Dynamics-informed priors (DIP) for neural mass modelling Authors: Alessia Caccamo, Dominic M. Dunstan, Mark P. Richardson, Alexander D. Shaw, Marc Goodfellow Published: 2026 ABSTRACT AbstractNeural mass models (NMMs) are important mathematical tools for inferring hidden neural mechanisms that generate healthy and pathological brain activities. A critical step in the inference process is parameter estimation, which calibrates NMMs based on measured neuroimaging data. While parameter estimation can be conducted via various approaches, one of the most influential methods is dynamic causal modelling (DCM). DCM adopts a Bayesian inference approach that relies on, and is sensitive to, the specification of prior parameter distributions reflecting a priori hypotheses about the causes of data. However, most parameters of NMMs encode neuronal properties that are not directly measurable. For this reason, in the absence of sufficient empirical data and well-founded prior beliefs, inference becomes increasingly susceptible to bias. Therefore, it was imperative to establish a comprehensive strategy for mapping model parameters to data. This study proposes a computational extension of DCM, named DCM with dynamics-informed priors (DIP-DCM), which adopts a genetic algorithm (GA) to map parameter values to model dynamics. Optimal sub-regions of the parameter space were subsequently selected and translated into groups of parameter priors for DCM. DIP-DCM was compared with the “standard” DCM inference and to the standalone GA, using two independent neuroimaging datasets. Results indicated that DIP-DCM models were the best predictors of data and captured key mechanistic signatures of psychiatric disease and pharmacological interventions. Overall, DIP-DCM handled local minima and explored diverse parameter regimes following trajectories informed directly by model dynamics and data. This study suggests that DIP-DCM is an advantageous route to parameter estimation when information is limited, enabling a data-driven derivation of parameter priors in exploratory studies, across different biological contexts and datasets. Introduction Large-scale cortical activity is widely recognised to emerge from the functional integration of distinct and spatio-temporally distributed neural events (Schomer & da Silva, 2017). While the causal relationships between these events remain inherently hidden and inaccessible to direct empirical observation, their collective effects manifest as signals measured via neuroimaging methods, such as electroencephalography (EEG), magnetoencephalography (MEG), and functional magnetic resonance imaging. Tools from dynamical systems theory, statistical inference, and computational analysis have become invaluable for mapping hidden neuronal processes and interactions to observable signals. Specifically, neural mass models (NMMs), with their long history dating back to 1956 (Beurle, 1956; Freeman, 1975; Lopes da Silva et al., 1974; Mountcastle, 1957), are often utilised to simulate cortical dynamics and to study emergent neural population phenomena. NMMs are abstract representations of local brain circuits and treat neural populations as single entities, or point processes. Therefore, neural population activity is described by average membrane potential and firing rate variables, and the population parameters reflect lumped neural properties, including synaptic strength, response time scales and rates of activity propagation. These models enable causal inference, the process that identifies the causes—encoded by the respective biophysical parameters—that underlie the observed empirical data and its modulation across different states of healthy functioning or neurological disease (David & Friston, 2003; Glomb et al., 2022). However, the problem of reconstructing causes from observed effects, that is, the inverse problem, does not generally offer a unique solution: multiple parameter regimes, under the same model, can generate the same data (Dunstan et al., 2023; Hartoyo et al., 2019; Hashemi et al., 2025). This behaviour is often referred to as parameter “non-identifiability” (Chis et al., 2016; White et al., 2016) or “degeneracy”, and it is an important ubiquitous property of real biological systems (Edelman & Gally, 2001; Marder & Goaillard, 2006; Prinz et al., 2004). Degeneracy refers to the presence of different mechanistic pathways for the same biological function and it is an inevitable consequence of natural selection. Notably, this “many-to-one” relationship contributes to the resilience and stability of biological functions, yet, it becomes problematic in the context of mechanistic modelling. Here, degeneracy can be influenced by several aspects including the model structure, the data used in the model fitting process, the number of free parameters and the parameter estimation strategy. The latter is of notable importance as parameter inference can be conducted using different techniques, and the chosen methodology is likely to affect the mapping between parameter values and observed data. Among the existing methods, dynamic causal modelling (DCM) (Friston et al., 2003) provides a well-established and principled framework for parameter inference in neuroimaging. Within DCM, model parameters are estimated using a computationally efficient variational Bayesian routine, known as Variational Laplace (VL) (Friston et al., 2003; Parr et al. T. & Friston, 2022; Zeidman et al., 2023). Data are modelled using a generative framework that defines prior expectations, or beliefs, about their underlying causes, along with a likelihood function that describes the probabilistic relationship between data and causes (Friston et al., 2003; Lee & Vanpaemel, 2018). Through a process of belief updating, parameter priors—the probability distributions assigned to parameters a priori—are updated into posterior distributions, used to derive the most probable explanations for the observed data. Therefore, parameter priors are a crucial requisite for solving the DCM inverse problem. However, as the inverse problem can display a “many-to-one” relationship between hidden causes and their manifestation, the DCM VL is predisposed to settle on one solution among many options, as pre-determined by the chosen parameter priors (Morita et al., 2010). Identifying the most suitable priors, however, can be challenging. For instance, in the context of NMMs, parameters are abstract representations of real neural population properties and do not directly correspond to measurable quantities (Coombes, 2023; Deschle et al., 2021; Dunstan et al., 2023; Ferrat et al., 2018; Goodfellow et al., 2022). Similarly, for novel models or datasets, it can be difficult to quantify and establish reasonable and objective beliefs about the model parameters. While these issues can be mitigated by conducting systematic explorations of the prior parameter space, these can become inefficient when parameter spaces are vast, and ineffective when the explored priors do not reflect the underlying generative process. Previous endeavours adopted alternative optimisation techniques for DCM, including a gradient-free Markov Chain Monte Carlo (MCMC) sampler and a Gaussian process optimisation (GPO), which used GPO-identified solutions to initialise a local gradient descent (Lomakina et al., 2015; Sengupta et al., 2015). Specifically, the GPO method improved parameter estimation accuracy compared with traditional variational Bayes, and enhanced computational efficiency compared with MCMC. However, this method was limited to providing information on the mode of the posterior distribution rather than approximating a posterior density. In addition, its application was restricted to haemodynamic equations and