An introduction to thermodynamic integration and application to dynamic causal models.

In generative modeling of neuroimaging data, such as dynamic causal modeling (DCM), one typically considers several alternative models, either to determine the most plausible explanation for observed data (Bayesian model selection) or to account for model uncertainty (Bayesian model averaging). Both procedures rest on estimates of the model evidence, a principled trade-off between model accuracy and complexity. In the context of DCM, the log evidence is usually approximated using variational Bayes. Although this approach is highly efficient, it makes distributional assumptions and is vulnerable to local extrema. This paper introduces the use of thermodynamic integration (TI) for Bayesian model selection and averaging in the context of DCM. TI is based on Markov chain Monte Carlo sampling which is asymptotically exact but orders of magnitude slower than variational Bayes. In this paper, we explain the theoretical foundations of TI, covering key concepts such as the free energy and its origins in statistical physics. Our aim is to convey an in-depth understanding of the method starting from its historical origin in statistical physics. In addition, we demonstrate the practical application of TI via a series of examples which serve to guide the user in applying this method. Furthermore, these examples demonstrate that, given an efficient implementation and hardware capable of parallel processing, the challenge of high computational demand can be overcome successfully. The TI implementation presented in this paper is freely available as part of the open source software TAPAS. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s11571-021-09696-9.

TAPAS: An Open-Source Software Package for Translational Neuromodeling and Computational Psychiatry.

Psychiatry faces fundamental challenges with regard to mechanistically guided differential diagnosis, as well as prediction of clinical trajectories and treatment response of individual patients. This has motivated the genesis of two closely intertwined fields: (i) Translational Neuromodeling (TN), which develops "computational assays" for inferring patient-specific disease processes from neuroimaging, electrophysiological, and behavioral data; and (ii) Computational Psychiatry (CP), with the goal of incorporating computational assays into clinical decision making in everyday practice. In order to serve as objective and reliable tools for clinical routine, computational assays require end-to-end pipelines from raw data (input) to clinically useful information (output). While these are yet to be established in clinical practice, individual components of this general end-to-end pipeline are being developed and made openly available for community use. In this paper, we present the Translational Algorithms for Psychiatry-Advancing Science (TAPAS) software package, an open-source collection of building blocks for computational assays in psychiatry. Collectively, the tools in TAPAS presently cover several important aspects of the desired end-to-end pipeline, including: (i) tailored experimental designs and optimization of measurement strategy prior to data acquisition, (ii) quality control during data acquisition, and (iii) artifact correction, statistical inference, and clinical application after data acquisition. Here, we review the different tools within TAPAS and illustrate how these may help provide a deeper understanding of neural and cognitive mechanisms of disease, with the ultimate goal of establishing automatized pipelines for predictions about individual patients. We hope that the openly available tools in TAPAS will contribute to the further development of TN/CP and facilitate the translation of advances in computational neuroscience into clinically relevant computational assays.

Computational Dissociation of Dopaminergic and Cholinergic Effects on Action Selection and Inhibitory Control.

BACKGROUND: Patients with schizophrenia make more errors than healthy subjects in the antisaccade task. In this paradigm, participants are required to inhibit a reflexive saccade to a target and to select the correct action (a saccade in the opposite direction). While the precise origin of this deficit is not clear, it has been connected to aberrant dopaminergic and cholinergic neuromodulation. METHODS: To study the impact of dopamine and acetylcholine on inhibitory control and action selection, we administered two selective drugs (levodopa 200 mg/galantamine 8 mg) to healthy volunteers (N = 100) performing the antisaccade task. The computational model SERIA (stochastic early reaction, inhibition, and late action) was employed to separate the contribution of inhibitory control and action selection to empirical reaction times and error rates. RESULTS: Modeling suggested that levodopa improved action selection (at the cost of increased reaction times) but did not have a significant effect on inhibitory control. By contrast, according to our model, galantamine affected inhibitory control in a dose-dependent fashion, reducing inhibition failures at low doses and increasing them at higher levels. These effects were sufficiently specific that the computational analysis allowed for identifying the drug administered to an individual with 70% accuracy. CONCLUSIONS: Our results do not support the hypothesis that elevated tonic dopamine strongly impairs inhibitory control. Rather, levodopa improved the ability to select correct actions. However, inhibitory control was modulated by cholinergic drugs. This approach may provide a starting point for future computational assays that differentiate neuromodulatory abnormalities in heterogeneous diseases like schizophrenia.

Inflexible social inference in individuals with subclinical persecutory delusional tendencies.

