Behavioural food choice paradigm, theta-burst stimulation protocol, and behavioral regressions.

(a) Example of decision stage. Participants were cued in advance about the type of decision required. Perceptual decisions required participants to choose the food item with the largest size while value-based decisions required participants to choose the food item they preferred to consume at the end of the experiment. Participants alternated between blocks of perceptual (blue) or value-based (red) choice trials (7-9 trials per task-block). (b) Regression results show that the larger the evidence strength, the more likely decision makers will respond accurately. Choice accuracy is only related to the evidence that is currently task-relevant (size difference SD for perceptual or value difference VD for value-based choice), not to the task-irrelevant evidence (RT is reaction time of current choice). (c) Similarly, we show that RTs are negatively associated only with the task-relevant evidence (and lower for perceptual choices overall, captured by regressor CH (1 = perceptual, 0 = value-based)). Consistent with previous findings, the results in (b) and (c) confirm that our paradigm can distinguish and compare evidence processing for matched perceptual- and value-based decisions. Error bars in (b) and (c) represent the 95% confidence interval range of the estimated effect sizes. * p < 0.05, ** p < 0.01, and *** p < 0.001. (d) Theta-burst stimulation protocol. After the fourth pre-TMS run, participants received continuous theta-burst stimulation (cTBS) over the left SFS region of interest (ROI) (area encircled and colored blue). cTBS consisted of 200 trains of 600 pulses of 5 Hz frequency for 50 s.

Study hypotheses.

Scenario 1: left SFS is causally involved in evidence accumulation. Theta-burst induced inhibition of left SFS should lead to reduced evidence accumulation (a), expressed as lower accuracy (a, 2nd row, left), slowing of RTs (a, 2nd row, right), and a reduction of DDM drift rate (b, right) without any effect on the boundary parameter (b, left). Since the neural activity devoted to evidence accumulation (area under the curve) should increase (c, left), we would expect higher BOLD signal in this case (c, right). Scenario 2: left SFS is causally involved in setting the choice criterion. Theta-burst induced inhibition of left SFS should lead to a lower choice criterion (d), expressed as lower choice accuracy (d, 2nd row, left), faster RTs (d, 2nd row, right), and a reduced DDM decision boundary parameter (e, left) without any effect on the DDM drift-rate (e, right). At the neural level, we should observe reduced BOLD activity due to the lower amount of evidence processed by the neurons (f, right), and reflected by the smaller area under the evidence-accumulation curve when it reaches the lower boundary (f, left).

Theta-burst stimulation over the left SFS affects choice behavior and selectively lowers the decision boundary for perceptual but not value-based choices.

(a) Choice accuracies/ consistencies and (b) response times (RTs) for perceptual (blue) and value-based (orange) decisions for different evidence levels during pre-cTBS (dark) and post-cTBS (light) stimulation periods. Error bars in (a) and (b) represent s.e.m. Consistent with previous findings, stronger evidence leads to more accurate choices and faster RTs in both types of decisions. Importantly, theta-burst stimulation significantly lowered choice accuracy selectively for perceptual, not value-based decisions (negative main stimulation effect for perceptual decisions and negative stimulation × task interaction; Supplementary Fig. 2c and see also Supplementary Fig. 2a for changes in choice accuracy across runs). Additionally, theta-burst stimulation also significantly lowered RTs in both choice types (negative main stimulation effect; Supplementary Fig. 2c and see also Supplementary Fig. 2b for changes in RTs across runs). (c) Theta-burst stimulation selectively decreased the decision boundary in perceptual decisions only (difference between estimated posterior population distributions; see Methods and Supplementary Fig. 5a for a detailed post-hoc analysis). All the other parameters, particularly (d) the drift rate (see also Supplementary Fig. 5b for post-hoc analysis) remain unaffected by stimulation. Error bars in (c) and (d) represent the 95% confidence interval range of the posterior estimates of the DDM parameters. * p < 0.05, ** p < 0.01, and *** p < 0.001.

Neural representation of accumulated evidence in the left SFS is disrupted after theta-burst stimulation, and is linked with behavior and neural computation.