required sufficiently smooth and convex objective functions, that is, the mathematical expressions that guide parameter estimation and quantify model performance. Thus, the GPO approach was considered less suitable for models of electrophysiological data (Lomakina et al., 2015), since they are more likely to display non-convexity and hence to be characterised by multiple local minima. Local minima are suboptimal solutions, or solutions that optimise the objective function strictly relative to nearby points. They are commonly encountered in the context of non-smooth objective functions, which are highly sensitive to small parameter perturbations. This sensitivity introduces irregularities, or “fractures”, in the parameter landscape, complicating the identification of global optima. Therefore, there is a need for novel strategies that guarantee robust parameter estimation, not only by ruling out poorly informed priors, but also by mitigating challenges posed by such complex relationships between objective functions and parameter values. To mitigate these limitations, the present study proposes a novel method which derives parameter priors directly from data by leveraging heuristic decision processes to shape and constrain subsequent DCM inversions. This method, termed DCM with dynamics-informed priors (DIP-DCM), adopted genetic algorithms (GAs) to identify multiple parameter configurations to be used as priors within the generative model space. These were named dynamics-informed priors (DIP), as they were shaped by, and learned from, the predicted observations generated by the model under different parameters. Notably, GAs are well-established global optimisation techniques employing the Darwinian principles of natural selection and evolution (Holland, 1992), thus offering a compelling conceptual link between evolutionary selection and selection of generative models. Moreover, GAs were previously found to be useful for parameter estimation in NMMs (Cona et al., 2011). This paper defines the methodological foundations of DIP-DCM and offers two distinct applications through the use of published M/EEG datasets (Biondi et al., 2022; Shaw et al., 2020). Ultimately, this approach can be utilised as an alternative strategy for choosing parameter priors for dynamic causal models, thereby broadening the applicability of DCM to a wider range of exploratory studies. This implementation is compatible with the DCM SPM12 environment and its use is suggested for DCM studies lacking strong prior hypotheses or theoretical assumptions about model parameters, or studies that would benefit from exploring broader parameter regimes at reasonable computational costs. Methods Datasets The neuroimaging datasets were from two independent studies. Study 1, Biondi et al. (2022): EEG data were recorded from 14 healthy participants using 64 electrodes. Recordings were performed in resting eyes-closed conditions, at baseline and following the administration of two anti-seizure medications (ASMs), lamotrigine (LTG) and levetiracetam (LEV), or placebo (PL). Recordings were conducted across 3 days of testing, with drug washout periods of ∼1–2 weeks, resulting in six sub-datasets, or experimental conditions, encompassing all participants (pre- and post-LTG, LEV, and PL). The EEG time series were segmented into 2-second epochs, with the number of epochs ranging from 6 to 72, depending on the participant. The time series were re-referenced to the common average and power spectral densities (PSDs) were computed for each epoch via Fast Fourier Transform for frequency bins from 2 to 45 Hz with 0.5 Hz resolution, as indicated in Biondi et al. (2022). PSDs were normalised to unit area, and averaged across epochs, channels, and subjects. The grand average PSDs were log-transformed, which enhanced the key differences in spectral power between pre-treatment and post-treatment (see Results 3.2). These log-transformed data were used in the modelling process (Supplementary 1.2). Study 2, Shaw et al. (2020): Two hundred and seventy-five channel cryogenic MEG were recorded from patients with schizophrenia and healthy controls during a visual task known to induce gamma oscillations. Data were obtained using the LCMV beamformer on a 4 mm grid. A virtual electrode was derived from the voxel in the visual cortex showing the biggest amplitude change in the gamma range between stimulus “on” and “off” conditions. Data included a total of 5498 and 6268 4-second epochs for patients and controls, respectively. PSDs were computed using Welch’s method (MATLAB pwelch) for frequency bins from 2 to 85 Hz with a Hamming window of 512 points and 50% overlap. PSDs were normalised to unit area, averaged across all epochs and log-transformed. Model fitting was performed on these log-transformed data. Neural mass model This study aimed to introduce a novel strategy for selecting parameter priors within the DCM framework as implemented in SPM12. To achieve this, a default SPM12 model was chosen. This was the “LFP” NMM (spm_fx_lfp function), which was considered appropriate for both datasets (Methods 2.1). Indeed, as reported in the published literature, the “LFP” model can generate a large repertoire of dynamics including high-frequency oscillations—compared with the traditional Jansen and Rit NMM from which it was derived (Jansen & Rit, 1995)—and it is suitable for modelling drug effects (Moran et al., 2007, 2013). Briefly, it is a mathematical model of a local cortical circuit comprising three neural populations, or masses, interconnected via synaptic interactions. As usual for NMMs, the activity of each mass is simplified to average post-synaptic potentials and firing rates. The post-synaptic potential produced in response to pre-synaptic firing is modelled by a second-order differential equation, which conveys a linear representation of synapses (“pulse-to-wave” or rate-to-potential conversion). The average firing rates are calculated from the post-synaptic potential as the output of a nonlinear “wave-to-pulse” sigmoid function. The model is described in depth in Moran et al. (2007) and in Supplementary 1, including details on parameters, their values, and biological meanings, and a schematic diagram (Figs. S1,S2; Tables S1–S4). Parameter estimation For each dataset and related experimental conditions, parameter estimation was conducted via three distinct approaches. (1) DCM, using default SPM12 settings. (2) A GA, which was adapted from Dunstan et al. (2023) and revised for SPM compatibility and for facilitating its integration with DCM models. (3) A novel hybrid combination of GA and DCM, named DCM with dynamics-informed priors (DIP-DCM). DIP-DCM will be the principal focus of this foundational paper. As is customary in DCM, generative models were specified as an observation function and parameter priors. The observation function, which remained unchanged across all three approaches, reflected the log-transformed PSD of the model, which was determined by the transfer function of the system under simplifying assumptions (Supplementary 1). The priors were the “standard” model priors, as defined in Moran et al. (2007). Parameters were log-scaled to ensure positivity (Friston et al., 2019) and specified as mean and variance of a Gaussian distribution. Parameter estimation was conducted via variational inversion, that is, variational Laplace (VL), by maximising the negative (variational) free energy (Zeidman et al., 2023). This was defined as the difference between model complexity and accuracy (Friston et al., 2006, Supplementary 2). However, it is important to note that while DCM studies typically give importance to effective connectivity parameters and use condition, or group, comparison matrices (typically