henhaobio
Sign in

NatureNeurosciencePeer-reviewed journal article

Brain-wide representations of prior information in mouse decision-making

Charles FindlingiD, Félix HubertiD, Luigi AcerbiiD, Brandon BensoniD, Julius BensoniD, Daniel BirmaniD, Niccolò BonacchiiD, E. Kelly BuchananiD, Sebastian A. BruijnsiD, Matteo CarandiniiD, Joana A. Catarino, Gaelle A. ChapuisiD, Anne K. ChurchlandiD, Yang DaniD, M. Felicia DavatolhaghiD, Eric E. J. DeWittiD, Tatiana A. EngeliD, Michele Fabbri, Mayo FaulkneriD, Ila FieteiD, Laura Freitas-SilvaiD, Berk GerçekiD, Kenneth D. HarrisiD, Michael HäusseriD, Sonja B. HoferiD, , Alexandre PougetiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

The neural representations of prior information about the state of the world are poorly understood. To investigate them, we examined brain-wide Neuropixels recordings and widefield calcium imaging collected by the International Brain Laboratory. Mice were trained to indicate the location of a visual grating stimulus, which appeared on the left or right with prior probability alternating between 0.2 and 0.8 in blocks of variable length. We found that mice estimate this prior probability and thereby improve their decision accuracy. Furthermore, we report that this subjective prior is encoded in at least 20% to 30% of brain regions which, remarkably, span all levels of processing, from early sensory areas (LGd, VISp) to motor regions (MOs, MOp, GRN) and high level cortical regions (ACAd, ORBvl). This widespread representation of the prior is consistent with a neural model of Bayesian inference involving loops between areas, as opposed to a model in which the prior is incorporated only in decision-making areas. This study offers the first brain-wide perspective on prior encoding at cellular resolution, underscoring the importance of using large scale recordings on a single standardized task.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 5 (18 Nov 2024) · CC BY 4.0

Mice infer stimulus-side priors from recent choices; prior signals are decodable pre-stimulus in 24.0% to 30.2% of 242 brain regions, including early visual and motor areas.

It is unresolved whether priors are merged with sensory evidence only in late decision areas or represented throughout the brain, as a distributed Bayesian network would imply, because earlier studies sampled few regions using different tasks. This work decodes a Bayes-optimal prior from International Brain Laboratory (IBL) Neuropixels recordings in 242 Allen atlas regions plus dorsal-cortex widefield imaging on one standardized task, mapping prior encoding brain-wide.

Key results

  1. ShownA Bayes-optimal prior is decodable from pre-stimulus activity at every processing level, from lateral geniculate nucleus (LGd) and primary visual cortex (VISp) through motor cortex and gigantocellular reticular nucleus to orbitofrontal and anterior cingulate cortex.

    Neuropixels: 30.2% of regions (73/242) significant at uncorrected p<0.05 (pseudosession test), e.g. one ventrolateral orbitofrontal session, 160 neurons, R^2 = 0.28; widefield: all dorsal cortical regions; merged maps, Benjamini-Hochberg false discovery rate 1%: 24.0% (58/242); Neuropixels vs widefield corrected R^2 Spearman R=0.52, p=0.0024, N=32 regions. · Decoding the prior during the inter-trial…; Assessing statistical significance; Figure legends

  2. ShownMice exploit the uncued block structure, beating chance but trailing a Bayes-optimal sampling observer on zero-contrast trials; choices are best explained by an action-kernel prior that exponentially weights recent actions.

    Psychometric shift: Wilcoxon t=15, p=2.0E-24, N=139 mice; zero-contrast accuracy 58.7% ± 0.4% (SEM) vs 61.1% ± 1.8% (SD) for a Bayes-optimal sampler (p=1.5e-8); recovery decay 4.97 vs 2.43 trials (p=0.001); action kernel best of 11 models (exceedance > 0.999), median decay 5.45 trials, near the action-kernel optimum (which loses 1.9% vs Bayes-optimal). · Mice use the prior to optimize their performance; Decoding the action kernel prior; Figure legends

  3. ShownIn Neuropixels sessions with usable video, prior signals survive removal of posture, movement and eye-position components, and decoded priors in both modalities track the animal's zero-contrast choices.

    DeepLabCut features decoded the prior in 38.0% (65/171) of sessions, yet neural R^2 for the video-residual prior tracked full-prior R^2 (Pearson R=0.89, p=8E-279, N=806 region-sessions); VISp/LGd eye-position residual R=0.8, p=7E-11, N=44; choice-vs-decoded-prior slopes significant in 17.8% of Neuropixels and 90.1% of widefield regions. · Decoding the prior during the inter-trial…; Proportion of right choices as a function of the…; Embodiment; Figure legends

  4. ShownNeural prior signals carry the temporal signature of the action-kernel strategy: action-kernel priors are decoded better than Bayes-optimal ones, decoded priors depend on at least 5 past actions, and neural timescales track behavioral action-kernel (not stimulus-kernel) timescales.

    Uncorrected R^2 higher for action-kernel than Bayes-optimal prior (Neuropixels t=2230, p=2.6E-30, N=242 regions; widefield t=13, p=4.1E-8, N=32); neural vs behavioral inverse decay constants: Neuropixels R=0.23, p=0.0037, N=164 sessions; widefield R=0.49, p=0.0006, N=46; stimulus-kernel versions n.s. (Neuropixels R=0.03; widefield R=-0.04). · Decoding the action kernel prior; Neural Decay Rate; Figure legends

  5. SuggestedGranger causality on region-wise decoded priors indicates bidirectional cortical-subcortical exchange of prior information, feedback to VISp from secondary motor, retrosplenial, prelimbic and anterior cingulate cortex, and more reciprocal loops than chance.

