Name: A Protocol For Uncovering Neural Mechanisms Of Neurotherapeutic Effects On Electroencephalography Using The Human Neocortical Neurosolver (Video) | JoVE Journal
Uploaded: 2026-05-19T13:50:00.000+00:00
Duration: 7 min 47 s
Description: Learn to use the Human Neocortical Neurosolver to link EEG biomarkers to neural circuit mechanisms. This computational protocol enables hypothesis-driven modeling of neurotherapeutic effects on brain dynamics.

ai_agent_IP

ai_agent

maintenance

instantTrialDurationMinutes

isBiopharma

enableNewMyJoveSectionDesign

educationFeatures

seoMetaApi

ai_copilot_institution_blacklist

ai_chat_institution_blacklist

freeAccessWindow

ssoRoleCapture

scheduledMaintenanceBanner

joveLabsFeatures

disable_autoplay_for_bots

extended_trial_duration

5 minutes

All code used to produce the results shown in this protocol can be found at: https://github.com/ntolley/hnn_jove. This work was supported by the Brown Biomedical Innovation to Impact Award, the National Institutes of Health (NIH; <a href="https://www.nih.gov/">https://www.nih.gov</a>; grant numbers U24NS129945 and P50MH109429), and the National Science Foundation (NSF; https://www.nsf.gov; grant number 2424101). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. This work used computational resources supported by the NIHS10 instrumentation grant S10OD036341 (High-Performance Compute Cluster for Brain Science) via the Center for Computation and Visualization (CCV) at Brown University.