(a) Left panel: Accumulated evidence (AE) simulation derived from the fitted DDM (left panel). Previous studies have illustrated how the accumulation-to-bound process convolved with the hemodynamic response function (HRF) results in BOLD signals; hence, the simulated AE provides a suitable prediction of BOLD responses in brain regions involved in evidence accumulation. Theta-burst stimulation selectively decreased AE for (a) perceptual (blue), not (b) value-based (orange) decisions (see Supplementary Fig. 6b for post-hoc analysis). We constructed a trialwise measure of accumulated evidence using RTs and evidence strength for our parametric modulator (see Methods). Individual ROIs extracted from the left SFS representing accumulated evidence across runs (right panels; see Methods) show that consistent with the DDM prediction, theta-burst stimulation selectively decreased BOLD response representing AE in left SFS during perceptual, not value-based decisions. Error bars in the left panels of (a) and (b) represent the 95% confidence interval range of the posterior estimates of the DDM parameters, while error bars in their respective right panels represent s.e.m. (c) Comparing pre- and post-cTBS contrasts of BOLD signals related to accumulated evidence, during perceptual decisions, show signal changes in left SFS (green) after theta-burst stimulation. Further contrasts comparing pre-post difference across both choice types (blue) confirm the selectivity of TMS effects for perceptual decisions. (d) To test the link between neural and behavioural effects of TMS, regression results show that after stimulation, BOLD changes in left SFS are associated with lower choice accuracy (left panel) for perceptual (PDM, blue) (negative left SFS × stimulation interaction) but not value-based choices (VDM, red), with significant differences between the effects on both choice types (difference-in-difference, DID, green, negative left SFS × stimulation × task interaction). On the other hand, cTBA-induced changes in left SFS activity are unrelated to changes in RT (right panel). Error bars in (d) represent the 95% confidence interval range of the estimated effect sizes. *p < 0.05, ** p < 0.01, and *** p < 0.001. (e) To test the link between neural activity and DDM computations, we included trialwise beta estimates of left-SFS BOLD signals as inputs to the DDM. Alternative models tested whether trialwise left-SFS (LD) activity modulates the decision boundary (α) (Model 1), the drift rate (β), or a combination of both (Models 3 and 4, see Methods and Supplementary Fig. 8 for more details). Model comparisons using the deviance information criterion (DIC, smaller values mean better fits) showed that Model 1 fits the data best, confirming that the left SFS is involved in selectively changing the decision boundary for perceptual decisions.

SFS-TMS-related changes in behaviour and neural computations are accompanied by increased functional coupling between the left SFS and occipital cortex.

(a) Psychophysiological interaction (PPI) analysis reveals an area in occipital cortex showing increased functional coupling with the left SFS during perceptual choices. (b) ROI analysis of individual PPI betas shows that aE-related functional coupling between the left SFS and OCC is selectively increased post stimulation during perceptual (left panel) but not value-based decisions (right panel). Error bars in (b) represent s.e.m. (c) Regression results testing the link between cTBS effects on left SFS-OCC functional coupling and behaviour. Increased SFS-OCC coupling is associated with lower choice accuracy (left panel) specifically for perceptual (PDM, blue, negative OCC × stimulation interaction) but not value-based choices (VDM, red). In addition, increased functional coupling is also associated with faster RTs (right panel) for perceptual (blue, negative OCC × stimulation interaction) and slower RTs for value-based choice (red, positive OCC × stimulation interaction). We corroborate these behavioral results with results at the computational level, where we show consistent associations of increased SFS-OCC functional coupling with lower boundary (Supplementary Fig. 14a,c), faster decision times (Supplementary Fig. 15a,c) and lower accumulated evidence (Supplementary Fig. 15b,c). Error bars in (c) represent the 95% confidence interval range of the estimated effect sizes. * p < 0.05, ** p < 0.01, and *** p < 0.001.

Domain-general and domain-specific regions involved in perceptual and value-based decisions.

(a) Domain-general regions. We found domain-general regions shared by both perceptual decision making (PDM) and value-based decision making (VDM), such as areas in the visual stream along the fusiform gyrus, cerebellar areas including the brainstem and motor areas such as premotor cortex and SMA (conjunction a p < 0.05, cluster-corrected, see Supplementary Table 1 for the complete list of regions). (b) Domain-specific regions. Comparing average decision-related activity between perceptual and value-based decisions revealed distinct brain activations. Blue-green represents significant neural activity for PDM > VDM decisions while red-yellow represents significant activity for VDM > PDM (see Supplementary Table 2). Among the active regions in the VDM > PDM contrast include the orbitofrontal cortex, the posterior cingulate cortex, and the media prefrontal cortex while the regions active in the PDM > VDM contrast include the frontal eye fields, the intraparietal sulcus and premotor cortex.

