Population-level temporal decoding and dynamical structure in frontostriatal circuits during decision making

Reviewer #1 (Public review):

Summary:

The authors explore how temporal information and decision-related dynamics are represented across FOF and ADS in rats. The authors used Neuropixels to record neurons simultaneously from FOF and ADS during a free-response auditory change-detection task. They then applied single-trial temporal decoding to estimate both the time elapsed since stimulus onset and the time remaining until movement initiation. When neurons in both FOF and ADS were sorted based on decoder weights, they showed ramping and transient bump-like dynamics aligned to stimulus onset. However, around the decision report, FOF showed a clearer ramping signal and stronger movement-aligned population reorganization than ADS. These results suggest that FOF and ADS share similar temporal dynamics during evidence evaluation, but that FOF undergoes a stronger reorganization near decision commitment.

Strengths:

(1) The authors recorded large-scale neural populations simultaneously from FOF and ADS, allowing direct and fair comparison between them in the same sessions.

(2) The free-response auditory change-detection task, which requires rats to evaluate sensory evidence over time and initiate a decision report, is suited to address the question. The behavioral results support that rats used sensory evidence to guide their choices.

(3) The authors used multiple approaches, including single-trial temporal decoding, decoder-weight PCA, PC loading trajectory, and population-geometry analyses, to explore the FOF and ADS dynamics. These methods provide converging evidence supporting that FOF and ADS share similar temporal dynamics during evidence evaluation but diverge around movement/decision commitment.

(4) The population-geometry analysis is quite strong and interesting because it compares epoch-specific neural subspaces and quantifies dimensionality and subspace alignment, showing stable subspaces during evidence evaluation and stronger subspace reorganization in FOF near movement initiation.

Weaknesses:

(1) The manuscript failed to include histological confirmation of probe placement.

(2) The direct FOF-ADS decoding comparison in fig 3f and 4f includes only 16 of 61 sessions because of imbalanced unit counts. While controlling for unit number is important, excluding most sessions may waste data. Restricting analyses to only 16 sessions questions the generalizability of the result.

(3) Fitted regression curves, and ideally confidence intervals, were missing from Figures 3c and 4c. Also, the confusion matrices in Figures 3a/b and 4a/b show a strong preference for predictions in the first and last time bins. The authors did not explain whether this reflects meaningful event-aligned neural activity or an endpoint artifact from decoding time as bounded discrete classes.

(4) The interpretation of the neuron groups defined by PCA on the decoder-weight matrix was confusing. The authors perform PCA on an N units by T time-bin matrix of LDA decoder weights, then group neurons according to their PC1 and PC2 scores. This is an interesting approach, but the current wording could make readers think that neurons at the extremes of PC1 or PC2 are necessarily the most important neurons for temporal decoding. In fact, these groups appear to represent neurons whose decoder-weight profiles project strongly onto the dominant weight-space patterns. They are not necessarily the neurons that contribute most strongly to decoding accuracy, nor are they necessarily the most common firing-rate dynamics in the raw neural population.

(5) Discussion is missing some needed context. First, given the causal role of ADS in evidence-accumulation-based choices (Yartsev et al., 2018), and its position as a key node that may integrate input from FOF (Brody & Hanks, 2016), the weaker decision-aligned transition in ADS compared with FOF should have been further discussed. If ADS contributes causally to the decision process, why does it show a much weaker population-state transition near decision commitment in the present data? Second, in DePasquale et al. (2024), more extensive choice vacillation was found in ADS, while greater choice certainty was found in FOF. Does this follow the same principle as the current manuscript, where FOF shows stronger reorganization near decision commitment compared to ADS?

(6) Current analyses do not fully exploit the simultaneous nature of the recordings. Apart from the comparison of decoding accuracy, most analyses could have been performed and compared based on the data collected independently from two regions.

(7) Figures 3-10 are hard to read and unpolished. Fonts are too small, and legends/labels are redundant.

Reviewer #2 (Public review):

Summary:

This work investigated differences in the temporal dynamics of neural populations in frontal orienting fields (FOF) and anterior dorsal striatum (ADS) in rodents during an auditory change detection task. The relative roles of these two regions have been studied previously and have been shown to play a role in the accumulation of evidence, with FOF converting this evidence into a categorical decision. By focusing on the temporal dynamics of neurons in these regions, the authors identified a subpopulation of neurons within FOF that displayed an abrupt ramping of activity near the time of decision commitment. Both FOF and ADS contained subpopulations exhibiting ramping activity aligned to stimulus onset. This is an interesting finding, suggesting that FOF contains a subpopulation of neurons that transforms accumulating evidence from other subpopulations in ADS and FOF into an action.

Strengths:

The conclusions of this paper are mostly well supported by data.

