Psilocybin’s effects on how the brain detects important visual information

Twenty-three participants [four females, 31 ± 4 years, 72 ± 15 kg (mean ± SD)] were recruited through word of mouth and social media advertising. To participate, individuals were instructed to contact the provided number via WhatsApp and subsequently received a phone call from the researchers. During the call, the researchers provided a brief explanation of the details and purpose of the experiment, as well as the inclusion and exclusion criteria. Participants were then given a full written explanation of the study and a copy of the informed consent form. To determine eligibility, all participants underwent a psychiatric interview to screen for exclusion criteria, which are detailed in the supplementary material and summarized below. After this screening, participants and researchers agreed on a date for the start of the experiment. Participants reported having 15 ± 13 prior experiences with serotonergic psychedelics, of which 3.1 ± 2.4 were considered challenging [mean ± SD]. All participants had normal or corrected-to-normal vision.

This study was conducted in accordance with the Declaration of Helsinki and approved by the Research Ethics Committee at the Universidad Abierta Interamericana (Buenos Aires, Argentina), protocol number 0-1068. All participants gave written informed consent and received no financial compensation for their participation in the experiment. All data was collected from the participants in natural settings, i.e. those chosen by the participants without intervention from the researchers. The researchers did not provide nor administered psilocybin to the participants, nor instructed them in any way concerning drug use.

Participants were required to have at least two prior experiences with a dose equal to or exceeding 3 g of dried psilocybin mushrooms. To participate in this research protocol, subjects volunteered to partake in a series of tests under the effects of psilocybin mushrooms and in the presence of four members of the research team. Subjects who consumed serotonergic psychedelics during the 15 days prior to the dosing day were not included in the study. The same applied for all psychoactive substances (including alcohol, caffeine, and tobacco) for a period of 24 h prior to the dosing day. To participate in the experiment, subjects declared their willingness to abstain from using psychedelics between measurement sessions. Drug use prior to the study was assessed by self-report; no biochemical tests were performed.

Subjects who fulfilled the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) criteria for the following disorders were excluded from the experiment: schizophrenia or other psychotic disorders, type 1 or 2 bipolar disorder (including first- and second-degree relatives), personality disorders, dissociative identity disorder, post-traumatic stress disorder, substance abuse or dependence in the past 5 years, depressive disorders, recurrent depressive episodes, obsessive-compulsive disorder, generalized anxiety disorder, dysthymia, panic disorder, bulimia or anorexia, as well as subjects with a history of neurological disorders. Pregnant women and subjects under psychiatric medication of any kind were excluded. Subjects exhibiting potential dysfunctional states as measured by the Depression Anxiety Stress Scale (with scores > 4 for depression, > 3 for anxiety, and >7 for stress) were subsequently reviewed by the clinical interviewer for confirmation.

Experimental conditions (high-dose vs. low-dose) were a priori unknown to participants and researchers as subjects implemented a self-blinding procedure, following the design introduced by Szigeti et al. (2021). The experiment was divided into two parts, one corresponding to the a high-dose condition (3 g of ground and homogenized -mixed to ensure uniform psilocybin distribution- dried psilocybin mushrooms in gel capsules, provided by the participants) and one corresponding to a low-dose control condition (0.5 g of ground and homogenized dried psilocybin mushrooms mixed with edible mushrooms to match the weight of the gel capsules between conditions). The low-dose served as an active control. Chemical analyses of Psilocybe cubensis indicate ~ 1% total psilocybin + psilocin content by dry weight (range ~0.9%–1.3%; Goff et al. 2024). Thus, the 3 g high-dose corresponds to ~ 30–35 mg of active alkaloids, consistent with a moderate dose, while the 0.5 g low-dose corresponds to ~ 5 mg, within a threshold-to-low range. Conditions were separated by an interval of 1 month to attenuate potential tolerance effects. During interviews conducted after completion of the experiment, we confirmed that participants implemented the self-blinding assisted by a third party of their choice. In all cases, experiments took place in the comfortable setting of a house and in the presence of the participant and the team of researchers. During the high-dose session, two participants opted to discontinue the eye-tracking task. Consequently, they were excluded from subsequent analyses of eye-tracking data, although no adverse effects were reported.