Theta-burst stimulation in left SFS selectively lowers choice accuracy for perceptual decisions, but RTs become faster after stimulation in both choice types.

(a) Observed accuracies and (b) mean RTs of individual participants for perceptual decision making, PDM (blue) and value-based decision making, VDM (orange-yellow) across pre-stimulation and post-stimulation runs. Error bars in (a) and (b) represent s.e.m. Theta-burst stimulation significantly lowered choice accuracy for PDM, not VDM, especially during the first post-stimulation run. On the other hand, RTs sped up in both perceptual and value-based decisions. Unsurprisingly, both accuracy levels and RTs began to recover during the second post-stimulation run for PDM. (c) In our regression analysis, comparing the pre-post PDM difference in accuracy with the pre-post VDM difference confirm the significant effect of theta-burst stimulation in selectively lowering choice accuracy (left panel) for perceptual decisions (green, negative stimulation × task interaction). In contrast, the effect is nonspecific for RTs (right panel). Error bars in (c) represent the 95% confidence interval range of the estimated effect sizes. * p < 0.05, ** p < 0.01, and *** p < 0.001.

Hierarchical Bayesian DDM.

Graphical representation of the hierarchical Bayesian DDM fitted to choice data. Clear circles represent latent variables while filled circles represent observed variables, i.e. choice data, y, and evidence, E, for each trial i. Choice data contain both accuracy and response times. The following equations show the distributions assumed for each of the latent variables in the model:

For the hyper-group or latent parameters at the highest level of the hierarchy (represented by μx and σx), we assumed flat uniform priors. The distributions for the following are:

The following model parameters are: α (decision threshold), κ (drift rate parameter scaling the evidence, E), and r (nondecision times).

The DDM disentangles the latent decision-relevant and decision-irrelevant processes observed with faster RTs.

(a) We derived a measure of decision times (DT, upper row) and (b) estimated non-decision times (nDT, lower row) from the DDM. We derived and estimated these parameters to test whether faster RTs in both perceptual decision making (PDM) and value-based decision making (VDM) after stimulation is due to the same or different latent processes in the DDM, and whether these processes are decision-relevant or not. These results show that different latent processes are driving faster RTs for PDM and VDM. Theta-burst stimulation significantly lowered decision times in perceptual (blue), not value-based (orange) decisions. In contrast, stimulation marginally but selectively decreased nondecision times for VDM, not PDM. A Post-hoc analysis confirms domain-specificity of lower nDT for VDM (Supplementary Fig. 5c). Taken together, these results explain what processes underlie observed faster RTs in both decisions. Error bars represent the 95% confidence interval range of the posterior estimates of the DDM parameters. * p < 0.05, ** p < 0.01, and *** p < 0.001.

Theta-burst stimulation in the left SFS reduced decision boundary for perceptual decisions.

(a) Bayesian poster probability distributions of DDM parameters (top row of panel) before (dark color) and after stimulation (light color) for perceptual (blue) and value-based (yellow-orange) decisions. Theta-burst stimulation has lowered the decision boundary for perceptual decision (PDM) not value-based decisions (VDM). (b) Drift-rate was unaffected after theta-burst stimulation, while (c) nondecision times decreased in VDM, not PDM. To test for a stimulation effect, post-hoc tests (middle row of panel) compare pre- and post-stimulation DDM parameters for each type of decision. The highest density interval (HDI) spans within the 95% interval (light color) and represents a null effect. We can statistically confirm the effect of theta-burst stimulation if the decision criterion (dashed vertical line) is outside the 95% HDI (dark color). In our post-hoc analysis, only the decision boundary during PDM had its criterion outside the 95% HDI and marginally for nondecision times. However, the criterion for the drift rate was well within the 95% HDI. Hence, these results show that theta-burst stimulation significantly decreased the decision boundary for PDM and nondecision times for VDM. To test whether stimulation is selective for only one decision domain, our post-hoc test (green, bottom row) compared the pre-post PDM differences with the pre-post VDM differences. In this analysis, both the criterion for the decision boundary and nondecision times were outside the 95% HDI.

Simulations of fitted model: Theta-burst stimulation in the left SFS reduced decision times and accumulated evidence.

