The Wong-Wang model (Deco et al., 2014) implemented as highly optimized C code (Schirner et al., 2021) was chosen to simulate whole-brain dynamics. This dynamic mean field model simulates the local regional neuronal activity as the result of a network composed of interconnected excitatory and inhibitory neurons coupled through NMDA and GABA receptor types. Details of the Wong-Wang model can be found in Deco et al. (2014). Briefly, brain dynamics are described by the following set of coupled non-linear stochastic differential equations:

where r_i^(E,I) denotes the firing rate of the excitatory (E) and inhibitory (I) populations, S_i^(E,I) identifies the average excitatory or inhibitory synaptic gating variables at local area, i, and I_i^(E,I) is the input current to the excitatory and inhibitory populations at local area, i. All parameters described in Supplementary Table 1 were set as in Deco et al. (2014), except those that are tuned during parameters optimization, as follows: parameter space exploration was performed for global coupling (G), which is a scaling factor denoting long-range coupling strength, and local parameters defining the strength of inhibitory (GABA) synapses (J_i), the strength of excitatory (NMDA) synapses (J_NMDA) and the strength of local excitatory recurrence (w₊). Thus, this model retains information on both global brain dynamics and local excitatory/inhibitory balance and is particularly interesting for the investigation of pathological conditions.

For each set of parameters combination, resting-state BOLD fMRI time-courses were simulated over 6 min length using a Balloon-Windkessel hemodynamic neurovascular coupling model (Friston et al., 2000) and the simFC was computed as described for the expFC (see section “Structural and functional connectivity”). Parameters were adjusted iteratively until the best fit, i.e., the highest correlation, between expFC and simFC was achieved (Supplementary Figure 1).

To investigate the impact of the cerebellum on brain dynamics generation, simulations were performed using three different combinations of connections and nodes (Figure 2):

For each of these three networks predictive power was evaluated as the mean PCC between expFC and simFC matrices in different clinical conditions (Healthy Controls, Alzheimer’s Disease, Frontotemporal Dementia, Amyotrophic Lateral Sclerosis).

To assess whether biophysical parameters derived from TVB differ according to the clinical condition, optimal model parameters were tested for normality (Shapiro-Wilk test) and differences between groups were assessed with non-parametric tests (Kruskal-Wallis across all groups and Mann-Whitney between each pair of groups) when they did not present a Gaussian distribution.

A multiple regression analysis was performed to investigate the relationship between individual scores of the 5 cognitive domains (attention, memory, language, logical-executive functions, visuospatial skills) and the optimal model parameters. Neuropsychological scores in each cognitive domain were considered as dependent variables while model parameters combined with age, gender, and group category were used as predictors in a backward approach. The regression algorithm automatically removed one or more predictors to identify which of them significantly (p < 0.05) explained neuropsychological scores variance in all subjects together.

Moreover, to assess the relevance of these parameters in discriminating between normal and pathological conditions, a discriminant analysis was performed using the group as the dependent variable and considering as independent variables, in turns: (i) model parameters alone, (ii) neuropsychological scores alone, and (iii) a combination of both. To visualize and assess the sensitivity and specificity of the best discriminative variables, receiving operating characteristics (ROC) curves and corresponding areas under the curve (AUC) were calculated.

Finally, a k-mean cluster analysis was performed to reconstruct subjects-specific excitation/inhibition profiles. The number of clusters was an input parameter, arbitrarily set equal to 4, matching the number of variables considered (i.e., the 4 model parameters). A frequency analysis of the clusters in terms of clinical groups enabled us to reach a deeper understanding of their respective excitatory/inhibitory balance profiles. This percentage of subjects belonging to one of the clusters in each clinical group was used to qualitatively describe the occurrence of specific patterns of excitatory/inhibitory balance in the population.

PCC obtained with the three networks were normally distributed (Shapiro-Wilk test), thus parametric tests were used to compare them between different conditions. First, to assess TVB predictive power for each clinical condition, a one-way ANOVA was performed between the PCC of each network across groups (Healthy Controls, Alzheimer’s Disease, Frontotemporal Dementia, Amyotrophic Lateral Sclerosis). Then, to assess the impact of each specific network on TVB predictive power, a multivariate general linear model (GLM) with Bonferroni correction was chosen to compare PCC values of the three networks within each group.

The biophysical parameters derived from TVB were compared between groups to assess whether, at group level, their value could differ according to the clinical condition. Both global and local biophysical parameters showed significant differences between groups (Table 2 and Figure 3A): Alzheimer’s Disease patients showed higher G and J_i compared to Healthy Controls and Frontotemporal Dementia (p < 0.05); Amyotrophic Lateral Sclerosis patients showed higher J_NMDA than Healthy Controls (p < 0.05); no statistically significant differences were found when comparing Healthy Controls and Frontotemporal Dementia with other groups.

Parameters used in backward regressions, significantly explained the variation of scores in different neuropsychological domains. The explained variance of each neuropsychological domain was progressively reduced by simplifying the regression model through the removal of one or more predictors and ranged from ∼20 to ∼8%. For each cognitive domain, a different combination of features was necessary to significantly (p < 0.05) explain a percentage of the variance (Table 3).

