Reply to: “No evidence of neural feature-specific pre-activation during the prediction of an upcoming stimulus” Authors: Gianpaolo Demarchi, Thomas Hartmann, Anne Hauswald, Juliane Schubert, Lisa Reisinger, Pavlos Topalidis, Kaja Benz, Fabian Schmidt, Quirin Gehmacher, Nathan Weisz Published: 2026 Introduction replying to O. Abdoun et al. Nature Communications 10.1038/s41467-026-73568-1 (2026) Reanalyzing open data from our previous work1. Abdoun et al. argue that there is no evidence for anticipatory preactivation of sound frequency-specific information in predictable tone sequences. Their primary argument is that the manner of sound transition—typically to an adjacent tone frequency—is confounded with a bias in decoding, where confusions between neighboring tone frequencies are more probable. According to their interpretation, this confound could independently contribute to above-chance decoding in the ordered sequence, even without invoking predictive processing. To support their claims, the authors present multiple lines of evidence, most notably a computational (a.k.a. theoretical null) and a simulation-based (a.k.a. empirical null) approach. The latter involves reshuffling epochs from random sequences to mimic ordered ones. This reordering procedure is the principal empirical support for Abdoun et al.’s claim: by systematically rearranging random sequences to resemble increasingly ordered ones and applying a decoding approach similar to that used in Demarchi et al., they obtained parametric changes in pre‑target decoding accuracy that resemble the empirical data. Because these random sequences lack predictive information about the upcoming sound, the authors conclude that our observed pattern cannot reflect anticipatory predictive activity. Although this simulation is thought-provoking and we applaud the authors for the idea, we note that the reshuffled data differ markedly from the real data—particularly after stimulus onset—and that the reordering process can inadvertently reuse training samples during testing, potentially inflating decoding accuracy. Despite these challenges, their critique raises valid considerations that merit discussion within the methodological constraints of the original study. However, the authors further claim that “there is no evidence of a feature-specific prediction-related neural activity in the human auditory system based on the open access dataset of Demarchi et al.” Yet, research also cited by the authors, such as work by ref. 2, introduced an alternative analytical framework using variations of the Demarchi paradigm that directly address the core issue raised by Abdoun et al. (see detailed discussion below). When applying both approaches, we have consistently reached similar conclusions2,3 Adapting the analysis approach of Schubert et al., we will first demonstrate that the original dataset also contains strong evidence of feature-specific neural preactivation. This challenges the blanket assertion made by Abdoun et al. Also, to move the discussion beyond theoretical debates, we present novel data that control for the confounding issue at the group level, providing compelling evidence for feature-specific neural predictive preactivation patterns in the auditory system. We conclude with a non-exhaustive list of unresolved issues, and lessons-learned by engaging in this valuable discussion with Abdoun et al. Results and Discussion Alternative approach to probe sound frequency specific neural preactivation patterns Aware of the issue described by Abdoun et al., Schubert et al.2,3 introduced an analytical approach specifically designed to address whether brain activity patterns indicate predictions of the most probable transition direction, assuming an ordered sequence. This was accomplished by training a classifier on ordered sound presentations using all off-diagonal (i.e., transition) trials and testing it on the on-diagonal (i.e., repetition) trials of both the ordered and random sequences. Because the test set comprises only identical self‑repetition trials, the concern raised by Abdoun et al. that decoding performance and the stimulus sequence of the test set may both be biased, does not apply. Here, we present an adaptation of the elegant Schubert et al. approach and apply it to both the real data and reordered sequences in the manner of Abdoun et al. This approach has recently been further developed and successfully applied4. Here, we outline the key elements of the method to aid clarity, while referring the reader to4 for a full account, including detailed implementation and additional considerations. In the first step, we train an LDA classifier to decode carrier frequency from both the ordered and random sequences, using only off-diagonal trials and excluding self-repetitions. This results in two models, each containing information about the carrier frequency presented at 0 ms training time. However, the model trained on the ordered sounds is expected to also encode information about the upcoming stimulus at 333 ms training time. We interpret this model as reflecting top-down predictive recruitment of carrier