Weaknesses:

(1) In the neural analysis, the authors use a technique in which the weights of a linear decoder are used to define a feature vector for each neuron. These weights are used to measure the overall contribution of a neuron in decoding time (from stimulus or decision commitment). Interpreting decoding weights in this way is technically not correct (Kriegeskorte and Douglas, "Interpreting encoding and decoding models"), as a large weight in a decoder is not necessarily indicative of a large effect. Weights in decoding models can become large in order to cancel noise. Alternative analyses, for instance, treating the time series of each neuron as a feature vector, could have supported the conclusions from this technique.

(2) In this same analysis, it appears that the abrupt change in response in FOF at the time of decision commitment is coming from a single subpopulation of about 130 neurons. In the example session (Figure 8J), there is a clear outlier (the neuron in the top right corner). A closer inspection of the single neuron responses in this group would strengthen the results to confirm the abrupt change in mean population response is not coming from a relatively small number of neurons and sessions.

(3) The significance of the dynamical motif corresponding to transient bumps was unclear. For example, when looking at Figure 6K-L, I do not see any neuron groups that exhibit a clear transient bump. I would characterize all groups as ramping, with some groups showing steeper ramps. It would be helpful if the figure displayed the fraction of variance explained by PC2 so that it would be clear how much variance the bump motif is contributing. Given that there was no discussion of the functional relevance of this second motif, interpretation of this result is unclear.

(4) The finding that FOF contains subpopulations which slowly ramp during the trial as well as a subpopulation which acts like a switch that abruptly turns on at the time of decision commitment is interesting and significant and presents several computational questions. For example, is this subpopulation a non-linear readout of the more slowly ramping populations? The approach based on constructing a feature vector for each neuron, projecting these vectors into a low-dimensional subspace, and partitioning into subpopulations is insightful and allowed distinguishing these different computational functions within a single region (FOF). However, I found this particular result to not be clearly stated and obscured by other seemingly less significant results (e.g., existence of the transient bump motif) and other less interpretable analyses (e.g., subspace re-alignment).

Reviewer #3 (Public review):

Summary:

This study investigates how frontostriatal circuits encode elapsed time and exhibit decision-related dynamics during an auditory change-detection task. Using population-level temporal decoding and analyses of low-dimensional neural dynamics, the authors compare activity in the frontal orienting field (FOF) and anterior dorsal striatum (ADS). The manuscript addresses an important question in systems neuroscience: how cortical and striatal circuits represent elapsed time and signal action initiation during decision-making.

The results suggest that FOF and ADS differ in how they represent decision-related information near decision commitment or behavioral report. In particular, FOF shows greater movement-aligned changes in temporal decoding and population geometry than ADS. These findings are potentially important because they may help clarify how cortical and striatal circuits contribute to timing, decision formation, and action initiation.

Strengths:

A major strength of the study is its use of population-level analyses to identify temporal structure and movement-aligned changes in neural dynamics. The analyses provide evidence that neural dynamics and low-dimensional population geometry change around the time of behavioral report, especially in FOF. This provides a useful population-level description of decision-related dynamics beyond what could be inferred from average firing rates alone.

Another strength is that FOF and ADS activity were recorded simultaneously during the same auditory change-detection task. This design strengthens the regional comparison by minimizing confounds related to session-to-session variability, including differences in task engagement, decision accuracy, or other behavioral variables across recordings. The simultaneous recordings therefore provide a strong basis for comparing temporal decoding and population dynamics between cortical and striatal circuits.

Weaknesses:

One limitation is that the physiological interpretation of the population-geometry analyses remains somewhat abstract. Concepts such as low-dimensional subspaces, subspace alignment, and subspace rotation are potentially powerful, but it is not always clear what specific changes in neural activity give rise to these effects. For example, it is difficult to tell whether changes in population geometry primarily reflect recruitment of different neurons, or changes in the dominant temporal profiles of the same neurons. This limits the physiological interpretability of the population-level findings.

A second limitation is that the mechanistic interpretation of the FOF-ADS difference remains underdeveloped. The observed differences could reflect an internally generated transition in frontostriatal dynamics, similar to the dynamical-regime and neural-mode transition described by Luo et al. (2025). Alternatively, they could reflect a circuit-readout process, analogous to the framework proposed by Stine et al. (2023), in which cortical activity drives threshold crossing in a downstream circuit, triggering orienting or motor signals that terminate the decision process. The current manuscript describes the regional differences clearly, but it does not fully discuss these mechanistic interpretations.

Finally, the strength of the evidence would be easier to evaluate if the manuscript more clearly reported the number of animals contributing to each major analysis and the consistency of the main effects across animals. Because many analyses are performed across sessions, the absence of this information makes it difficult to assess whether the key findings are robust across animals or could be influenced by one or a small number of animals.