A habenula-insular circuit encodes the willingness to act

The study investigates how humans decide whether to initiate an action at all, rather than which action to choose or when to execute it. The authors define willingness-to-act as the probability of acting given current contextual factors (reward magnitude, reward probability, and broader environmental richness/poverty). They highlight that whereas neural mechanisms for action selection are well studied, the circuitry for deciding to act versus withhold action is less understood, despite relevance to adaptive behavior and clinical symptoms like apathy and impulsivity. Prior work implicates frontal cortical regions (vmPFC, anterior insula, dACC, sgACC) and subcortical structures (NAc, LDT, VP, VTA, habenula, PPN) in motivated behavior, but their circuit-level interactions for initiating voluntary action remain unclear. Building on prior evidence that subcortical networks (including basal forebrain, striatum, substantia nigra, habenula, and PPN) predict timing of self-initiated actions, the present study focuses on the antecedent decision of whether to act, using an experimental paradigm where participants could choose to do nothing and continue watching a movie or accept effortful opportunities for potential reward. The central hypothesis is that specific contextual features shape willingness-to-act, which is encoded in a cortico-subcortical circuit linking anterior insula and habenula and ultimately influences motor implementation via the nigrostriatal pathway.

The authors review evidence that frontal cortical regions (vmPFC, anterior insula, dACC, sgACC) contribute to motivated behavior and valuation. Subcortically, regions such as NAc, LDT, VP, VTA, habenula (HB), and PPN are implicated in motivational drive and action initiation. Prior human 7 T fMRI work showed early predictive activity for self-initiated actions in the striatum, SN/VTA, BF, HB, and PPN, with BF mediating environmental context effects on when to act. Other animal and human studies link SNc activity to gating and invigorating movement initiation, and the habenula to avoidance, negative outcomes, and impulsivity control via influence on midbrain dopamine and serotonin systems. Resting-state studies show HB coupling with the salience network (including anterior insula). Together, these studies motivate testing how cortical and subcortical nodes coordinate to compute willingness-to-act and translate it into action.

Participants: 25 healthy adults (7 males, 18 females; ages 18–40). One excluded for infrequent responding; two excluded from fMRI for excessive motion, leaving n=24 for behavior and n=22 for fMRI ROI analyses. Compensation included £15/hour plus performance bonus (£3–£7). Ethics approved by Oxford CUREC; informed consent obtained. Task: Inside a 7 T scanner, participants watched a Planet Earth episode with intermittent offers to act. Each 2 s offer was a colored rectangle containing 1–21 dots. Color encoded reward magnitude (low 5p, medium 10p, high 20p; color-to-magnitude mapping counterbalanced). Dot count encoded reward probability (linear increase; 21 dots = certain reward). Participants could accept by button press, interrupting the movie to perform an effort task; or reject and continue watching. Upon acceptance, participants exerted 50% of their individual MVC on a dynamometer for a duration sampled from N(3.5 s, 0.5 s). Success made them eligible for the probabilistic monetary reward indicated by the offer; failure yielded no reward. Feedback followed for 2 s. Rejections yielded 0p with 2 s feedback, with movie uninterrupted for an equivalent duration. ITI sampled from N(4.5 s, 0.5 s). Total 216 trials, divided into six 36-trial blocks. Environmental richness manipulated by block: rich blocks had higher ratios of high magnitude/probability offers (50% high; 33% mid; 16% low), poor blocks the inverse (16% high; 33% mid; 50% low). Block type alternated and was announced at block onset. Behavioral modeling: Willingness-to-act per trial defined as the predicted probability of accepting the offer given contextual factors. A generalized linear mixed-effects model (logistic link) estimated effects of current reward magnitude, probability, block type, their interactions, previous-trial expected value, previous response (act/no-act), and elapsed total time (fatigue covariate). Subject-level models were also fit to compute individual trial-by-trial willingness-to-act via the logistic function using participant-specific coefficients. MRI acquisition: Siemens 7 T scanner. Functional EPI: 1.5 mm isotropic, multiband factor 3, TR 1962 ms, TE 20 ms, flip angle 66°, GRAPPA 2, 96 axial slices, AP phase encoding, FOV covering whole brain with -30° angulation. Pre-saturation whole-brain EPI acquired. Structural T1 MP-RAGE: 0.7 mm isotropic, TR 2200 ms, TE 3.02 ms, TI 1050 ms, GRAPPA 2. Fieldmap: 2 mm isotropic, TR 620 ms, TE1 4.08 ms, TE2 5.10 ms. Physiological (cardiac/respiratory) recorded. Preprocessing: FSL pipeline: normalization, 3 mm FWHM smoothing, high-pass filter (100 s), motion correction (MCFLIRT), brain extraction (BET), and multi-step registration (task EPI→pre-sat EPI; EPI→structural with BBR and fieldmap; structural→MNI with FNIRT). Physiological noise modeled (PNM). High-motion timepoints scrubbed (5 ± 3% across participants). Whole-brain GLM (GLM1): Voxelwise regressors included stimulus onset, willingness-to-act (parametric), response (act/no-act), effort magnitude, outcome onsets, reward outcome (rewarded/not), and non-convolved nuisance regressors for instantaneous distortions at response/effort and outcome onsets. First-level per session, then FLAME 1+2 mixed-effects at group level. Cluster correction Z > 3.1, P < 0.001. ROI time-course analyses: A priori subcortical ROIs: dorsal striatum (DS), nucleus accumbens (NAc), midbrain dopaminergic system (MidD; SN/VTA), pedunculopontine nucleus (PPN), habenula (HB), and basal forebrain (BF), defined from atlases and transformed to native space. Functional ROIs for anterior insula and SMA defined as 1.5 mm radius spheres around activation peaks. Time-series extracted, normalized, up-sampled 20×, spline-interpolated, epoched from -2 to +8 s around offer onset. OLS GLMs fit at each time step:

