The evidence reviewed here supports a shift from a linear dopamine-centric model to a network model in which reward, aversion, and behavioral control emerge from interactions among partially dissociable but strongly coupled systems, including dopaminergic, peptidergic, endocannabinoid, histaminergic, metabolic, glial, and cortical–brainstem regulatory processes (Koob and Volkow, 2016; Lüscher and Janak, 2021; Volkow and Blanco, 2023; Nestler, 2025). In this view, dopamine remains indispensable for many forms of reinforcement learning and incentive motivation, but its effects are gated, amplified, constrained, or reweighted by parallel systems that encode stress, arousal, interoception, inflammatory tone, and executive control (Kahnt and Schoenbaum, 2025; Robinson and Berridge, 2025). Accordingly, addiction is better conceptualized not as a hyperdopaminergic disorder per se, but as a failure of coordinated network regulation in which the balance between approach-promoting and avoidance-promoting processes becomes progressively biased toward compulsive drug seeking. This formulation accommodates a wider range of findings than classic dopamine models, including persistent drug seeking under punishment, stress-induced relapse, blunted responsiveness to natural rewards, and the prominent role of non-dopaminergic interventions in modulating addiction-like behavior (Koob and Volkow, 2016; Kim and Picciotto, 2023; Valentino et al., 2024).

A central implication of this network perspective is that neuromodulatory systems should not be treated as parallel “add-ons” to dopamine, but as interacting control layers with partially distinct computational roles. Orexin links arousal and internal state to motivational vigor and cue responsiveness; dynorphin and CRF bias the system toward negative affect and negative reinforcement; endocannabinoids regulate synaptic gain and plasticity; histamine modulates attentional allocation and behavioral flexibility; ghrelin and GLP-1 embed reward processing within metabolic and interoceptive state; and glial mechanisms shape the permissive inflammatory and homeostatic environment in which these signaling systems operate (Koob, 2021; Klausen et al., 2022; Valentino et al., 2024). These systems converge on overlapping hubs—including the ventral tegmental area, nucleus accumbens, prefrontal cortex, extended amygdala, lateral habenula, and now potentially the SVTg—where they can alter both immediate motivational output and the longer-term plasticity that stabilizes addictive behavior (Lüscher and Janak, 2021; Grillner, 2025; Zichó et al., 2025). Thus, the most useful organizing principle is not transmitter identity alone, but how each system contributes to network functions such as salience allocation, reward prediction, stress recruitment, action selection, and behavioral flexibility.

Within this framework, the SVTg–lateral habenula axis represents one of the most provocative recent additions because it offers a potential upstream mechanism for regulating dopaminergic and aversion-related circuitry at the brainstem level (Grillner, 2025; Zichó et al., 2025). The available data suggest that the SVTg is a GABAergic brainstem structure with reward-related properties that has been shown to inhibit the lateral habenula, and to promote place preference, reduce anxiety-like behavior, and be engaged by reward and reward-predicting cues in rodents (Zichó et al., 2025). Conceptually, this positions the SVTg as a candidate integrative node through which cortical, limbic, and midbrain signals may bias motivational behavior toward approach or avoidance. However, this interpretation remains preliminary. First, the evidence is overwhelmingly preclinical and heavily dependent on optogenetic, viral-tracing, and cell type-specific manipulations performed under highly controlled conditions. Such methods are powerful for causal dissection but may exaggerate circuit specificity relative to physiological dynamics. Second, claims of translational conservation remain inferential: although anatomical markers suggest homologous structures in rat, macaque, and human brainstem, there is currently no direct functional evidence that the human SVTg plays an analogous role in reward regulation, relapse, or affective disorders (Grillner, 2025; Zichó et al., 2025). Third, the degree to which this circuit contributes to addiction, rather than more general reward learning or anxiety regulation, remains unresolved. The SVTg therefore should be regarded as a compelling hypothesis-generating discovery rather than an established clinical target.

