Abstract
Understanding how the brain transforms peripheral sensory inputs into higher-level representations, and how these contribute to perception and behavioral performance, is a central question in sensory neuroscience. However, in human olfaction, the temporal evolution of neural odor codes and their functional significance remain poorly characterized, especially at early stages. To address which odor features define early neural responses and how these relate to olfactory function, we recorded EEG from male and female participants as they inhaled diverse odors. Participants also completed standardized tests of odor detection, discrimination, and identification, along with questionnaires. Time- and frequency-resolved decoding and representational similarity analysis revealed that early theta activity encodes low-level physicochemical properties of odor molecules, with encoding peaking at 370 ms. Critically, the fidelity of this early theta coding to odor physicochemical properties selectively correlated with participants’ trait-level odor discrimination ability, but not with other olfactory measures. In contrast, delta-band representations of pleasantness emerged later (peaking at 980 ms), linked only to trait-level odor affective reactivity, as measured by questionnaires. These results suggest that earlier theta-band representations reflect a distinct functional role from the later-emerging delta-band activity and are associated with olfactory performance. Extending these findings, separate EEG recordings during a task involving odor discrimination showed that early theta decoding accuracy was significantly higher on correct than incorrect trials, indicating that theta-band coding accounts for trial-by-trial performance fluctuations. Collectively, our study demonstrates that early theta-band representations of low-level odor features—prior to perceptual representations—are already functionally relevant to odor-guided behavior.
Significance Statement
The olfactory system rapidly transforms the physicochemical characteristics of odor molecules into perceptually and behaviorally relevant signals. However, the nature of neural representations at each processing stage—and whether these representations influence olfactory performance—remains poorly understood. Using electroencephalography during odor stimulation, we show that individuals with superior odor discrimination ability more precisely encode physicochemical odor features within the first 300 ms. Neural coding in the same time window also predicts trial-by-trial performance in a task involving odor discrimination. In contrast, later brain activity reflects perceived pleasantness and does not relate to olfactory ability. These findings suggest that early-stage neural representation of low-level odor features plays a unique role in shaping odor-guided behavior.
Introduction
The transformation of the physical properties of external stimuli—such as the wavelength of light or the structure of odor molecules—into perceptual and cognitive representations is a fundamental aspect of sensory processing. In the olfactory system, odor molecules are initially encoded through differential binding to hundreds of olfactory receptors, each tuned to specific physicochemical features of odors (DeMaria and Ngai, 2010). Odor information is then transmitted to the olfactory bulb (OB), onward to primary, secondary, and higher-order olfactory regions, all within the first few hundred milliseconds after inhalation (Stadlbauer et al., 2016). During this rapid cascade, odor representations are thought to evolve from early physicochemical-based encodings into higher-level perceptual representations. However, the temporal dynamics of this transformation in humans remain unclear.
In rodents, studies have shown that neural activity patterns in the OB and downstream piriform cortex (PCx)—the main region of the primary olfactory cortex—reflect the physicochemical properties of odors (Pashkovski et al., 2020). Moreover, representations in the PCx correspond more closely with perceptual properties than those in the OB, substantiating the transformation toward perceptual representation (Pashkovski et al., 2020). Consistent with this, functional magnetic resonance imaging (fMRI) studies in humans showed that the anterior PCx, which receives direct input from the OB, encodes physicochemical features of odors, whereas the downstream posterior PCx represents perceptual features (Gottfried et al., 2006; Fournel et al., 2016). Regarding temporal dynamics, a recent electroencephalography (EEG) study in humans has shown that the correlation of neural activity patterns with perceived pleasantness, a primary dimension of olfactory experience, peaks at ∼600 ms after odor onset (Kato et al., 2022). In contrast, earlier neural activity lacks such correlations, suggesting that initial representations may reflect lower-level features of odors. However, this hypothesis remains largely unexplored.
Another key question concerns the behavioral relevance of these early-stage representations. In rodents, neural activity patterns in response to different odors become increasingly differentiated between 100 and 400 ms after stimulus onset (Rennaker et al., 2007; Miura et al., 2012; Gschwend et al., 2015; Bitzenhofer et al., 2022; Poo et al., 2022). Moreover, greater odor separation in both the OB and PCx is associated with superior behavioral odor discrimination (Miura et al., 2012; Gschwend et al., 2015), suggesting that early neural representations support olfactory performance.
Similarly, in humans, both scalp EEG and intracranial recordings show that neural separation between odors increases over time, peaking between 100 and 500 ms (Jiang et al., 2017; Kato et al., 2022; Dikeçligil et al., 2023a). Intracranial recordings from the PCx have further revealed the separability of odor representations correlates with behavioral odor identification accuracy (Dikeçligil et al., 2023b). Structural features of early olfactory regions—such as OB volume and the microstructure of the olfactory tract projecting to the PCx—have also been linked to individual differences in olfactory performance (Hummel et al., 2015; Echevarria-Cooper et al., 2022), further highlighting the role of early-stage processing in shaping behavior. Nevertheless, few human studies have directly investigated the behavioral relevance of early odor coding. Moreover, in both rodent and human studies, it remains unclear whether and how early-stage neural representations of physicochemical odor features contribute to odor-guided behavior.
We therefore asked two questions in humans: (1) Does early-stage brain activity represent physicochemical odor properties? (2) What is the behavioral relevance of such an early representation? In Experiment 1, we examined neural representations of odors using time-resolved EEG decoding followed by representational similarity analysis (RSA; Fig. 1D, orange arrow). To explore the behavioral relevance of these representations, we assessed whether odor-evoked neural patterns were associated with participants' trait-level olfactory ability and olfaction-related responsiveness, as measured by tests and questionnaires (Fig. 1D, green arrows). Because early representations correlated with odor discrimination ability, Experiment 2 investigated whether trial-by-trial decoding accuracy predicted behavioral performance in a task requiring odor discrimination. Together, these experiments clarify how early-stage neural odor representations are structured and how they relate to behavioral olfactory performance in the human brain.
Materials and Methods
Ethical approval
This study was approved by the Ethics Committee of the University of Tokyo and was performed in accordance with the Declaration of Helsinki. The experiments were conducted after obtaining written informed consent from all participants.
Experiment 1
Participants
This study included 41 participants (age, 18–28 years; 21 women) who received monetary compensation. Among them, eight participants with insufficient valid trials (<20 trials for at least one odor; see below, EEG preprocessing, for details) and one who could not complete the experiment owing to their physical condition were excluded. Consequently, data from 32 participants (age, 19–28 years, 17 women) were included in the final analysis. The inclusion criteria were as follows: right-handed individuals with an Edinburgh Handedness Inventory (Oldfield, 1971) score >60; native Japanese speakers (self-report); absence of olfactory, respiratory, psychiatric, or neurological disorders; no history of traumatic brain injury; no metals in the body; nonsmokers; not pregnant; and not taking any medication (self-report). The participants were instructed not to eat or drink, except water, for 2 h before the experiment.
Odor stimuli and delivery
Nine odors with a variety of pleasantness and perceptual qualities and without apparent trigeminal sensations, as reported in our previous study (Kato et al., 2022), were used: allyl caproate (Ally; 10%; Tokyo Chemical Industry), fructone (Fru; 10%; Tokyo Chemical Industry), citral (Cit; 1%; Tokyo Chemical Industry), linalool (Lin; 1%; Santa Cruz Biotechnology), vanillin (Van; 8%; Tokyo Chemical Industry), acetophenone (Ace; 1%; Tokyo Chemical Industry), alpha-pinene (Pin; 10%; Sigma-Aldrich), cyclodithalfarol (Cyc; 1%; Tokyo Chemical Industry), and 4-pentanoic acid (4Pe; 10%; Tokyo Chemical Industry). The odors were dissolved in propylene glycol (kindly provided by the T. Hasegawa). An odorless condition was included as a control, in which air was presented instead of odorous stimuli. The stimuli were presented using a computer-controlled setup with a customized olfactometer (OL022; Burghart Messtechnik), as described previously (Kato et al., 2022). In brief, odors embedded in a stream of odorless air (flow rate at 7.5 L/min), which were humidified and heated to ∼36°C, were delivered monorhinally to the right nostril while participants were performing velopharyngeal closure breathing. The odor onset time (time = 0) was defined as the time point at which the odor reached the nostril, which was measured using a photoionization detector (mini-PID Model 200 B, Aurora Scientific), as described previously (Kato et al., 2022).
Experimental procedures
The study was performed over 2 d, with EEG recording and odor rating sessions on the first day and olfactory test and questionnaire sessions on the second day (Fig. 1A).
