Predictive Feedback, Early Sensory Representations, and Fast Responses to Predicted Stimuli Depend on NMDA Receptors
Curation statements for this article:-
Curated by eLife
Evaluation Summary:
In this study, the authors investigated how predictions modulate performance using a combination of pharmacological experiments, high-density EEG, Bayesian modeling, and machine learning. This is an interesting study with a complex set of analyses. The detailed assessment and interpretation of all the findings could be strengthened by providing a more unified and hypotheses-driven approach.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (eLife)
Abstract
Learned associations between stimuli allow us to model the world and make predictions, crucial for efficient behavior (e.g., hearing a siren, we expect to see an ambulance and quickly make way). While there are theoretical and computational frameworks for prediction, the circuit and receptor-level mechanisms are unclear. Using high-density EEG, Bayesian modeling, and machine learning, we show that inferred “causal” relationships between stimuli and frontal alpha activity account for reaction times (a proxy for predictions) on a trial-by-trial basis in an audiovisual delayed match-to-sample task which elicited predictions. Predictive β feedback activated sensory representations in advance of predicted stimuli. Low-dose ketamine, an NMDAR blocker, but not the control drug dexmedetomidine, perturbed behavioral indices of predictions, their representation in higher-order cortex, feedback to posterior cortex, and pre-activation of sensory templates in higher-order sensory cortex. This study suggests that predictions depend on alpha activity in higher-order cortex, β feedback, and NMDARs, and ketamine blocks access to learned predictive information.
SIGNIFICANCE STATEMENT We learn the statistical regularities around us, creating associations between sensory stimuli. These associations can be exploited by generating predictions, which enable fast and efficient behavior. When predictions are perturbed, it can negatively influence perception and even contribute to psychiatric disorders, such as schizophrenia. Here we show that the frontal lobe generates predictions and sends them to posterior brain areas, to activate representations of predicted sensory stimuli before their appearance. Oscillations in neural activity (α and β waves) are vital for these predictive mechanisms. The drug ketamine blocks predictions and the underlying mechanisms. This suggests that the generation of predictions in the frontal lobe, and the feedback pre-activating sensory representations in advance of stimuli, depend on NMDARs.
Article activity feed
-
Evaluation Summary:
In this study, the authors investigated how predictions modulate performance using a combination of pharmacological experiments, high-density EEG, Bayesian modeling, and machine learning. This is an interesting study with a complex set of analyses. The detailed assessment and interpretation of all the findings could be strengthened by providing a more unified and hypotheses-driven approach.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)
-
Reviewer #1 (Public Review):
1. The authors are interested in understanding how NMDAR blockade (via Ketamine) affects the relationship between content-predictions and behavior. To do so, the authors need to establish, first, how behavior, including not only reaction times (RTs) but also accuracy, and the inverted efficiency score (IES, which quantifies the relationship between accuracy and RT), change as a function of (1) predictions regardless of drug manipulation and of (2) drug manipulations regardless of predictions. In other words, did predictions increase overall accuracy, shorten RT, and reduce IES in the pre-drug condition, regardless of drug manipulation? did Ketamine (compared to DEX and to pre-/post-drug) affect overall accuracy, RT, and IES, regardless of predictions? After having established these effects in isolation, the …
Reviewer #1 (Public Review):
1. The authors are interested in understanding how NMDAR blockade (via Ketamine) affects the relationship between content-predictions and behavior. To do so, the authors need to establish, first, how behavior, including not only reaction times (RTs) but also accuracy, and the inverted efficiency score (IES, which quantifies the relationship between accuracy and RT), change as a function of (1) predictions regardless of drug manipulation and of (2) drug manipulations regardless of predictions. In other words, did predictions increase overall accuracy, shorten RT, and reduce IES in the pre-drug condition, regardless of drug manipulation? did Ketamine (compared to DEX and to pre-/post-drug) affect overall accuracy, RT, and IES, regardless of predictions? After having established these effects in isolation, the authors can analyze the interactions between prediction and drug manipulations on different behavioral measures and understand whether the effects reported are specific to RT.
2. The authors are interested in understanding how NMDAR blockade (via Ketamine) affects the relationship between content-predictions and neural oscillations in the alpha and beta bands. The authors focused on cue-locked induced activity (baseline-corrected) during the delay window, after the predictive auditory cue and before the to-be-predicted visual stimulus. From the current results, it is unknow whether and how both ketamine and the control drug DEX affect resting-state/ongoing neural oscillations, i.e., in the baseline period. Are the results due to a change in resting-state oscillations or to oscillatory processes induced by predictive cues?
3. The authors state that delay-period (pre-image) alpha power correlated with predictions. To fully understand this correlation, it is important to establish, first, whether, in the pre-drug condition and regardless of prediction manipulation, pre-image alpha power (as opposed to pre-cue alpha power), is related to any behavioral measures (accuracy or RT) using a standard approach (e.g., binning by alpha power; in line with previous studies), rather than the more complex HDDM. Do predictions or drug manipulation modulate this existing relationship between behavior and alpha power? Are these results specific to any ROI?
