Brain-wide dynamics linking sensation to action during decision-making

The study investigates how the brain transforms ambiguous sensory input into action during decision-making, focusing on where and how evidence integration, motor planning and execution are coordinated across brain areas and neural population dimensions. Prior work shows gradual accumulation of sensory evidence into a decision variable in several areas (notably frontal-premotor and posterior parietal cortex and striatum), yet recent findings reveal widespread encoding of sensory, choice and action signals across the brain in trained animals. This raises questions about where sensory inputs are converted into integrated, task-relevant representations that guide action and whether specific regions intrinsically specialize in evidence integration or whether learning sculpts this computation. The authors use a task that temporally separates sensory processing from movement-related activity and brain-wide recordings to delineate how integrated sensory evidence is converted to preparatory activity and ultimately action, and to assess the parallelism or segregation of these transformations across regions and activity subspaces.

Prior studies reported evidence accumulation signals in multiple areas: frontal-premotor cortex, posterior parietal cortex, and downstream striatum. However, large-scale recordings have shown broader, distributed encoding of sensory inputs, choices, and actions across the brain in trained animals, complicating the mapping from sensation to action. Theoretical and empirical work on intrinsic timescales suggests hierarchical differences across cortex that might shape integration durations, yet it is unclear if these intrinsic dynamics determine task-related integration times. Motor control research shows preparatory activity preceding action that resides in a movement-null subspace orthogonal to execution dynamics, mainly studied in motor/premotor cortex and primates. This study integrates these lines of work by examining brain-wide population dynamics linking evidence accumulation to movement preparation and execution, and by testing the roles of learning and intrinsic timescales.

Subjects: Head-fixed, food-restricted mice trained on a noisy visual change detection task. Stimuli: Drifting grating with temporal frequency (TF) fluctuating every 50 ms around 1 Hz (σ = 0.25 octaves). Mice were rewarded for licking upon a sustained increase in TF; rewards were time-limited to promote prompt responses. Baseline monitoring period (3–15.5 s) required immobility; early licks or wheel movement aborted trials, isolating sensory processing from movement. Behavioural monitoring: High-speed videography (face and pupil) and running wheel movement tracking. Neural recording: Dense silicon electrode (Neuropixels) recordings across the brain: 15,406 units total (12,772 curated stable units) from 51 brain regions in 15 trained mice (114 sessions, 167 probe insertions; 50,997 trials). Regions spanned cortex, thalamus, basal ganglia, hippocampus, midbrain, cerebellum and hindbrain. Encoding analysis: Single-cell Poisson GLMs on trial-by-trial spiking with predictors for stimuli (TF fluctuations), task events (change epochs), and behaviour (lick preparation/execution, movements). Cross-validated nested tests held out predictors to assess their unique contributions and identify neurons encoding TF (during no-movement baseline), lick preparation (−1.25 s to lick), and lick execution (videography-based onset). Pulse-triggered responses: Identified fast (>1 s.d.) and slow (<−1 s.d.) TF pulses during baseline without movements. Aligned neural activity to pulses; quantified peak latency and response duration (half-peak width). Validated against GLM-derived TF kernels. Assessed response facilitation to a second fast pulse at varying inter-pulse intervals (0–0.4 s) by comparing the second pulse response to the first. Behavioural modeling: Distinguished integration vs outlier detection strategies. Computed psychophysical kernel (stimulus preceding early licks) and fit an exponential decay to estimate integration timescale (τ). Compared with an outlier-detection agent. Tested super-additivity of two fast pulses on lick probability vs independent effects. Fit a leaky integrator model (τ ≈ 0.25 s) to predict early lick timing and hit reaction times, compared to a non-integrating model. Learning control: Recorded from untrained mice exposed to identical stimuli with random rewards (6,215 units; 45 sessions; 6 mice) to test whether non-visual TF encoding emerges with learning. Intrinsic timescales: Estimated autocorrelation time constants of neural activity during inter-trial, no-stimulus, no-movement periods for each neuron/region; compared to fast-pulse response durations across regions and between trained vs untrained mice. Population dynamics: Decomposed each region’s activity around lick onset into movement and movement-null subspaces. Movement dimensions were defined to maximize similarity with activity in orofacial motor/premotor nuclei controlling licking; movement-null dimensions were orthogonal. Quantified subspace occupancy before and after lick onset. Assessed the disproportionate contribution of TF-responsive neurons to projections in each subspace. Tested alignment between responses to fast TF pulses and preparatory activity, and alignment of pulse responses with the first movement-null vs movement dimensions across regions.

