Computational Modeling of Catecholamines Dysfunction in Alzheimer’s Disease at Pre-Plaque Stage

RML equations

The RML architecture consists of four computational modules, two simulating part of the medial prefrontal cortex: the dorsal anterior cingulate cortex (dACCAct and dACCBoost) and two simulating the brainstem catecholamine nuclei (VTA and LC). Cortical and subcortical modules are reciprocally connected in order to generate a control loop that optimizes both catecholamines release (subcortical) and decision-making operations (cortical) while the RML interacts with the external environment (Fig. 1a). This machinery allows the optimization of neuromodulation supporting the cognitive and behavioral efficacy of the RML while it executes a task.

Fig.1

a) Overview of the RML model. This consists of two interacting systems: a state-action selection system (dACC_Boost and dACC_Act), based on reinforcement learning processes, and a parameter modulation system via catecholamine release (VTA and LC). The RML model can be connected to an external neural component (e.g., the model FROST representing the fronto-parietal network) that is targeted by the LC output (NE), while the entire system (RML + external component) interacts with the environment. Modified from [25]. b) The FROST model [28] simulates a fronto-parietal network for visuo-spatial WM (F, prefrontal neurons; P, parietal neurons). The RML-FROST integrated system can perform visuo-spatial WM tasks by optimizing prefrontal activity (F) via LC activity. c) Plot showing the gain field modulation by LC on one prefrontal (F) neuron, during a delayed matching-to-sample task. A stimulus is presented inside the receptive field of the neuron, followed by a delay period, and then by a second stimulus (target) in the same spatial position.

Here we describe the RML functioning in its discrete-time implementation. All the free parameters of the model, reported in Table 1, were set as in Silvetti et al. [25] (see also this work for the dynamical version of the model). The software used for the simulations can be downloaded from the RML GitHub repository: https://github.com/AL458/RML. Data from human studies were extracted from the figures of their relative articles by means of WebPlotDigitizer (https://automeris.io/WebPlotDigitizer/).

We now describe the functioning and equations of each RML component.

Control over other brain areas

The RML uses the LC-based control signal (NE) to optimize performance of other modules simulating other brain regions. The dACC_Act module implements a decision-making function in a domain-independent way, i.e., the activation of state/action channels can encode any decision process in other brain areas outside dACC. In addition, since optimization of any brain region improves behavioral performance, the dACC_Boost can use LC signals to modulate any cortical area to optimize its performance. For example, we simulated a visuo-spatial working memory (WM) task (Simulation 1, below) by connecting the RML to an independently developed and published neural model that represents the fronto-parietal network for visual space representation.

Simulations

We administered to the RML both a visuo-spatial WM task (Simulation 1, Fig. 2a), and a decision-making task involving different levels of effort engagement (Simulation 2, Fig. 2b), in both a condition simulating the VTA degeneration and a control condition. We then compared the simulation results at both the behavioral and the neural levels with those deriving from human studies (see below). To mimic standard experimental paradigms as closely as possible, we repeated each simulation only 20 times (i.e., 20 simulated participants). This verified that the model could generate results with a large effect size (p-values, but not effect sizes, can be improved arbitrarily by running more simulated participants).

Fig.2

a) Delayed matching-to-sample task (Simulation 1). This task was administered to the RML-FROST system (Fig. 1b, c). The task difficulty (WM load) was simulated by varying from 1 to 30 s the delay between sample and target presentation (the figure only shows the case with 10 s). b) Effort task (Simulation 2). The RML chooses between two options (blue and red squares), a high effort option (thick black arrow moving the lever) leading to a large reward (HR option) and a low effort option (thin black arrow moving the lever) leading to a small reward (LR option).