denoted as A and B matrices), these are not of importance here because the model represents a single cortical column (Supplementary 1.1, Fig. S1). This reflects a deliberate simplification: by restricting the model to a single node, this work focuses on investigating the performance of different parameter estimation strategies, before extending to multi-region architectures. The GA framework utilised the NSGA-II variant from the MATLAB gamultiobj function (Deb, 2001). Differently from DCM, it searched within closed parameter intervals, spanning biologically realistic values (Supplementary 1.1), rather than adopting parameterised probability densities. Details of the GA are provided in Supplementary 2 alongside differences in GA implementation between the present study and Dunstan et al. (2023). Briefly, parameter optimisation was conducted by adopting selection, crossover, and mutation operators that allow the parameter values to evolve across generations. Parameters were optimised by minimising two objective functions, J1 and J2, tailored to the observation of two peaks in the (log) PSDs. It should be noted that this genetic algorithm can provide point estimates of the parameters but does not define priors. Therefore, free energy cannot be evaluated. In this context, a meaningful metric is the residual difference between model predictions and observed data, quantified as the root mean square error (RMSE) between the power spectrum of the model and of the data for different frequencies: J1=∑ω∈Ω1(log(y(ω))−y^(ω))2, J2=∑ω∈Ω2(log(y(ω))−y^(ω))2, where y^ is the log PSD of the model (generated as per Supplementary 1.2–3), log(y) is the log PSD of the data, ω is the frequency, and Ω1, Ω2 are two different frequency ranges. Given that the two studies yielded qualitatively different spectra, the chosen frequencies were Ω1=[6.5,12.5] Hz, Ω2=[11.5,30] Hz for Study 1 (Biondi et al., 2022) and Ω1=[10.7,65.8] Hz, Ω2=[35.6,85] Hz for Study 2 (Shaw et al., 2020). These objectives were informed by empirical observations, ensuring that the alpha and beta bands (in Study 1 and Study 2) and the gamma bands (in Study 2) were well represented. Specifically, the two studies sought to identify the mechanisms underlying changes in beta and gamma features across experimental conditions, making it essential that these features were reproduced by the model. For this reason, the frequency bands defining the two objectives were selected to overlap within the beta and gamma bands for Study 1 and Study 2, respectively. Other frequencies were considered less important for these studies and were thus excluded from the objectives. In doing so, the algorithm was prevented from selecting dynamics that primarily fit components deemed less relevant. In this approach, contributions of other frequencies emerge naturally via the evolutionary processes of the algorithm, bypassing the need for frequency-specific weights, which may not always be generalisable or easy to define a priori. All these considerations were important to generate appropriate priors for the DCM VL (see DIP-DCM below). However, it is also possible for users to define custom objective functions tailored to specific domain requirements. The GA yielded sets of parameters, analogous to maximum likelihood estimates, which were used to generate a Pareto front of solutions, that is, a set of points in the 2D objective space where no solution can be improved without worsening the other. From the Pareto front one solution was selected, that is, the solution with smallest Euclidean distance between the two objectives in the objective space. The GA was repeated 1000 times under randomised initial conditions sampled via Latin Hypercube (LH) within the specified bounds (Dunstan et al., 2023). As for the DCM-VL, all GA optimisations were conducted on the log-scale parameters that had similar numerical magnitues. These parameters were only exponentiated and re-scaled during the evaluation of the forward model to compute the model prediction. The DIP-DCM approach combined the above GA and DCM into a “two-step” strategy. First, the GA explored the parameter space globally, offering a mapping between parameter values and model dynamics. Second, DCM was used for local “exploitation” of the neighbourhood of points close to the parameter estimate identified by GA. Specifically, among the 1000 GA parameter sets, the top m sets were selected based on their model dynamics, as scored by RMSE. Each parameter in the set was used as the mean of a Gaussian distribution that assumed unit variance (i.e., all diagonal elements of the covariance matrix were equal to one—rather than equal to the “standard” variances—as all estimates were considered equally plausible). This is an uninformative value that is larger than the variances of the “standard” priors. Moreover, utilising the same value of variance across all parameters allowed all parameters to be sufficiently updated, and to generate more robust inferences (Supplementary 3; Figs. S4,S5). The parameter estimates from the GA were used to form the dynamics-informed priors defining m distinct generative models. Each model entered the VL routine to reach a stable prediction, recovering m variational posteriors and the associated free energy values. The m posterior densities were formalised into a single distribution as indicated in Appendix A, which is a process similar to Bayesian model averaging (Penny et al., 2010). However, the traditional Bayesian model averaging implies weighting the distributions by the respective model evidence. This was not necessary here because all the distributions had the same weight (further evidenced in Supplementary 3). Moreover, by averaging across multiple realisations (at least 400), DIP-DCM reduced variance in the estimation, contributing to more statistically robust predictions as demonstrated in Supplementary 4 (Fig. S6). Nevertheless, since the number of selected priors and inversions is a tunable parameter within this framework, users can tailor it based on the specific goals and constraints of their study. Study 1 utilised m=500 and Study 2 utilised m=400 . A summary of DIP-DCM is provided in Figure 1 and further details are in Appendix A. DIP-DCM flow chart. Dynamic causal modelling with dynamics-informed priors (DIP-DCM) involves two steps. First, a genetic algorithm is utilised to select parameter regimes associated with appropriate model dynamics. The genetic algorithm is repeated for a number n of realisations. Second, the information learned is applied to construct parameter priors, which enter the DCM variational inference routine. This process generates a number m of posterior parameter distributions that are averaged into a single distribution. For an in-depth description, refer to Methods 2.3. Abbreviations: GA, genetic algorithm; DCM, dynamic causal modelling. Inference on the effects of treatments and pathology Inference was conducted at the group level, and all calculations are formally defined in Appendix B. Group-level effects, with groups being pre-treatment and post-treatment in Study 1 or control and schizophrenia in Study 2, were estimated as the mean absolute difference between 105 draws from the corresponding posterior parameter distributions of DCM and DIP-DCM and the empirical distributions of GA point estimates. Plausibility of non-zero differences were inferred from the 95% Bayesian Credible Intervals (BCIs) of the difference distributions. Intervals that excluded zero were interpreted as analogous to statistical significance at the α=0.05 level. These effects were streamlined by interpreting their practical importance on the basis of Cohen’s d standardised effect sizes (Sullivan & Feinn, 2012), with standardisation