<ol>
	<li>Gribkoff VK, Kaczmarek LK. The need for new approaches in CNS drug discovery: Why drugs have failed, and what can be done to improve outcomes. Neuropharmacology. 2017;120:11-19.</li>
	<li>Loo SK, Lenartowicz A, Makeig S. Research review: Use of EEG biomarkers in child psychiatry research&#8212;current state and future directions. J Child Psychol Psychiatry. 2016;57(1):4-17.</li>
	<li>McLoughlin G, Makeig S, Tsuang MT. In search of biomarkers in psychiatry: EEG-based measures of brain function. Am J Med Genet B Neuropsychiatr Genet. 2014;165(2):111-121.</li>
	<li>Douglas RJ, Martin KAC. Neuronal circuits of the neocortex. Annu Rev Neurosci. 2004;27:419-451.</li>
	<li>Harris KD, Shepherd GMG. The neocortical circuit: themes and variations. Nat Neurosci. 2015;18(2):170-181.</li>
	<li>Murakami S, Okada Y. Contributions of principal neocortical neurons to magnetoencephalography and electroencephalography signals. J Physiol. 2006;575(3):925-936.</li>
	<li>Sherman MA, et al. Neural mechanisms of transient neocortical beta rhythms: Converging evidence from humans, computational modeling, monkeys, and mice. Proc Natl Acad Sci U S A. 2016;113(33):E4885-E4894.</li>
	<li>Shin H, et al. The rate of transient beta frequency events predicts behavior across tasks and species. eLife. 2017;6:e29086.</li>
	<li>De Pieri M, et al. Pharmaco-EEG of antipsychotic treatment response: A systematic review. Schizophrenia. 2023;9(1):85.</li>
	<li>Hyun J, Baik M, Kang U. Effects of psychotropic drugs on quantitative EEG among patients with schizophrenia-spectrum disorders. Clin Psychopharmacol Neurosci. 2011;9(2):78-85.</li>
	<li>Jobert M, et al. Guidelines for the recording and evaluation of pharmaco-EEG data in man: The International Pharmaco-EEG Society (IPEG). Neuropsychobiology. 2012;66(4):201-220.</li>
	<li>Mandema JW, Danhof M. Electroencephalogram effect measures and relationships between pharmacokinetics and pharmacodynamics of centrally acting drugs. Clin Pharmacokinet. 1992;23:191-215.</li>
	<li>Leiser SC, Dunlop J, Bowlby MR, Devilbiss DM. Aligning strategies for using EEG as a surrogate biomarker: A review of preclinical and clinical research. Biochem Pharmacol. 2011;81(12):1408-1421.</li>
	<li>Wilson FJ, Danjou P. Early decision-making in drug development: The potential role of pharmaco-EEG and pharmaco-sleep. Neuropsychobiology. 2016;72(3-4):188-194.</li>
	<li>Klumpp H, Shankman SA. Using event-related potentials and startle to evaluate time course in anxiety and depression. Biol Psychiatry Cogn Neurosci Neuroimaging. 2018;3(1):10-18.</li>
	<li>Proudfit GH, et al. Depression and event-related potentials: Emotional disengagement and reward insensitivity. Curr Opin Psychol. 2015;4:110-113.</li>
	<li>Luck SJ, et al. A roadmap for the development and validation of event-related potential biomarkers in schizophrenia research. Biol Psychiatry. 2011;70(1):28-34.</li>
	<li>Salisbury DF, Collins KC, McCarley RW. Reductions in the N1 and P2 auditory event-related potentials in first-hospitalized and chronic schizophrenia. Schizophr Bull. 2010;36(5):991-1000.</li>
	<li>Kang E, et al. Atypicality of the N170 event-related potential in autism spectrum disorder: A meta-analysis. Biol Psychiatry Cogn Neurosci Neuroimaging. 2018;3(8):657-666.</li>
	<li>Modi ME, Sahin M. Translational use of event-related potentials to assess circuit integrity in ASD. Nat Rev Neurol. 2017;13(3):160-170.</li>