    Significant directed pairs among prior-encoding regions averaged 71.6% (widefield) and 35.9% (Neuropixels) across sessions, exceeding 1000-pseudosession nulls; reciprocated pairs (A->B->A) 37.7% (widefield) and 10.8% (Neuropixels), above chance; coarse Cosmos-level Neuropixels edges at Bonferroni-corrected p < 0.05. · Decoding the prior during the inter-trial…; Granger causality; Figure legends

Methods, data and reproducibility

Cross-validated Lasso (Neuropixels) or ridge (widefield) decoders predicted a Bayes-optimal stimulus-side prior from pre-stimulus activity, tested against pseudosession nulls from resampled stimulus sequences and pooled per region by Fisher's method; behavioral models were compared by Bayesian cross-validation, and Granger causality tracked prior flow.

  • IBL 2024 public Brain Wide Map release: Neuropixels 1.0 recordings during the visual decision task with 0.2/0.8 block priors · mouse · n = 699 recordings; 459 sessions; 139 mice
  • Region-level Neuropixels decoding set after trial, session and unit quality control (at least 5 units per region-session) · mouse (regions defined by Allen Common Coordinate Framework) · n = 242 regions; 2289 region-sessions; 414 sessions
  • Widefield calcium imaging of dorsal cortex (layers 2/3) during the same task; 51 sessions passed the trial-count criterion · mouse · n = 52 sessions, 6 mice; 32 regions; 1539 region-sessions
  • DeepLabCut video features (licking, whisking, wheel, nose and paw position) for Neuropixels embodiment controls · mouse · n = 171 sessions; 806 region-sessions
  • Pupil-position tracking in sessions recording VISp or LGd, used for eye-position controls · mouse · n = 44 of 53 sessions
  • Joint DeepLabCut plus eye-position feature set for combined embodiment decoding · mouse · n = 124 sessions; 660 region-sessions
  • Stimulus-side decoding results reused from the companion Brain Wide Map analysis (regions with at least two sessions) · mouse · n = 201 regions
Methods & tools
Neuropixels 1.0 probes, Widefield calcium imaging of dorsal cortex (15 Hz), scikit-learn Lasso, Ridge, LassoCV, RidgeCV and LogisticRegression, DeepLabCut pose estimation (plus a separate pupil tracker), spectral_connectivity (Eden-Kramer lab) for Granger causality, psychofit toolbox (cortex-lab) for neurometric/psychometric fits, Pseudosession null-distribution testing, Fisher's method and Benjamini-Hochberg false discovery rate correction, Adaptive Metropolis-Hastings with Gelman-Rubin diagnostic; Bayesian model selection, BEST Bayesian test with a hierarchical mouse/session model, Allen Common Coordinate Framework, Cosmos atlas and Swanson maps, Bayes-optimal, action-kernel and stimulus-kernel behavioral models
Software & environment
Python decoding with scikit-learn (Lasso, Ridge, LassoCV, RidgeCV, LogisticRegression), DeepLabCut plus a separate pupil pose estimator, spectral_connectivity for Granger analysis and psychofit; no package versions or containers stated.
Experimental conditions
Neuropixels 1.0 probes (one or two insertions per session), widefield calcium imaging of dorsal cortex at 15 Hz, and two video cameras for pose and eye tracking; no computing hardware or runtimes reported.
Reproducibility gaps
Surgical, training, spike-sorting and widefield acquisition protocols are deferred to earlier IBL papers and a PhD thesis; model derivations are in Supplementary Information; only 171 of 459 Neuropixels sessions qualified for video controls, widefield none.
Limitations (6)
  • Significance of action-kernel prior decoding cannot be tested region by region, because a valid null would need an exact generative model of the animal's choices; region maps therefore use the Bayes-optimal prior, and choice-linked tests rely on synthetic action-kernel sessions. · stated by the authors
  • Neuropixels sampling is sparse: about 10 recordings per region are needed to recover the reported significance, yet the median is 6, so encoding fractions are likely underestimated and few significant sessions per region preclude robust regional claims. · stated by the authors
  • Widefield's finding that every dorsal cortical region encodes the prior may reflect higher signal-to-noise or calcium signal from axons originating outside each area; decoding scores also scale with widefield pixel count (R=0.82) and are confounded by Neuropixels unit and session counts. · stated by the authors
  • In Neuropixels, region-level Granger links do not survive multiple-comparison correction, likely because only about 10% of directed region pairs were recorded simultaneously, mostly in two or fewer sessions; significance appears only after coarse Cosmos-atlas pooling. · stated by the authors
  • Because the prior is a simple Bernoulli variable, discriminating linear probabilistic population codes from sampling-based schemes is difficult; results are shown only to be compatible with two linear-code requirements. · stated by the authors
  • Widefield data come from only 6 mice and have no video, so the all-region widefield encoding and the strong widefield choice links are not controlled for posture or gaze, unlike the Neuropixels embodiment analyses. · noted in this summary

Read this if you study how expectations shape perception and choice, or need a drift-robust pseudosession framework for decoding slow latent variables from brain-wide Neuropixels and widefield data.

Citations
120
FWCI
53.23
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISIshizu '24Liu '25Gastrell '26Quax '21Langlois '25Zhang '26
78% semantic matchNeuroscience
Kotaro Ishizu, S. Nishimoto, Yutaro Ueoka+1 more·Nature Communications·May 22, 2024· 17 citations