It has been suspected that abnormalities in social inference (e.g., learning others' intentions) play a key role in the formation of persecutory delusions (PD). In this study, we examined the association between subclinical PD and social inference, testing the prediction that proneness to PD is related to altered social inference and beliefs about others' intentions. We included 151 participants scoring on opposite ends of Freeman's Paranoia Checklist (PCL). The participants performed a probabilistic advice-taking task with a dynamically changing social context (volatility) under one of two experimental frames. These frames differentially emphasised possible reasons behind unhelpful advice: (i) the adviser's possible intentions (dispositional frame) or (ii) the rules of the game (situational frame). Our design was thus 2 × 2 factorial (high vs. low delusional tendencies, dispositional vs. situational frame). We found significant group-by-frame interactions, indicating that in the situational frame high PCL scorers took advice less into account than low scorers. Additionally, high PCL scorers believed more frequently that incorrect advice was delivered intentionally and that such misleading behaviour was directed towards them personally. Overall, our results suggest that social inference in individuals with subclinical PD tendencies is shaped by negative prior beliefs about the intentions of others and is thus less sensitive to the attributional framing of adviser-related information. These findings may help future attempts of identifying individuals at risk for developing psychosis and understanding persecutory delusions in psychosis.

Switch costs in inhibitory control and voluntary behaviour: A computational study of the antisaccade task.

An integral aspect of human cognition is the ability to inhibit stimulus-driven, habitual responses, in favour of complex, voluntary actions. In addition, humans can also alternate between different tasks. This comes at the cost of degraded performance when compared to repeating the same task, a phenomenon called the "task-switch cost." While task switching and inhibitory control have been studied extensively, the interaction between them has received relatively little attention. Here, we used the SERIA model, a computational model of antisaccade behaviour, to draw a bridge between them. We investigated task switching in two versions of the mixed antisaccade task, in which participants are cued to saccade either in the same or in the opposite direction to a peripheral stimulus. SERIA revealed that stopping a habitual action leads to increased inhibitory control that persists onto the next trial, independently of the upcoming trial type. Moreover, switching between tasks induces slower and less accurate voluntary responses compared to repeat trials. However, this only occurs when participants lack the time to prepare the correct response. Altogether, SERIA demonstrates that there is a reconfiguration cost associated with switching between voluntary actions. In addition, the enhanced inhibition that follows antisaccade but not prosaccade trials explains asymmetric switch costs. In conclusion, SERIA offers a novel model of task switching that unifies previous theoretical accounts by distinguishing between inhibitory control and voluntary action generation and could help explain similar phenomena in paradigms beyond the antisaccade task.

Inhibition failures and late errors in the antisaccade task: influence of cue delay.

In the antisaccade task participants are required to saccade in the opposite direction of a peripheral visual cue (PVC). This paradigm is often used to investigate inhibition of reflexive responses as well as voluntary response generation. However, it is not clear to what extent different versions of this task probe the same underlying processes. Here, we explored with the Stochastic Early Reaction, Inhibition, and late Action (SERIA) model how the delay between task cue and PVC affects reaction time (RT) and error rate (ER) when pro- and antisaccade trials are randomly interleaved. Specifically, we contrasted a condition in which the task cue was presented before the PVC with a condition in which the PVC served also as task cue. Summary statistics indicate that ERs and RTs are reduced and contextual effects largely removed when the task is signaled before the PVC appears. The SERIA model accounts for RT and ER in both conditions and better so than other candidate models. Modeling demonstrates that voluntary pro- and antisaccades are frequent in both conditions. Moreover, early task cue presentation results in better control of reflexive saccades, leading to fewer fast antisaccade errors and more rapid correct prosaccades. Finally, high-latency errors are shown to be prevalent in both conditions. In summary, SERIA provides an explanation for the differences in the delayed and nondelayed antisaccade task. NEW & NOTEWORTHY In this article, we use a computational model to study the mixed antisaccade task. We contrast two conditions in which the task cue is presented either before or concurrently with the saccadic target. Modeling provides a highly accurate account of participants' behavior and demonstrates that a significant number of prosaccades are voluntary actions. Moreover, we provide a detailed quantitative analysis of the types of error that occur in pro- and antisaccade trials.

Generative models for clinical applications in computational psychiatry.

Despite the success of modern neuroimaging techniques in furthering our understanding of cognitive and pathophysiological processes, translation of these advances into clinically relevant tools has been virtually absent until now. Neuromodeling represents a powerful framework for overcoming this translational deadlock, and the development of computational models to solve clinical problems has become a major scientific goal over the last decade, as reflected by the emergence of clinically oriented neuromodeling fields like Computational Psychiatry, Computational Neurology, and Computational Psychosomatics. Generative models of brain physiology and connectivity in the human brain play a key role in this endeavor, striving for computational assays that can be applied to neuroimaging data from individual patients for differential diagnosis and treatment prediction. In this review, we focus on dynamic causal modeling (DCM) and its use for Computational Psychiatry. DCM is a widely used generative modeling framework for functional magnetic resonance imaging (fMRI) and magneto-/electroencephalography (M/EEG) data. This article reviews the basic concepts of DCM, revisits examples where it has proven valuable for addressing clinically relevant questions, and critically discusses methodological challenges and recent methodological advances. We conclude this review with a more general discussion of the promises and pitfalls of generative models in Computational Psychiatry and highlight the path that lies ahead of us. This article is categorized under: Neuroscience > Computation Neuroscience > Clinical Neuroscience.