	<li>Horvath A, et al. EEG and ERP biomarkers of Alzheimer&#8217;s disease: A critical review. Front Biosci (Landmark Ed). 2018;23:183-220.</li>
	<li>Malver LP, et al. Electroencephalography and analgesics. Br J Clin Pharmacol. 2014;77(1):72-95.</li>
	<li>Preskorn SH, et al. Normalizing effects of EVP-6124 on event-related potentials and cognition: A randomized trial in schizophrenia. J Psychiatr Pract. 2014;20(1):12-24.</li>
	<li>Schwertner A, et al. Effects of subanesthetic ketamine on visual and auditory event-related potentials in humans: A systematic review. Front Behav Neurosci. 2018;12:70.</li>
	<li>Visser S, et al. Dose-dependent EEG effects of zolpidem provide evidence for GABAA receptor subtype selectivity in vivo. J Pharmacol Exp Ther. 2003;304(3):1251-1257.</li>
	<li>Okoroafor F, et al. Neurophysiologic biomarkers of invasive neuromodulation therapy for epilepsy. Neuromodulation: Technol Neural Interface. 2026;29(3):360-375.</li>
	<li>Maki-Marttunen T, et al. Biophysical psychiatry&#8212;how computational neuroscience can help understand the complex mechanisms of mental disorders. Front Psychiatry. 2019;10:534.</li>
	<li>Murray JD, Demirtas M, Anticevic A. Biophysical modeling of large-scale brain dynamics and applications for computational psychiatry. Biol Psychiatry Cogn Neurosci Neuroimaging. 2018;3(9):777-787.</li>
	<li>Cagney DN, et al. The FDA NIH biomarkers, endpoints, and other tools (BEST) resource in neuro-oncology. Neuro Oncol. 2018;20(9):1162-1172.</li>
	<li>Cecchi M, et al. Validation of a suite of ERP and QEEG biomarkers in schizophrenia. Schizophr Res. 2023;254:178-189.</li>
	<li>Dura-Bernal S, et al. NetPyNE, a tool for data-driven multiscale modeling of brain circuits. eLife. 2019;8:e44494.</li>
	<li>Linden H, et al. LFPy: a tool for biophysical simulation of extracellular potentials. Front Neuroinform. 2014;7:41.</li>
	<li>Neymotin SA, et al. Human Neocortical Neurosolver (HNN), a new software tool for interpreting MEG/EEG data. eLife. 2020;9:e51214.</li>
	<li>Sanz Leon P, et al. The Virtual Brain: a simulator of primate brain network dynamics. Front Neuroinform. 2013;7:10.</li>
	<li>Hamalainen M, et al. Magnetoencephalography&#8212;theory, instrumentation, and applications. Rev Mod Phys. 1993;65(2):413-497.</li>
	<li>Ikeda H, Wang Y, Okada YC. Origins of the somatic N20 and high-frequency oscillations evoked by trigeminal stimulation in the piglets. Clin Neurophysiol. 2005;116(4):827-841.</li>
	<li>Okada YC, Wu J, Kyuhou S. Genesis of MEG signals in CNS structure. Electroencephalogr Clin Neurophysiol. 1997;103(4):474-485.</li>
	<li>Bush PC, Sejnowski TJ. Reduced compartmental models of pyramidal cells. J Neurosci Methods. 1993;46(2):159-166.</li>
	<li>Jones SR, et al. Neural correlates of tactile detection. J Neurosci. 2007;27(40):10751-10764.</li>
	<li>Jones SR, et al. Quantitative analysis of MEG mu rhythm. J Neurophysiol. 2009;102(6):3554-3572.</li>
	<li>Law RG, et al. Thalamocortical mechanisms regulating beta events. Cereb Cortex. 2022;32(4):668-688.</li>
	<li>Fernandez Pujol C, Blundon EG, Dykstra AR. Laminar specificity of the auditory perceptual awareness negativity: A biophysical modeling study. PLoS Comput Biol. 2023;19(6):e1011003.</li>
	<li>Kohl C, Parviainen T, Jones SR. Neural mechanisms underlying auditory evoked responses. Brain Topogr. 2022;35(1):19-35.</li>
	<li>Lankinen K, Ahveninen J, Jas M, Raij T, Ahlfors SP. Neuronal modeling of cross-sensory visual evoked magnetoencephalography responses in the auditory cortex. J Neurosci. 2024;44(17):e1119232024.</li>