enabling direct mapping between studies, contexts, and measurement scales. As convention, d=0.2 indicated a non-trivially small effect, d=0.5 a medium effect, and d=0.8 a strong effect. Within this context, d was indicative of the effect magnitude of pharmacological intervention or pathology. All work was performed using MATLAB (Mathworks Ltd, USA, R2023b) and the SPM12 toolbox. Results To illustrate the challenges associated with parameter estimation, free energy landscapes were derived as indicated in Appendix C. Figure 2 provides a specific example of the effects of parameter priors on the search for the most plausible DCM. This example explores the range of plausible prior means for parameter G1 , which represents the magnitude of connectivity between pyramidal and stellate cell populations of the model. Notably, this is a lumped parameter that would be difficult to measure or calculate directly in real biological systems. Therefore, choosing an accurate prior is also difficult. The “LFP” NMM (Methods 2.2; Moran et al., 2007) was fitted to the PSD of EEG data recorded in baseline conditions (pre-placebo dataset from Study 1, Methods 2.1), both in log space. The mean of the parameter distribution was varied in log space in the range [−1,0.85], which included the “standard” prior, log(G1)=0 (Moran et al., 2007; Appendix C) and the remaining model parameters were set to their “standard” prior. Variational free energy (Methods 2.3) was calculated from each parameter setting estimated through the standard variational inversion, and the information was utilised to construct a free energy profile, as shown in Figure 2. Small perturbations in the parameter value were accompanied by changes in the model dynamics (Fig. 2, between a1 and a2), which created discontinuities in the free energy landscape and isolated basins of convergence (sets of points that would converge to the local or global optimum within the range explored). Notably, when the inversion was initiated from the standard prior, the posterior estimate of G1 did not correspond to the global maximum free energy, which remained unexplored and required an alternative prior distribution. This illustrates that, depending on the choice of parameter prior, variational inference methods relying on gradient descent techniques can be susceptible to converging on suboptimal solutions, despite the existence of superior solutions elsewhere in the parameter space. These findings illustrate that the parameter landscape contains multiple local minima. This motivates the use of approaches such as DIP-DCM, which can explore a wider range of parameter configurations and potentially help identify informative prior settings, as investigated in the following sections. Free energy profile of a connectivity parameter. The mean of the Gaussian distribution of parameter G1 (Moran et al., 2007 model) was varied in log space across the biologically plausible range [−1,0.85]. All remaining parameters were set to the “standard” priors (Supplementary 1, Table S1). For each prior setting, the standard DCM variational inversion estimated the posterior parameters from which the associated variational free energy was calculated. Variational free energy values were scaled between 0 and 1, with the maximum free energy set to 1. The x-axis displays the value of the prior mean for the parameter. “Standard prior” is the predefined prior mean for the parameter. “VL estimate” corresponds to the posterior mean (expectation) of the parameter after estimation using variational Laplace (VL). a1,a2: spectra of model corresponding to the colour-coded parameter value, plotted against the data (Biondi et al., 2022, pre-placebo). The global optimum is in blue and the posterior estimate updated from the standard prior is in red. Simulations Parameter recovery analyses were conducted on synthetic data (“ground truth”), to investigate the accuracy of DCM and DIP-DCM in parameter estimation (Fig. 3). “Ground truth” parameter recovery on synthetic data. (A) RMSE mapping illustrating two regions of the parameter space where the model has equal dynamics. Parameter bounds are those used in DIP-DCM. (B) Parameter recovery using DCM with “standard” priors, DCM with a different prior mean, and DIP-DCM. Model fitting was conducted on the synthetic data corresponding to the parameter set “1” (True 1). DCM can only recover one of the two true parameters, either “1” or “2”, depending on the prior mean. DIP-DCM can recover both. Abbreviations: DIP-DCM, dynamic causal modelling with dynamics-informed priors; DCM, dynamic causal modelling. All parameters were fixed at values chosen from model fits to data, except the parameters under test. In the example presented here, the free parameters were R1 and Di . The model was simulated from the parameter set “1” (R1=−0.11 , Di=0.34 ; Fig. 3A), and the RMSE was calculated between these synthetic data at “1” and the model simulated at all other parameter combinations across a 2002 grid spanning the bounds of R1 and Di used in the GA step of DIP-DCM. Notably, the landscape highlights the existence of two local minima: one at the known parameter value (R1=−0.11 , Di=0.34 ) and one at a different value of R1 (R1=0.72 , Di=0.34 ). Therefore, as often observed in models of biological systems, the “ground truth” is not a single value. Supplementary 5 illustrates another example of this phenomenon, and further examples in a related model can be found in Dunstan et al. (2025). DIP-DCM and variants of the standard DCM were used to recover this “ground truth” that includes the parameter sets “1” and “2”. The DCM variants included DCM using “standard” priors or different prior means (Fig. 3B), but also DCM with larger variances and with priors derived from the DIP-DCM bounds, where the prior mean is in the middle of the interval and the variance spans 99% of it (these are shown in Supplementary 5; Fig. S7). Notably, DIP-DCM is the only method that can recover the “true” multi-modal distribution. Conversely, DCM can recover one mode of R1 (or generate a posterior distribution that includes one mode), either R1=−0.11 at “1” or R1=0.72 at “2”, contingent on the proximity of the prior to “1” or “2”. These results show the advantages of using a quantification of the fitness landscape to inform priors within the DCM framework, as in DIP-DCM, to mitigate biases. From Figure 3, it can also be suggested that parameter correlations can be complex, and different regions of the parameter space—where distinct parameter values produce the same data—can display positive or negative parameter correlations. The posterior correlation structures from DCM and DIP-DCM were quantified in Supplementary 6 (Fig. S8). Results indicate that the two approaches yield different estimates of parameter correlations, which are expected to vary depending on the location in parameter space. Ultimately, these simulations demonstrate the face validity of DIP-DCM, which can successfully recover the “ground truth” parameter distribution in synthetic data. The following analysis will seek to apply it to and validate its performance on real-world data. Study 1 Dynamics-informed priors outperform standard priors, yielding more plausible generative models The dataset (Methods 2.1) was collected in healthy participants at baseline and following the administration of LTG, LEV, or placebo. Compared with placebo, exposure to LTG and LEV increased the spectral power in the beta band (13–30 Hz; LTG vs placebo, p-value = 0.014 and LEV vs placebo, p-value = 0.011, by Mann–Whitney U-test; Fig. 4A, B), consistent with prior research on drug-naïve patients with epilepsy and other conditions (Cho et al., 2012; Dini et al., 2023). For each experimental condition (Methods 