Dynamic belief updating in mice, rats, humans and models

Projektinformationen

DYNABUP

ID Finanzhilfevereinbarung: 101107298

DOI

10.3030/101107298

EK-Unterschriftsdatum 28 April 2023

Startdatum 1 Mai 2024

Enddatum 30 April 2027

Finanziert unter

• Marie Skłodowska-Curie Actions (MSCA)

Gesamtkosten

Keine Daten

EU-Beitrag

€ 276 681,60

Koordiniert durch

ECOLE NORMALE SUPERIEURE
France

Periodic Reporting for period 1 - DYNABUP (Dynamic belief updating in mice, rats, humans and models)

Berichtszeitraum: 2024-05-01 bis 2026-04-30

Zusammenfassung vom Kontext und den Gesamtzielen des Projekts

Every day, humans and other animals must decide on the basis of information that is incomplete, noisy and changing. To do so, the brain accumulates successive pieces of evidence over time and revises its beliefs as new observations arrive. Two properties of the world make this difficult: the environment is stochastic, so a surprising observation may simply reflect noise, and it is volatile, so the same observation may instead signal that the situation itself has changed. Resolving this tension requires the brain to weight each incoming piece of evidence according to how reliable and how relevant it is. How mammalian brains achieve this — and whether they achieve it in the same way across species — is the question at the heart of DYNABUP.
The project addresses this question with a deliberately interdisciplinary strategy that combines experimental psychology and computational modelling with rodent neurobiology. Behavioural tasks are inherited from cognitive psychology, decisions are formalised with Bayesian models drawn from probability theory, and molecular tools are used to interrogate the underlying circuits. This integration of behavioural-science methods with cellular neurobiology is not incidental: it is what makes the human–animal comparison meaningful, because it allows behaviour in the two species to be described in the same mathematical terms rather than by analogy.
The action pursued three objectives: (1) to identify cross-species similarities and differences in probabilistic inference and their underlying computations; (2) to determine how cortical responses are gain-modulated by different forms of uncertainty; and (3) to test whether the locus coeruleus noradrenergic (LC-NA) brainstem system causally regulates this process. The outgoing phase at Boston University provided access to high-throughput rodent behaviour, cortical recording techniques and molecular tools; the returning phase at ENS-PSL provides the human cognitive neuroscience and computational modelling environment in which the cross-species comparison is completed.
The strategic context is twofold. First, rodents represent the large majority of animals used in European research, and translational programmes in neuropsychiatry routinely assume that rodent cognition generalises to humans — an assumption that is rarely tested directly. Establishing which cognitive dimensions translate, and whether residual differences are superficial (parameters) or profound (algorithms), is therefore a precondition for both better science and better-optimised animal use, in line with Directive 2010/63/EU and the 3Rs. Second, several neuropsychiatric conditions have been reframed as disorders of inference, and many involve neuromodulatory dysfunction; characterising these mechanisms with cellular precision has direct relevance for future diagnostic and therapeutic work.

Arbeit, die ab Beginn des Projekts bis zum Ende des durch den Bericht erfassten Berichtszeitraums geleistet wurde, und die wichtigsten bis dahin erzielten Ergebnisse