	<li>Kaplan L, et al. Modeling cortical dynamics using HNN [poster presentation]. Presented at: Society for Neuroscience Annual Meeting; San Diego, CA, USA; 2025.</li>
	<li>Diesburg DA, Wessel JR, Jones SR. Biophysical modeling of ERP generation. J Neurosci. 2024;44(20).</li>
	<li>Bonaiuto JJ, et al. Laminar dynamics of beta bursts. Neuroimage. 2021;242:118479.</li>
	<li>Ferrante M, Blackwell KT, Migliore M, Ascoli GA. Computational models of neuronal biophysics. Curr Med Chem. 2008;15(24):2456-2471.</li>
	<li>Geerts H, et al. Quantitative systems pharmacology for neuroscience drug discovery. CPT Pharmacometrics Syst Pharmacol. 2020;9(1):5-20.</li>
	<li>Geerts H, et al. Computational neuroscience and systems pharmacology. J Pharmacokinet Pharmacodyn. 2024;51(5):563-573.</li>
	<li>Kappenman ES, Luck SJ. ERP components: brainwave recordings. Oxford Handbook ERP Components. 2012;1:3-30.</li>
	<li>Goncalves PJ, et al. Training neural density estimators. eLife. 2020;9:e56261.</li>
	<li>Papamakarios G, et al. Normalizing flows for probabilistic modeling. J Mach Learn Res. 2021;22(57):1-64.</li>
	<li>Pastore M, Calcagni A. Measuring distribution similarities. Front Psychol. 2019;10.</li>
	<li>Tolley N, et al. Estimating parameters in neural models with SBI. PLoS Comput Biol. 2024;20(2):e1011108.</li>
	<li>Gramfort A, et al. MEG and EEG analysis with MNE-Python. Front Neuroinform. 2013;7:267.</li>
	<li>Sliva DD, et al. Transcranial stimulation and EEG perception. Front Psychol. 2018;9:2117.</li>
	<li>Thorpe RV, et al. Distinct neocortical mechanisms underlie human SI responses to median nerve and laser evoked peripheral activation. bioRxiv. 2021; Available at: https://doi.org/10.1101/2021.10.11.463545.</li>
	<li>Delorme A, Makeig S. EEGLAB toolbox. J Neurosci Methods. 2004;134(1):9-21.</li>
	<li>Oostenveld R, et al. FieldTrip software. Comput Intell Neurosci. 2011;2011:156869.</li>
	<li>Puce A, H&#228;m&#228;l&#228;inen MS. A review of issues related to data acquisition and analysis in EEG/MEG studies. Brain Sci. 2017;7(6):58.</li>
	<li>Tejero-Cantero A, et al. sbi: toolkit for simulation-based inference. J Open Source Softw. 2020;5(52):2505.</li>
	<li>Lee S, Jones SR. Distinguishing mechanisms of gamma frequency oscillations in human current source signals using a computational model of a laminar neocortical network. Front Hum Neurosci. 2013;7:869.</li>
	<li>Szul MJ, et al. Beta burst waveform motifs. Prog Neurobiol. 2023;228:102490.</li>
	<li>Hashemi M, et al. Bayesian virtual epileptic patient. Neuroimage. 2020;217:116839.</li>
	<li>Billeh YN, et al. Systematic integration of structural and functional data into multi-scale models of mouse primary visual cortex. Neuron. 2020;106(3):388-403.</li>
	<li>Borges FS, Moreira JV, Takarabe LM, Lytton WW, Dura-Bernal S. Large-scale biophysically detailed model of somatosensory thalamocortical circuits in NetPyNE. Front Neuroinform. 2022;16:884245.</li>
	<li>Markram H, et al. Reconstruction of neocortical microcircuitry. Cell. 2015;163(2):456-492.</li>
	<li>Hagen E, N&#230;ss S, Ness TV, Einevoll GT. Multimodal modeling of neural network activity: computing LFP, ECoG, EEG, and MEG signals with LFPy 2.0. Front Neuroinform. 2018;12:92.</li>
	<li>Geerts H. Of mice and men in CNS drug discovery. CNS Drugs. 2009;23(11):915-926.</li>
	<li>Nestler EJ, Hyman SE. Animal models of neuropsychiatric disorders. Nat Neurosci. 2010;13(10):1161-1169.</li>
</ol>