EEG recording session
Before the EEG recording session, the participants underwent training until they could perform velopharyngeal closure breathing, respond to the experimental task within the specified time limit, and maintain fixation without blinking. The training lasted for 5–15 min.
The EEG recording session consisted of 10 subsessions, each with 30 trials: three for each of the nine odors and three for the odorless condition. Therefore, data from 30 trials were acquired across the subsessions for each odor and odorless condition. The order of the trials was pseudorandom. At the beginning of each subsession, the participants inserted a nosepiece into their right nostril to receive the olfactometer airflow. Each trial consisted of three phases: fixation, response, and blinking (Fig. 1B). During the fixation phase, the participants were instructed to focus on a white cross and refrain from blinking. The olfactometer valve switched the airflow from the base air to the stimulus air for 1 s during the fixation phase. The stimulus onset time was jittered to minimize odor-related expectancies. The response phase started 3 s after the stimulus offset, during which the participants rated odor pleasantness on a scale of 1 (unpleasant) to 5 (pleasant). They were instructed to report 0 if they did not sense any odor, and these trials were removed from the analyses. To mitigate the influence of motor-related confounding factors in the decoding analysis, a delayed-response paradigm was used. Ratings were provided by adjusting randomly presented numbers using key presses (Shibata et al., 2016; Kato et al., 2022). The response phase lasted until a response or a maximum of 7 s. Each recording subsession lasted ∼9 min, with a 1–3 min break between them.
Odor rating session
The odor rating session, where odor intensity and pleasantness ratings were obtained, began 15–30 min after the EEG recording session. The odor delivery conditions (e.g., flow rate and interstimulus interval) were the same as those used in the EEG recording session. A total of 30 trials were conducted, with three trials for each of the nine odors and three for the odorless condition, presented in a pseudorandom order. Ratings for both intensity and pleasantness were provided by clicking one of the six buttons presented on a computer display. The buttons were labeled as “odorless,” “weak,” “a little weak,” “neither,” “a little strong,” and “too strong” for intensity rating and “odorless,” “unpleasant,” “a little unpleasant,” “neutral,” “a little pleasant,” and “pleasant” for pleasantness rating. The intensity ratings were converted to a 6-point scale ranging from 1 (odorless) to 6 (too strong). Pleasantness ratings were converted to a 5-point scale ranging from 1 (unpleasant) to 5 (pleasant), with trials rated as “odorless” excluded. The ratings were then averaged across the trials.
Olfactory test session
To assess the participants' olfactory ability, three standardized olfactory tests were conducted in the following order: the Sniffin' Sticks Threshold test with 2-phenylethanol (Burghart Messtechnik; Hummel et al., 1997; Croy et al., 2009), the Sniffin' Sticks Discrimination test (Burghart Messtechnik; Hummel et al., 1997), and the Open Essence test (FUJIFILM Wako Chemicals; Okutani et al., 2013). All tests were conducted according to the manufacturer’s instructions. In brief, threshold was assessed by a 7-reversal initially ascending single staircase procedure using a dilution series of 2-phenylethyl alcohol, beginning at 4% with successive 1:2 dilution ratios. The threshold score was calculated as the average of the last four of the seven staircase reversal points. The possible scores ranged from 1 (least sensitive) to 16 (most sensitive). The discrimination test consisted of triangle tests for 16 pairs of odors, with the discrimination score representing the number of correctly answered trials. The possible scores ranged from 0 to 16. The Open Essence test allows odor identification assessment using odors familiar to Japanese people (Okutani et al., 2013). The test included 12 odors, and for each odor, six choices (i.e., four for odor names, “not sure,” and “odorless”) were provided. The score was the number of correct answers, and the possible range was 0–12 points.
In the original papers validating the tests, the scores for normosmic groups were 11.88 for the threshold test (average across 87 participants, mean age 41.6 years; Croy et al., 2009), 12–13 for the discrimination test (estimated from a graph, average across 24 participants, 18–32 years; Hummel et al., 1997), and 11 for the Open Essence test (median across 95 participants without allergies; Okutani et al., 2013). The olfactory test session took ∼45 min.
Questionnaire session
After the olfactory test session, participants completed the following questionnaires, which assess attitudes toward odors in daily life: the Japanese versions of the Affective Impact of Odor Scale (AIO-J; Wrzesniewski, 1999), Odor Awareness Scale (OAS-J; Smeets et al., 2008), and Chemosensory Pleasure Scale (CPS-J; Zhao et al., 2019), which took ∼15 min in total. Japanese versions were prepared using back-translation. The AIO-J is an 8-item questionnaire designed to assess the effect of odors on food preferences, locations, cosmetics/healthcare products, and people. The score was determined as the mean rating of all items (ranging from 0 to 3), with higher scores indicating a greater impact of odors on the preference for the aforementioned topics. The intersubject mean score reported in the original study of US college students (n = 116) was 1.79 (Wrzesniewski, 1999). The OAS-J is a 30-item questionnaire, with each item scored on a scale of 1 to 5, to measure awareness of odors in the environment. The original OAS comprises 32 questionnaire items, with the score calculated as the sum of all items. However, the OAS-J omitted two items that were not suitable in the Japanese context. Scores were determined as the mean rating across the remaining 30 items (ranging from 1 to 5); higher scores indicated greater odor awareness. The intersubject mean score reported in the original study of Dutch students (n = 525) was 113.65 (Smeets et al., 2008), which corresponds to 3.55 in our scoring method. The CPS-J is a 12-item questionnaire that quantifies the hedonic capacity to enjoy smells and tastes. The original study did not report the CPS scores of their participants (Zhao et al., 2019). The CPS-J score was computed as the mean rating of the items (ranging from 0 to 5); higher scores indicated higher hedonic capacity.
EEG acquisition
EEG signals were acquired at a sampling rate of 2,048 Hz using 64 active Ag-AgCl electrodes (BioSemi ActiveTwo; BioSemi) placed according to the international 10–20 system. In addition, two mastoid and two electrooculogram (EOG) electrodes were used, as previously reported (Kato et al., 2022).
EEG preprocessing
EEG data were analyzed using EEGLAB (version 13.6.5b), an open-source toolbox for EEG data analysis (Delorme and Makeig, 2004)—along with custom-written MATLAB (R2023a) scripts, as previously reported (Kato et al., 2022), unless otherwise stated. Briefly, line noise was removed, and bad electrodes were identified and interpolated using PREP (Bigdely-Shamlo et al., 2015). Independent component analysis (ICA) was performed, and artifactual independent components (ICs) were automatically defined using MARA (Bell and Sejnowski, 1995; Winkler et al., 2011). The rank of the input data for ICA was adjusted for the number of interpolated electrodes (i.e., [64 scalp electrodes + 6 external electrodes − number of interpolated electrodes]). Following ICA, ICs regarded as artifacts were removed, and bad electrodes identified by PREP were interpolated again. The data were then high-pass filtered at 0.2 Hz [passband edge, 0.2 Hz; cutoff frequency (−6 dB), 0.1 Hz; transition band width, 0.2 Hz; filter order, 33,792] and segmented into 2,820 ms epochs, beginning at 760 ms before stimulus onset. Contrary to our previous olfactory event-related potential study (Kato et al., 2022), we did not use a low-pass filter or temporal downsampling in the current study, as we employed time–frequency decomposition afterward. Subsequently, trials with an absolute amplitude exceeding 100 μV in the vertical EOG (top minus bottom EOG) were rejected to eliminate remaining artifacts. Eight participants with insufficient valid trials (<20 trials for at least one odor) were excluded from the analyses. For the remaining participants (n = 32), the preprocessing steps are summarized as follows: The mean and standard deviation (SD) of the removed ICs across participants were 34.7 ± 5.5, and the number of interpolated electrodes was 9.3 ± 5.5. The number of rejected trials due to the absolute amplitude exceeding 100 μV in the vertical EOG was 2.3 ± 4.9 trials. The intersubject mean number of remaining trials per odor was 28.9 ± 1.5 trials.
Time–frequency decomposition
To obtain spectral power and phase angles, the Morlet wavelet transform of single trials was performed using the “newtimef” function of EEGLAB. The analysis was conducted over a frequency range of 1–40 Hz in 30 logarithmically spaced steps at a temporal resolution of 50 Hz. The number of wavelet cycles increased linearly from 1 cycle at 1 Hz to 20 cycles at 40 Hz. After truncating the boundaries, time–frequency decomposed data for −200 to 1,000 ms after odor onset were obtained, resulting in 60 time points for each frequency in each trial. To visualize power modulation, spectral powers were averaged across odors, converted into decibel (dB) changes relative to the mean of the entire time window in the odorless condition, and compared with that of the odorless condition (Fig. 2A–C). To visualize the degree of event-related phase synchronization across trials, intertrial phase coherence, which takes a value between 0 (no synchronization) and 1 (perfect synchronization; Busch et al., 2009), was computed for each odor and odorless condition. After Fisher's z-transformation, the values for the odor conditions were averaged across odors and compared with those for the odorless condition (Fig. 2D–F).