Simulation 1: Working memory impairment

In this simulation, the RML was administered a visuo-spatial delayed matching-to-sample task (Fig. 2a), in both normal and VTA-damaged conditions (DA levels reduced by 60%). In order to simulate this task, we integrated the RML with a recurrent neural network simulating visuo-spatial WM functions: The FROntal-Striatal-Thalamic model (FROST; Fig. 1b, c, [28]; see [25] for a mathematical description of this model and how it is connected to the RML). The FROST model is a recurrent neural network with frontal neurons (F) and parietal neurons (P), reciprocally interconnected and simulating visuo-spatial WM. The RML model is connected to the FROST model so that the dACC module can select one of the items retained in the F network, while the LC modulates F neurons activity (Fig. 1b, c). During the delayed matching-to-sample task (Fig. 2a), a pattern of visual items (stimuli) was presented inside the receptive fields of P neurons then a target item was presented after a delay. During the delay, spatial information was retained in F neurons thanks to the recurrent F-P activity. The RML read out the representations in the F layer of the FROST model and made a decision on whether the target matched or not one of the previously presented stimuli. Importantly, the RML also played a role of optimizer of the FROST performance as a function of memory load. This optimization process was implemented by gain modulation of F neurons by the NE signal afferent from the LC module (orange circuit in Fig. 1b, c), which increased the coding precision of spatial locations retained in F neurons. This mechanism simulates the effect of NE on prefrontal neurons [29]. One of the main features of the FROST model is that it can simulate the decrease of memory precision as a function of memory load, so that the higher the number of items to be retained in memory, the lower is the precision of their representation in its F cortical neurons (Fig. 1b). In this simulation, we also implemented the detrimental effect of the time delay on memory precision by decreasing the activity from the frontal cortex F as a function of time. We implemented this by making the gating signal A from Equation S20 in Silvetti et al. [25] decaying as a function of time: Ȧ=-jA , where j = 10^–4 is a decay constant.

The delayed matching-to-sample task implied different memory loads, namely a template of 4 items to be retained for 1, 10, or 30 s (Fig. 3a). We used a block design where we administered three blocks of 70 trials, each with one specific memory load. In 50% of trials, the probe matched with the template.

The statistical analysis of the results was conducted by a repeated measure 3×2 ANOVA (memory load by DA lesion). We estimated three design matrices related to three dependent variables, the first related to accuracy, the second related to reaction times (RTs), and the third related to the LC boosting by the dACC_Boost module. The results were qualitatively compared with experimental results from both human controls and patients from [30]. Plot data from [30] were extracted through WebPlotDigitizer from Fig. 5 of the original article.

Fig.3

Neural and behavioural results from Simulation 1 (a–c), compared with experimental data from humans (d, e), related to the performance of MCI and control participants in a delay memory test. Blue plots indicate human or simulated controls, while red plots indicate human or simulated patients. a) LC module activation (arbitrary units, a.u.) as a function of task difficulty (delay duration). Error bars indicate s.e.m. b) Response accuracy as a function of difficulty (±s.e.m). c) RTs as a function of task difficulty (±s.e.m). d) Response accuracy in human participants (from [30]; number of recalled items±s.e.m) in a similar WM task. e) RTs in human participants (±s.e.m) from the same data set [30].

Simulation of VTA lesion

The VTA lesion was implemented by multiplying the output of the VTA module in the RML by a number in the range of (0, 1), for example a 60% VTA degeneration implied to multiply the VTA output (DA and DA_B signals) by 0.4. For Simulation 1 (WM impairment), we used the dynamical version of the RML [25] and simulated the cortical-subcortical activity with a time resolution of 10 ms. In Simulation 2 (motivational impairment) we used the discrete version of the RML.

The reinforcement meta-learner model: A qualitative description

We investigated the consequences of dopaminergic neurons loss on catecholamine release, and thence on behavior, by using the Reinforcement Meta-Learner (RML) computational model [25]. This model describes the dynamical interactions between the catecholamine brainstem nuclei (LC and VTA) and the dorsal anterior cingulate cortex (dACC), simulating the optimal regulation of catecholamines release as a function of environmental conditions. The architecture of the model pivots on the biologically grounded idea that behavioral adaptation can be seen as a problem of optimization [33, 34] whose objective is to maximize the long-term reward while minimizing costs. In order to do so, animals need to learn not only how to optimally control their behavior but also how to control the internal variables that influence brain processes. For example, a foraging rat makes decisions about its movements in the environment and, at the same time, controls internal variables linked to activity engagement, to how much energy to spend in action performance (cost/benefit trade-off), and whether to take into account or not eventual changes in the surrounding environment (plasticity/stability trade-off) [35]. Learning to optimally control these internal variables is defined as meta-learning [36, 37].