More broadly, the literature across neuromodulatory systems reveals a recurring tension between mechanistic richness in animal models and limited translational traction in humans. Orexin is a clear example. Preclinical evidence strongly implicates orexin in cue-induced reinstatement, high-effort reward seeking, and stress-related relapse, particularly through OX1R-mediated mechanisms. Yet this mechanistic plausibility has not yet translated into robust clinical evidence that orexin antagonism substantially modifies addictive behavior in humans. The relatively modest and heterogeneous effects observed clinically, together with the complex phenotype of narcolepsy, argue against simplistic interpretations in which orexin is treated as a singular “relapse transmitter” (Koob, 2021). Similar caution applies to histamine: the preclinical data support a role in attentional gain, inhibitory control, and modulation of corticostriatal function, but direct evidence for clinically meaningful effects of histaminergic manipulation in addiction remains sparse. In both cases, the animal literature likely captures genuine mechanisms, but the human relevance of these mechanisms may depend on disease stage, substance class, sleep state, comorbidity, or baseline circuit dysfunction—variables that are often minimized in preclinical designs but are central in clinical populations.

The endocannabinoid and metabolic systems further illustrate why an integrated network model is preferable to a transmitter-centered one. The endocannabinoid system is well positioned to modulate motivational salience because it regulates glutamatergic and GABAergic transmission across corticolimbic circuits and is highly sensitive to developmental timing, environmental context, and prior drug exposure. However, the translational literature remains complicated by major sources of heterogeneity, including cannabis potency, route of administration, polysubstance use, age at exposure, and premorbid psychiatric vulnerability. Adolescent exposure studies are especially informative mechanistically, but their clinical interpretation is limited because they cannot fully disentangle causal drug effects from pre-existing liability. Similarly, the literature on GLP-1 is promising but still immature. GLP-1 receptor agonists reduce alcohol and drug intake in multiple preclinical models, and early human studies suggest potential effects on craving and abstinence, yet the mechanisms remain uncertain and the clinical signal is still modest and incomplete (Klausen et al., 2022). Importantly, the available review evidence emphasizes that GLP-1 effects may not be explained solely by reduced reward, but may also involve satiety, aversion, stress-linked negative feedback, and habenula-related mechanisms; moreover, proposed dopamine transporter effects may be species dependent and not reproduced in humans (Klausen et al., 2022). These issues are not peripheral details: they mean that GLP-1 should currently be considered a state-dependent network modulator with translational potential, not a validated anti-addiction mechanism.

Glial mechanisms are particularly important for any cohesive network model because they introduce slower, permissive, and homeostatic forms of regulation that cannot be captured by neuron-only accounts. Astrocytic glutamate transport, microglial activation, cytokine signaling, and trophic modulation all have the capacity to reshape reward circuitry by altering extracellular glutamate tone, synaptic plasticity, and inflammatory set points (Lacagnina et al., 2017; Skupio et al., 2020; Nestler, 2025). In conceptual terms, glia may determine whether neuromodulatory perturbations remain transient or become stabilized into relapse-prone states. Yet here too the translational gap is substantial. Most evidence comes from rodent models using experimental manipulations that are difficult to map directly onto human addiction, and the available human markers—PET ligands for neuroinflammation, peripheral inflammatory markers, or magnetic resonance spectroscopy measures—remain indirect, nonspecific, and variably reproducible. The field therefore faces a methodological asymmetry: glial processes are increasingly central to mechanistic theories, but they are still difficult to quantify in living humans with enough specificity to guide intervention.

At the computational level, the review also highlights a mismatch between the complexity of the biology and the simplicity of many prevailing models. Classical reward-prediction-error accounts remain highly useful, but they cannot fully explain punishment-resistant drug seeking, prolonged negative reinforcement, or learning about reward features that go beyond scalar value (Kahnt and Schoenbaum, 2025). Likewise, polygenic risk approaches capture only a modest fraction of addiction vulnerability and remain difficult to interpret mechanistically in the absence of richer intermediate phenotypes (Lüscher and Janak, 2021; Volkow and Blanco, 2023). What is needed are models that explicitly integrate multiple control layers: fast dopaminergic prediction signals, slower stress and anti-reward processes, even slower glial and epigenetic stabilization, and top-down cortical modulation of action selection. Such models are conceptually attractive, but they are currently underdetermined by the data because few studies simultaneously measure behavior, circuit dynamics, and molecular or glial states within the same individuals or across comparable preclinical and clinical paradigms.