The discriminative power of TVB parameters and neuropsychological scores is reported for all comparisons (Healthy Controls vs. Alzheimer’s Disease, Healthy Controls vs. Frontotemporal Dementia, Healthy Controls vs. Amyotrophic Lateral Sclerosis, Alzheimer’s Disease vs. Frontotemporal Dementia, Alzheimer’s Disease vs. Amyotrophic Lateral Sclerosis, Frontotemporal Dementia vs. Amyotrophic Lateral Sclerosis) in Table 4. TVB parameters always yielded a poorer discriminant power (about 70%) than that offered by neuropsychological scores (about 90%). When neuropsychological scores were complemented by TVB values as joint independent variables, the discriminative power increased in all classifications reaching 100% when distinguishing between Alzheimer’s Disease and Healthy Controls, and between Frontotemporal Dementia and Amyotrophic Lateral Sclerosis. To visualize all these results, ROC curves are reported in Figure 4.

Each of the four clusters identified with the k-means analysis was characterized by a different combination of values of TVB-derived biophysical parameters, as reported in Figure 5A and Supplementary Figure 2. Considering the biophysical meaning of each parameter derived from the simulation, we can describe the k-means clusters as follows:

All groups were represented in more than one clusters, as shown in Figure 5B. The distribution of subjects belonging to different condition across clusters revealed that: cluster 1 was more frequent in Healthy Controls (20%) and Amyotrophic Lateral Sclerosis (26.7%) than in Alzheimer’s Disease and Frontotemporal Dementia (both 6.7%); cluster 2 was less present in Healthy Controls (13.3%) than in pathological conditions (Alzheimer’s Disease = 33.3%, Frontotemporal Dementia = 33.3%, and Amyotrophic Lateral Sclerosis = 20%); cluster 3 was mostly composed of Healthy Controls subjects (46.7%), somewhat frequent in Frontotemporal Dementia (33.3%) and hardly present in Amyotrophic Lateral Sclerosis (6.7%) and Alzheimer’s Disease subjects (0%); cluster 4 was the most frequent in Alzheimer’s Disease (60%) and Amyotrophic Lateral Sclerosis (46.7%) and less frequent in Healthy Controls (20%) and Frontotemporal Dementia (26.7%).

TVB simulation performed both in Healthy Controls and pathological conditions led to good fit values between the expFC and simFC (Table 2). No differences were found between PCC values of each network across clinical groups, but significant differences were found in each group comparing PCCs of the three networks. For all the pathological groups, PCC values obtained with the embedded cerebro-cerebellar subnetwork were significantly higher (p < 0.01) than those obtained with the whole-brain network, while PCCs of the cortical subnetwork were significantly higher (p < 0.003) than those of the whole-brain network in Alzheimer’s Disease and Frontotemporal Dementia (Table 2 and Figure 3B). No differences were found between the three networks in Healthy Controls.

Parameters derived from TVB simulations using the Wong-Wang model yield information on global brain dynamics and on the local excitatory/inhibitory balance. In particular, the global scaling factor G denotes the strength of long-range connections, and higher global coupling means a greater weighting of the global over the local connectivity. The remaining three parameters define the balance between excitation and inhibition in the simulated network: J_NMDA represents the strength of excitatory synapses in the network, J_i denotes the strength of inhibitory synapses and w₊ the strength of recurrent excitation.

Our results revealed that different clinical groups were characterized by specific TVB parameters providing new clues for the interpretation of the dysfunctional mechanisms in local microcircuits.

To date, patterns of altered FC in Alzheimer’s Disease from a resting-state networks perspective have been reported mainly in the default mode network (DMN) (Hohenfeld et al., 2018; Ibrahim et al., 2021), although a wider involvement has been suggested by our group in previous work (Castellazzi et al., 2014). Our data demonstrated that G and J_i were higher in Alzheimer’s Disease patients compared to Healthy Controls suggesting that Alzheimer’s Disease subjects were characterized by increased global coupling and overinhibition. This increased G-value in our Alzheimer’s Disease group could be interpreted as a compensatory mechanism counteracting altered cerebral connectivity, but it might also underly the hypersynchrony typically characterizing disrupted networks in patients (Castellazzi et al., 2014; Maestú et al., 2019). Furthermore, our data shows an increased inhibition suggesting that GABAergic dysfunction plays a role in Alzheimer’s Disease pathology, which is in line with the novel hypothesis that GABAergic remodeling might be an important feature of neurodegeneration (Bi et al., 2020).

It is worth noting that Alzheimer’s Disease showed higher G and J_i also compared to Frontotemporal Dementia patients, strengthening the tenet that the pathophysiological mechanisms underlying the two diseases are different. Indeed, Frontotemporal Dementia showed G and J_i values similar to Healthy Controls, consistent with the similarity of cortical neural synchronization in Frontotemporal Dementia and Healthy Controls (Nardone et al., 2018).