Two days before the dosing day for each condition, participants completed a set of self-reported questionnaires designed to assess various psychological traits. These included the State-Trait Anxiety Inventory (Spielberger 1983), the Big Five Inventory (John et al. 1991), the Tellegen Absorption Scale (Tellegen and Atkinson 1974), and the Short Suggestibility Scale (Kotov et al. 2004). Additional self-report scales were administered both before dosing and immediately after the acute effects, including the State-Trait Anxiety Inventory (Spielberger 1983), the Positive and Negative Affect Schedule (Sandin 1999), and the Psychological Well-being Scale (Luna et al. 2020). Before dosing, participants also completed assessments on expectations (EXP) and contextual factors related to prior psychedelic experiences (Haijen et al. 2018). Following the acute effects, participants completed the Mystical Experience Scale (MEQ) (MacLean et al. 2011) and the Altered States of Consciousness Scale (5D-ASC) (Studerus et al. 2010). In addition, participants rated subjective drug effects using a visual analog scale (VAS) at four time points during the dosing day, beginning 1 h after ingestion and continuing at hourly intervals until the effects subsided. A comprehensive description of all questionnaires used in the study is provided in the supplementary material.

A total of 100 images were randomly subsampled from the Object and Semantic Images and Eye-tracking (OSIE) dataset (Xu et al. 2014), which originally comprised 700 images. The original OSIE dataset was deliberately designed with the objective of mitigating the biases that are commonly observed in existing image collections. In contrast to traditional datasets, which often feature a single dominant object in a central position, the images in OSIE are characterized by the presence of multiple dominant objects distributed across the scene. The images are diverse, spanning a wide range of object categories with strong semantic relevance. This design supports detailed analyses of attentional dynamics by incorporating both pixel-level details and object- and semantic-level attributes. The selected images had a resolution of 800 × 600 pixels and were displayed against a uniform grey background to ensure consistent viewing conditions. The same set of images was presented in both the high-dose and the low-dose conditions, ensuring that any potential influence of low-level visual features (e.g. luminance, color) was balanced across conditions. Nevertheless, individual images were not spatially homogeneous in terms of their low-level visual features (i.e. some regions could have higher luminance, color, etc.), which could have increased the saliency of specific regions independently of their semantic content.

The measurements described below were conducted as part of a broader study examining the effects of psychedelics on perception, creativity, language, and music production. Findings related to these domains will be presented in future reports. Prior to the task described in this study, participants completed another eye-tracking task involving free exploration of artworks from various periods and styles, with the results of this experiment detailed in a previous publication (Muller et al. 2023).

For the eye-tracking tasks, participants were seated 60 cm from a centrally positioned monitor, with their heads stabilized using a custom-made chin rest to minimize movement. The eye tracker was positioned between the participant and the screen. The display measured 31 × 17.5 cm and had a resolution of 1920 × 1080 pixels. During the presentation of visual stimuli, gaze coordinates were recorded in both spatial and temporal dimensions using the Gazepoint GP3 HD portable eye tracker, which operates at a temporal resolution of 150 Hz and achieves a visual angle accuracy of 0.5°–1.0°. Stimulus presentation and eye-tracking control were managed by a custom Python script developed using the open-source PsychoPy library (https://www.psychopy.org/↗).

The task reported in this study began 90 min after participants consumed capsules containing either a low or high dose of psilocybin. An overview of the experimental procedure is provided in Fig. 1. Each trial (T) commenced with a fixation cross displayed for 1 s, followed by a stimulus presentation lasting 3 s, during which participants were free to observe the image. The same stimuli were used across all experimental conditions, with their order randomized for each participant and condition. A standard five-point calibration was conducted after every 20th block to ensure that the eye tracking detected fixations accurately throughout the duration of the experiment.

Due to insufficient data quality, i.e. more than 30% of fixations either missing or recorded outside the screen for over 30% of the stimuli (suggestive of lack of engagement with the task), data from six participants was excluded from the present analysis. Furthermore, for each participant and condition, only images with less than 30% of fixations outside the boundaries were analyzed. This threshold corresponds to more than 2 SDs above the mean proportion of outside-screen fixations and applying it resulted in the exclusion of only a small and comparable proportion of trials across conditions. All fixations beyond the image boundaries were excluded from further analyses. No significant differences were found in the number of excluded fixations across experimental conditions.

A velocity-based algorithm (Engbert and Kliegl 2003) was employed to automatically identify fixations with minimum durations of 150 ms following the recording of gaze position coordinates (Manor and Gordon 2003). The mean horizontal (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${x}_i$\end{document}) and vertical (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${y}_i$\end{document}) positions of each fixation, along with the time of the ith fixation (t_i) were then determined, with positions expressed in pixels and time in seconds. For each image, participant, and condition, the following metrics were calculated: the total number of fixations within the image (N), the mean distance between fixations (ds), and the mean time between fixations (dt). The following formulas were applied to compute these measures:

These measures were subsequently averaged across all images, yielding a mean value for each participant under each dosing condition.

