Thesis defense – Margaux Giraudet

Adaptation to volatile environments is a fundamental, everyday challenge that relies predominantly on thalamocortical pathways. In particular, mediodorsal thalamus (MD) and ventroanterior-ventrolateral (VAVL)–to–secondary motor cortex (M2) projections are well positioned to shape action selection, yet their computations during adaptive choice remain unclear.

During my thesis, I combined a self-initiated two-lever task in head-fixed mice, focal MD and VAVL lesions, longitudinal two-photon imaging of MD and VAVL to M2 axons, and a reinforcement-learning race model to determine how thalamocortical signals govern preparation, vigor, and action selection under uncertainty. Mice chose left (L1), right (L2), or near-simultaneous (L1&L2) presses; outcomes followed probabilistic (80/20) contingencies and double presses never yielded reward. A Rescorla–Wagner rule updated action values (Q1, Q2), which drove independent accumulators with a finite co-selection window (δt) that defines double-press events. The model captured choice fractions, reward rate, and reaction times (RTs), isolating two latent variables: ΔQ (value difference) that biases which action wins, and ΣQ (value sum) that controls initiation vigor.

Targeted MD lesions selectively impaired de novo rule learning without disrupting reversal or gross motor output. Learning rates and the value-to-rate mapping were preserved, but δt widened specifically after MD lesion, producing more L1&L2 presses. Because double presses never yield reward, a wider δt mechanically reduced performance, compressed ΔQ, and diminished rate separation, thereby promoting ambiguous co-selection despite intact value updating. In contrast, VAVL lesions spared learning, indicating circuit specificity. Indeed, if both thalamic nuclei anatomically project to M2, our connectomic experiments demonstrated topographic segregation: MD preferentially targets medial M2, whereas VAVL innervates M2 more broadly, supporting thalamocortical functional heterogeneity.

To link these computations to thalamocortical signaling, we imaged MD and VAVL axons in M2 across learning in behaving mice. Clustering of MD bouton activities revealed a specific ensemble in M2 that remain stable over days and showed robust pre-press modulation. Critically, this pre-press MD-M2 signal was not a generic correlate of performance (it did not scale with success rate) but was tightly coupled to RT and to double-press control. Trials binned by RT showed that stronger pre-press activity preceded faster responses, explaining substantial RT variance. Its relationship to selection was context-dependent: under long pre-press delays, greater pre-press activity predicted fewer double presses (an anti-double-press bias), whereas under short delays its predictive power approached chance.

This framework unifies behavior, computation, and circuit dynamics. The race model formalizes how ΔQ and ΣQ jointly shape selection and speed; δt widening after MD lesion pinpoints a temporal-coincidence gating failure that degrades performance by promoting co-activation; and imaging identifies a pre-press thalamocortical state that links preparation to both RT and double-press control. This mechanistic chain parsimoniously explains why MD lesions alter learning despite preserved motor abilities and value updating.

Taken together, these findings indicate that MD-M2 circuit specifically conveys a control signal before the press that both accelerates RT and, under long delays, biases choices away from double-pressing toward the target lever. MD lesions selectively increase double-press, as expected from an altered pre-press signal that weakens the anti-co-selection bias at long delays. Overall, our data suggest that MD-M2 circuit controls cortical preparatory states and enforces temporally precise, single-action exclusive selection under uncertainty.