrimination tasks. Combining GABA measurements with functional brain connectivity demonstrates that training on a target detection task involves local connectivity and disinhibition of visual cortex, while training on a feature discrimination task involves inter-cortical interactions that relate to suppressive visual processing. Our findings provide evidence that learning optimizes perceptual decisions through suppressive interactions in decision-related networks.


MRS acquisition
We chose to utilize a short-echo, full signal intensity semi-LASER sequence to achieve lower apparent T2 relaxation, minimal J-coupling evolution and smaller chemical shift displacement errors relative to the PRESS and STEAM sequences 1 . In addition, the adiabatic refocusing pulses in the semi-LASER provided minimal signal loss, high B1+ insensitivity and localization against the varying destructive interferences throughout the brain at ultrahigh field. This MRS sequence has been extensively tested and resulted in high quality spectra ( Supplementary Figures 1,2,3) across high and ultra-high field magnetic fields at different MRI centers 1-5 . We used VAPOR 6 water suppression and outer volume suppression 5 .
A dielectric pad (BaTiO3, 14.5 x 12.5 cm 2 ) was placed over the left occipito-parietal cortex to increase B1 efficiency in the regions where the MRS voxels were placed 7 . First and second order shims were adjusted for each voxel separately using FASTMAP (fast, automatic shimming technique by mapping along projections) with echo-planar imaging readout 8 .
Acquisition parameters were optimized for each voxel by determining the appropriate transmit voltage (flip angle calibration) and flip angle (VAPOR calibration), to maximize readout and water suppression respectively 7 .
Two non-suppressed water spectra were acquired: one for eddy current correction and reconstruction of the phased array spectra (the RF pulses of the VAPOR scheme were turned off, NT = 2, TR = 5.010 s, TE = 36 ms, number of dummy scans = 2, spectral bandwidth = 6 kHz, data points = 2048) and one for use as reference for metabolite quantification (VAPOR and OVS schemes turned off in order to eliminate magnetization transfer effects, NT = 2, TR = 5.010 s, TE = 36 ms, number of dummy scans = 2, spectral bandwidth = 6 kHz, data points = 2048). The reconstruction of the phased array spectra included weighting the spectra based on the sensitivity of each receive element at the VOI and correcting for the different constant phase shift terms of the complex spectra prior to the summation. Single scan spectra summed from 32 channels were corrected for frequency and phase variations induced by subject motion and then summed before further analyses.

Behavioral data analysis
To take into account individual variability in performance, we estimated behavioral improvement as the difference in mean performance (i.e. mean accuracy per 200 trials) between the first training block and the training block with maximum performance per participant (85% of the participants achieved maximum performance during the last two MRS measurements), divided by performance in the first training block.