(a) Bayesian posterior probability distributions (top row) of decision times and (b) accumulated evidence before (dark color) and after (light color) for perceptual (blue) and value-based (yellow-orange) decisions. Post-hoc tests (middle row) comparing pre- and post-stimulation revealed that theta-burst stimulation significantly decreased both decision times and accumulated evidence for perceptual, not value-based decisions. However, post-hoc tests comparing the pre-post PDM differences with the pre-post VDM differences showed that the criterion is marginally inside the 95% HDI.

Neural representations of accumulated evidence across the whole brain for PDM and VDM.

(a) Conjunction between areas representing accumulated evidence (AE) during perceptual (PDM) and value-based decisions (VDM) reveals activations in visual and parietal areas, such as the cuneus, postcentral gyrus, and lingual gyrus (see Supplementary Table 3). (b) Representations of accumulated evidence for PDM and VDM. Particularly, BOLD responses associated with AE in perceptual decisions are seen in the left SFS (see Supplementary Table 3). (c) Negative conjunction contrasts shared by both PDM and VDM represent the efficiency of evidence accumulation (i.e., the inverse of accumulated evidence). These contrasts reveal activations in occipital and parietal areas (see Supplementary Table 3). Particularly, parietal areas have been previously implicated in the efficiency of accumulating evidence. (d) Negative contrasts for each choice domain reveal activations in parietal and occipital areas (see Supplementary Table 4). (e) Contrasts comparing the neural representation of accumulated evidence between value-based and perceptual decisions revealed activations in areas such as the ventromedial prefrontal cortex (vmPFC) and the nucleus accumbens (see Supplementary Table 4).

Neural-HDDM Alternatives.

To test whether the left SFS is mechanistically involved in the latent decision-relevant processes, we included trial-by-trial left SFS neural betas as an input to our hierarchical drift diffusion model. We compared our neural HDDM with the HDDM without neural inputs. We additively incorporated trialwise left SFS and used a scale parameter, mc,s, to account for its modulatory effects. The following equations show the distributions assumed for each of the latent variables in the model:

Particularly, (a) Model 1 incorporates trialwise left SFS betas into the decision boundary. (b) Model 2 incorporates trialwise left SFS betas into the drift rate. (c) Model 3 includes trial-by-trial betas in both decision thresholds and the drift rate, but with separate scale parameters. (d) Model 4 includes trial-by-trial betas in both latent parameters but with one common scale parameter.

Reanalysis of latent DDM parameters using the neural-HDDM confirmed the results of lower decision boundary in PDM and lower nondecision times in VDM.

We used the winning neural-HDDM model where the left SFS is modulating the decision threshold (Supplementary Fig. 8a). (a) Bayesian posterior probability distributions (top row) of all DDM parameters show similar effects with the DDM without neural inputs. Post-hoc tests (middle and bottom rows) comparing pre-post perceptual decision making (PDM) differences with pre-post value-based decision making (VDM) differences show that theta-burst stimulation selectively reduced the decision boundary for perceptual, not value-based decisions. In particular, the criterion is well outside the 95% HDI. (b) The mean drift rate remained unaffected even with our neural-HDDM model (the criterion is within the 95% HDI). (c) Post-hoc tests also revealed lower nondecision times specifically for VDM, not PDM after stimulation (the criterion is outside the 95% HDI).

Reanalysis of decision times and accumulated evidence using the neural-HDDM provide improvements in model evidence and clearer statistical inference.

Using the winning neural-HDDM model (Supplementary Fig. 8a), we derived measures for (a) decision times and (b) accumulated evidence and tested whether improvements in model evidence reflect improvements in statistical inference. Post-hoc tests comparing pre-post PDM difference revealed a main stimulation effect in both DT and AE for PDM (middle row, criteria outside the 95% HDI). Further post-hoc analysis (bottom row) comparing the pre-post PDM difference with the pre-post VDM difference show that decision times are now marginally close to the border of the 95% HDI while accumulated evidence is now outside the 95% HDI.

Average brain activity that is common for both types of choice (conjunction between PDM and VDM trials)

Average brain activity that is distinct for both types of choice.

Average brain activity that represents evidence accumulation for both types of choice

Average brain activity that represents efficiency of evidence accumulation for both types of choice

Differences-in-differences results for choice accuracy/consistency and response times