N.T. and S.R.J. are co-inventors on a pending patent application related to methods for parameter inference in neural circuit models described in this work. The remaining authors declare no conflicts of interest.

Computational neural modeling of EEG biomarkers may allow deeper insight into how CNS therapeutics reconfigure neural circuits and provide predictions on the biological processes underlying therapeutic effects. The workflow presented here demonstrates how a commonly measured EEG biomarker, auditory ERPs, together with biophysical modeling using the HNN, can be used as a window into the mechanisms by which a drug impacts neural activity. By linking macroscale EEG measurements to underlying cellular and circuit-level processes, this protocol provides a structured and hypothesis-driven framework for mechanistic interpretation. Importantly, the approach is not restricted to ERPs and can be extended to investigate other local EEG signals, including low-frequency neural oscillations40,63 and transient spectral events7,47,64, thereby broadening its applicability across electrophysiological biomarkers and experimental paradigms.
Compared to other frameworks for neural modeling of EEG, HNN offers a balance of model complexity and computational efficiency that is particularly advantageous for iterative hypothesis testing. For example, The Virtual Brain enables simulation of large-scale brain networks that generate spatiotemporal EEG signals34,65. However, to achieve whole-brain modeling, neural activity is represented using reduced mathematical formulations, which eliminate detailed cellular features such as pyramidal neuron morphology and limit the ability to directly link model parameters to cellular mechanisms of drug action. Conversely, large-scale morphologically and physiologically detailed models can simulate EEG signals with high biological realism66,67,68,69, but at a substantial computational cost, often requiring several hours of computation to simulate only a few seconds of neural activity. This computational burden can limit accessibility and slow the iterative process required for hypothesis generation and testing. HNN occupies an intermediate position (Figure 2), enabling simulation of localized neocortical circuits with sufficient biological detail to generate cell- and circuit-level predictions while maintaining computational efficiency (i.e., simulations on the order of seconds), making it well suited for integration into experimental workflows.
Despite these advantages, several limitations should be considered when applying EEG and biophysical neural modeling to study brain disease and drug mechanisms. The biophysical cell and circuit properties that generate EEG signals do not capture the full spectrum of biological processes affected by pharmacological interventions. For example, systemic or immunological responses may not directly influence EEG signals and therefore may not be reflected in the modeled outputs. In addition, mechanistic hypotheses are often derived from animal studies, which may not fully translate to human brain function, particularly in neuropsychiatric disorders where clinical outcomes are based on behavioral and cognitive assessments70,71. Another important challenge is distinguishing between acute and chronic pharmacological effects. While acute drug-receptor interactions are relatively well characterized, the long-term adaptations induced by sustained drug exposure are less well understood and may not be fully captured in current modeling frameworks. Furthermore, the HNN model represents a single localized canonical neocortical network, whereas neurotherapeutics and CNS diseases often exert distributed effects across multiple brain regions. Although influences from other regions can be approximated through changes in the timing and strength of exogenous inputs, direct empirical characterization of these upstream or downstream circuits is often limited, which constrains model interpretation.
Parameter degeneracy represents a fundamental challenge in all biophysical neural models, as multiple parameter configurations can produce similar model outputs. In this protocol, SBI is used to address this issue by estimating distributions of parameters that generate ERP waveforms consistent with empirical data (Figure 10). This approach enables quantification of uncertainty in model parameters, providing a more robust framework for mechanistic interpretation than single-point estimates. However, for computational tractability, SBI is applied to a limited subset of parameters corresponding to hypothesized drug mechanisms, and assumptions regarding non-estimated parameters can influence the resulting network dynamics. Expanding inference to larger parameter spaces can be achieved using approaches such as sequential neural posterior estimation, which iteratively refines parameter estimates and enables exploration of higher-dimensional parameter distributions52 (&#62;10 dimensions). In addition to probabilistic inference, incorporating independent experimental constraints can further reduce parameter uncertainty and improve the specificity of model predictions. Because EEG signals primarily reflect coordinated activity across cortical layers, complementary techniques such as invasive laminar electrophysiology&#8212;including measurements of cell spiking, LFP, and CSD&#8212;provide valuable information for constraining model solutions and refining mechanistic hypotheses.
Successful application of this protocol depends on careful execution of several critical steps. After identifying an ERP biomarker and installing the modeling framework (Steps 1&#8211;2), the primary requirement is achieving successful outcomes at each stage of the workflow (Figure 3). In Steps 3&#8211;5, this involves selecting and refining hypothesized parameters that can be manually tuned or optimized to achieve a close fit between simulated and empirical pre-treatment and post-treatment ERPs. If a satisfactory fit cannot be obtained, alternative parameters should be explored and iteratively tested. While repeated failures are unlikely given prior demonstrations of HNN&#8217;s ability to reproduce ERP features, persistent failure may indicate the need to modify the default network model or incorporate additional biophysical detail. Step 6 requires careful configuration of SBI, including appropriate selection of parameter ranges, summary statistics, and training parameters to ensure accurate estimation of parameter distributions. Upon successful completion of Step 6, the protocol yields both model-based predictions and associated uncertainty estimates. Step 7 is critical for validating these predictions, although the specific validation strategies depend on available experimental modalities. Potential validation approaches include laminar electrophysiological recordings to assess layer- and cell-specific spiking activity and LFP/CSD signals7, layer-resolved MEG/EEG measurements, magnetic resonance spectroscopy or positron emission tomography for assessing neurotransmitter systems, and diffusion tensor imaging for evaluating structural connectivity such as thalamocortical pathways.
Troubleshooting and customization are integral to adapting the protocol to different datasets and experimental contexts, particularly in Steps 3&#8211;6 where model parameters are fit to empirical data. Parameter optimization (Steps 4&#8211;5) may fail to converge to a high correlation (Corr &#62; 0.95), in which case several adjustments can be made. These include modifying optimizer hyperparameters (e.g., increasing population size in the CMA-ES solver to improve robustness, with increased computational cost), refining scaling and smoothing parameters (e.g., testing smoothing values between 5 and 60 ms), and expanding the range of exogenous drive parameters or introducing additional drives to better capture waveform features. In some cases, optimized simulations may achieve high correlation while failing to capture lower-amplitude ERP features such as the P1 component; this can be addressed by applying stricter loss thresholds or weighting specific time windows to emphasize these features during optimization. For SBI (Step 6), failure of PPCs indicates that the simulated waveforms do not adequately reproduce the empirical data (Supplementary Figure 1). In such cases, the prior parameter distributions should be revised by expanding parameter ranges or including additional parameters, and the size of the training dataset may need to be increased. Additional improvements can be achieved by modifying summary statistics or selecting alternative SBI architectures. Finally, when validation in Step 7 fails, the default HNN network may require modification to include additional or alternative circuit elements. The modular design of HNN supports such extensions, allowing modification of synaptic connectivity and cellular properties through the GUI, and more advanced structural changes through the Python interface. For example, prior work has modified the default model to incorporate more detailed interneuron connectivity in frontal cortex46, resulting in new testable predictions. The open framework of HNN facilitates sharing and reuse of expanded models, supporting continued refinement and validation across experimental contexts.