Time- and frequency-resolved decoding analysis
We performed a pairwise decoding analysis for each of the 36 possible odor pairs for each time–frequency point (60 time points × 30 frequency steps). Both the real and imaginary parts of the time–frequency decomposed data from 64 scalp electrodes were used as features, resulting in 128 features (64 for real parts and 64 for imaginary parts) for each time–frequency step. This approach, called complex spectrum decoding, utilizes phase and power information simultaneously and provides high and stable decoding accuracies (Higgins et al., 2022; section 4.1.2; “Information content available to complex spectrum decoding”).
Decoding models were individually constructed for each participant using a nested cross-validation (CV) procedure with an
For each participant and each odor pair, the decoding accuracy was measured as the percentage of correctly classified test data across the outer CVs. The grand average across-odor-pair decoding accuracy shown in Figure 3A was calculated by first obtaining the across-odor-pair-mean accuracy for each participant and then averaging them across participants. When focusing on the delta and theta bands, decoding accuracies across the three frequency steps within each band (i.e., delta, 1.0–1.3 Hz; theta, 3.6–4.6 Hz) were averaged. Hereinafter, the decoding accuracy at theta/delta band refers to this averaged accuracy. The significance of decoding performance was tested against the chance level (50%) using a one-sided, one-sample Student's t test.
Weight map
To visualize the contribution of each electrode to odor decoding in the theta and delta bands, we constructed a weight map for each frequency band at the time point where the decoding accuracy was highest (430 ms for theta and 470 ms for delta). First, we transformed the weight vectors from the classifier according to Haufe et al. (2014) for each frequency step and odor pair. Subsequently, the absolute values of the weights corresponding to the imaginary and real parts of each electrode were averaged. These values were then averaged across three frequency steps within each of the theta and delta bands and across odor pairs. The weight map for each participant was averaged to create an intersubject mean weight map shown in Figure 3C.
To examine whether the topographic patterns of the weight maps differed between frequency bands, topographic analysis of variance (TANOVA) was used (Murray et al., 2008). In this analysis, we scaled the weight map for each frequency band such that the mean and SD of the weight values across the electrodes for each map were 0 and 1, respectively. The difference in the topographic pattern was then quantified by computing a global map dissimilarity (GMD) value, which was the root mean square of the differences in weights for all electrodes. Finally, statistical significance was evaluated using a nonparametric permutation test in which the actual GMD value was compared against the null distribution of GMD values generated by shuffling conditions (i.e., frequency bands) within participants (number of randomization runs, 10,000).
Representational similarity analysis
To assess the temporal dynamics of how the chemical and perceptual characteristics of odors are represented in the brain, we conducted a RSA (Kriegeskorte et al., 2008; Kriegeskorte and Kievit, 2013; Nili et al., 2014) in a time-resolved manner, focusing on the theta and delta bands. In RSA, the representational structure of a given entity (e.g., odor) in a given space (e.g., neural space) is defined as a representational dissimilarity matrix (RDM) consisting of the pairwise distances of all samples in that space. Similarities between different representational structures were then examined as correlations between RDMs. This study compared the representational structures of nine odors in the neural space with those in the physicochemical and perceptual spaces. Correlations between RDMs were computed for each participant, time point, and frequency band using Spearman’s rank correlation. The RDM based on the alternate frequency band was partialed out to assess the neural representations specific to each band (Fig. S4C–F). In addition, pleasantness and physicochemical RDMs were further partialed out from each other: specifically, pleasantness was partialed out to examine representation of physicochemical properties, and Chem_127s (see below, Physicochemical RDM) was partialed out to examine representation of pleasantness (Fig. 4B–D). Following Fisher's z-transformation of the partial correlation coefficients, a group-level inference was made using a one-sided, one-sample Student's t test.
Neural RDM
The RDM for the neural space (neural RDM) was constructed based on the accuracies of pairwise decoding at the theta and delta bands and at each time point, assuming that a higher decoding accuracy suggests a higher dissimilarity of neural response patterns between the odor pair (Grootswagers et al., 2017).
Physicochemical RDM
The chemical properties of odorous molecules can be described using thousands of physicochemical descriptors. However, the specific descriptors that are important for the coding of olfactory neurons are not comprehensively understood. To reduce dimensionality, previous studies have used principal component analysis (PCA) and used prominent PCs to characterize odors (Khan et al., 2007; Mandairon et al., 2009; Snitz et al., 2013; Secundo et al., 2014; Fournel et al., 2016). In the current study, we extracted PCs from a dataset consisting of the scores of 5,217 physicochemical descriptors for a large odor set used in a previous study (Snitz et al., 2013; 1,358 odors including all nine odors used in the current study; Fig. S1B). The scores for the physicochemical descriptors were computed using Dragon (ver. 7.0, Talete; Tetko et al., 2005) based on molecular structures obtained from PubChem (http://pubchem.ncbi.nlm.nih.gov/search/). Because the ranges of scores varied across descriptors, before applying PCA, we normalized each descriptor using the mean and SD. We defined the first 127 components (Chem_127s), which covered 95% of the original variance, as the metric of the odor physicochemical characteristics. Chemical RDM was separately constructed based on the first PC (Fig. 4B, Chem_1) and the rest of the PCs (Fig. 4C, Chem_126s), as the first PC exhibited remarkably higher variance (32.4% of the original variance; Figs. S1B, S3A) than did other PCs, and has been the focus of previous olfactory studies (Khan et al., 2007; Mandairon et al., 2009; Secundo et al., 2014). The distances between the odors were computed using the Euclidean distance metric.
Pleasantness and intensity RDMs
Pairwise odor dissimilarities were quantified as pairwise Euclidean distances of the pleasantness rating obtained during the EEG recording and intensity rating obtained after the EEG recoding.
Correlation analysis of individual differences
To explore the behavioral relevance of early odor coding, we examined intersubject correlations between odor decoding accuracy (i.e., odor separation; Fig. 5A) and individual olfactory characteristics, as measured by standardized olfactory tests (i.e., threshold, discrimination, and identification), and olfaction-related questionnaires (i.e., AIO, OAS, and CPS; Fig. 1Db). Decoding accuracies, averaged across all odor pairs in the theta and delta bands, were used, and Pearson’s correlation coefficients were computed in a time-resolved manner.
Furthermore, for the time windows, frequency bands, and variables showing significant partial correlations in the RSA (Fig. 4B–D), we assessed whether the fidelity of pleasantness or physicochemical coding was associated with individual olfactory characteristics. This was done by calculating Pearson's correlations between z-transformed partial correlation coefficients derived from the RSA and the scores on the three olfactory tests and three questionnaires.
In addition, for Chem_1 in the theta band at 210–400 ms, where both the degree of odor separation and fidelity of physicochemical coding showed a significant correlation with odor discrimination ability (Fig. 5C,D, right), whether separation or fidelity had a greater impact on discrimination ability was analyzed using a partial correlation. In this analysis (Fig. 5G), the correlation between the degree of odor separation and odor discrimination ability was examined while controlling for the fidelity of physicochemical coding and vice versa.
Statistical significance test
All statistical tests were performed using MATLAB functions (Statistics and Machine Learning Toolbox, R2023a), with the number of observations equal to the number of participants, unless otherwise stated.
A cluster-based nonparametric test was used for multiple testing corrections (Maris and Oostenveld, 2007). In particular, for the time–frequency map of decoding accuracies (Fig. 3A, Fig. S2A), adjacent time and frequency points exceeding the threshold (uncorrected p < 0.05, one-sided Student's t test) were clustered. For each cluster, the cluster-level test statistic was defined as the sum of the t values within a cluster. To estimate the null distribution of the test statistics, the decoding accuracies of each participant were randomly multiplied by +1 or −1, followed by a one-sided Student’s t test to define clusters of time–frequency points using the same cluster-forming threshold and cluster-level test statistic described above. This step was repeated 10,000 times; for each repetition, the largest cluster-level statistic was used to create a null distribution. The multiple comparison adjusted p value for each cluster was derived from its ranking in the null distribution. The significance level was set at p < 0.05 (cluster-level, one-sided).