frequency-specific patterns. The classifier derived from the random sequence, by contrast, is trained only on sequences that lack predictable transitions and captures bottom‑up, stimulus‑driven preactivation. To align more closely with4, we emphasize that the bottom-up classifier captures anticipatory, stimulus-driven preactivation signals, while the top-down classifier reflects abstract computations of the most probable next tone given the internal model of the ordered sequence. For testing on the self-repetition trials, we evaluate the classifier’s ability to decode the most likely stimulus at 0 ms time-point under the assumption of an ordered sequence. Self‑repetition trials are relabeled as the expected next tone (e.g., A/A is relabeled A/B), ensuring that predictive recruitment can be quantified without transitional bias. We apply the same relabeling to the random condition, deviating from the control analysis proposed by Abdoun et al. (“Decoding the most likely stimulus instead of the presented one”). Consequently, when the trained classifiers are applied to the test set, the top-down model (trained on ordered sequences) draws its informative pattern from negative training times, whereas the bottom‑up model (trained on random sequences) derives its relevant information from positive training times when tested in a time‑generalized manner (see Fig. 1A for a schematic). Dual‑classification approach for probing sound‑frequency predictions. A Diagram of the method. Top‑down and bottom‑up linear classifiers are trained on off‑diagonal trials from ordered and random sequences, respectively. Self‑repetition trials are relabeled as the most probable next tone and used for testing, ensuring that transitional bias is minimized. B Time‑generalized decoding results for the original dataset (left) and for reshuffled data (right). Matrices show t‑values for the bottom‑up (top row) and top‑down (bottom row) classifiers; contours mark significant clusters, and the vertical line denotes stimulus onset. Robust pre‑stimulus and post‑stimulus effects are seen only in the original data, indicating that the new approach overcomes the bias identified by Abdoun et al. Collectively, this dual‑classifier framework makes explicit the complementary contributions of bottom‑up, anticipatory patterns and top‑down, expectation‑driven representations. Time‑generalized decoding matrices obtained from this approach can be compared across entropy levels for the interpretable window of −333 ms to +333 ms, given that identical self‑repetition trials are used for testing. The general rationale of this approach is depicted in Fig. 1A. Using the regression‑based statistical framework introduced by Demarchi et al., we found robust prediction effects for both classifiers (Fig. 1B, left panel). During the pre-stimulus period, bottom-up and top-down neural patterns reveal increased preactivation effects for more ordered sequences. This finding corroborates and extends the main conclusions of Demarchi et al. Furthermore, for top-down neural patterns in particular, a strong positive post-stimulus effect was observed, indicating the activation of predicted carrier-frequency patterns in response to a self-repeating tone. In more ordered sequences, these events are more surprising, and this pronounced post-stimulus effect may reflect a violation of expectations. To ensure that the bias identified by Abdoun et al. does not affect this analysis, we re‑ran the analysis on artificially reordered random sequences using the code provided by Abdoun et al. The effects in the reshuffled data were markedly different (Fig. 1B, right panel) and, in part, opposite to those in the original data, particularly for the top‑down patterns. These findings strengthen the case that our novel approach addresses the concern raised by Abdoun et al. and that the Demarchi dataset contains clear evidence of feature‑specific, prediction‑related neural activity. Novel data demonstrating feature-specific anticipatory auditory predictions To substantially advance the discussion, novel experiments are required that empirically disentangle predictive processes from the bias introduced by transitions between neighboring tones. As a first step along these lines, we conducted a pilot study with twenty‑seven healthy volunteers (18 female, sex/gender info was self-reported; mean ± s.d.: 23.26 ± 2.72 years; subjects recruited within a course and compensated with credits). The task was closely modelled on1 but incorporated key modifications to eliminate the impact of decoding bias. Participants completed four continuous blocks of passive sound listening, each containing 4 000 tones. Two blocks served as baselines for the decoding approach: one comprised a fully randomized sequence of four pure tones logarithmically spaced between 200 Hz and 2000 Hz, and the other comprised eight such tones. Tones were presented every 333 ms and 25% were self‑repetitions, as in ref. 1. The remaining two blocks were “ordered” sequences: within each block, the order of tones was fixed, but the specific order was randomized for each participant. In the