Finally, Amyotrophic Lateral Sclerosis patients were characterized by an increased J_NMDA, which is in line with the cortical hyperexcitability frequently reported in this pathological condition (Pradhan and Bellingham, 2021).

TVB-derived biophysical parameters combined with age, gender, and group category significantly explained the variance of neuropsychological scores both in Healthy Controls and pathological groups. This suggests that the levels of excitation, inhibition and global coupling are associated with cognitive performance in the different clinical conditions.

The clinical relevance of TVB parameters is further highlighted by results of the discriminant analysis. The discriminative power of neuropsychological tests alone was always higher than the one obtained with TVB parameters alone. However, when neuropsychological measures were combined with TVB parameters, the discriminative power improved, reaching in some cases 100% of accuracy. Importantly the performance of our classification was satisfactory not only to distinguish Healthy Controls from patients, but also to differentiate patients belonging to different neurodegenerative conditions. This opens an interesting perspective for the development of new diagnostic tools combining TVB parameters with neuropsychological scores in future machine learning approaches.

For each subject, TVB predicted the optimal parameters, providing a subject-specific description of the excitatory/inhibitory balance that could be analyzed at group level (as discussed above) or used to establish cluster membership in a data-driven approach.

While at group level there were noticeable differences in excitatory/inhibitory parameters, when considering single-subjects’ profiles, K-means clusters were not group specific; this suggests an heterogeneous excitatory and inhibitory balance across subjects that could be exploited for future personalized interventions. Considering the biophysical meaning of TVB parameters, cluster 1 was mainly characterized by overexcitation and was more frequent in Healthy Controls and Amyotrophic Lateral Sclerosis. As well as being consistent with the fact that hyperexcitability is a well-known feature of Amyotrophic Lateral Sclerosis patients (Brunet et al., 2020; Pradhan and Bellingham, 2021), the presence of some Healthy Controls subjects in this cluster is not surprising. Indeed, the effect of aging on the glutamatergic system is currently under investigation (Segovia et al., 2001), and even if glutamate is mostly reported to decrease with age (Roalf et al., 2020), age-related effects on glutamatergic release and uptake processes and NMDA receptors activation could be differentially modulated in some healthy subjects. Even in cluster 2 we found some Healthy Controls, which presented not only high excitation but also high global coupling strength. This high G-value could be due to an increased strength of long-range connectivity or increased synchrony between nodes. It is worth noting that cluster 2 was more common in pathological conditions than in Healthy Controls, and this is in line with the frequent observation of hyperexcitation and hypersynchrony in Alzheimer’s Disease (Henstridge et al., 2019; Maestú et al., 2019, Maestú et al., 2021), Frontotemporal Dementia (Hughes et al., 2018; Benussi et al., 2019), and Amyotrophic Lateral Sclerosis (Dukic et al., 2019; Pradhan and Bellingham, 2021). Cluster 3 was the most related to Healthy Controls, while the number of patients was marginal (0 for the Alzheimer’s Disease group). This cluster is mainly characterized by high recurrent excitation; interestingly, alterations of this property are less explored as potential mechanisms in clinical conditions. Only a few network models have been developed to account for the influence of recurrent excitation and to explore its changes in pathologies. Strong self-excitation was shown to be required in network models to achieve satisfactory simulations of decision making and working memory tasks (Wong and Wang, 2006), in line with our evidence of high recurrent excitation in Healthy Controls. Moreover, in a previously proposed computational model applied to Alzheimer’s Disease the variation of local recurrent excitation has been suggested as a brain mechanism employed to compensate for alterations induced by other types of synapse loss (Bachmann et al., 2020). Unfortunately, nothing is known about recurrent excitation in network models of Frontotemporal Dementia or Amyotrophic Lateral Sclerosis, but the presence of Frontotemporal Dementia and Amyotrophic Lateral Sclerosis patients in cluster 3 prompts to explore the effect of this parameter in clinical conditions other than Alzheimer’s Disease. Finally, cluster 4 was mostly associated with Alzheimer’s Disease patients, followed by Amyotrophic Lateral Sclerosis, Frontotemporal Dementia and Healthy Controls. While convergent findings are increasingly supporting the role of a GABA function increase in Alzheimer’s Disease (Li et al., 2016; Govindpani et al., 2017; Bi et al., 2020), GABAergic dysfunction is mostly described as an overall decrease of cortical inhibition in Frontotemporal Dementia and Amyotrophic Lateral Sclerosis. Our results suggest the possibility of an increased GABAergic activity not only in Alzheimer’s Disease patients, as already observed in literature, but also in subsets of subjects affected by other neurodegenerative conditions.

It is important to point out that Frontotemporal Dementia patients appeared to be the most distributed between clusters, without a main cluster membership, and this evidence reflects the heterogeneity of our Frontotemporal Dementia cohort, which is in line with the wide spectrum of neurotransmitters deficits which has been already observed in Frontotemporal Dementia (Murley and Rowe, 2018).