The translational consequences of this analysis are substantial. First, the limited efficacy of many candidate treatments should not be interpreted simply as evidence that the underlying biology is wrong. In several cases, negative or weak clinical findings may reflect problems of target engagement, inadequate patient stratification, endpoint selection, or the fact that a mechanism is only relevant in a subset of individuals or at a specific stage of the addiction cycle. The literature on CRF1 antagonists is instructive: despite strong preclinical rationale and evidence of endocrine target engagement, human laboratory studies have yielded largely negative or equivocal behavioral results, and no convincing clinical efficacy has yet emerged for major stress-related psychiatric disorders or addiction (Koob, 2021). This does not necessarily invalidate the broader stress-framework, but it does show that mechanistic plausibility alone is insufficient for therapeutic translation. Second, many interventions are still evaluated in diagnostically broad populations without biomarkers that identify the circuit dysfunction they are intended to modify. This is a major obstacle for neuromodulatory therapies such as rTMS or deep brain stimulation, whose effects are likely to depend critically on whether the targeted network is actually dysregulated in a given patient (Lüscher and Janak, 2021; Valentino et al., 2024). Third, implementation constraints matter. Even compelling biological advances may fail clinically if they are difficult to deliver chronically, poorly tolerated, too expensive, or not adaptable to diverse health systems and comorbidity profiles.

These translational limitations also underscore important differences between preclinical and clinical evidence. Animal models excel at causal specificity but often simplify the phenotype, typically isolating one substance, one stressor, one behavioral endpoint, and one developmental window. Human addiction, by contrast, is shaped by polysubstance use, social stress, psychiatric comorbidity, developmental heterogeneity, sex differences, sleep disturbance, and structural determinants of health (Volkow and Blanco, 2023; Valentino et al., 2024). This difference in dimensionality has methodological consequences. A manipulation that robustly reduces reinstatement in rodents may correspond to only a small effect on relapse risk in humans because the human phenotype is sustained by multiple partially redundant processes. Conversely, clinically important variables such as access to treatment, adherence, socioeconomic instability, or trauma exposure are seldom modeled adequately in basic studies. Progress will therefore require tighter reverse translation, more ecologically valid behavioral phenotyping, and multimodal longitudinal studies that can bridge circuit-level hypotheses with real-world clinical trajectories.

An additional unifying but underdeveloped dimension is time. The systems reviewed in this article operate on partially distinct temporal scales: dopamine on phasic and tonic timescales, orexin and stress systems across arousal states, endocannabinoid and glial processes across slower homeostatic and inflammatory windows, and metabolic signals across feeding, circadian, and endocrine rhythms. Most addiction studies, however, sample only one of these temporal layers. As a result, the field still lacks robust models of how short-timescale incentive signaling becomes embedded within longer-timescale stress and homeostatic dysregulation. This limitation is not merely theoretical. It may explain why acute laboratory measures often correlate weakly with long-term clinical outcomes and why interventions that change short-term craving do not necessarily alter relapse trajectories. Future work should therefore treat temporal organization as a core feature of addiction neurobiology rather than as a secondary consideration.

Taken together, the literature supports a cohesive network model in which addiction reflects a progressive failure of flexible valence regulation across interacting neuromodulatory, glial, metabolic, and cortical–brainstem systems. The value of this model lies not in replacing one master transmitter with several others, but in explaining how multiple systems with distinct functions and timescales converge on shared behavioral phenotypes: exaggerated cue-driven approach, diminished valuation of natural rewards, impaired avoidance learning, stress-triggered relapse, and reduced executive flexibility. At the same time, the review also makes clear that the empirical basis for this model is uneven. Some nodes and mechanisms are well supported in preclinical work but weakly validated in humans; some translational candidates are biologically plausible but clinically unproven; and several of the most integrative ideas still outpace the available methods. A critical next step is therefore not simply to expand the list of implicated systems, but to test explicit interaction models across species, scales, and clinical contexts. Only then can systems-level theories of reward and addiction support precision, circuit-informed interventions that are both mechanistically sound and clinically meaningful.