• GLM2.1: BOLD ~ willingness-to-act + response + constant (for HB and other ROIs; also for SMA/insula when comparing WTA vs response effects)
• GLM2.2: BOLD ~ reward magnitude + probability + block (rich/poor) + previous expected value + previous response + total time + current response + constant (to test encoding of individual components) Significance assessed with leave-one-out peak selection and Holm–Bonferroni correction. Connectivity analyses: PPI analyses tested modulation of functional coupling by psychological variables:
• GLM2.3: Insula–HB coupling modulated by willingness-to-act.
• GLM2.4: HB–MidD coupling modulated by interaction of willingness-to-act and act/no-act; repeated with SNc- and VTA-specific ROIs. Structural equation modeling (SEM): Network models linking AI→HB→MidD→striatum→SMA (with BF and PPN inputs to MidD) were fit using lavaan (ML estimation). Model comparisons via AIC contrasted the hypothesized directionality against alternatives with reversed or altered flows.

Behavioral:

• Participants accepted 45.81 ± 2.79% of offers on average.
• Mixed-effects logistic regression: higher reward magnitude increased acceptance (β = 2.07 ± 0.26, χ²(1) = 48.64, P < 0.001); higher reward probability increased acceptance (β = 3.53 ± 0.26, χ²(1) = 167.46, P < 0.001); poor vs rich blocks increased acceptance (β = −0.85 ± 0.18, χ²(1) = 16.33, P < 0.001; coded such that poor blocks raised willingness). Previous-trial expected value tended to decrease acceptance (β = −0.16 ± 0.08, χ²(1) = 3.40, P = 0.065). Previous response (act/no-act) had no effect (β = −0.06 ± 0.17, P = 0.70). Neuroimaging: act/no-act encoding (stimulus-locked):
• BOLD correlated with act/no-act in DS (t(21) = 9.78, P < 0.001, d = 2.08), MidD (t(21) = −3.52, P = 0.008, d = 0.75), PPN (t(21) = −3.09, P = 0.01, d = 0.66), HB (t(21) = 3.16, P = 0.01, d = 0.67), BF (t(21) = −3.91, P = 0.004, d = 0.83); not in NAc (t(21) = −1.22, P = 0.24). MidD negative relation reflects stimulus-locked contrast; movement-locked analyses showed expected SNc activation at movement onset (Supplementary). Willingness-to-act (WTA) encoding:
• HB tracked trial-by-trial WTA controlling for actual decisions (t(21) = 4.26, P = 0.002, d = 0.90). No other a priori ROI showed this (all P > 0.12).
• Whole-brain analysis: WTA correlated with activity in anterior insula (peak Z = 5.16; MNI 43, 16, 1) and SMA/caudal ACC (peak Z = 5.50; MNI −6, −2, 54), cluster-corrected P < 0.001. Anterior insula did not encode the binary act/no-act decision directly, whereas SMA did.
• Time-course comparisons: WTA explained anterior insula BOLD better than SMA (paired t(21) = 3.01, P = 0.007, d = 0.64). Act/no-act decisions explained SMA BOLD better than anterior insula (paired t(21) = 10.42, P < 0.001, d = 2.22). Two-way RM-ANOVA showed strong interaction (F(1,21) = 84.66, P < 0.001, ηp² = 0.80). Component encoding:
• Anterior insula tracked individual contextual components: reward magnitude (t(21) = 2.83, P = 0.03, d = 0.60), reward probability (t(21) = 6.02, P < 0.001, d = 1.28), block type poor>rich (t(21) = 2.83, P = 0.03, d = 0.60), and previous-trial expected value (t(21) = 2.73, P = 0.02, d = 0.58). In HB, only reward magnitude correlated with BOLD (t(21) = 3.71, P = 0.005, d = 0.79).
• HB–WTA relationship was stronger on trials where participants withheld action despite high-value offers, consistent with a role in impulse control (paired t(21) = −4.93, P < 0.001, d = 1.05; Supplementary). Connectivity and network modeling:
• PPI: WTA modulated functional coupling between anterior insula and HB (t(21) = −2.29, P = 0.03, d = 0.49); no significant modulation for HB–SMA (P = 0.61).
• PPI: HB–MidD coupling was moderated by the interaction of WTA and act/no-act (t(21) = −2.09, P = 0.049, d = 0.44). Subdivision analysis: significant for HB–SNc (t(21) = −2.56, P = 0.018, d = 0.55) but not HB–VTA (t(21) = 1.49, P = 0.15); stronger coupling modulation for SNc than VTA (paired t(21) = 3.36, P = 0.003, d = 0.72).
• SEM: A hypothesized cortico–subcortical–cortical model (AI→HB→MidD→striatum→SMA; with BF and PPN to MidD) fit the data better (AIC = 376,817) than alternatives with reversed or altered directions (AICs: 534,500; 377,651; 534,141). All specified path coefficients were significantly different from zero (Supplementary Table S4). Overall: Contextual features shape willingness-to-act; HB uniquely tracks WTA among subcortical ROIs; anterior insula encodes both WTA and its component determinants; SMA translates WTA into action. WTA emerges via an AI–HB circuit and is conveyed to motor implementation through the nigrostriatal pathway, particularly via SNc.

