Reversal Learning Exposes Different Cognitive Signatures Across Mouse Models of Neurodevelopmental Disorders
Adaptive decision-making does not fail in only one way. In this preprint, reversal learning tasks paired with reinforcement-learning modeling separated distinct cognitive phenotypes across three mouse models of neurodevelopmental disorders, even when standard behavioral readouts looked similar at first glance.
The work centers on a simple but clinically relevant problem: many genetic disorders linked to neurodevelopment disrupt flexible behavior, yet broad task scores often blur the underlying deficit. A mouse may respond too slowly, stop engaging, or persist with an outdated choice strategy, and those behaviors do not mean the same thing biologically. By layering deterministic and probabilistic reversal learning onto an automated operant platform, the investigators tried to split those components apart and assign them to latent learning processes rather than treating them as one undifferentiated impairment.
Deterministic Reversal Learning Brought Out the Strongest Deficits in Mecp2 Haploinsufficiency
Among the three genotypes tested, Mecp2+/− mice showed the clearest disruption during acquisition of deterministic reversal learning. The impairment was heterogeneous, which matters because it suggests uneven expression of the phenotype rather than a uniform failure across the cohort. These mice also showed reduced task engagement and increased premature responses, a combination that points toward both motivational and inhibitory-control components of dysfunction.
A single behavioral score would have flattened those distinctions. Reduced engagement can suppress apparent performance even when the underlying learning machinery is not the only problem, while premature responding can reflect impulsivity, poor action timing, or difficulty maintaining a response policy long enough to benefit from feedback. In a clinical frame, this looks less like one global “cognitive deficit” and more like a compound syndrome in which initiative, response control, and rule updating are all strained to different degrees.
MECP2 Duplication Mice Kept Probabilistic Learning Intact
The MECP2-Tg1 mice showed only mild impairment in deterministic reversal learning and preserved performance in probabilistic reversal learning. That contrast is important because probabilistic reversal learning places a different burden on the animal: instead of simply following a stable rule, the subject must interpret noisy feedback and adapt when reward contingencies become unreliable. Preservation of that ability suggests that not all forms of flexibility are equally vulnerable in this model.
The biological implication is straightforward. Extra MECP2 copy number did not erase adaptive choice altogether; it shifted the margin of difficulty. Deterministic reversal, where the rule changes cleanly and the correct response should switch decisively, appears more sensitive to the genotype than a task that includes uncertainty by design. For translational neuroscience, this kind of dissociation matters because it argues against assuming that every assay of “cognitive flexibility” measures the same circuit or the same learning algorithm.
Latent Learning Signals Emerged in Kmt2c Haploinsufficiency Despite Preserved Behavior
The most revealing finding came from Kmt2c+/fs mice. Overall behavioral performance was largely preserved, which on its own could have suggested a near-normal phenotype. Reinforcement-learning modeling told a different story. These mice showed increased learning from unrewarded outcomes and faster forgetting of unchosen options, effects that were not obvious in the raw behavior.
That pattern is the kind of computational signature that conventional task scoring often misses. If an animal updates too strongly from negative feedback, it can become overresponsive to occasional omission and destabilize choice policy. Faster forgetting of unchosen options implies a weaker internal record of alternatives not selected on the previous trial, which could narrow future decision-making and reduce the capacity to maintain a broad choice set. In practice, the animal may still look competent on average while operating through a noisier and less stable internal algorithm.
Why Computational Modeling Changes What the Task Can See
Reversal learning tasks are often treated as measures of behavioral flexibility, but the present framework shows that flexibility contains separable elements. Engagement determines whether the animal participates consistently enough for the data to be meaningful. Premature responding reflects action timing and inhibitory control. Sensitivity to reward and nonreward shapes the learning rate. Memory for unchosen options affects how much earlier decisions continue to influence later ones. When these pieces are modeled explicitly, different genotypes can diverge even when they finish with similar headline performance.
This is where the work becomes especially relevant for neurodevelopmental research. Mecp2 haploinsufficiency, MECP2 duplication, and Kmt2c deficiency do not collapse into a single “NDD behavior” category. Each model expresses a different arrangement of deficits, and those differences may mirror how patient groups with distinct genetic etiologies show overlapping diagnoses but different cognitive profiles in daily life. A battery that captures only success or failure risks missing the biologically informative structure of the deficit.
A Framework Built for Disentangling Heterogeneous Phenotypes
The study’s main contribution is methodological as much as biological. Automated reversal learning combined with reinforcement-learning analysis created a system that can distinguish between overt behavioral impairment and hidden computational abnormalities. That approach is especially useful in disorders where the phenotype is patchy, variable, or masked by compensatory behavior.
One sentence captures the practical consequence: a mouse can perform “well” and still learn in the wrong way. In Kmt2c+/fs mice, the covert learning signature was visible only when the data were decomposed into latent variables. In Mecp2+/− mice, the phenotype was more overt, with engagement and premature responding compounding reversal failure. In MECP2-Tg1 mice, the deficit was milder and task-dependent, with probabilistic learning spared.
Because this is a preprint and has not yet undergone peer review, the results should be read as an early mechanistic framework rather than a settled template. Still, the message is unusually crisp for a behavioral genetics study: if investigators want to understand cognitive dysfunction across neurodevelopmental models, they will need to measure not just whether an animal changes course, but how it updates belief, how it values unrewarded outcomes, and how much of the past it keeps in reserve when a choice has not been made.