four‑tone ordered block each participant received a random permutation of the four frequencies (e.g., 2000 Hz → 431 Hz → 200 Hz → 928 Hz), and similarly for the eight‑tone block. As a result, across participants the ordered frequencies were randomly distributed across positions, preventing any consistent mapping between sequence position and neighboring frequency. MEG recording, preprocessing, and decoding followed the procedures described in ref. 2. When classifiers were trained on the random sequences and tested on the subject‑specific ordered sequences (mirroring the original analysis), we observed robust pre‑stimulus prediction effects for both the four‑ and eight‑tone conditions (Fig. 2A). Thus, despite removing the systematic bias, classification accuracy increased with sequence regularity during the anticipatory period, replicating the pattern seen in our earlier work. Novel evidence for feature‑specific anticipatory auditory predictions. A Pre‑stimulus statistics of the decoding accuracy from classifiers trained on random sequences and tested on ordered sequences, following the Demarchi et al.1 approach. Now, the ordered sequences were subject‑specific and randomly assigned across participants to minimize sequence–frequency bias. Left panel: the classical four‑tone design; right panel: an eight‑tone improved design. B Slopes from regressing log-frequency on serial position for the randomized ordered sequences. Slopes clustered near zero, particularly for the eight-tone condition, indicating that the classical bias was effectively removed. The green band denotes the region of practical equivalence (ROPE, ± 0.1 around zero) and the blue band the 89% Highest Density Interval (HDI). While the four-tone slopes remained dispersed with several values outside the ROPE, the eight-tone slopes were narrowly centered within it, suggesting stronger evidence for practical equivalence to zero. To verify that randomization eliminated the linear relationship between log-frequency and serial position, we regressed the log-frequency of each ordered sequence against its reordered position and computed the slope (Fig. 2B). The classical ordered sequences yielded positive slopes (+0.33 for the four-tone and +0.14 for the eight-tone design, in log-frequency space). In the randomized sequences, slopes clustered around zero, far from the classical values. We computed 95% Highest Density Intervals (HDIs) together with the probability of the parameter lying within a Region of Practical Equivalence (ROPE) around zero5,6. For the four-tone condition, the distribution of slopes was wide and bimodal, with only 19% of the posterior mass falling inside the ROPE (Pr(ROPE) = 0.19), and the 89% HDI extended well beyond it, indicating inconclusive evidence for practical equivalence. In contrast, for the eight-tone condition, slopes were narrowly centered around zero, with 89% of the posterior mass inside the ROPE (Pr(ROPE) = 0.89) and the 89% HDI almost entirely contained within it, providing strong evidence that increasing sequence length effectively controls for decoding bias across participants. Further considerations and recommendations Abdoun et al. raise a valid concern that generally applies to cases where a trained classifier exhibits a bias that is systematically amplified by transitions to the next neighboring stimulus in more ordered sequences. Nevertheless, our alternative analysis demonstrates strong neural preactivation patterns in the Demarchi dataset that cannot be accounted for by the Abdoun et al. critique (Fig. 1). Going beyond this dataset, we provide new data that at the group level decouples the decoding bias from the manner the stimulus transitions are regulated: each participant received a randomized ordered sequence, so the probability of landing on neighboring tones was minimized within subjects and averaged out across the group. The results clearly strengthen the argument that feature-specific predictive preactivations exist in the auditory system (Fig. 2). Looking ahead, engaging with Abdoun et al.’s critique has highlighted several lessons for study design and analysis. To reduce decoding bias, future experiments would benefit from incorporating more carrier frequencies than those used in Demarchi et al. (see, e.g., ref. 7) and from varying the transition matrices across participants or even within participants (as in ref. 8). For within-subject manipulations, faster stimulation rates (> 3 Hz) may be worth exploring. When eight tones are used, transition slopes lie almost entirely within a narrow range near zero; with four tones the average slope is similarly low, but the distribution includes some large values. In such cases, increasing the sample size beyond what is typically used would help, though we note that the issue raised by Abdoun et al. is already mitigated at the group level in our sample of 27. A drawback of randomizing transition matrices is that it produces jumpy stimulus sequences. In natural environments, stimuli typically exhibit temporal autocorrelation, characterized by structured dependencies across time. For sound frequency, this is