The RML is built on the hypothesis that the mammalian brain performs meta-learning by optimizing the release of neuromodulators [38] as a function of task demands and expected rewards. Although this idea could be extended to any neuromodulator, for parsimony we decided to model the control of catecholamines (DA, and NE), a class of neuromodulators having strong and immediate influence on cognition and behavior. In our model, the release of catecholamines is regulated by a cortical-subcortical macro-circuit including the brainstem catecholamine nuclei (VTA and LC), the dACC, and their reciprocal connections [39–41]. The RML architecture (Fig. 1a) is based on two interacting loops connecting four computational modules: dACC_Act, dACC_Boost, VTA, and LC. An internal loop manages the interaction between the brainstem nuclei VTA and LC and the dACC modules (Fig. 1a, orange and red bidirectional arrows), whereas an external loop manages the interaction between the cortical modules and the external environment. The RML is aimed at minimizing both the costs of neuromodulators release (internal loop), and the costs of motor actions (external loop; e.g., the metabolic cost of climbing a stair) while optimizing performance (i.e., maximizing reward). In other words, we can describe the RML as an optimizer whose objective is to maximize the reward discounted by costs.

Although dACC_Act and dACC_Boost work in parallel and have the same objective of maximizing net reward, they are responsible respectively for value-based action selection (dACC_Act part of the external loop in Fig. 1, Equation 1,2) and the modulation of LC and VTA activity (that is, the “boost”; dACC_Boost is part of the internal loop in Fig. 1a, Equation 3, 4). Catecholamines affect the dynamics of the dACC modules in real time, i.e., while the model is interacting with the environment. In particular, they modulate the amount of cognitive and physical effort (by LC module, Equation 2) [42, 43] that the model exerts to execute a task, its plasticity for knowledge updating (learning rate, by the LC module, Equation 5a), and the magnitude of reward signals (by VTA module, Equation 6a,b). In agreement with recent experimental findings, the RML architecture assumes that, while DA influences effort-based decision-making, NE is directly involved in effort regulation [44]. The RML can be integrated with other computational models, e.g., to a WM model (see Simulation 1 and Fig. 1b, c), so that the efferent signal from the LC can modulate the dynamics of other simulated brain regions and optimize their performance [29] (Fig. 1a, b, orange arrows). Importantly (Simulation 1), the cortical-subcortical system modeled by the RML is strongly involved in the management of the WM functioning in primates. Indeed, neurophysiological findings have shown that LC neuromodulatory signal critically affects WM by improving coding precision in prefrontal neurons [45, 46], while a deficit in NE transmission in prefrontal neurons leads to WM impairment [47, 48]. Moreover, neuroimaging findings demonstrated an increase of dACC activity as a function of effort in WM tasks [49, 50]. Finally, it is worth stressing that the RML implements WM optimization as a form of cognitive effort, assuming that the VTA influences WM performance indirectly, by influencing the dACC function, which, in turn, modulates LC activity. For these reasons, and because LC activation is a major biological marker of both cognitive and physical effort [51, 52], the RML architecture is well suited to simulate how VTA degeneration interacts with LC-dependent effort in both physical and cognitive tasks.

The equations regulating the RML functioning are presented and discussed in the “Materials and Methods” section above (see [25]).

Simulation of two behavioral tasks

We disrupted the DA transmission in the RML to simulate VTA degeneration during the early stages of AD progression. While manipulating DA levels, we investigated the RML performance and dynamics in two domains that are particularly sensitive to the early stages of AD progression, i.e., WM and motivation to exert effort to get a reward. We investigated these domains by administering to the RML both a visuo-spatial WM task (Fig. 2a) and a decision-making task involving different levels of effort engagement (effort, Fig. 2b), while manipulating different levels of VTA degeneration. All the parameters and the task features used in these simulations are the same as those used in Silvetti et al. [25]. Simulation results were compared with data from human patients in early stages of AD progression.