WP1 — Cross-species comparison of probabilistic inference. The originally proposed change-detection variant of the flash task was abandoned early in the action for two reasons: a study with a closely related design was published between award and start, and the task implementation available in the host laboratory lacked nose fixation during stimulus presentation. Detailed analysis showed that, without fixation, rats adopted "degenerate" strategies — committing after one or two flashes, orienting towards the chosen port and ignoring later evidence — while still reaching reward rates above 75%.
I therefore re-designed the nose ports, re-built the shaping procedure around enforced nose fixation, and moved from a 24/7 live-in facility to experimenter-initiated daily training. Psychophysical integration kernels confirmed that rats then integrated evidence over a protracted temporal horizon. In parallel, I introduced a manipulation of the strength of evidence carried by each sample (flash brightness), yielding a novel task in which evidence reliability varies within a trial. A large behavioural dataset was acquired (N = 32 rats). I derived a mathematically explicit model of choice based on the normative solution of the task, parameterised with side biases, inattention, sequential effects and distinct sources of computational variability. Three properties emerged: protracted integration; spontaneous down-weighting of weaker evidence; and a clear signature of inference noise — imprecision arising when beliefs are updated — previously documented in humans but not isolated in animals, where sensory noise and inattention were assumed to be the only drivers of choice variability.
WP2 — Cortical recordings. This WP did not progress as planned. Inspection of GCaMP expression in the transgenic rat line revealed substantial inter-individual variability, with too few animals expressing sufficiently for high-throughput widefield imaging. An alternative route, intracerebroventricular injection of jGCaMP8m in P1–P2 rat pups, was implemented after an ethics protocol amendment and gave promising but still variable results; effort was redirected to WP1 to protect the action's core deliverable. The available dual-electrode telemetric EEG system yielded poor-quality signals with no detectable task-evoked responses. I instead piloted high-density (30-electrode) EEG in anaesthetised mice and rats, which proved promising and now forms the basis of a cross-species alignment method for my future research programme, though funding constraints in the host lab prevented it from progressing beyond pilot stage.
WP3 — Causal role of the LC-NA system. Optogenetics was replaced by chemogenetics for throughput reasons (a single laser limited manipulations to ~3 animals/day), with the added benefit of allowing direct comparison with human pharmacology. Excitatory DREADDs (PRS×8-hM3Dq) were expressed in the LC of trained rats (N = 15); transduction was verified histologically with anti-TH, and DCZ-evoked pupil dilation under anaesthesia served as a functional positive control. Tonic LC-NA activation slowed reaction times and produced a small increase in choice accuracy, with a steeper psychometric slope. Complementary pharmacological manipulations (atomoxetine, atipamezole) gave a coherent picture. Fit of the computational model (WP1) to this dataset will assist in identifying which part of the decision process is regulated by NA signaling. A collaboration was initiated for a parallel human atomoxetine study (Hôpital Sainte-Anne, Paris); the ethics protocol is approved and data collection should begin in early 2027.
Training, collaboration and outputs. I extended my animal expertise from mice to rats (handling, surgery, perfusion, breeding, perinatal surgery), supervised six undergraduate students, restarted and coordinated lab meetings, and established external collaborations for a methodological review on behavioural shaping with C. Constantinople, K. Kuchibhotla and K. Miller (to appear on the corresponding preprint). Two outputs have appeared to date: a preprint on shaping thses principles (PsyArXiv, 2025) and a co-authored paper on the regulation of the decision threshold by the locus coeruleus (Neuropsychopharmacology, 2026).

Fortschritte, die über den aktuellen Stand der Technik hinausgehen und voraussichtliche potenzielle Auswirkungen (einschließlich der bis dato erzielten sozioökonomischen Auswirkungen und weiter gefassten gesellschaftlichen Auswirkungen des Projekts)

The principal advance is a behavioural paradigm that extends the classical Poisson-clicks family of tasks to variable evidence reliability. In the canonical version, all evidence samples are equally informative; here, each sample carries a manipulated strength, so the animal must not only accumulate but also weight. This single modification unlocks a question that had been essentially inaccessible in animals, and it does so within a task that dozens of laboratories already run — which should ease adoption.
Second, the accompanying computational model separates sources of decision variability that had remained conflated in animal work. Demonstrating that rats' choices carry a signature of inference noise, alongside sensory noise and inattention, aligns the animal literature with the framework developed in human decision-making and removes one of the standing objections to treating rodent and human choice behaviour as commensurable. Because the same model can be fitted to both species, cross-species comparison becomes quantitative rather than qualitative.
Third, the LC-NA experiments compare, side-by-side and within a single task and modelling framework, three routes to elevated noradrenergic signalling (selective chemogenetic activation, reuptake inhibition, alpha-2 antagonism). The literature currently attributes several distinct functions to noradrenaline — sequential effects, perceptual gain, speed–accuracy trade-off — but these were established in unrelated settings. This design allows them to be arbitrated against one another.
Fourth, the pilot high-density EEG recordings in rodents establish the feasibility of a cross-species "reference signal" that can be recorded with comparable spatial coverage in both rodents and humans, addressing a well-known bottleneck in translational neuroscience: human and rodent methods are complementary rather than commensurable.
Needs for further uptake. Realising the full value of these results requires: (i) completion of the human behavioural dataset in the calibrated task version during the returning phase, and joint modelling of both species — the central deliverable now in progress; (ii) publication of the empirical manuscript with open deposition of data, analysis code and model implementations, so that other groups can fit the framework to their own datasets; (iii) dedicated funding to move high-density rodent EEG from pilot to behaving animals, which was the binding constraint during the outgoing phase; and (iv) delivery of the approved human pharmacological study (early 2027), which will close the loop between rodent chemogenetics and human pharmacology.