The Stochastic Early Reaction, Inhibition, and late Action (SERIA) model for antisaccades.

The antisaccade task is a classic paradigm used to study the voluntary control of eye movements. It requires participants to suppress a reactive eye movement to a visual target and to concurrently initiate a saccade in the opposite direction. Although several models have been proposed to explain error rates and reaction times in this task, no formal model comparison has yet been performed. Here, we describe a Bayesian modeling approach to the antisaccade task that allows us to formally compare different models on the basis of their evidence. First, we provide a formal likelihood function of actions (pro- and antisaccades) and reaction times based on previously published models. Second, we introduce the Stochastic Early Reaction, Inhibition, and late Action model (SERIA), a novel model postulating two different mechanisms that interact in the antisaccade task: an early GO/NO-GO race decision process and a late GO/GO decision process. Third, we apply these models to a data set from an experiment with three mixed blocks of pro- and antisaccade trials. Bayesian model comparison demonstrates that the SERIA model explains the data better than competing models that do not incorporate a late decision process. Moreover, we show that the early decision process postulated by the SERIA model is, to a large extent, insensitive to the cue presented in a single trial. Finally, we use parameter estimates to demonstrate that changes in reaction time and error rate due to the probability of a trial type (pro- or antisaccade) are best explained by faster or slower inhibition and the probability of generating late voluntary prosaccades.

mpdcm: A toolbox for massively parallel dynamic causal modeling.

BACKGROUND: Dynamic causal modeling (DCM) for fMRI is an established method for Bayesian system identification and inference on effective brain connectivity. DCM relies on a biophysical model that links hidden neuronal activity to measurable BOLD signals. Currently, biophysical simulations from DCM constitute a serious computational hindrance. Here, we present Massively Parallel Dynamic Causal Modeling (mpdcm), a toolbox designed to address this bottleneck. NEW METHOD: mpdcm delegates the generation of simulations from DCM's biophysical model to graphical processing units (GPUs). Simulations are generated in parallel by implementing a low storage explicit Runge-Kutta's scheme on a GPU architecture. mpdcm is publicly available under the GPLv3 license. RESULTS: We found that mpdcm efficiently generates large number of simulations without compromising their accuracy. As applications of mpdcm, we suggest two computationally expensive sampling algorithms: thermodynamic integration and parallel tempering. COMPARISON WITH EXISTING METHOD(S): mpdcm is up to two orders of magnitude more efficient than the standard implementation in the software package SPM. Parallel tempering increases the mixing properties of the traditional Metropolis-Hastings algorithm at low computational cost given efficient, parallel simulations of a model. CONCLUSIONS: Future applications of DCM will likely require increasingly large computational resources, for example, when the likelihood landscape of a model is multimodal, or when implementing sampling methods for multi-subject analysis. Due to the wide availability of GPUs, algorithmic advances can be readily available in the absence of access to large computer grids, or when there is a lack of expertise to implement algorithms in such grids.

Inversion of hierarchical Bayesian models using Gaussian processes.

Over the past decade, computational approaches to neuroimaging have increasingly made use of hierarchical Bayesian models (HBMs), either for inferring on physiological mechanisms underlying fMRI data (e.g., dynamic causal modelling, DCM) or for deriving computational trajectories (from behavioural data) which serve as regressors in general linear models. However, an unresolved problem is that standard methods for inverting the hierarchical Bayesian model are either very slow, e.g. Markov Chain Monte Carlo Methods (MCMC), or are vulnerable to local minima in non-convex optimisation problems, such as variational Bayes (VB). This article considers Gaussian process optimisation (GPO) as an alternative approach for global optimisation of sufficiently smooth and efficiently evaluable objective functions. GPO avoids being trapped in local extrema and can be computationally much more efficient than MCMC. Here, we examine the benefits of GPO for inverting HBMs commonly used in neuroimaging, including DCM for fMRI and the Hierarchical Gaussian Filter (HGF). Importantly, to achieve computational efficiency despite high-dimensional optimisation problems, we introduce a novel combination of GPO and local gradient-based search methods. The utility of this GPO implementation for DCM and HGF is evaluated against MCMC and VB, using both synthetic data from simulations and empirical data. Our results demonstrate that GPO provides parameter estimates with equivalent or better accuracy than the other techniques, but at a fraction of the computational cost required for MCMC. We anticipate that GPO will prove useful for robust and efficient inversion of high-dimensional and nonlinear models of neuroimaging data.