Simulation 1: Working memory impairment

WM impairment is one of the earliest signs of AD. In this simulation we show how VTA degeneration can determine WM impairment by disrupting the functioning of the dACC, LC and VTA system. To reproduce WM dysfunctions the RML model was connected to the FROST model, a recurrent neural network simulating the fronto-parietal circuits involved in visuo-spatial WM [28] (Fig. 1b, c). As we show in Fig. 3a, the LC output increased as a function of delay duration (blue plot; main effect of memory load on LC output in controls: F(2,19) = 3,91, p = 0.0286), keeping the accuracy high (Fig. 3b, blue plot) and RTs low across different working memory loads (Fig. 3c, blue plot), as it happens also in healthy humans [30].

In case of VTA degeneration, both the RML-FROST behavior and its neural dynamics are impaired. At the behavioral level, this causes poor accuracy due to the necessity of high NE levels for this task (Fig. 3b, red plot; lesion main effect: F(1,19) = 22.15, p < 0.0001). Also the RTs worsened with VTA degeneration (Fig. 3c, red plot; lesion main effect: F(1,19) = 15.76, p < 0.0001). This was due to the consequent disruption of VTA-dACC-LC interaction, with the effect, at the neural level, of downregulating LC activity (Fig. 3a, red plot; main effect of DA lesion on LC output: F(1,19) = 19.37, p < 0.0001). This happened especially for high memory loads (lesion x memory-load interaction: F(2,38) = 4.93, p = 0.0125). In Fig. 3d and 3e, we show also that the simulated behavioral results are in agreement with those from human patients with mild cognitive impairment due to AD exhibiting a significant group effect and delay effect [30].

Simulation 2: Motivational impairment in effort engagement tasks

In this task, the RML could choose between high effort-high reward and low effort-low reward options. Although this experimental paradigm takes inspiration from rodent studies [31, 32, 44], where effort levels were manipulated in terms of physical effort, here it must be interpreted as a task involving indifferently either physical or cognitive effort, because, in the RML, the central neuro-computational mechanisms basing effort-based decision-making are shared across different effort modalities. Both human and nonhuman animal studies indicate that catecholamines are crucial to energize behavior and choose effortful actions when a large reward is available [31, 32]. Simulation results from this experimental paradigm are compared with those from patients performing a verbal short-term memory task (digit span).

Simulation 2a

Figure 4 shows both behavioral and neural results from the RML, as a function of progressive VTA degeneration, and compares them with those from human volunteers executing a task involving cognitive effort (digit span). At the behavioral level, when the VTA was intact, the RML exhibited an optimal behavior, preferring to exert an effort to get a large reward (Fig. 4a, 0% VTA degeneration). With the progression of the VTA degeneration, behavioral performance remained at first stable and then, when degeneration approached 50%, it dropped quickly. The percentage of trials where the RML refused to execute the task had similar dynamics. During the early VTA degeneration, the model kept a good level of commitment, followed by a rapid emergence of apathetic behavior (many refusals to engage) when the degeneration progressed (Fig. 4a, inner plot). The same performance pattern has been found in AD patients tested with a digit span task [53] every six months: here performance remained stable during the very early stages of the disease (first two assessments, at 0 and 6 months) and became significantly poorer from the third assessment (from 12 months) (Fig. 4b). At the neural level, this behavioral pattern emerges from the MPFC-brainstem adaptation to the VTA degeneration. Figure 4c shows an increase of LC activity, peaking at about 40% of VTA lesion, and dropping with further VTA degeneration (inverted U dynamics). This LC response was due to a compensatory mechanism where, during the initial stages of VTA degeneration, the dACC_Boost compensated the drop of reward signal from the VTA by increasing the boosting signal promoting the release of both NE and DA. This compensatory mechanism can be interpreted in terms of the classic physiology concept of functional reserve. Afterwards, LC activity decreased, as the residual VTA output (conveying reward signal) was not enough to overcome the intrinsic cost of catecholamine boosting anymore. It is worth noting that with a severe VTA degeneration the LC output dropped below the baseline corresponding to an intact VTA, leading to a chronic downregulation of NE release, despite the LC was not lesioned. This result closely reproduces the LC activation (measured by pupil dilation [44, 54]) during a digit span task in human participants affected by different disease severities, showing a significant group effect of different cognitive impairments [55] (Fig. 4d).