To analyze the visual saliency, we first converted each image into a saliency map using DeepGaze II, a computational model of visual attention based on deep neural networks trained to predict human fixations on natural images. The model was pretrained on the SALICON dataset, which contains 10 000 diverse images annotated with mouse-tracking data, and subsequently fine-tuned on the MIT1003 dataset, composed of 1003 natural scene photographs viewed by human participants under eye-tracking for 3 s. Validation was conducted via 10-fold cross-validation on the MIT1003 dataset to ensure the model generalizes well to unseen images within the same domain (Linardos et al. 2021). The generated saliency maps reflect the likelihood of a region capturing visual attention. Each recorded fixation was then mapped to its corresponding saliency value from the respective image's saliency map. The average saliency for each trial was calculated, and a final saliency value for each participant and experimental condition was obtained by averaging across all images. To examine how participants' focus on salient regions evolved over time, we divided the 3-s viewing period into five equal intervals of 600 ms each and calculated the average saliency of the fixations within each subinterval.

To quantify the predictability of fixation patterns in high saliency regions, we computed the Shannon entropy of fixation distributions within high-saliency areas. The saliency map was first binarized using a threshold, which was systematically varied in subsequent analysis to assess the robustness of the results across different saliency levels. The definition of this threshold and the range of values explored are detailed in the following paragraph. Connected clusters of the binary map were identified and treated as individual regions with high visual salience (see Fig. 4a). Each fixation was assigned to a high salience cluster based on its spatial location. The Shannon entropy was then calculated for the resulting distribution of fixations across clusters, using the formula:

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${p}_i$\end{document} represents the proportion of fixations within the ith cluster relative to the total number of fixations within the clusters and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $k$\end{document} is the total number of clusters. Thus, higher entropy values indicate a more uniform distribution of fixations across the high salience clusters of the image.

This analysis only included images with more than one cluster to ensure the entropy calculation accurately reflected variability in fixation patterns. The analysis was performed across a range of thresholds to ensure the results were not biased by the chosen value. The lower bound of the explored range was defined as the threshold at which at least 70% of the images contained more than one cluster, while the upper bound corresponded to the point where the average number of clusters started to decline, indicating a loss of information about salient regions. For each threshold, entropy was compared between dosing conditions.

Resting state EEG data was recorded using a 24-channel mobile system (mBrainTrain LLC, Belgrade, Serbia; http://www.mbraintrain.com/↗) paired with an elastic electrode cap (EASYCAP GmbH, Inning, Germany; www.easycap.de↗). Twenty-four Ag/AgCl electrodes were positioned according to the standard 10–20 system at the following locations: Fp1, Fp2, Fz, F7, F8, FC1, FC2, Cz, C3, C4, T7, T8, CPz, CP1, CP2, CP5, CP6, TP9, TP10, Pz, P3, P4, O1, and O2. Reference and ground electrodes were placed at FCz and AFz, respectively. The wireless EEG DC amplifier (weight: 60 g; dimensions: 82 × 51 × 12 mm; resolution: 24-bit; sampling rate: 500 Hz; passband: 0–250 Hz) was securely attached to the back of the electrode cap, between electrodes O1 and O2. EEG signals were digitized and transmitted via Bluetooth to a notebook operated by the experimenter, who was positioned behind the participant. Participants were instructed to undertake 5-min recordings with their eyes closed, both 15 min prior to and 2 h after the administration of the dose. During these recordings, the subjects were asked to keep their eyes fully closed, relax all muscles, and avoid any physical tension.

EEG data was preprocessed using the Python MNE library (Gramfort 2013). A band-pass filter with a frequency range of 0.5–90 Hz was applied to the data to remove low-frequency drift and high-frequency noise. Additionally, a notch filter was employed to eliminate power line noise at 50 Hz and its harmonics. The filtered data were then segmented into fixed-length epochs of 2 s for subsequent analysis. Automatic epoch rejection was conducted using MNE's autoreject algorithm (Jas et al. 2017), which identified and excluded epochs characterized by excessive noise. A manual inspection was conducted to ensure the quality of the data. Independent component analysis (ICA) was then applied to the cleaned epochs to identify components that were associated with artifacts such as eye blinks and muscle activity. This was subsequently corroborated by visual inspection. The remaining bad channels were interpolated, and the data was re-referenced to the average of all channels. Participants were excluded from EEG analyses when one of the conditions resulted in fewer than 50 valid epochs (<100 s). Three were removed, leaving 20 participants (≥75 epochs per condition). ICA removed on average 3.45 components in the low-dose condition and 2.80 in the high-dose condition. This resulted in 20 subjects for subsequent analysis, with a mean of 119 ± 22 epochs.