2.1), log PSDs were modelled via GA, DCM, and DIP-DCM (Methods 2.3) to identify the mechanisms underlying the increase in beta oscillations following ASM administration. For these comparisons, the acronym DCM is used to refer to a DCM adopting “standard” priors. DIP-DCM consistently outperforms GA and DCM. Spectra were calculated from the Biondi et al. (2022) dataset (Methods 2.1). (A) Study design. (B) Average data and models generated via DCM with “standard” priors (b1), GA (b2) and DIP-DCM (b3). (C) RMSE was calculated between data and models generated by DCM (red, n = 1), GA (turquoise, n = 500), and DIP-DCM (navy blue, n = 500). RMSE was calculated between 6.5 and 30 Hz, in line with the objective functions of the GA (formulae are reported in Methods 2.3). This gives an indication of the goodness of fit of the model, and specifically, that the GA estimates are best used as informative priors for DCM, since the RMSE is substantially reduced when the GA estimates enter the DCM inversion within the DIP-DCM approach. Abbreviations: DIP-DCM, dynamic causal modelling with dynamics-informed priors; DCM, dynamic causal modelling; GA, genetic algorithm; RMSE, root mean square error; PL, placebo; LTG, lamotrigine; LEV, levetiracetam. The performance of GA, DCM, and DIP-DCM was evaluated by calculating the root mean square error (RMSE) between model and respective data for posterior means (DCM, DIP-DCM) or point estimates (GA) produced by each inversion (Fig. 4B). Specifically, the GA produces 500 independent point estimates (one for each run of the optimisation), DIP-DCM returns 500 posterior means and variances, and the standard DCM returns a single posterior characterised by a mean and a variance. Therefore, this figure utilises information on the spectral responses corresponding to each of these posterior means or point estimates to calculate the respective RMSE, which here provides a common metric of goodness of model fit for the three approaches. It can be noted that the GA produced parameterisations associated with a broad range of model dynamics (or spectral fits) as evidenced by the RMSE calculated against data (Fig. 4C). Some of these dynamics were similar to those of DCM (Fig. 4B, C, b1, b2). However, 87.3% of these were inferior to those of DCM. While the GA explores different parameter sets, it may not converge on parameter regions that yield optimal dynamics. Crucially, combining GA with DCM via DIP-DCM improved all model calibrations (Fig. 4B, b3), with 99.7% improvement from GA (decrease in paired RMSE data from turquoise to blue, Fig. 4C) and from DCM alone (burgundy, Fig. 4C). This result indicates that the GA may identify suboptimal parameters, which, however, lie in the basin of convergence of the global optimum of the free energy landscape. Therefore, GA-derived parameters are best used as informative priors for VL within DIP-DCM, amplifying the advantages of the two methodologies. Similarly, these results show that DIP-DCM can reach parameter regions associated with lower errors than those accessible through a single DCM inversion. Additional comparisons with intermediate initialisation strategies are provided in Supplementary 3 (Figs. S3–S5; Table S5). DIP-DCM generates reliable parameter inferences DCM, GA, and DIP-DCM generated different parameter distributions. Exemplar posterior parameter distributions generated by DCM and DIP-DCM and the empirical distribution of GA point estimates (Methods 2.3) are shown in Figure 5A. For specific parameters, such as Te, both GA and DIP-DCM retrieved multi-modal distributions (Fig. 5A) that display more than one peak compared with DCM (as it assumes normality). Multi-modality can introduce interpretational complexity, yet it also provides a richer understanding of the alternative parameter regimes, or generative mechanisms, supported by the same data. Thus, DIP-DCM can give information on the most dominant parameter values, but it can also highlight alternative parameter trajectories. DIP-DCM generates inferences on ASM mechanisms with superior specificity and sensitivity. (A) Representative examples of parameter distributions (posterior distributions for DCM and DIP-DCM and the empirical distribution of point estimates for the GA) for the excitatory time constant parameter Te. All parameter densities are normalised to 1 and displayed within the same range. (B) Mechanistic effects were obtained as the difference between the parameter distributions of post-treatment and pre-treatment conditions, and selected based on 95% Bayesian Credible Intervals and statistical power (Cohen’s d≥0.2 ), as per Methods 2.4. The y-axis units of the parameter effects vary depending on the parameter, that is, for Te and Ti the unit is seconds, He, R1, and R2 are unitless. Abbreviations: DCM, dynamic causal modelling; GA, genetic algorithm; DIP-DCM, dynamic causal modelling with dynamics-informed priors; PL, placebo; LTG, lamotrigine; LEV, levetiracetam. The posterior parameter distributions of DCM and DIP-DCM and the empirical distribution of GA point estimates were used to make mechanistic inferences and derive the effects of placebo, LTG, and LEV, as indicated in Methods 2.4. Effects were calculated as the absolute difference between mean parameter values at baseline and following pharmacological intervention. Subsequently, effects were selected based on 95% BCI and by considering a sufficient statistical power (Cohen’s d ≥0.2 ). These effects are reported in Figure 5B. While all methods consistently indicated that placebo does not produce detectable effects (Fig. 5B), discrepancies emerged with respect to the ASMs. DCM predicted a single LTG effect, namely, an increase in the inhibitory synaptic time constant parameter (Ti). Conversely, GA and DIP-DCM both identified an increase in Ti but also a decrease in the parameter Te, representing the excitatory synaptic time constant, and an increase in parameter R1. Moreover, DCM did not predict any LEV effects, similarly to the placebo condition, despite the observed increase in beta power (Fig. 4B). On the contrary, GA revealed an increase in Ti and a decrease in Te, while DIP-DCM identified a single LEV-specific effect, that is a selective decrease in Te. These effects are discussed in Discussion 4.1–2. DIP-DCM provides the best trade-off between reliable inference and computational efficiency A potential drawback of global search heuristics such as GA is that they can be computationally expensive compared with the DCM variational inference. Notably, DIP-DCM optimised the trade-off between accuracy and computational cost, outperforming the standalone GA (Fig. 6A). The performance of DIP-DCM was further assessed as indicated and shown extensively in Supplementary 4. In the first step of DIP-DCM, parameters evolve over a number of generations set to be 500. Therefore, this number was reduced to determine the minimum required to generate models with suitable dynamics. Results suggested that a minimum of 150 generations were necessary to maintain a stable model fitness as scored by RMSE, while maintaining computation times moderate (Fig. 6A). Computation times were ≈1 minutes for one full DIP-DCM inversion on a personal computer (Intel Core i7 (13th gen), 12 CPU cores). However, when iterating over more than 400 priors, models were run on a server or on a supercomputer (ISCA, University of Exeter). The individual DIP-DCM model fits can be executed at comparable per-model speeds, yet their execution can be parallelised. Therefore, the cumulative computation time across all models will differ depending on the parallelisation approach the user wishes to adopt, or on the number of priors selected by the user. Efficiency of DIP-DCM. The performance of DIP-DCM