4. The authors discuss the relationship between alpha oscillations and excitability. I wonder if the authors mean physiological excitability (i.e., neuronal ensemble firing activity) or something else. This is unclear from the reference cited (32-33) since they refer to studies showing a relationship between alpha power and perceptual reports, not physiological excitability. The authors could consider citing here the invasive studies reporting a link between alpha and neuronal excitability (Chapeton et al., 2019; Haegens et al., 2011; Watson et al., 2018; Bollimunta et al., 2008, 2011; Lundqvist et al., 2020; Dougherty et al., 2017; van Kerkoerle et al., 2014). Additionally, I wonder if the authors could elaborate more on how increased excitability may be related to a decreased SNR. This is particularly important to understand a strong statement the authors made: "These results suggest that ketamine blocks predictions by reducing frontal SNR." What are the neurophysiological mechanisms behind this process?
5. The analysis of the ERP is subpar. The authors focus only on an arbitrary channel and a single limited time window. Accordingly, null effects on ERP should be interpreted with caution and statements like: "This was not due to feedforward sensory processing, as all three sounds generated similar auditory ERPs (Figure S3a)." are unwarranted. To improve this analysis, the selection of channels could be participant-specific (e.g., pick the electrodes with strongest early ERP component for each participant) and the statistical testing could be run on the entire time course, for example, using a cluster permutation test across the time dimension. The authors need to address how predictions and drug manipulations affect ERP time course, focusing especially on very early components (<100 ms) which are more likely to reflect feedforward activity. I want to highlight that the authors assume that the ERP is a direct measure of feedforward activity: however, accumulating evidence suggests that the ERP in part reflects prestimulus/baseline oscillatory activity (baseline-shift, Nikulin et al 2007, Iemi et al 2019). Accordingly, any changes of prestimulus oscillatory activity might results in changes in the ERP. Please be cautious when equating ERP to feedforward activity. Were both visual and auditory ERPs analyzed?
6. I found the statistical results difficult to follow and incomplete. The authors refer to two analysis: regression, and repeated measures ANOVA. why were two types of tests used? What are the repeated measures in the ANOVA? is the regression single-trial? I wonder if the authors could build a more complete statistical model including all possible tests (for example, before/after drug, drug1/2, prediction condition, and interactions), thus addressing the issue of running multiple tests on the same data that is present in the current analysis. Additionally, the description of the statistical results is incomplete. Please include all statistical information for each test: e.g., whenever you report a p value, also report the corresponding statistics. Specifically, when you use ANOVA, please include F statistics, degrees of freedom (with Huynd-Feldt correction if necessary), and, critically, run post-hoc comparisons bonferroni-corrected for multiple comparisons (report t statistics and p values of relevant comparisons). While the bar plots are very helpful, they lack statistical significance lines that are necessary to understand the relevant comparisons. Additionally, please report null effects throughout the paper with an appropriate statistic allowing to test for the null hypothesis (i.e., Bayes factor Analysis). This is especially important to understand whether the effects reported are frequency-specific or specific to a region of interest.
I found some statements unclear.
- "This was not due to low accuracy as subjects' average accuracy was 77.8% under ketamine (85.7% without ketamine, and 81.0% under DEX)." Was the difference significant? Can this statement be substantiated with a statistical test (e.g., Bayes factor)? Can you control for accuracy differences in the RT analysis?
- "ketamine average modified observer's assessment of alertness/sedation (OAA/S) score of 4.85 compared to 3.33 under DEX (5, awake - 1, unresponsive))." Can this statement be substantiated with a statistical test?
- "The interaction effect confirmed that, under ketamine, delay period alpha power at the RF electrode cluster was similar across sounds (Figure 2e)." I don't understand what the significant interaction effect means if the alpha power was similar across prediction conditions. Please always explain what the significant main and interaction effects mean.
- When writing: "We ran a regression analysis of prediction (HP, MP, NP) X drug condition (before drug, under drug; drug either ketamine or DEX).": please spell out the dependent variable.It's necessary that the authors are clear about whether the effects they report are specific to a certain frequency band, region of interest in sensor or source space. Please report null effects. Why are some analysis presented in sensor space and other in source space?
-
Reviewer #2 (Public Review):
The study targets the behavioral and neural dynamics and pharmacological routes by which multisensory predictions (auditory to visual), are formed and used to efficiently match corresponding sensory inputs. The approach combines computational modelling of behavioral data, EEG, and ketamine administration as a model of NMDA receptor blockage.
Ketamine blocks clear effects of predictability seen in behavior (response times), and interferes with the pathways (frontal alpha increase, beta power connectivity from frontal to sensory areas) which are identified as carrying out the predictive processing.
The study design is clear and well suited to assess the questions at hand. The number of techniques applied and their level of sophistication is impressive.
I think that the interpretability and message of the paper …Reviewer #2 (Public Review):
The study targets the behavioral and neural dynamics and pharmacological routes by which multisensory predictions (auditory to visual), are formed and used to efficiently match corresponding sensory inputs. The approach combines computational modelling of behavioral data, EEG, and ketamine administration as a model of NMDA receptor blockage.