Fig.4

Model behaviour and neural dynamics (±s.e.m.) in the execution of the Effort Task as a function of a progressive VTA lesion (left column), and comparison with human data (right column). a) At the behavioral level, the percentage of simulated optimal choices (HR) first decreased slowly and then dropped abruptly with the VTA damage progression. The inner plot shows the percentage of “Stay” choices (refusal to engage in the task): this remained low during the early stages of VTA degeneration, and increased quickly afterwards. b) Behavioral performance of AD patients at different assessment time points. This is in agreement with the simulation in panel ‘a’ (data from [53]) where performance remains good during the early disease stages and then it drops quickly. c) At the neural level, simulated LC response (arbitrary units, a.u.) followed an inverted U shape when VTA was progressively lesioned. With low VTA damage, the dACC_Boost module compensated by promoting catecholamine release. As LC was not lesioned, this resulted in an over-activation that compensated VTA degeneration and kept behavioral performance stable (a, b); with a more severe VTA lesion, the LC under-activated, leading to abrupt performance drop (a, b). d) The same pattern was found in patients with progressive cognitive impairment [55] (data averaged over the effort dimension from panel f; MCI, mild cognitive impairment). Pupil dilation is used as a proxy for LC activation, see main text). e) LC simulated activation as a function of task difficulty (number of digits), in three different simulated populations: controls, MCI, and Early-AD. f) Pupil dilation (a proxy of LC activity) as a function of task difficulty (number of digits), in human control and patient groups (data from [55]).

Influence of VTA degeneration on LC modulation

Simulations 1 and 2 show how VTA degeneration disrupts the cortical-subcortical processes regulating catecholamine release as a function of environmental demand. In particular, the model suggests that VTA degeneration alters the control of LC in a nonlinear way. In the early stages of VTA degeneration, the dACC promotes catecholamine release, resulting in a partial VTA compensation and a higher (than normal) level of NE. LC temporary compensation defers performance drop (as described above, Fig. 4a, b) and it also counteracts the tendency of refusing to engage in the task (Fig. 4a), which can be interpreted as a resilience to apathy. With the progression of the VTA degeneration, NE level undergoes an “inverted U shape” dynamics leading to an initial increase of the LC activation followed by a chronic depression (Fig. 4c). The RML model allows an interpretation of this phenomenon as due to the continuous updating of the optimal point in the tradeoff between catecholamine upregulation (directed to preserve behavioral performance) and catecholamine downregulation (directed to save the cost of catecholamine release); however, in humans also the progressive degeneration of the LC itself may play a role, as proved in patients and AD mouse models [56–58]. The initial upregulation of catecholamines was found in patients at initial stages of AD, when the neuropsychological deficit emerges from clinical testing but is still not evident in everyday life, as patients can maintain a normal performance at the cost of a higher effort [55]. Moreover, the “inverted U shape” LC compensation, together with the progressive LC neurodegeneration, could explain why experimental data show either an increase or a decrease of NE levels in AD patients [59]. Indeed, based on our simulations, the LC output could augment or decrease depending on the disease progression [60]. These results are also in line with recent evidence from middle-aged cognitive normal adults indicating that task-evoked pupillary responses, revealing an LC-noradrenaline involvement, could be associated with AD polygenic risk scores (AD-PRSs) providing information for early screening of genetically at-risk individuals. In particular, in a digit span task, individuals with normal cognitive functions but high AD-PRSs exhibited a greater pupil dilation (effort) in a high (9-digit) cognitive load condition [61]. Figure 5 summarizes these processes [60].