For the time-series data (RSA and intersubject correlation; Figs. 4B–D, 5C–G; Figs. S2D, S4C–F, S5B–E), the same procedure was used, except that the clusters were defined based on adjacent time points. The null distributions of test statistics were constructed by randomly multiplying rho values by +1 or −1 for RSA (Fig. 4B–D; Figs. S2D, S4C–F) and by shuffling the scores of olfactory tests or questionnaires across participants before calculating correlations with decoding accuracies (Fig. 5C,G, light green; Fig. S5B) or rho values (Fig. 5D–F,G, dark green; Fig. S5C–E) for intersubject correlation analysis.
The specific method for statistical testing, multiple testing correction, and the alpha level used for each analysis are explained in the figure legends, where the corresponding results are presented.
Experiment 2
Participants
This study included 50 participants (age, 18–29 years, 26 women), who received monetary compensation. Among them, three participants with insufficient valid trials (<60 trials for at least one odor, see above, EEG preprocessing, for details), one participant who could not complete the experiment owing to their physical condition, and three participants who could not memorize odor labels in the odor learning session were excluded. Consequently, 43 participants (age, 18–29 years, 22 women) were included in the final analysis. The inclusion criteria were the same as those used in Exp. 1. The participants were instructed not to eat or drink, except water, for 2 h before the experiment.
Odor stimuli and delivery
Two odors, Fru (10%) and Lin (10%), were presented (Fig. S1). These odors were selected based on pilot experiments so that discrimination performance in the two-alternative forced-choice (2AFC) task would be in the range of 80–90%. Odor preparation and delivery were conducted in the same manner as in Exp. 1.
Experimental procedures
The experiment consisted of odor memorization, EEG recording, and odor rating sessions conducted in 1 d.
Odor memorization session
Before the odor memorization session, the participants were trained in velopharyngeal closure breathing until the airflow via the nostril could not be detected by a breathing sensor.
The purpose of the odor memorization session was for participants to memorize pairings of alphabetical labels with odors (“odor A” and “odor B,” labels were counterbalanced across participants). This allowed participants to perform a 2AFC task during the EEG session. The session consisted of two parts: learning and testing (Fig. S6A). In the learning part, each odor was alternately presented twice (A→B→A→B), each for 1 s, together with a visual cue indicating the label (i.e., “A” or “B”; Fig. S6A, left). The interval between the odor presentations was 9 s, during which a fixation mark was displayed. After the learning part, the instruction “test session will start now” was displayed for 3 s, and the test part began. In each trial of the test part, an odor was presented for 1 s, and then, two boxes labeled “A” and “B” were displayed, with their presented side being randomized across trials. The participants reported which odor was presented by clicking on one of the two boxes with a mouse. Soon after the response, feedback was displayed, showing the participant’s answer and the correct label (Fig. S6A, right). The next trial began 13 s after the odor offset in the preceding trial. The odors were presented in a pseudorandom order, wherein presenting the same odor more than four times in a row was avoided. The test was terminated when the rates of correctly identifying true positives for both odors exceeded 70% (five out of seven true positive trials for each odor) or 40 trials had been performed without reaching the 70% threshold. In the second group, the entire procedure was repeated after a break of a few minutes. The intersubject mean and SD of the number of trials needed to exceed the threshold was 13.4 ± 6.4. Three participants who did not exceed the threshold after the second test part were excluded from the study.
EEG recording session
As in Exp. 1, participants practiced the experimental task prior to the EEG recording session. The EEG recording session consisted of six subsessions. Each subsession consisted of 32 trials, 16 for each odor, which were presented in a pseudorandom order in which we avoided presenting the same odor more than four times in a row. Therefore, data from 96 trials for each odor were acquired across all subsessions. At the beginning of each subsession, the participants inserted a nosepiece into their right nostril, through which the olfactometer airflow was presented. Subsequently, to remind the participants of the odor labels, each odor was presented once with a visual cue indicating the odor label. After that, the visual message “test session will start now” was displayed for 3 s, and the first trial was started.
Each trial consisted of three phases: fixation, response, and blinking (Fig. 6A). The timing and procedure for the fixation and blinking phases were the same as for Exp. 1. During the response phase, boxes with odor labels were displayed, and the participants reported which odor was presented by clicking on one of the two boxes with a mouse. The presented side of the labels was randomized across trials so that the participants would not prepare for hand movement during the fixation phase. The response phase lasted for a maximum of 7 s or until a response, and if the response phase was <7 s, the blinking phase followed, where a white circle was displayed. The participants were allowed to blink during the response and blinking phases. The total duration of the response and blinking phases was 7 s. Each recording subsession lasted ∼10 min, with a 1–3 min break between them.
Odor rating session
Approximately 15–30 min after the EEG recording session, the participants evaluated the odor intensity and pleasantness. A total of 20 trials (10 for each odor) were presented in pseudorandom order, where presenting the same odor more than four times in a row was avoided. The procedures were the same as those used in the odor rating session in Exp. 1. Data from one participant could not be obtained due to time constraints.
EEG acquisition
Performed in the same manner as in Exp. 1.
EEG preprocessing
This was performed in the same manner as in Exp. 1. Three participants with insufficient valid trials (<60 trials for at least one odor) were excluded from the preprocessing stage. For the remaining participants (n = 43), the intersubject mean and SD were as follows: The number of removed ICs was 31.8 ± 6.9, the number of interpolated electrodes was 6.7 ± 4.4, the number of rejected trials due to the absolute amplitude exceeding 100 μV in the vertical EOG was 4.3 ± 11.3 trials, and the number of remaining trials per odor was 93.4 ± 6.1 trials.
For the 24 participants included in the decoding analysis, the number of removed ICs was 32.3 ± 6.8, the number of interpolated electrodes was 5.9 ± 3.8, the number of rejected trials due to the absolute amplitude exceeding 100 μV in the vertical EOG was 2.2 ± 4.4 trials, and the number of remaining trials per odor was 94.1 ± 3.8 trials.
Time–frequency decomposition
This was performed in the same manner as in Exp. 1.
Decoding analysis
In this analysis, we included only participants who performed the experimental task significantly better than chance (binominal test, p < 0.05) and had a sufficient number of incorrectly (≥10) answered trials for each odor (n = 24; age, 18–29 years, 12 women). Decoding analysis was performed using data from correctly answered trials as training data and both correctly and incorrectly answered trials as test data. For correctly answered trials, leave-one-trial-out was used for the outer loop CVs so that the decoding accuracy could be evaluated for each trial separately. When incorrect trials were used as test data, all correct trials were used as training data. The other procedures were the same as in Exp. 1. We focused on the theta band (3.6–4.6 Hz), where the contribution to odor discrimination was suggested in the Exp. 1 (Fig. 5C,D,G). The decoding accuracy was calculated separately for correctly and incorrectly answered trials. In Figure S6C, the decoding accuracy averaged across all trials is plotted.
To evaluate whether accuracy in the correct trials was higher than that in the incorrect trials, a one-sided paired Student’s t test was used (Fig. 6F). This analysis was performed at time points where significant decoding accuracy was observed in correct trials (170–820 ms; Fig. 6E, green line).
Statistical significance test
Performed in the same manner as in Exp. 1. For the multiple testing correction, the clusters were defined based on adjacent time points, and the null distributions of the test statistics were constructed by randomly multiplying the decoding accuracy (Fig. 6E) or its difference (Fig. 6F) by +1 or −1.
Code accessibility
Code supporting this study is available at a dedicated Github repository: https://github.com/Touhara-lab/JNeuro_Kato2025. The datasets generated and analyzed during the current study have been deposited in Zenodo (DOI: https://doi.org/10.5281/zenodo.17336415).
Results
Odors modulated oscillatory power and phase activities at the delta and theta bands
Data from 32 participants aged 19–28 years (including 17 women) who underwent EEG recording, odor rating, olfactory test, and questionnaire session (Fig. 1A) were analyzed in Experiment 1 (Exp. 1). During the EEG recording session, nine structurally and perceptually diverse monomolecular odors (Fig. S1) or clean air were presented using a computer-controlled olfactometer, and the participants rated the pleasantness of the odors (Fig. 1B). The odor ratings showed that the nine odors had substantially different pleasantness levels, while their perceptual intensities were similar (Fig. 1C). The EEG signals were decomposed into time and frequency components using the Morlet wavelet transform. Consistent with previous findings (Huart et al., 2012, 2013), odor-induced oscillatory activity was observed mainly in the delta and theta bands within 1,000 ms after odor presentation (Fig. 2A–C). Although phase analysis was not reported in previous studies, an increase in intertrial phase coherence was also observed in these frequency and time ranges (Fig. 2D–F).