The power spectral density (PSD) was computed for each participant and experimental condition using the MNE implementation, applying the multitaper method. The minimum and maximum frequencies were set to 1 and 40 Hz, respectively. PSDs were averaged across epochs and channels to obtain a single value per 0.5 Hz and then logarithmically scaled. To assess dosing condition differences while correcting for multiple comparisons, we performed a cluster-based permutation test across frequencies using MNE's permutation_cluster_1samp_test (5000 permutations, P < .05 at the cluster level). Additionally, PSD data was aggregated into specific frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–40 Hz) and averaged across epochs. This information was used to plot topomaps for each frequency band in each condition, and topomaps for the statistical test between conditions, identifying channels showing significant differences.

The complexity of broadband signals was assessed using the Lempel-Ziv lossless compression algorithm (Schartner et al. 2017, Timmermann et al. 2019, Cavanna et al. 2022, Pallavicini et al. 2021). This algorithm divides a binary string into non-overlapping, unique substrings. The greater the diversity within the string, the larger the number of substrings. The total count of these substrings is referred to as the Lempel–Ziv complexity (LZc). The initial step involved computing the instantaneous signal envelope for each epoch and channel, using the Hilbert transform. Following this, Z-score normalization and binarization through a median split were performed. The median split ensured an equal proportion of 1 and 0 s across all channels, thereby mitigating biases arising from unbalanced sequences. Finally, LZc was calculated for the binarized signal of each epoch and channel, and the values were averaged across epochs to obtain a single value per channel.

Questionnaire results were compared between the dosing conditions using Student's paired t-test, as implemented in Python's Scipy library (https://scipy.org↗). We reported uncorrected P-values, and highlighted instances where P-values remained significant after applying the Benjamini–Hochberg false discovery rate (FDR) correction at α = 0.05. Corrections were not applied to instances where statistical independence could not be assumed, such as entropy results derived from multiple thresholds. Effect sizes were calculated using Cohen's d for parametric tests and rank-biserial correlation (RBC) for non-parametric tests, with a detailed summary provided in Supplementary Table S1. Estimating the appropriate sample size was challenging due to the limited prior research on eye-tracking metrics in psychedelic studies. We estimated a sample size of 23 participants based on an expected medium to large effect size (d = 0.6), which is commonly observed in studies of psychedelics like psilocybin, with statistical power set to 0.8 and a significance level of 0.05, in a two-tailed, within-subjects design. To complement frequentist analyses, Bayesian statistics were employed to assess the relative evidence for the null and alternative hypotheses. We computed the Bayes factor (BF10) favoring the alternative hypothesis, using Python's pingouin library (https://pingouin-stats.org↗) (Held and Ott 2018), with BF10 < 1/3 interpreted as supporting the null hypothesis. Non-parametric Wilcoxon signed-rank tests were used to compare eye-tracking measures and VAS scores across dosing conditions.

The total number of fixations (N), average distance between fixations (ds), and average time between fixations (dt), were computed for each stimulus and averaged over all stimuli for each subject and dosage condition, as detailed in the Methodology section. To assess differences between dose conditions, we performed non-parametric paired Wilcoxon signed-rank tests (Fig. 2b). The corresponding uncorrected (N = 3) P-values are reported in the figure. The analysis revealed a significant effect of dosing, with the metric ds exhibiting lower values under the high-dose condition. Effect sizes based on RBC are reported in Supplementary Table S2. As an additional control, we also examined the percentage of fixations outside the screen boundaries, which did not differ between high-dose and low-dose conditions. The corresponding plot is shown in Supplementary Fig. S2.