reflected in smooth pitch transitions or spectral content changes in speech, as well as melodic and harmonic progressions in music. Consequently, many scientific questions regarding neural preactivation patterns may require retaining the confound described by Abdoun et al. In such cases, we recommend employing the analysis approach outlined by Schubert et al.2 or the extension described above. If the original Demarchi et al. analysis is applied—for instance, to compare conditions within participants (e.g., attended versus non-attended ordered sequences) or across groups—it is crucial to also assess the decoding bias of the trained classifier. For within-participant comparisons, using the same classifier across conditions is appropriate, provided the transitional structure of the critical experimental conditions remains consistent. In group experiments, it is essential to examine and compare temporally resolved decoding bias between groups (e.g., see ref. 9 If no significant bias differences are found between groups, then differences identified using the original Demarchi et al. analysis can be attributed to altered neural processing in response to the ordered stimulus sequence. We thank Abdoun et al. for drawing attention to a limitation in our original analysis. Such contributions are invaluable for advancing the field and deepening our collective understanding of neural prediction processes. While some issues remain—e.g., selective interpretation of findings for example, questions about how certain findings are interpreted—we agree that the concern they raised is important for the specific combination of design and analysis used in the original manuscript. Their critique prompted us to refine our analytical approaches and to conduct a new study, both of which provide evidence for frequency‑specific predictive preactivation in the auditory system. More broadly, our work underscores that while theoretical arguments and simulations can highlight potential issues, the broader questions about predictive processing in the auditory system must ultimately be resolved empirically through carefully designed experiments. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Reporting Summary Peer Review File Acknowledgements We want to thank Mr. Manfred Seifter for the support during the data collection for both studies. The authors acknowledge the computational resources provided by Salzburg Collaborative Computing (SCC), funded by the Federal Ministry of Education, Science and Research (BMBWF) and the Land Salzburg. KB was supported by the ÖAW P26-AW2605-P. LR was funded by FWF Grant 10.55776/PAT446622 and by the Land Salzburg Grant 20204-WISS/225/288/4-2021. FS received financial support from the Land Salzburg (20102/F2400537-FPR). P.T., J.S., and Q.G. were supported by the FWF Grant W 1233-B. Q.G. was also supported by the FFG BRIDGE1 project 871232. Author contributions G.D. and N.W. conceived the studies and designed the overall research framework. G.D. performed most of the experimental work, data analysis, and visualization. N.W. contributed to study design and provided critical guidance throughout the project. T.H. and L.R. contributed to specific experimental components and assisted with data acquisition. G.D. and N.W. drafted the manuscript. T.H., A.H., J.S., L.R., P.T., K.B., F.S., Q.G., and N.W. contributed to data interpretation and critically revised the manuscript. N.W. supervised the project and acquired funding. Peer review Peer review information Nature Communications thanks Leonardo Bonetti and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Data availability The down sampled version of the data of the original study is available at Zenodo (10.5281/zenodo.3268713). The original non-resampled data (ca. 300 Gb) is available from the corresponding author. For the pilot study, we provide a minimum working MAT file on the GitHub repository (github.com/gdemarchi/Markov2.0), alongside with all the script for the new analysis of the old data, and the pilot data. Raw data (for the subjects in the pilot dataset) can be provided anytime upon request. The pilot data will be released fully with a separate publication. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Change history 7/27/2026 In the version of this article initially published, a peer review file was not included and is now available in the HTML version of the article. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-73567-2. References Associated Data Supplementary Materials Reporting Summary Peer Review File Data Availability Statement The down sampled version of the data of the original study is available at Zenodo (10.5281/zenodo.3268713). The original non-resampled data (ca. 300 Gb) is available from the corresponding author. For the pilot study, we provide a minimum working MAT file on the GitHub repository (github.com/gdemarchi/Markov2.0), alongside with all the script for the new analysis of the old data, and the pilot data. Raw data (for the subjects in the pilot dataset) can be provided anytime upon request. The pilot data will be released fully with a separate publication.