Experimental design, stimulus characteristics, and analysis flow of Experiment 1 (Exp. 1). A, Experiment schedule. B, Timeline of the experimental task for electroencephalogram (EEG) recording. Odors were presented to the participants’ right nostrils using an olfactometer, following the recommendations for acquiring olfactory event-related potentials. Participants were then asked to rate odor pleasantness. Ratings were provided by pressing keys to adjust the initially presented numbers, which were randomly selected from 0 to 5. C, Rated pleasantness and intensity of each odor averaged across participants (n = 32; mean ± SEM). Pleasantness (1, unpleasant; 5, pleasant) was rated during the EEG recording, whereas intensity (1, odorless; 6, too strong) was rated after the EEG recoding. Colors correspond to odor identity, consistent with the color coding used in Figure 6 and Figures S1 and S6. Abbreviations of odors are shown at the bottom. D, Overall analysis flow: First, time- and frequency-resolved pairwise decoding of odors was conducted to identify time–frequency points where individual odors exhibit distinct EEG patterns. In the top panel, the left image shows an example of time–frequency map of the decoding accuracy for an odor pair, and the right image shows decoding accuracy of all odor pairs for a specific time and frequency point. Next, the nature of information coded in neural activities was assessed through representational similarity analysis (RSA, orange arrow), correlating neural representational dissimilarity matrices (RDMs, top panel) with physicochemical and pleasantness RDMs (left panel). Finally, intersubject correlation (green arrows) was assessed to investigate the association between participants' olfaction-related characteristics (right panel) and neural separation (a), as well as the fidelity of neural coding to odor properties (b).
Odor-induced EEG power and intertrial phase coherence. Spectral power (A–C) and intertrial phase coherence (D–E) of odor-induced EEG recorded in Exp. 1. To visualize the changes induced by odor, the differences between the odor and odorless conditions (odor > odorless) were quantified as t values for both power and phase coherence and plotted according to the color bar. Time–frequency maps at five representative electrodes (A and D) and topographic maps (B, C, E, and F) at two representative time–frequency points (theta, 4.1 Hz; delta, 1.0 Hz; both at 400 ms after odor onset) are shown. The red circles in the insets indicate the locations of the selected electrodes. A, anterior; P, posterior; L, left hemisphere; R, right hemisphere. In the topographic map, all 64 electrodes are marked with black dots.
Delta and theta bands differently contribute to odor separation
To identify the time and frequency points exhibiting distinct activity patterns in response to different odors (i.e., odor separation), time- and frequency-resolved decoding analyses were performed (Figs. 1D, top, 3). Because both power and phase could code sensory information (Schyns et al., 2011; Ng et al., 2013; Wallroth et al., 2018), a complex spectrum decoding method was used, which can simultaneously take into account power and phase by using the real and imaginary parts of the time–frequency decomposed EEG (see Materials and Methods, Time- and frequency-resolved decoding analysis for details; Higgins et al., 2022). For clarity, throughout the Results section, we report the onset, offset, and peak time points of significant effects, with the peak defined as the time of the maximum t value or correlation. However, because time–frequency analysis inherently integrates information across neighboring intervals, the reported onset and offset should be interpreted with caution. Accordingly, our interpretation of temporal aspects is primarily based on the peak time point.
The across-odor-pair-mean accuracies were significantly higher than chance level (50%) within the delta and theta bands (1.00–7.65 Hz, −200–1,000 ms, cluster-size corrected; tmax(31) = 8.56, p = 5.77 × 10−10, d = 1.51, at 390 ms and 3.57 Hz; Fig. 3A). Significant accuracy prior to odor onset in the low-frequency bands is likely attributable to the temporal smoothing inherent in time–frequency analysis (Cohen, 2014). We confirmed that the significant cluster did not extend earlier than −220 ms relative to odor onset (Fig. S2A).
Time- and frequency-resolved pairwise decoding of odors. A, Grand mean decoding accuracies averaged across 36 odor pairs. The black contour line indicates statistical significance (one-sided, one-sample Student's t test, p < 0.05, cluster-based permutation correction). B, Time courses of decoding accuracies for theta (3.6–4.6 Hz; green line) and delta (1.0–1.3 Hz; yellow line) bands, extracted from A. Shaded areas indicate 95% confidence intervals across participants. C, Topographic maps of classifier weights for the delta and theta bands at the latencies where decoding accuracies were at their maximum (430 ms for theta and 470 ms for delta). Absolute values of weights, averaged across all odor pairs, are shown. The location of electrodes is indicated with black dots. Topographic patterns significantly differed between theta and delta bands (p = 1.14 × 10−2, topographic analysis of variance). D, Decoding accuracies of each odor pair at specified time points after odor onset. The top and bottom triangular matrices correspond to the delta and theta bands, respectively. Abbreviations of odors are shown on the left and bottom.
Notably, when examining the accuracy of each odor pair, while most pairs achieved their maximum accuracy in the theta band (>3 Hz), some pairs exhibited their highest accuracy in the delta band (<3 Hz; Fig. S2B,C). This may be because delta and theta bands encode different features of odors, with each band supporting the neural separation of specific odor pairs. Therefore, in subsequent analyses, we focused separately on the theta (3.6–4.6 Hz) band, consisting of the three frequency steps centered around the decoding peak, and the delta band (1.0–1.3 Hz), consisting of the three lowest frequency steps. Although these ranges are narrower than those typically referred to as delta and theta, they were intentionally selected to minimize signal overlap between the bands and to better dissociate their respective neural representations (Cohen, 2014). While our definition of the theta band includes what is conventionally classified as both high delta and low theta, we use the term “theta” throughout this paper for the sake of simplicity.
The decoding accuracies averaged within each band are shown in Figure 3B (across-odor-pair-mean) and Figure 3D (each odor pair). Decoding accuracies differed across odor pairs, and the patterns of accuracy across odor pairs also differed between the theta and delta bands (Fig. 3D). This further indicates that odor separation in the theta and delta bands is based on distinct odor characteristics. To examine whether the signal sources contributing to the decoding differed between these frequency bands, classifier weight maps at time points corresponding to the peak decoding accuracy for each band (theta, 430 ms; delta, 470 ms) were constructed (Fig. 3C). Given that different topographies indicate different underlying signal sources (Vaughan, 1982; Michel et al., 2004), we used a comparison of the weight maps to address this question. TANOVA (Murray et al., 2008) revealed a significant difference in the topographic patterns of the weight maps constructed for the theta and delta bands (p = 1.14 × 10−2; Fig. 3C). Together, these results suggest that the theta and delta bands carry different facets of odor information, each encoded in a distinct configuration of signal sources.
Early theta represents physicochemical properties, and late delta represents both physicochemical properties and pleasantness
To investigate the nature of the information encoded within the theta and delta bands, a time-resolved RSA was performed (Figs. 1D, orange arrow, 4A; Kriegeskorte et al., 2008; Kriegeskorte and Kievit, 2013). For each participant, RDMs were constructed to quantify the neural dissimilarity between odors. These neural RDMs were then compared with the RDMs constructed in the odor space, specifically in the physicochemical and pleasantness spaces (Fig. 4B–D, top panels; individual pleasantness RDMs are shown in Fig. S4A). The pleasantness RDM was derived from pleasantness ratings collected during the EEG recording. The neural RDM was computed based on the accuracy of pairwise odor decoding at each time point and frequency band (Fig. 3D). Note: Figure 3D illustrates the group-level RDM; analyses used individual participant RDMs.
RSA correlating neural and odor spaces. A, Analysis scheme. Neural RDMs were constructed for each time point in both the theta and delta bands. Partial correlations were computed between the neural RDMs and each of the physicochemical or pleasantness RDMs. To mitigate the influence of spectral leakage, the RDM of the alternate frequency band was partialed out. Additionally, to investigate representations specific to physicochemical properties or pleasantness, either pleasantness RDM or physicochemical RDM constructed based on first 127 principal components (Chem_127s) were partialed out. For results without this partialing out, see Figure S4. B–D, Top panels show the RDMs of corresponding odor properties (see Fig. S4A for pleasantness RDMs of individual participants). Bottom panels show time course of z-transformed partial correlation coefficients between the delta (orange lines) or theta (green lines) bands and Chem_1 (B), Chem_126s (C), and pleasantness (D). Bottom horizontal lines indicate statistical significance (one-sided, one-sample Student's t test, p < 0.05, cluster-based permutation correction). Shaded areas indicate 95% confidence intervals across participants. Arrowheads and adjacent numbers indicate the time points of peak t values.