was assessed in terms of computational efficiency considering that one of the drawbacks of the GA is the computational cost. The latter could be reduced by reducing the number of generations. The DIP-DCM approach that uses GA-derived estimates as prior means was also compared with the use of randomly sampled priors (sampled via Latin Hypercube). (A) Reduction in the number of generations used in the first step of DIP-DCM. RMSE is the root mean square error between spectra of model and data, runtime is for a single inversion. Data are shown as mean ± S.E.M. The log PSD of the model was fitted to the log PSD of the placebo condition, which serves as the critical baseline for all comparisons. (B) Exemplar pre-placebo spectra of the model plotted against data. The log PSDs of the model were obtained from randomly sampled priors (LH, b1) or dynamics-informed priors (DIP, b2). Abbreviations: LH, Latin hypercube; GA, genetic algorithm; gen, generations. Moreover, it was also important to understand whether improvements similar to those of DIP-DCM could be achieved by seeding prior means at random via Latin Hypercube (LH), rather than generating priors via GA. Prior covariance and bounds for LH-based models were identical to those of the DIP-DCM approach. While the LH-based approach was substantially faster, it failed to generate appropriate model dynamics (Fig. 6A, B). Specifically, LH-based models exceeded the fitness threshold established by GA (Fig. 6A), resulting in a majority of implausible models (Fig. 6, b1), compared with DIP-based models (Fig. 6, b2). Taken together, these results demonstrate that DIP-DCM generates optimal calibrations efficiently, requiring minutes to achieve accurate parameter estimation. Moreover, these results show that the GA can be terminated earlier once the prior is appropriately positioned within the parameter space, allowing the DCM variational inference to be leveraged for rapid convergence. Study 2—Application of DIP-DCM to MEG data Study 1 (Results 3.2) indicated that different parameter estimation approaches can give rise to different mechanistic inferences for the same NMM and data. While dynamic causal models rely on a single normally distributed prior, DIP-DCM takes into account additional parameter regimes and produces more informative predictions. To corroborate these findings, analyses were repeated on the Shaw et al. (2020) dataset (Fig. 7A, Methods 2.1), providing a practical demonstration of the wider applicability of DIP-DCM. Application of DIP-DCM to MEG data. (A) Study design. (B) Data and model spectra generated by DCM (b1) and DIP-DCM(b2). Data are shown as mean. Since all model spectra are very similar, the standard deviation is not shown because it is too small to be visible. (C) Representative examples of parameter distributions obtained via DCM (c1) and DIP-DCM (c2). All parameter densities are normalised to 1 and displayed within the same range. (D) Effects were obtained as the difference between the parameter distributions of post-treatment and pre-treatment conditions, and selected based on 95% Bayesian Credible Intervals and statistical power (Cohen’s d≥0.2 ), as per Methods 2.4. The y-axis units of the parameter effects vary depending on the parameter, that is, for Te the unit is seconds, and R2 is unitless. Abbreviations: DIP-DCM, dynamic causal modelling with dynamics-informed priors; Ctl, control; Scz, schizophrenia. Power spectra were modelled via DCM and the hybrid DIP-DCM workflow optimised for efficiency (see Fig. 6 and Supplementary 4). DIP-DCM yielded model dynamics that closely resembled data across the entire broadband spectrum (2−85 Hz), compared with DCM with “standard” priors (Fig. 7B, panels b1, b2). This result is consistent with conclusions from Results 3.2. Mechanistic inferences were obtained as per Methods 2.4, and examples of posterior distributions are shown in Figure 7C (DCM with “standard” priors in c1 and DIP-DCM in c2). Notably, the posterior distributions of parameter Te (Fig. 7C) were largely uni-modal compared with those of Study 1. For this parameter, both DIP-DCM and the standard DCM identified similar effects, indicating that impairments in visual processing and gamma oscillations in patients with schizophrenia are attributed to the speed of decay of excitatory activities (Fig. 7D). However, DIP-DCM also identified an additional effect, namely an increase in parameter R2, which is indicative of higher firing thresholds and, therefore, of reduced responsiveness to sensory inputs. Similarly, in Study 1, DIP-DCM detected key effects—such as the decrease in parameter Te following LEV administration—that DCM failed to identify. Taken together, these results indicate that DIP-DCM can help make disease-specific inferences with increased sensitivity. Discussion This paper compares the performance of three parameter estimation approaches to illustrate the challenge of parameter estimation in NMMs, where large and often degenerate parameter spaces hamper the inference process. While parameter priors in Bayesian inference are often informed by expert knowledge, this approach can bias the parameter estimation by overlooking alternative parameter configurations. Moreover, in cases where parameter priors remain unclear or not represented among the candidate options, standard approaches may be limited. To mitigate these challenges, the present study introduces DIP-DCM, a dynamics-informed approach for DCM, designed to navigate complex parameter landscapes more effectively and to help draw inferences that are pure reflections of the information encoded in the data. Reliability and validity of parameter inferences in DIP-DCM This study demonstrates that different choices in the modelling process lead to different inferences. Specifically, the choice of priors in DCM significantly impacts the validity of the posterior parameter estimates and the resulting inferences, as expected. Typically, the challenge of finding suitable priors is addressed via Bayesian Model Reduction (BMR) (Friston et al., 2016), empirical Bayes (Friston et al., 2015), or Parametric Empirical Bayes (PEB) (Zeidman et al., 2019). A less common alternative involves deriving priors from a subset of individuals, effectively serving as a form of model re-training to narrow the parameter space for the remaining subset (Lee & Vanpaemel, 2018). However, these approaches may introduce bias by “overfitting” the model to the characteristics of the group, potentially reducing the applicability of inferences to other cohorts. BMR generally starts with uninformative priors to compute the full model posteriors before proceeding with computations for the reduced models. However, PEB requires specifying priors for the individual-level analysis to compute the posteriors that are used as empirical priors for the group level. However, there may be cases where these hierarchical or empirical approaches cannot be considered. For example, where the priors for the individual-level analysis are also difficult to determine and where analyses are conducted on grand mean averaged data directly as, for example, in Tewarie et al. (2024). Moreover, there may be cases where subject-level modelling is challenged by limited data. Furthermore, while PEB and BMR provide useful solutions for hierarchical prior estimation and for nested model comparison, within the DCM framework they leverage Gaussian assumptions. Therefore, these approaches may not naturally extend to complex parameter spaces with different geometry, as those illustrated in this work. While there may be other methods that enable navigation of complex parameter spaces, previous work suggested that genetic algorithms are effective strategies (Dunstan et al., 