• Behavioural evidence for temporal integration: Early-lick-triggered stimulus average decayed with τ ≈ 0.27 s, longer than an outlier-detection agent. Two fast pulses within ~0.25 s increased lick probability beyond independent effects. A leaky integrator model with τ ≈ 0.25 s better predicted early lick times and single-trial reaction times.
• Widespread encoding of sensory evidence and actions: Lick execution was encoded broadly, with ≥50% of neurons responsive across the brain. A substantial fraction encoded lick preparation. Sparse TF-responsive neurons (≈5–45% depending on region) tracked baseline TF fluctuations without movements, distributed across many regions including frontal cortex (MOs, ACA, mPFC, FRP, ORB, MOP), basal ganglia (CP, GPe, SNr), hippocampus (DG, CA1, CA3, SUB), midbrain (MRN, APN, SCm, NPC), thalamus, and cerebellum (Lob4/5, SIM, CENT3, CRUS1/2, DCN), but absent in orofacial motor/premotor nuclei and medulla.
• Propagation and timescales: Early visual areas (LGd, VISp, SCs) had earliest, brief pulse responses tracking TF. Non-visual regions exhibited delayed, prolonged responses (half-peak widths hundreds of ms), especially frontal-motor cortex, basal ganglia, cerebellum, midbrain, and specific thalamic nuclei, supporting integration across multiple samples.
• Neural integration signature: Facilitation of the second fast-pulse response peaked at 0.2–0.4 s in most frontal, basal ganglia, cerebellar, midbrain regions; minimal facilitation in most early/higher visual areas. Facilitation magnitude correlated with single-pulse response duration across regions.
• Change-period dynamics: Outside early visual system and hippocampus, TF-responsive populations showed ramping activity whose slope scaled with change magnitude; early visual areas (e.g., SCs) showed step-like sustained responses (signalling change without integration). TF-responsive neurons in MOs ramped earlier and stronger than non-responsive neurons.
• Learning dependence: Untrained mice showed TF encoding mainly in visual and limited midbrain areas (SCs, LGd, LP, VISp, APN, SCm), but not in frontal-motor cortex, cerebellum, striatum, or MRN. Thus, non-visual TF encoding and integration emerged with learning.
• Intrinsic timescales dissociate from integration times: Intrinsic autocorrelation timescales (measured during inter-trial rest) did not predict fast-pulse response durations across regions or neurons and were similar in trained vs untrained mice, indicating integration timescales are shaped by task experience rather than intrinsic regional dynamics.
• Link between evidence and preparation: In regions capable of integration (frontal cortex, basal ganglia, cerebellum, midbrain, certain thalamic nuclei), population vectors for TF-pulse responses aligned with preparatory activity before lick; no alignment in non-integrating areas (e.g., SCs). TF-responsive subpopulations were recruited earlier than non-responsive populations before hit licks; earlier recruitment correlated with longer integration timescales.
• Orthogonal population dynamics: Preparatory activity resided in movement-null subspace across regions and transitioned abruptly into movement subspace at lick onset. Movement-null activity peaked then sharply decreased after lick onset, indicating a reset of integration.
• TF-responsive subpopulation dominance: TF-responsive neurons contributed disproportionately to preparatory activity within movement-null subspace in premotor regions (frontal cortex: ACA, MOs, MOP, ORB, mPFC; cerebellum: Lob4/5, SIM, DCN; basal ganglia: CP, SNr/GPi, GPe; midbrain: MRN, NPC, SCm; thalamus: VAL, VB). Their contribution collapsed to chance after lick onset, despite continued stimulus change, consistent with cessation of accumulation after commitment.
• Dimensional alignment: Responses to sensory pulses aligned with the first movement-null dimension, not with the first movement dimension, across regions beyond early visual system, explaining how evidence can be integrated widely without directly driving movement.

The findings show that sensory evidence is encoded sparsely but widely across the brain and that learning transforms transient sensory responses into sustained, integrative representations in premotor circuits. Evidence accumulation directly engages movement-preparatory dynamics in a movement-null subspace, aligning perceptual decision-making computations with motor control frameworks. The preparatory activity emerges earliest in regions with longer evidence-integration timescales (frontal cortex, basal ganglia, cerebellum) and transitions abruptly at action onset into an orthogonal movement subspace across the brain, indicating a coordinated, brain-wide state switch rather than a strictly serial transformation localized to particular areas. The dissociation between intrinsic timescales and integration durations suggests experience-dependent shaping of multi-regional loops to match task demands. Together, these results address the core question of how ambiguous sensory inputs are transformed into choice and action by revealing a distributed, parallel, and learning-dependent sensorimotor transformation implemented across shared population subspaces.

This work establishes a brain-wide framework linking sensory evidence integration to motor preparation and execution. Learning recruits sparse neural subpopulations throughout premotor circuits that integrate evidence over hundreds of milliseconds and drive preparatory activity in a movement-null subspace, which collapses at movement onset as activity transitions to movement-execution subspace. Integration timescales are independent of intrinsic regional dynamics, indicating task experience sculpts the computation. The study unifies decision-making and motor-control concepts by showing that evidence accumulation and movement preparation occupy a shared subspace across many regions, enabling global yet controlled sensorimotor transformations. Future research should identify the circuits generating the internal trigger for movement initiation and test the generality of these principles in tasks with multiple sensorimotor contingencies.

Generalization beyond the studied paradigm remains to be demonstrated; the authors note that it is unclear how broadly the observed principles extend to tasks involving multiple sensorimotor contingencies. The specific brain regions generating the internal trigger that transitions activity from movement-null to movement subspace were not identified and require future causal investigation. Additionally, conclusions are based on correlational recordings in head-fixed mice performing a specific visual change detection task, which may limit direct extrapolation to other behaviors or species.