As in a previous olfactory study (Fournel et al., 2016), dimensionality of physicochemical space was reduced using PCA (Fig. S3A). We defined the first 127 principal components (PCs), which correctively accounted for 95% of the variance in the originally 5,127-dimensional physicochemical space, as the metric of whole physicochemical space. Among these PCs, physicochemical RDMs were separately constructed using the first PC (Fig. 4B, Chem_1) and the remaining 126 PCs (Fig. 4C, Chem_126s). This approach was adopted because PC1 explained a substantial proportion of variance in the original physicochemical space (Fig. S3A) and has been a focus in previous studies (Khan et al., 2007; Mandairon et al., 2009; Secundo et al., 2014).
RDMs based on the theta and delta bands were significantly correlated, possibly attributed to spectral leakage during the wavelet transform [Fig. S2D; see Chapter 13 in Cohen (2014)]. Thus, the RDM based on the alternate frequency band was partialed out to assess the neural representations specific to each band. In addition, as pleasantness and physicochemical RDMs exhibited weak yet significant correlations (Fig. S3B), they were further partialed out from each other (Fig. 4A, right) to identify representation specific to each aspect. Results without this partialing out are shown in Figure S4C–E.
We observed distinct associations between the theta (Fig. 4B–D, green lines) and delta (Fig. 4B–D, orange lines) bands with the physicochemical properties and pleasantness of odors. The theta band showed a significant correlation solely with Chem_1 (80–640 ms, tmax(31) = 4.85, p = 1.64 × 10−5, d = 0.86, at 370 ms). In contrast, the delta band exhibited significant correlations with Chem_126s (420–1,000 ms, tmax(31) = 3.75, p = 3.65 × 10−4, d = 0.66, at 790 ms). Additionally, the delta band showed a significant correlation with pleasantness (720–1,000 ms, tmax(31) = 2.77, p = 4.74 × 10−3, d = 0.49, at 980 ms). Note that, as shown in Figure S4E, the emergence of pleasantness occurred earlier when physicochemical properties were not partialed out, replicating our previous study (Kato et al., 2022). As a control, we also performed RSA based on intensity ratings and found that neither the delta nor theta band showed a significant correlation (Fig. S4F).
These findings underscore a clear distinction in the encoding of information between the theta and delta bands and highlight the dissociation between the neural representation of physicochemical properties and pleasantness. Specifically, the theta band predominantly encodes physicochemical properties (Chem_1) that are unrelated to pleasantness during the early time period (80–640 ms). In contrast, the delta band encodes physicochemical properties (Chem_126s) distinct from those represented in the theta band during the 420–1,000 ms period and encodes pleasantness (720–1,000 ms), which is not correlated with physicochemical properties.
Neural odor separation correlates with odor discrimination ability
To investigate the behavioral relevance of neural odor coding, we examined correlations between neural measures, olfactory abilities assessed by a standardized test, and olfaction-related traits assessed by questionnaires. It should be noted that the EEG data in Exp. 1 were recorded during odor pleasantness ratings rather than during olfactory ability assessments; nonetheless, we adopted this approach to allow a single EEG recording session to be used for efficiently exploring associations with a broad range of independently measured olfaction-related traits.
Odor detection thresholds and discrimination ability were assessed using the Sniffin' Sticks test kit (Hummel et al., 1997; Croy et al., 2009) while odor identification was measured with the Open Essence test, which is culturally adapted for Japanese participants (Okutani et al., 2013). In addition to ability measures, we also assessed participants’ self-rated attitudes toward odors in daily life, using Japanese versions of the AIO-J (Wrzesniewski, 1999), OAS-J (Smeets et al., 2008), and CPS-J (Zhao et al., 2019). Score distributions for each test and questionnaire are shown in Figure 5B and Figure S5A, respectively.
Intersubject correlation between neural odor coding and olfactory abilities. A, Schematic illustration of two neural coding measures: odor separation, assessed by decoding accuracy (horizontal axis), and fidelity to odor properties, assessed by RSA (vertical axis). Each dot represents a single odor, and similar colors indicate similar odor properties. Odors are more spatially dispersed in panels with high separation and more aligned with color similarity in panels with high fidelity. B, Histograms showing the distribution of scores on olfactory tests. C, Time courses of Pearson's correlation coefficients between the across-odor-pair-mean decoding accuracy (odor separation) in the delta (yellow lines) and theta (green lines) bands and the score of odor identification (left), threshold (middle), and discrimination (right) tests. The inset shows a scatterplot between the odor discrimination score and decoding accuracy in the theta band at the peak latency of their correlations (410 ms). The least squares regression line is depicted, accompanied by its 95% confidence interval (shaded area). Results for olfaction-related questionnaires (AIO-J, OAS-J, CPS-J) are shown in Figure S5B. D–F, Time courses of Pearson's correlation coefficients between the score of odor identification (left), threshold (middle), and discrimination (right) scores and z-transformed RSA-derived rho values in theta band for Chem_1 (D), in the delta band for Chem_126s (E), and pleasantness (F). Gray shaded areas indicate the time window of interest where significant partial correlations were observed in the RSA (Fig. 4B–D). The inset shows a scatterplot between the odor discrimination score and z-transformed rho value of RSA with Chem_1 in the theta band at peak latency of their correlations (290 ms). The least squares regression line is depicted, accompanied by its 95% confidence interval (shaded area). Results for olfaction-related questionnaires are shown in Figure S5C–E. G, Partial correlation between the odor discrimination score and across-odor-pair-mean decoding accuracies (odor separation, light green) or z-transformed rho value of RSA with Chem_1 (coding fidelity, dark green) in the theta band while partialing out each other. The gray shaded areas indicate a time window of interest, where significant correlations were observed in Figure 5D, right panel. In all the panels, the bottom horizontal lines indicate statistical significance (one-sided, p < 0.05, cluster-based permutation correction). The arrowhead and adjacent number indicate the time point of peak correlation.
Regarding neural measures, we used two complementary indices: mean decoding accuracy across odor pairs (Fig. 1Da) and the RSA-derived correlation coefficients (Fig. 1Db). These measures were chosen because they reflect distinct aspects of neural odor coding. As shown in Figure 5A, higher decoding accuracy indicates that different odors are more distinctly represented in the brain (odor separation). In contrast, higher RSA correlation coefficients suggest that the pattern of decoding accuracy across odor pairs aligns with specific odor properties, such as physicochemical characteristics (fidelity of coding).
First, to assess whether odor separation in the theta and delta bands relates to participants’ olfactory-related traits, intersubject correlations were computed between each olfactory measure and the mean decoding accuracy across odor pairs at each time–frequency point. Odor discrimination ability was significantly and positively correlated with the decoding accuracy in the theta band during 170–740 ms (rmax = 0.44, p = 5.67 × 10−3, at 410 ms; Fig. 5C, right, green line) but not in the delta band (Fig. 5C, right, orange line). In contrast, odor detection threshold, identification ability, and self-rated attitudes toward odors showed no significant correlations with decoding accuracy in either the delta or theta band (Fig. 5C, left and middle; Fig. S5B). These results suggest that the degree of odor separation in the theta band is specifically associated with odor discrimination ability, but not with other olfactory abilities or attitudes related to odors.
Physicochemical coding in theta correlates with odor discrimination ability
To further investigate the behavioral relevance of the specific information encoded in the brain, we next computed intersubject correlations between RSA-derived correlation coefficients (i.e., fidelity of coding; Fig. 5A) and olfactory measures. In this analysis, the partial correlation coefficients derived from RSA (Fig. 4B–D) served as measures of the fidelity of neural coding to the corresponding odor properties. We focused on fidelity to Chem_1 in the theta band (80–640 ms), Chem_126s in the delta band (420–1,000 ms), and pleasantness in the delta band (720–1,000 ms), where significant partial correlations were observed in RSA (Fig. 4B–D).
We found that coding fidelity to Chem_1 in the theta band significantly correlated with odor discrimination score during 210–400 ms after odor onset (rmax = 0.499, p = 1.81 × 10−3, at 290 ms; Fig. 5D, right), but not with identification or threshold scores (Fig. 5D, left and middle). In contrast, fidelity to Chem_126s and to pleasantness in the delta band did not correlate with any of the olfactory ability measures (Fig. 5E,F). Similarly, attitudes toward odors did not correlate with fidelity to physicochemical features (Fig. S5C,D). However, two attitude measures, AIO-J and CPS-J, showed significant correlations with coding fidelity to pleasantness in the delta band at later latencies. Specifically, participants who are more emotionally responsive to odors (AIO-J) or experience greater chemosensory pleasure (CPS-J) in daily life exhibited higher fidelity in pleasantness coding (AIO-J: 860–1,000 ms, rmax = 0.461, p = 3.98 × 10−3, at 980 ms; CPS-J: 820–1,000 ms, rmax = 0.406, p = 4.85 × 10−3, at 980 ms; Fig. S5E, left and right). These results demonstrate a functional dissociation between early theta-band coding of low-level features and later delta-band coding of pleasantness: Chem_1 coding in theta activity is selectively associated with odor discrimination ability, whereas pleasantness coding in delta activity reflects individual differences in affective responsiveness to odors in daily life.