The findings demonstrate that deciding whether to act is influenced by immediate opportunity value (reward magnitude and probability) and environmental richness, and that these influences are neurally instantiated in a specific cortico–subcortical–cortical circuit. Anterior insula encodes both the integrated willingness-to-act and its constituent contextual elements, positioning it as a cortical hub integrating multimodal information about opportunities and environment. The habenula tracks trial-by-trial willingness-to-act independent of the ultimate choice and shows stronger engagement when high-value offers are resisted, consistent with its role in impulse control and inhibitory influence over dopaminergic midbrain. SMA shows stronger representation of the final act/no-act decision and a later time-to-peak, consistent with translating deliberative states into motor execution. Connectivity analyses indicate that willingness-to-act modulates functional coupling between AI and HB, and that the translation into action involves HB coupling to SNc within the nigrostriatal pathway. SEM supports a model wherein contextual information is integrated in AI, passes to HB to gate dopaminergic influences on striatum, culminating in SMA where the binary act/no-act decision is implemented. These results address the research question by delineating how context shapes willingness-to-act and by mapping the flow of information across brain regions critical for turning valuation into action initiation or restraint. The work has implications for understanding disorders of motivation, such as apathy and impulsivity, and suggests the habenula as a key node linking contextual evaluation to motor gating.

This study introduces and operationalizes willingness-to-act as a context-dependent probability of initiating an action and identifies its neural circuitry using 7 T fMRI. Behaviorally, reward magnitude, probability, and environmental richness robustly modulate willingness-to-act. Neurally, the anterior insula encodes both willingness-to-act and the component contextual factors; the habenula uniquely tracks willingness-to-act among subcortical ROIs and shows coupling with SNc as action decisions unfold; SMA encodes the ensuing binary act/no-act decision. PPI and SEM converge on a cortico–subcortical–cortical pathway in which contextual information is integrated in AI, transmitted to HB, and then conveyed via the nigrostriatal pathway to SMA for implementation. Future research should (i) dissociate medial vs lateral habenular contributions with higher-resolution or invasive methods, (ii) parse the roles of neuromodulatory systems (dopaminergic, serotonergic, cholinergic) in shaping willingness-to-act, (iii) test generalization across tasks and real-world settings, and (iv) investigate clinical populations (e.g., depression, apathy) where willingness-to-act may be disrupted.

• Spatial resolution limits in 7 T fMRI prevented dissociation of medial vs lateral habenula subdivisions; analyses combined them into a single HB ROI.
• Although movie content was uncorrelated with task events and a control experiment indicated no confounding effect of interest-in-the-movie, unmodeled variability in the movie remained as noise.
• Sample size was modest (n=25; n=22 for fMRI), potentially limiting detection of smaller effects and generalizability.
• Task context involved low-stakes monetary rewards and a specific effort modality (grip at 50% MVC), which may limit generalization to other effort domains or higher-stakes, real-world decisions.
• The study relies on BOLD correlations and SEM for effective connectivity; causal inferences about directionality are constrained without perturbation methods.