MRS data analysis
Eddy-current correction and reconstruction of the phased array spectra was applied using inhouse scripts. Water residual signal was removed using a Hankel singular value decomposition (HLSVD) MATLAB routine 9 . LC-Model 10 was used to quantify metabolite concentrations in the range of 0.5 to 4.2 ppm (Fig. 2c, Supplementary Figures 1,2) using optimal initialization parameters.
The model spectra of aspartate (Asp), ascorbate/vitamin C (Asc), glycerophosphocholine Analysis) for carrying out the density matrix formalism. Simulations were performed with the same RF pulses and sequence timings as that on the 7T system in use 11 .
We followed the same macromolecule inclusion procedure as Bednarik et al. 12 .
Macromolecule spectra acquired from the occipital cortex from 3 healthy volunteers, using an inversion recovery sequence (TR=3 s, TE=36 ms, inversion time TI=0.685 s), were included in the LCModel basis set. The residual signal of the methylene of tCr at 3.93 ppm was removed by post processing and the high-frequency noise was suppressed using a Gaussian filter (σ=0.05 s) before including the macromolecule spectrum into the LCModel basis set.
We referenced metabolite concentrations to the sum of the concentrations of Creatine (Cr) and Phosphocreatine (PCr), that is total Creatine (tCr). In particular, for each MRS voxel, we normalized GABA/tCr in each training block to GABA/tCr in the baseline block (Fig. 3). We computed GABA/tCr change for each participant as the difference between GABA/tCr in the training block with maximum performance and GABA/tCr in the baseline block. We chose tCr as reference for two main reasons. First, tCr concentration was measured in the same spectrum, concurrently with GABA, while water concentration was estimated from a different scan. Using another metabolite acquired in the same spectrum as reference accounts for the possibility of small but relevant changes in neuronal density and spectral data quality that might be expected during periods of task activity 13 . Second, referencing metabolites to tCr has been shown to have better reproducibility compared to other referencing methods 14 and has been widely used as a reference metabolite in MRS studies 15,16 . Our control analysis (see 'Control analyses') confirmed that tCr concentration did not change significantly during training across voxels or tasks, suggesting that our results are specific to GABA changes and are not driven by changes in tCr concentration. Finally, we replicated our findings (Supplementary Figure 4) using absolute GABA quantification (GABA referenced to water) to ensure that our results were not driven by the chosen reference (e.g. 3,17,18 ).
Only data without lipid contamination, GABA CRLB values smaller than one standard deviation above the mean and GABA values per block within two standard deviations from the mean across participants were included in further steps of MRS related analyses. That is, OCT data for 6 participants (2 for SN, 4 for FD) and PPC data for 6 participants (4 for SN, 2 for FD) were excluded due to high CRLB values. Thus, data from 18 participants were included for further analysis for the SN and 22 participants for the FD task. To account for variability in tissue composition within the MRS voxel across participants, we conducted whole brain tissue-type segmentation of the T1-weighted anatomical scan using SPM12.2 (SPM segment) and calculated percentage of gray matter (GM) and white matter (WM) voxels in each of the MRS voxels.

rs-fMRI data pre-processing
We pre-processed the resting-state fMRI (rs-fMRI) data using SPM12.2 (http://www.fil.ion.ucl.ac.uk/spm/software/spm12/) following the optimized pipeline described in recent work 19 . Data were excluded from one participant with incomplete data acquisition. We first processed the T1-weighted anatomical images by applying brain extraction and segmentation (SPM segment We modeled the pre-processed data in a first-level analysis model (SPM first-level analysis) using an autoregressive AR(1) model to treat for serial correlations and regressing out the signal from CSF, WM, the motion parameters (translation, rotation and their squares and derivatives) and the signal from noise components (i.e. components overlapping with ventricles or brainstem) 21 .

Functional connectivity analysis
We computed functional connectivity measures (connectivity between MRS voxels, temporal coherence within each MRS voxel) based on the following method. We computed the overlap across participant MRS voxels for OCT and PPC separately and created a group MRS mask that included gray matter voxels present in at least 50% of the participants' MRS voxels. For each participant, we extracted the average time course of the gray matter voxels within each MRS mask. We then applied a 5th order Butterworth band-pass filter, between 0.01 and 0.08 Hz, to remove effects of scanner noise and physiological signals (respiration, heart beat) 22 .
We computed the functional connectivity between the OCT and the PPC MRS voxels as the

MRS Data quality controls
We provide data quality metrics and statistics showing that data quality is highly similar across MRS voxels and participant groups. In particular, we considered: a) differences in that it is not appropriate to reject spectra due to changes in CRLBs over time.

Control analyses
We conducted the following control analyses that corroborated our results showing mean changes in GABA with training and correlations with behavioral improvement.
First, to ensure our results were not simply driven by GABA measurements at baseline, we tested a linear mixed effects model on the training blocks only (i.e. excluding the baseline block; LME model for OCT GABA with Task and training MRS Block as fixed effects). This analysis showed a significant interaction between Task and MRS blocks (Task x Block: F(1,83)=4.97, p=0.03) and a significant main effect of MRS block (F(1,83)=4.06, p=0.05).
These analyses suggest that the learning-dependent GABA changes we observed were due to training rather than simply differences in GABA between the training blocks and the baseline.
Second, we demonstrated that the learning-dependent changes we observed in GABA levels could not be simply due to the order with which the MRS voxels were acquired during training. For OCT GABA, the Task x Block interaction remained significant (Task