To further clarify the theta-band effects, where both separation and fidelity were associated with discrimination ability, we examined which neural measure explains more. Specifically, we computed intersubject partial correlations between the fidelity to Chem_1, degree of odor separation, and odor discrimination ability, considering their correlation with each other. We only examined the time window where the coding fidelity to physicochemical properties showed a significant correlation with odor discrimination ability (i.e., 210–400 ms; Fig. 5D, right). We found a significant partial correlation between odor discrimination ability and coding fidelity to Chem_1, but not with the degree of odor separation (230–320 ms, rmax = 0.424, p = 8.74 × 10−3, at 290 ms; Fig. 5G). These results suggest that, during the 230–320 ms poststimulus period in the theta band, Chem_1 coding fidelity—not odor separation—was specifically associated with individual differences in odor discrimination ability.
Early theta-band coding is linked to olfactory task performance at the single-trial level
Exp. 1 demonstrated that individuals with greater ability to discriminate odors exhibited greater neural odor separation in the theta band, as well as higher fidelity of neural coding to physicochemical properties during early latencies. These findings suggest that neural odor representations in this time window may be functionally relevant to discrimination ability. Notably, this aligns with rodent studies showing that early neural separation supports odor discrimination performance (Miura et al., 2012; Wang et al., 2019). However, it is important to note that the EEG data in Exp. 1 were acquired during a pleasantness rating task rather than a task explicitly requiring odor discrimination. Since neural coding may vary depending on task demands, it remains to be tested whether neural–behavioral associations are observed under different task conditions. Moreover, because Exp. 1 focused on interindividual differences, it did not address whether trial-by-trial variations in neural coding predict behavioral outcomes.
To address these questions, Experiment 2 (Exp. 2) examined neural odor coding during 2AFC task, which has been used in both animal and human studies to investigate odor discrimination (Uchida and Mainen, 2003; Kepecs et al., 2007; Miura et al., 2012; Chapuis et al., 2013; Cormiea and Fischer, 2023). In this task, participants were presented with one of two odors and asked to determine which of the two it was. Because the task required participants to distinguish between the two odors, we hypothesized that, within the time window and frequency band previously associated with odor discrimination, neural coding would differ between correct and incorrect trials. Specifically, if early theta-band odor coding contributes to discrimination, then decoding accuracy—when the classifier is trained on correctly answered trials—should be higher for correct than for incorrect trials (Fig. 6B).
Trial-by-trial analysis of EEG data during the 2AFC task (Exp. 2). A, Timeline of the experimental task used during EEG recording. Before the EEG recording session, participants memorized the labels (“odor A” or “odor B”) for two odors (Fig. S6A). These odors were presented using the same olfactometer as in Exp. 1. During the EEG recording, participants were prompted to indicate which odor was presented by clicking a box labeled as A or B, with the presented side of the boxes randomized across trials. B, Schematic depiction of the working hypothesis described in the text. Top part: Example sequence of odor presentation and participant’s response. This example shows some correctly (circle) and incorrectly (X mark) answered trials. Bottom part: Hypothesized distribution of neural activity for odors A and B in feature space with decision boundary classifying the odors (diagonal dotted lines). According to the hypothesis, in the time and frequency band where neural coding of odors contributes to odor discrimination, neural activities of odor A and B will be correctly separated by the boundary in correctly answered trials (left), whereas they will not be so in incorrectly answered trials (right). Deep blue, odor A; light blue, odor B; circles, neural coding of odors in correctly answered trials; X marks, neural coding of odors in incorrectly answered trials. C, The percentage of correctly answered trials in a discrimination task during EEG recording for each participant. Red dots indicate participants who were included in the subsequent analysis (n = 24). D, Intensity (left; 1, odorless; 6, too strong) and pleasantness (right; 1, very unpleasant; 5, very pleasant) ratings for each odor (n = 23; mean ± SEM). The data from one participant could not be obtained due to time constraints. Corresponding data for all participants are shown in Figure S6B. E, Time course of decoding accuracies and 95% confidence intervals (shaded areas) for correct (green line) and incorrect (gray line) trials in the theta band. The bottom horizontal line of the corresponding color indicates statistical significance (one-sided, one-sample Student's t test, p < 0.05, cluster-based permutation correction). Arrowheads and adjacent numbers indicate the time points of peak t values. F, Time course of decoding accuracies of correct trials in the theta band after subtracting those of incorrect trials, along with 95% confidence intervals (shaded areas). The bottom horizontal line indicates statistical significance (one-sided, paired Student's t test, p < 0.05, cluster-based permutation correction). The gray shaded areas indicate a time window of interest, where significant decoding accuracy was observed in Figure 6E. The arrowhead and adjacent number indicate the time point of peak t values.
We used two odors, Fru and Lin (Fig. S1), both of which are monomolecular, perceptually unfamiliar, and do not naturally evoke specific objects or their names (e.g., orange). Since testing the hypothesis required a sufficient number of both correct and incorrect trials, we selected an odor pair with a moderate level of discrimination difficulty based on a pilot study. Furthermore, for the main analysis, participants were included only if they performed significantly better than chance in the experimental task (binomial test, p < 0.05) and had an adequate number of incorrectly answered trials (≥10). The mean correct response rate on the experimental task was 87.6 ± 13.7% for all participants (n = 43) and 81.9 ± 11.3% for the subset of participants who met the aforementioned criteria (n = 24; Fig. 6C). Figure 6D shows results of sensory ratings conducted after EEG recordings for participants included in main decoding analysis (for all participants, see Fig. S6B).
The decoding analysis was conducted in the same manner as in Exp. 1, with two key differences: only trials with correct behavioral responses were used to train the decoder, and a leave-one-trial-out cross-validation approach was employed to assess decoding accuracy on a trial-by-trial basis. The time–frequency map of the across-trial mean decoding accuracy (Fig. S6C) was consistent with the results of Exp. 1 (Figs. 3A, S2B, Fru vs Lin). In the theta band, the decoding accuracy significantly exceeded the chance level for both correctly (170–820 ms; tmax(23) = 7.01, p = 1.92 × 10−7, d = 1.43 at 490 ms; Fig. 6E, green line) and incorrectly answered trials (300–700 ms; tmax(23) = 4.97, p = 2.49 × 10−5, d = 1.02 at 530 ms; Fig. 6E, gray line). However, when decoding accuracies were compared, they were significantly lower in incorrectly answered trials than in correctly answered trials at 190–320 ms (tmax(23) = 3.59, p = 7.73 × 10−4, d = 0.73 at 260 ms; Fig. 6F). The low decoding accuracy observed when applying a decoder trained on behaviorally correct trials to behaviorally incorrect trials suggests that the neural coding patterns underlying incorrect responses differed from those associated with correct responses. This time window overlaps with the period during which the fidelity of theta-band coding to physicochemical properties correlated with odor discrimination ability in Exp. 1 (Fig. 5D,G).
Taken together, these results suggest that theta-band odor neural coding between 190 and 320 ms account not only for interindividual differences in odor discrimination ability (Exp. 1) but also explain trial-by-trial variations in olfactory task performance involving odor discrimination (Exp. 2).
Discussion
The brain rapidly transforms sensory inputs into neural representations that support perception and behavior. In olfaction, how this transformation unfolds over time—especially before perceptual representations emerge—remains unclear. In this study, we identified distinct spatiotemporal oscillatory patterns in the delta and theta bands, each encoding different aspects of odor information (Fig. 3). Early theta-band activity primarily represented the physicochemical properties of odors, while later delta-band activity reflected both physicochemical features and perceived pleasantness (Fig. 4). Further, individuals with greater affective responsiveness to odors in daily life showed higher fidelity in pleasantness coding in the delta band (Fig. S5), whereas those with superior odor discrimination ability showed higher fidelity in physicochemical coding in the theta band (Fig. 5). Moreover, in a 2AFC task, correct trials showed higher decoding accuracy in the early theta band than incorrect trials, suggesting that early-stage neural coding contributes to trial-by-trial behavioral performance (Fig. 6).