As described in the Methods section, normative saliency maps were generated using DeepGaze II to quantify the visual salience associated with the subjects' fixations (Fig. 3a). The average saliency values across all images were compared between the high-dose (HIGH) and low-dose (LOW) conditions. The high-dose condition showed significantly greater saliency compared to the low-dose condition. To further examine temporal dynamics, the 3-s viewing period was divided into five equal 600 ms intervals. Three of these five intervals showed significantly greater saliency in the high-dose condition compared to the low-dose condition: t2 (P = .026), t3 (P = .035), and t4 (P = .026) (Fig. 3b). Both uncorrected and FDR-corrected P-values are provided in Supplementary Table S3. Effect sizes based on RBC are reported in Supplementary Table S2.

The calculation of Shannon Entropy of fixation distributions across high-saliency visual clusters, as described in the Methods section, is illustrated in Fig. 4a. The number of clusters detected in the binarized saliency maps varied as a function of the threshold, as shown in the left panel of Fig. 4b. To ensure a robust calculation of entropy, we selected a range of thresholds (marked by red lines). Green vertical lines indicate all thresholds within this range where entropy values were significantly different between conditions, with higher entropy consistently observed for the high-dose psilocybin (HIGH) condition compared to the low-dose (LOW) condition. The right panel of Fig. 4b illustrates this result for a specific threshold within the selected range.

We compared the logarithmic PSD between dose conditions, averaged across all channels, using a cluster-based permutation test (Fig. 5a). Two significant clusters were identified: one spanning 3.5–11.5 Hz (P = .017) and another spanning 13.0–20.0 Hz (P = .037), both showing reduced spectral power in the high-dose condition relative to the low-dose condition. These clusters correspond to the upper delta and theta to low alpha range, and the high alpha to beta range, respectively.

We tested whether the overall intensity of subjective effects (total VAS) related to global EEG power across all frequency bands, implementing an Analysis of Variance (ANOVA) model including EEG power in all bands and dose as predictors. This analysis was conducted for total VAS scores obtained 1, 2, 3, and 4 h after psilocybin administration. The models yielded significant results for t = 1 h [F(6,36) = 3.53, P = .01], 3 h [F(6,36) = 3.41, P = .012], and 4 h [F(6,36) = 10.57, P = .000001], with non-significant results for 2 h [F(6,36) = 1.32, P = .29]. Next, we implemented band-specific ANOVA models for each time point, which are summarized in Supplementary Tables S7–S10. While some models yielded significant effects of EEG power on the total VAS scores, the interactions were non-significant. This result indicates lack of evidence of a modulatory effect of psilocybin on the association between EEG power and VAS scores.

Figure 5b displays the topographic distribution of z-scored spectral power for the delta, theta, alpha, and gamma bands, as well as the broadband LZc, for both dose conditions. Additionally, topographic maps of the t-values derived from the statistical tests comparing conditions are also displayed, for the same frequency bands and complexity measure. As shown in the last column of Fig. 5b, increased LZc was found in the high-dose condition across all channels.

To investigate the relationship between neural oscillatory activity and visual exploration behavior, we first implemented an ANOVA model including delta, theta, alpha, beta, and gamma EEG power and dose as predictors, providing a unified test of the main effects and interactions of interest. This omnibus model was significant [F(6,36) = 2.52, P = .039] for the prediction of the average saliency, and non-significant for the prediction of the Shannon entropy of the high saliency regions [F(6,36) = 1.18, P = .371]. Subsequently, we implemented band-specific ANOVA models. We found that only the model for delta was significant [F(3,24) = 4.1, P = .017], with a significant effect of the EEG power [F(1,24) = 6.13, P = .019] and a borderline P-value for the interaction between EEG power and dose [F(1,24) = 4.08, P = .051]. The other models did not yield significant results with threshold P = 0.05 (F and P-values for all models are provided in Supplementary Table S6). Finally, we computed correlations between the power band with a significant effect on saliency (delta) and the average saliency for both doses (Fig. 6). This analysis revealed significant negative correlations only in the high-dose condition between delta power and average saliency across all fixations. Correlation values between EEG spectral power and the saliency metric, including corrected and uncorrected P-values for both conditions, are provided in Supplementary Tables S4 and S5.

We performed correlation analyses to examine whether individual differences in expectancy of absorption were associated with the observed changes in visual salience processing. We calculated Spearman's rank correlations between EXP-Absorption scores and our primary outcome measures (Average Saliency and Cluster Entropy) separately for the high-dose and low-dose conditions. The analysis revealed no significant correlations, indicating that although participants reported higher EXP of absorption prior to the high-dose session, these EXP did not predict the degree of salience processing or gaze entropy between salient regions observed during the task. The results of this analysis are shown in. Supplementary Table S11