Ketamine blocks clear effects of predictability seen in behavior (response times), and interferes with the pathways (frontal alpha increase, beta power connectivity from frontal to sensory areas) which are identified as carrying out the predictive processing.
The study design is clear and well suited to assess the questions at hand. The number of techniques applied and their level of sophistication is impressive.
I think that the interpretability and message of the paper can be made stronger and easier to assess. Especially the large number of results presented makes it difficult for the reader to capture the main message.Hypothesis-based approach:
The paper is extremely rich in experimental techniques and analyses (RT, drift diffusion model, evoked responses, time-frequency, connectivity, causality, decoding), and I think the message would clearly benefit from a more reduced and unified approach.
Currently the analyses are introduced successively in the results section, but not all are spelled out a priori. This makes it difficult to distinguish a priori from post-hoc analyses, and also dilutes the interpretability of the results.
Please spell out all a-priori hypotheses in the introduction, and clearly label exploratory or post-hoc analyses as such.
Some reduction might also be beneficial, for example, the connectivity analyses based on specific anatomical regions are treated too briefly to be fully appreciated, and might be better presented in more detail in a separate paper.
Clarity:
The result section is lacking some bits of information, especially since it appears before the methods.
Importantly, please state early on whether the ketamine administration was a within or between subject manipulation.
I suggest to provide a table with all experiments, number of participants, and trials carried out to give a better overview.Also, please give a more detailed description of the stimuli and statistical models (see below), as well as the rationale for the drift diffusion model.
Statistical models:
The results section would be clearer if one omnibus statistical model was spelled out and used for all comparisons. Given the bayesian statistics applied to the drift diffusion model, why not apply a Bayesian approach all together?
The authors use the terms linear regression and correlation, but report ANOVA results without stating in the results section whether the predictability factor was modeled as parametric. It should be discussed whether a 3-level manipulation really qualifies as a numeric regressor.
Baseline effects, for instance general RT or alpha power differences between the treatment groups are currently not consistently reported. The above omnibus model would include them, and allow for a more systematic reporting thereof. More general group differences should also be discussed in light of the specificity of the proposed mechanisms.
Especially the null-effects, crucial for the interpretation that ketamine blocks predictions, need to be justified by more than an insignificant p-value.
For instance: p. 10, l.11: 'The interaction effect confirmed that, under ketamine, delay period alpha power at the RF electrode cluster was similar across sounds'Please consistently report F-values for ANOVAs, degrees of freedom, as well as Bayes Factors to show the absence of an effect. If interactions are interpreted (often the case to show that a given effect disappears under ketamine), the null-effect under treatment should be explicitly tested. An interaction could also result from diverging effect sizes.
Drift diffusion model:
The use of a drift diffusion model is not sufficiently well introduced, especially with respect to the accumulation of evidence.
In the paradigm, no more evidence accumulation happens in the delay period, after the auditory stimulus ended, which is in line with the finding of an effect on bias, but not drift rate. To me, the accumulation of information occurs rather across trials, which is currently captured in the causal power analyses. Please clarify this aspect.How strong is the correlation between transitional probabilities and causal power?
Behavioral results:
p. 6, l.10: 'This was not due to low accuracy as subjects' average accuracy was 77.8% under ketamine (85.7% without ketamine, and 81.0% under DEX).'
Was the difference in accuracy significant? Why would it not matter for the results?
Also, were there overall RT differences under ketamine?Time-frequency results:
Overall, not all tested contrasts are equally well justified, or based on prior hypotheses. Especially, alpha power is only analyzed in a right frontal cluster (based on a significant correlation with predictability), but then later, different regions are addressed in source space, and also different frequency bands.
The procedure for the permutation cluster-based testing needs to be described in more detail (p. 29, ll. 12) - why is it that all found clusters have the same number of electrodes? The reference (6) seems wrong?
All following analyses are performed on the right frontal electrodes. This should be explicitly taken into account in the discussion.
Discussion:
The discussion is very brief, and provides simply a confirmatory assessment of the results, plus a quick and slightly vague outlook on implications for depression.
Please provide a more in-depth assessment of the results, and possible limitations of the findings.
For instance, discuss the specificity of the blocking of NMDA receptors by ketamine, and address general group differences in RT, accuracy, or EEG measures in this respect.The lateralization of the alpha response should also be discussed.
I am admittedly not an expert on neuropharmacological routes of predictability, but an assessment of the novelty and comparability of this approach compared to previous studies (Corlett, 2016; Weber, 2020) should be added.
-
Reviewer #3 (Public Review):
The authors performed several analyses to pinpoint whether and how top-down prediction can modulate performance. They used a combination of pharmacological experiments, high-density EEG, Bayesian modeling, and machine learning. In particular, the behavioural+pharmacological results seem to show strong and consistent effects. To investigate the mechanisms underlying this effect, the authors present several EEG analyses. However, given the analyses' complexity, some of the preprocessing steps and the results' interpretations are not fully clear. For this reason, in its current form, it is difficult to understand the implications of these findings.
-
-