Early olfactory representations reflect low-level features
Early sensory processing, which primarily handles peripheral input, does not necessarily encode stimuli in ways that directly support perception or recognition. For example, neurons in the early visual cortex are tuned to features such as orientation and spatial frequency, whereas higher-order visual regions represent information more closely aligned with percepts, such as objects and faces (DiCarlo and Cox, 2007; DiCarlo et al., 2012). In olfaction, prior rodent studies have shown that both the OB and PCx encode the physicochemical properties of odors within the first few hundred milliseconds following inhalation, with PCx activity being more closely linked to perception (Pashkovski et al., 2020). Similarly, human fMRI studies have suggested that the anterior PCx primarily encodes chemical structure, while the posterior PCx reflects perceptual dimensions (Gottfried et al., 2006; Fournel et al., 2016). Despite these insights, the temporal dynamics of this transformation remain understudied. Our findings suggest that physicochemical features of odors are represented during the early stages of processing—prior to the emergence of odor pleasantness representations—as confirmed by analyses that statistically controlled for the shared variance between these features (Fig. 4B). Together with prior work, these results provide a clearer view of the temporal evolution of human olfactory processing.
Neural coding of physicochemical properties explains odor discrimination ability
To explore the behavioral relevance of early neural coding, we examined correlations between neural measures and participants' olfactory traits. Among the three olfactory abilities assessed by standardized tests and three psychological scales on olfaction-related attitudes, only odor discrimination ability was significantly correlated with theta-band neural odor separation, peaking at 410 ms (Fig. 5C), as well as with the fidelity of neural coding to physicochemical properties in the theta band, peaking at 290 ms (Fig. 5D). Partial correlation analyses further revealed that coding fidelity, rather than the degree of odor separation, was the stronger predictor of discrimination performance, with the effect peaking at 290 ms (Fig. 5G).
While cortical representations of chemical features have been previously demonstrated in both humans and rodents (Gottfried et al., 2006; Fournel et al., 2016; Pashkovski et al., 2020), it has remained unclear whether these low-level representations merely serve as precursors to perceptual representations or have a direct impact on odor-guided behavior. Our results address this gap by suggesting that early-stage neural representations of physicochemical odor properties are directly related to behavioral performance.
Of note, this correlation emerged even though the EEG task was not an odor discrimination task, suggesting that these neural representations may reflect trait-like processing capacities that extend beyond the specific demands of the task. One possible explanation is that early neural representations reflect stable individual characteristics arising from structural features of the olfactory system. Prior studies have reported associations between olfactory abilities and morphological measures such as volume of OB (Hummel et al., 2015) and the microstructure of the olfactory tract (Echevarria-Cooper et al., 2022). Alternatively, the odor pleasantness rating task employed in Exp. 1 may involve cognitive processes that overlap with those required for odor discrimination, thereby indirectly recruiting similar neural mechanisms.
Early neural coding supports trial-by-trial task performance
Exp. 2 extended our findings to the within-participant level using a 2AFC task, similar to those used in rodent studies on neural correlates of odor discrimination. Our results showed that theta-band activity patterns differed between correct and incorrect trials, peaking at 260 ms (Fig. 6E,F). This parallels rodent data showing that early OB and PCx activity predicts trial-by-trial behavioral performance on odor discrimination (Miura et al., 2012; Gschwend et al., 2015). Notably, this peak latency overlaps with the period when theta-band Chem_1 coding fidelity was associated with odor discrimination ability in Exp. 1 (Fig. 5D,G), suggesting a shared neural mechanism underlying both inter- and intraindividual variability in odor discrimination.
Interestingly, decoding accuracy in later time window (>320 ms) did not differ between correct and incorrect trials (Fig. 6E,F), implying that odor information persisted but became less accessible for guiding behavior—perhaps due to a limited temporal window for readout. Similar dissociations between neural representation and behavioral performance have been reported in rodents (Miura et al., 2012), underscoring the importance of early-stage coding in odor discrimination.
While these results highlight the behavioral relevance of early neural coding, several caveats should be noted. First, although 2AFC tasks are commonly used to assess odor discrimination in animal and human studies (Uchida and Mainen, 2003; Kepecs et al., 2007; Miura et al., 2012; Chapuis et al., 2013; Cormiea and Fischer, 2023), similar paradigms have been labeled as identification tasks in some human research (Iravani et al., 2021; Dikeçligil et al., 2023a). Typical identification tasks involve choosing a verbal label (e.g., “chocolate” or “orange”) from a fixed list, engaging more semantic processes than the task in Exp. 2. Still, our task likely involved cognitive demands beyond pure perceptual discrimination. Similarly, the discrimination test used in Exp. 1, though standardized to assess odor discrimination ability (Hummel et al., 1997), involved comparing three sequential odors, demanding working memory. Thus, although both experiments implicate early theta-band activity as a neural correlate of odor discrimination, potential confounds from other cognitive processes call for cautious interpretation.
Dual oscillatory channels code low-level odor features
We found that distinct aspects of odor physicochemical properties were encoded in different frequency bands. The first principal component of chemical space (Chem_1) was represented in the theta band (peak at 370 ms; Fig. 4B), whereas a composite of the remaining components (Chem_126s) was represented in the delta band (peak at 790 ms; Fig. 4C). These findings suggest distinct oscillatory channels for encoding low-level odor features.
We propose two hypotheses to explain this observation. The first is a sequential model, in which Chem_1 is represented early and leads to Chem_126s and perceptual representations—a process mirroring hierarchical progression. This aligns with studies showing OB and PCx activity gradually transforms into perceptual codes (Pashkovski et al., 2020). The second is a parallel model, in which Chem_1 and Chem_126s are independently processed in distinct circuits—one supporting odor discrimination (theta) and the other supporting hedonic evaluation (delta). Such segregation is well known in vision (e.g., object identity vs spatial location; Mishkin et al., 1983), and recent rodent neuroanatomical studies suggest a similar organization in olfaction (Chen et al., 2022). Whether such parallel streams exist in the human olfactory system remains an open question.
Limitations
First, because odor physicochemical properties are highly multidimensional and intercorrelated, we could not determine the specific features encoded by theta and delta activity. Ideally, peripheral input to the central olfactory system should be defined based on the physicochemical properties recognized by olfactory receptors—an area that remains poorly understood and requires further investigation. Second, our analysis was limited to the first 1,000 ms after odor onset. However, prior human studies indicate that odor discrimination can involve multiple inhalations over several seconds (Bowman et al., 2012). Thus, discrimination in humans likely depends on neural processing across multiple stages and extended timescales, with early neural coding representing only part of the process. Third, odors were presented to the right nostril while participants used a mouth-breathing technique. Although widely used in olfactory EEG studies for its temporal precision (Kobal and Hummel, 1988; Hummel and Kobal, 1992; Becker et al., 1993; Tateyama et al., 1998; Masago et al., 2001; Iannilli et al., 2015; Kato et al., 2025), this approach may elicit neural activity that differs from natural birhinal sniffing, as suggested by previous studies (Lascano et al., 2010; Zelano et al., 2016; Arabkheradmand et al., 2020; Dikeçligil et al., 2023b). Fourth, time–frequency decomposition inherently introduces temporal smoothing, which can blur the precise timing of statistically significant effects. Consequently, precise onset and offset times should be interpreted with caution.
Conclusions
In summary, our study reveals distinct spectrotemporal dynamics in the neural representations of odor physicochemical properties and pleasantness, as well as their behavioral relevance. Coding of physicochemical features emerged first in the theta band (peak at 370 ms), followed by additional chemical dimensions in the delta band (peak at 790 ms), and pleasantness representations in the delta band (peak at 980 ms). Critically, early theta-band activity predicted individual differences in odor discrimination ability via its fidelity to physicochemical feature coding (peak at 290 ms) and also tracked trial-by-trial performance fluctuations during a task involving odor discrimination (peak at 260 ms). These findings demonstrate that neural representations preceding perceptual ones are functionally linked to behavioral olfactory performance.
Footnotes
We thank Rumi Iwasaki and Shinobu Kaneko for their help in recruiting participants and conducting experiments, Prof. Masashi Sugiyama for his advice on machine learning techniques, and all the members of Touhara Laboratory for their discussion. This work was supported by the Grant-in-Aid for JSPS Fellows to M.K. (23KJ0377), Grant-in-Aid for Scientific Research on Innovative Areas from Japan Society for the Promotion of Science to M.O. (21H05808, 23H04335, and 25H00998), and JST-Mirai program to K.T. (JPMJMI17DC and JPMJMI19D1).
The authors declare no competing financial interests.
This paper contains supplemental material available at: https://doi.org/10.1523/JNEUROSCI.0203-25.2025
- Correspondence should be addressed to Masako Okamoto at masakookamoto3{at}gmail.com.