Fig.5

Schema summarizing the interaction between dACC modules and catecholamine modules in the RML, during the progressive degeneration of VTA. Coronal brain sections indicating the possible presence of diffusive degeneration. a) In normal conditions, dACC modules make decisions and optimize catecholamines release to maximize performance (i.e., reward). b) In case of mild VTA degeneration, the dACC can compensate the VTA loss (DA: dopamine), by upregulating the boosting signal (functional reserve, ticker blue downward arrows). This compensates only partially the VTA degeneration (due to neural cells loss), but increases strongly the LC (NE: norepinephrine) output as the LC is not impaired, at least during early stages of the disease. This keeps behavioural performance close to normal, at least for easy tasks, but with a higher recruitment of catecholamines. c) When the VTA degeneration is severe the cost for compensation is too high, and therefore the dACC decreases the boosting signal (thinner downward arrows), causing a downregulation of both LC and an even stronger VTA output decrease. This causes the behavioural deficits described in Figs. 2 and 3, including the WM impairment due to the loss of LC function. d) Chronic catecholamines down-regulation leads to severe cognitive impairments and possibly to a lower clearance of Aβ peptide in LC target areas, thus contributing to the disease progression toward the plaque stage.

The simulations also show how the downregulation of both DA and NE following severe VTA degeneration causes the emergence of apathy, with a decreased willingness to perform a task (Fig. 4a). The model explains the emergence of apathy as related not only to the loss of the VTA function but also to the disruption of catecholamine control including NE and due to a primary VTA deficit (Fig. 5c, d).

The progressive catecholaminergic dysfunction associated with the degeneration of VTA and LC, and the resulting behavioral effects, agree with the experimental observations in rodent models [62–64]. Here experimental lesions in VTA produce altered locomotor activity, deficits in passive avoidance learning, and aggressive behavior, with all these effects often related to the size of the VTA lesion [62–64]. However, it should be considered that the behavioral alterations resulting from a focused damage of the VTA might differ from those resulting from a progressive degeneration involving a subset of dopaminergic VTA neurons as in AD. Moreover, a progressive degeneration of VTA neurons could gradually spread to other vulnerable brain areas such as LC, as demonstrated in the Tg2576 AD mouse model [20, 56].

Chronic LC downregulation as a risk factor of neurodegeneration

The simulations described above show how VTA degeneration might affect the modulation of catecholamine nuclei by the midfrontal cortex, up to a chronic downregulation of both DA and NE. In the long term, catecholamine downregulation might trigger a cascade of events leading to a stronger manifestation of the disease features. In particular, chronic depression of LC neurons activity contributes to producing abnormal (hyperphosphorylated) tau proteins in LC neurons [18] themselves, which can aggregate into neurofibrillary tangles [14, 17] and lead to permanent LC degeneration. The persistent lower NE produced by the loss of LC neurons has been associated with a lower clearance of Aβ peptide in LC target areas [65, 66], including amygdala (Amg) [67], hippocampus (Hip) [68], and PFC [69, 70]. This leads to an accumulation of Aβ plaques, leading to the AD-typical diffuse cortical atrophy [71] and a worsening of clinical symptoms (Fig. 5d). Aside from contributing to tauopathy and Aβ plaque formation, the decreased level of DA and NE might directly cause dysfunctions in the VTA and LC target areas [72–74], like Amg, Hip, and PFC, which are critically involved in AD development.

In the long run, the chronic malfunctioning of the VTA-LC system might also affect the functioning of NbM which receives dopaminergic input from VTA and noradrenergic input from LC, and provides the principal source of acetylcholine for PFC, Amg, and Hip [75–77]. At the early stages of pathology, most NbM neurons are unaffected by tauopathy and many of the affected neurons remain alive [78]. In this case, the AD symptoms might be mainly due to the processes discussed above, related to the effects of the VTA-LC malfunctioning on the cortical and limbic areas where they project. The prolonged malfunctioning of the VTA-LC system might increase the vulnerability to tauopathy and neurofibrillary degeneration of the cholinergic component of the NbM. Once developed, the tauopathy in these components might lead to the degeneration of the cholinergic axons, leading to the worsening of symptoms [79].