2023). Therefore, DIP-DCM is proposed as an alternative or orthogonal approach for guiding the parameter inference process, which deliberately seeks to relax strong assumptions about the structure of priors. First, the priors for the group-level analysis are objectively derived from data via GA, which explores the model dynamics under different parameter regimes. Second, a multi-start approach is used and can generate multimodal posteriors that are no longer Gaussian (compared with the DCM implementation of the above-mentioned approaches). Comprehensive benchmarking of DIP-DCM against all existing Bayesian approaches will be addressed in future studies. This paper compared DIP-DCM with a “standard” DCM and benchmarked it against intermediate initialisation strategies (Supplementary 3). These analyses demonstrate that utilising a wider variance allows better model fits, even for single DCM inversions. Conversely, using the tighter (“standard”) variances substantially biased the inferences, even when used within the DIP-DCM approach. Moreover, a single inversion still sacrifices the specificity of the inferences that would otherwise be obtained from multiple inversions. Ultimately, priors are difficult to ascertain and this work demonstrates that the DIP-DCM approach can correctly identify suitable priors, the true posterior estimates, and the known effects of anti-seizure medication, for example, changes in excitatory response time scales following LEV administration. Notably, rather than generating inferences that are fundamentally different from those of a standard DCM, DIP-DCM can identify additional effects with increased sensitivity, generating precise and testable hypotheses. This approach helps prevent biases introduced by a specific set of priors—which may lead to predictions that are not universally applicable—by considering a range of plausible outcomes. Thus, as the model parameters are often not uniquely identifiable from the data, their distributions may exhibit multiple peaks, or modes, each corresponding to a different parameter regime (Gábor & Banga, 2015). A detailed validation of the posterior estimates is provided in Table 1 for key parameters. Within this study, parameters Te and Ti—representing the time scales of excitatory and inhibitory synaptic dynamics—are of particular importance. Different receptors are known to have different kinetics, determining the duration and nature of post-synaptic responses. While there is no direct mapping between parameter values and receptor types in this NMM, plausible receptor dynamics can be hypothesised based on the duration of the post-synaptic potential (PSP). Ultimately, the multi-modality of Te and Ti may reflect the dynamic nature of synaptic transmission and network activity. Each mode could correspond to distinct receptor kinetics, synaptic terminal properties, or features of excitatory post-synaptic currents (EPSCs) and inhibitory post-synaptic currents (IPSCs), as outlined in Table 1. Table 1.Posterior parameter modes generated by DIP-DCM in Study 1.ParameterPosterior modeSupporting evidenceTi, representing the inhibitory synaptic time constant, or τ.Two modes DIP, LTG: ≈8.4 ms, and ≈27 ms vs ≈10 ms in DCMInhibitory post-synaptic current (IPSC) kinetics: fast IPSCs have a decay time constant of <10 ms (Bartos et al., 2002); GABAergic synaptic properties in layers II-IV, [6;10] ms (Gupta et al., 2000).Slow IPSCs have decay time constants of tens of milliseconds (Banks et al., 1998).For fast decaying IPSC τ is three to five times smaller than τ of slowly decaying IPSCs, depending on the brain region (Nusser et al., 2001; Sceniak & Maciver, 2008).Inhibitory time constants are normally longer than excitatory time constants (Lemaréchal et al., 2022).Te, representing the excitatory synaptic time constant, or τ.Three modes DIP, LTG: ≈1.8 ms, ≈5.2 ms, and ≈10 ms vs ≈5.6 ms in DCM. One mode DIP, LEV: ≈8.7 ms vs ≈7.6 ms in DCMAMPA-mediated current rise times = 0.29±0.04 ms, and decay = 2±0.8 ms (in rat cortical neurons) at excitatory synapses where the pre-synapse is a pyramidal neuron and the post-synapse is an interneuron (Angulo et al., 1999). Note that the excitatory post-synaptic potential (EPSP) kinetics can be conserved across mammalian species (human, rodent; Wilbers et al., 2023).AMPA receptors respond to pre-synaptic inputs and activate with a time scale in the order of 10−1 ms, and deactivate in milliseconds (Pampaloni & Plested, 2022). τAMPA≈1−10 ms (Moreno-Bote & Parga, 2004).Rise time = [0.2;0.8] ms, decay time = [1.3;2] ms, in cultured rat cortical neurons (Kleppe & Robinson, 1999). τ≈16±5.3 ms in layer V pyramidal cells and in the neocortex of guinea pigs (Kim & Connors, 1993; Koch et al., 1996).AMPA receptors have fast and slow components to mediate different gating modes and modal switching, over the time scale of 5–10 ms. A low open probability mode with short open periods, and a high open probability mode with longer open periods (Prieto & Wollmuth, 2010).Time scales are not fixed. Time constants can change dynamically as the network activity changes (Bernacchia et al., 2011).R1, representing the non-linear gain.One mode DIP, LTG: ≈0.4 ms vs ≈0.96 ms in DCM.Changes in sensitivity of neurons to inputs without enhancing or suppressing overall activity, resembling intrinsic excitability properties. It can be linked to receptor ligand availability. No empirical measurements available. Posterior parameter modes generated by DIP-DCM in Study 1. The estimated parameters led to inferences presented in detail in Discussion 4.2. Acting as an intermediary between GA and DCM, DIP-DCM recovered key ASM-induced changes in model parameters, suggesting that both ASMs cause faster decays in excitatory activities, while LTG additionally slows the decay of inhibitory activities. Moreover, results suggest that LTG increases the parameter R1, indicating that small fluctuations in membrane potential produce larger changes in output firing rates. These inferences are in line with existing experimental evidence suggesting that LTG exerts a bidirectional modulation of both excitatory and inhibitory activity, while LEV appears to selectively affect excitatory activity (Greenhill & Jones, 2010; Margineanu & Klitgaard, 2003; Meehan et al., 2011). Ultimately, the results support several important conclusions: (1) GA and DCM could produce qualitatively similar model dynamics, yet lead to different mechanistic inferences, which highlights the inherent challenges of parameter estimation in neural mass models. (2) DCM with “standard” priors is insufficiently sensitive to detect non-trivially small effects, such as the decrease in Te following LTG and LEV administration. (3) Global optimisations such as DIP-DCM prevent such type II errors, considering that the absence of an effect does not equate to the evidence of its absence. (4) DIP-DCM demonstrates both sensitivity and specificity, as evidenced by the effects of LEV on parameter Te and by the effects of schizophrenia on parameter R2. Plausibility of DIP-DCM hypotheses on the effects of anti-seizure medication DIP-DCM reconciled inferences on ASM effects across the tested approaches and identified group-level mechanistic fingerprints of ASMs for this NMM, which were supported by existing data. It is well established that LEV modulates presynaptic mechanisms involved in glutamate release at excitatory synapses, leading to changes in AMPA receptor-mediated currents (Carunchio et al., 2007). LEV binds to the presynaptic vesicle protein SV2A (Lynch et al., 2004; Sugaya et al., 2010), which has a modulatory role in transmitter release (Meehan et al., 2011). Consequently, LEV is unlikely to cause a substantial blockage of glutamatergic transmission (Meehan et al., 2011). The modulatory role of LEV aligns with inferences from DIP-DCM, indicating moderate effects on excitatory synapses. Moreover, the pharmacodynamics seem to involve a reduction in the release of vesicular content, which could explain the faster decay kinetics of EPSCs hypothesised from DIP-DCM. This is consistent with a theoretical framework where a reduction in neurotransmitter release may result in more rapid clearance, leading to faster decay of excitatory currents, or vice versa (Eggers & Lukasiewicz, 2006). This conclusion is also in line with studies conducted in epileptic rats showing that LEV increases the slope of field EPSPs (fEPSPs) (Salaka et al., 2022). Taken together, these results indicate that LEV only acts on excitatory neurons, supporting the hypothesis introduced by Margineanu and Klitgaard (2003). These data also suggest that LEV may not simply suppress excitation but fine-tune and stabilise synaptic transmission towards healthier levels (Salaka et al., 2022). However, different ASMs are known to act via different mechanisms (Rogawski & Löscher, 2004). ASMs are generally thought to modify the dynamic balance between excitation and inhibition, by either acting on excitation, inhibition, or on both concurrently. Differently from LEV, LTG has been suggested to increase the ratio between the global background synaptic excitation and inhibition (Greenhill & Jones, 2010). As indicated by DIP-DCM, this could plausibly occur through increasing the time scale separation between the excitatory and inhibitory synaptic currents. Nonetheless, earlier studies indicated a lack of observable changes in spontaneous EPSC and IPSC decays measured via patch-clamp following LTG administration (Cunningham & Jones, 2000). This highlights the complexity of relating cell-level recordings to properties of NMMs. Patch-clamp experiments, as those performed by Cunningham and Jones (2000), may not accurately extrapolate population-level activity. In addition, this patch-clamp experiment considered excitatory and inhibitory events in isolation, which may be inappropriate when investigating the effects of pharmacological agents that act across excitatory and inhibitory systems. Thus, if these results are ever to be tested empirically, it may be more appropriate to measure global background synaptic excitation and inhibition and synaptic plasticity as in Greenhill and Jones (2010) and Salaka et al. (2022). Notably, the pronounced increase in inhibition-to-excitation ratio found by Greenhill and Jones (2010) was paralleled by changes in neuronal excitability, which is supported by the changes in population sensitivity (parameter R1 ) inferred by DIP-DCM. Taken together, these results illustrate how DIP-DCM can be applied across different experimental contexts providing hypotheses that can be readily tested. Accelerated global search across dynamic causal models This study highlights the advantages of using a hybrid DIP-DCM strategy, which enhances search efficiency while guaranteeing solution quality. Other two-phase hybrid optimisations—where global methods are used for broad sampling and local methods for rapid convergence—were shown to be very effective at mitigating degeneracy and improving inference robustness across a wide range of optimisation problems. Hybrid optimisations, combining GAs and gradient descent-based (GD) methods, have proven useful for challenging problems with local minima outside the neuroimaging and DCM context (Gan et al., 2012), demonstrating superior performance in parameter estimation for biological network models (Rodriguez-Fernandez et al., 2006). Additionally, similarly to DIP-DCM, they have been shown to outperform GA alone in terms of efficiency and solution quality by positioning solutions in regions that facilitate convergence (D’Angelo & Palmieri, 2021). Performance remained superior even with a reduced number of generations, mirroring findings from the present study, where DIP-DCM achieved superior results for a third of the generations required by GA. Moreover, these results suggest that the DIP-DCM approach is more advantageous than the use of GA alone, which can retrieve suboptimal solutions and is more computationally expensive. When the GA solutions are used as priors for DCM, within the DIP-DCM framework, the trade-off between inference accuracy and computational efficiency is found to be optimal. Furthermore, findings seem to resonate with similar strategies used within the DCM context, such as the massively parallel DCM (mpDCM) for haemodynamic models (Aponte et al., 2016). By adopting the concept of “thermodynamic integration”, mpDCM computes multiple MCMC chains simultaneously at varying “temperatures”. Temperatures affect the trade-off between exploration and exploitation by adjusting the relative influence of the prior and likelihood in the sampling process, thereby guiding the trajectory from prior to posterior (Aponte et al., 2022; Frässle et al., 2021). This approach facilitated parameter exploration, preventing chains from being trapped in local optima, and accelerated convergence by allowing chains to exchange information through mechanisms inspired by evolutionary principles. While DIP-DCM differs from mpDCM in its mechanisms for parameter exploration and information exchange between search pathways, both methods share common goals. They improve convergence to solutions and adopt principled approaches to manage computational costs. Finally, although alternative inference methods to variational Bayes have been used in the DCM framework for modelling electrophysiological data—such as MCMC (Sengupta et al., 2015)—to the best of our knowledge this study provides a tool where the DCM for M/EEG is driven by global optimisations. Robustness considerations and future directions It can be noted that there is a trade-off between capturing the complexity of the “fitness landscape” and accurately quantifying uncertainty. Thus, it is important to account for accurate uncertainty quantification in hybrid parameter estimation methods such as DIP-DCM, and future work will be required to fully address this. In this work, and specifically in Study 2, the posterior parameter distributions estimated via DIP-DCM showed a decreased variance relative to those estimated by DCM. It could be argued that this may be due to VL initialisation close to regions of parameter space that minimise the error between model and data. Conversely, in Study 1, the posterior parameter distributions estimated via DIP-DCM showed an increased variance relative to those estimated by DCM. Thus, it is not necessarily the case that DIP-DCM shrinks variances because the priors were preconditioned on the same data used in the first step of the hybrid DIP-DCM approach (i.e., by using a GA). In principle, if the GA was trained on a different dataset or on part of the same dataset, the priors would be less tightly coupled to the observations used for parameter inference. However, in this case, the prior may no longer be a good description of the data. It is envisaged that future research will formally assess the extent to which DIP-DCM affects uncertainty quantification, particularly regarding how it shapes posterior variance. Conclusions To conclude, this study illustrates the existence of multiple calibrations, and associated inferences, for the same NMM and M/EEG data, and proposes a dynamics-informed framework for DCM as an alternative to traditional, local, inversion routines. This framework accurately estimates model parameters and the associated uncertainty by leveraging global optimisations, maintaining moderate computational costs, and suggesting potential benefits for a wider range of applications. Supplementary Material