This discussion about the long-term effects of the catecholamine-frontal dysfunctions on a wider cortical-subcortical network is important for two reasons. Firstly, it provides a system-level perspective linking early catecholamine dysfunction to the typical histological findings in AD, together with the mechanistic description of cognitive impairment progression. This system-level perspective could be critical to develop better diagnostic approaches leading to address the challenges in AD research related to the lack of effective preventive strategies and reliable biomarkers for early diagnosis of the disease [80–82]. Secondly, it is critical to devise new therapies. AD is currently incurable and existing treatments produce only a modest and transient amelioration of symptoms [83]. To date, acetylcholinesterase inhibitors and memantine are the only drugs approved for its management. These drugs provide a symptomatic improvement but have a minor impact on the disease progression. The extensive investigation of the molecular and cellular pathogenesis in AD over the past few decades has provided significant progress in the understanding of the disease, stimulating the development of a number of novel strategies that seek to modify the disease progression. The major developments in this direction are the amyloid and tau-based therapeutics [84]. Aside from these approaches, the increased understanding of the mechanisms underlying AD pathogenesis and progression could also lead to the development of complementary therapeutic strategies. In this line, the results obtained in this article suggest that therapies involving also drugs enhancing DA transmission in early stages of the disease could provide better results with respect, for example, to drugs acting on norepinephrine alone that have been shown to give mixed results [17]. This proposal agrees with data supporting a role of dopaminergic drugs in AD therapy [85]. In particular, electrophysiological studies performed on AD patients showed positive effects of dopaminergic drugs on cortical synaptic plasticity mechanisms and on cognitive performances, suggesting the use of these drugs in the treatment of the disease [86–88].

In addition, the results of the simulations run with the model also suggest the investigation of the possible benefits of employing psychological interventions alongside drug-based treatments. In this respect, we have seen that the model introduces a subjective factor weighting the objective relevance of both rewards and costs received by the patient in daily life activities. This suggests the possibility of psychological interventions directed to lower the perceived subjective estimation of costs and increase the subjective estimation of gains to increase the probability of decisions in favor of coping and engagement. In turn this would lead to harvest the physiological benefits discussed above produced by enhanced catecholamine release. This view agrees with recent proposals to employ psychological interventions to treat depression and apathy symptoms as a way to contribute to prevent and treat AD [89, 90]

The hypothesis proposed in this paper agrees with the system-level perspective to the study of the brain according to which different behaviors are produced by the interplay of different parts of the brain rather than by specific areas working in isolation [91–94]. This view is in line with recent works proposing to study AD as a system-level disorder where various protective, age-related, and disease-promoting factors interact to produce the core mechanisms underlying the disease [80, 95, 96]. Recently, we have shown how this system-level approach could be critically used to explain in new ways the origin of different cortical-cerebellar-basal ganglia dysfunctions related to Tourette’s syndrome [97], ADHD [98], and Parkinson’s disease [99]. Similarly, in this article we have argued that a system-level approach could lead to view the AD early pathogenesis and development under a different perspective.

Concluding remarks

Both behavioral and neural data simulated by the model were consistent with recent experimental observations in patients and, thanks to the anatomo-functional plausibility of the RML, it was possible to formulate specific hypotheses on the neural mechanisms linking the VTA impairment to the disease progression and to the altered behavior. In this way it was possible to associate the neural-scale to the behavioral-scale features making the model suitable for: 1) quantify system-level dynamics through computer simulations, and bind them to the neural processes underlying AD pathogenesis at pre-plaques stage; 2) contribute to formulate a novel, integrated, view of the pathogenesis of AD, which might help to devise new computer simulation-based early diagnostics, for example by means of computational phenotyping techniques [100], and then new therapeutic interventions targeting the very early stages of the disease.