Central nervous system (CNS) therapeutic development faces unique challenges, with approval rates lower than in other disease areas, highlighting the need for innovative methodological approaches1, particularly those that can uncover treatment-related effects on brain dynamics. A well-established approach to study the effect of therapeutics on neural activity is electroencephalography (EEG)2,3. EEG provides a signature of in vivo circuit-level brain dynamics and offers strong translational potential from rodent models to human trials, as the neural circuitry generating EEG signals shows homology across species4,5,6,7,8. In pharmaceutical development, EEG can serve multiple roles, including providing translational readouts between animal and human studies, evaluating drug safety, guiding compound selection, informing dose-response relationships, assessing proof-of-mechanism in early clinical phases, and enabling clinical trial stratification and cohort enrichment9,10,11,12,13,14. Despite these advantages, the interpretation of EEG signals remains a major challenge, particularly when attempting to link observed changes to underlying neural mechanisms.
A robust EEG biomarker used in CNS drug discovery is the event-related potential (ERP). ERPs reflect time-locked sensory-evoked brain activity and have been widely used to study neurodevelopmental and neuropsychiatric disorders, including depression15,16, schizophrenia17,18, autism spectrum disorder19,20, and Alzheimer’s disease21. ERPs are also used to assess treatment effects and dose ranges on brain circuits22,23,24,25, where normalization toward healthy responses may indicate therapeutic efficacy26. However, a key limitation of ERPs and other EEG biomarkers (e.g., brain oscillations) is that their associations with disease states or drug effects are largely correlational. While statistical analyses can identify relationships between biomarkers and outcomes, they do not provide mechanistic insight into how specific neural circuit elements generate these signals. The causal contributions of specific cell types and circuit mechanisms therefore remain unclear. Understanding the cellular and circuit origins of EEG signals could considerably enhance their value by linking observed signatures to underlying physiology27,28. In this manuscript, the term EEG “biomarker” refers to measurable changes in EEG signals following therapeutic intervention, consistent with the Food and Drug Administration–National Institutes of Health Biomarkers, EndpointS, and other Tools (FDA–NIH BEST) framework definition29, rather than implying formal qualification for a specific clinical use30.
While invasive electrophysiological recordings can provide detailed cell- and circuit-level insights, these approaches are largely restricted to animal models and are difficult to translate directly to human studies. Alternative approaches, such as inverse modeling techniques, can estimate source activity from EEG signals but often lack explicit mechanistic representations of underlying neural circuitry. Biophysical simulations offer a complementary framework by modeling the physical processes through which neural circuits generate measurable EEG signals31,32,33,34 (Figure 1). Compared to purely statistical biomarker analyses or inverse methods without mechanistic grounding, biophysical modeling enables direct testing of hypotheses linking neural circuit dynamics to observed electrophysiological signals.
<img alt="EEG biomarker identification process; inverse modeling; mechanistic hypothesis; parametric fitting diagram." class="xfigimg" src="/files/ftp_upload/70618/70618fig1.jpg" data-altupdatedat="1788282455000" data-alt="Figure 1"> 
Figure 1. Biophysical modeling to develop and test mechanistic hypotheses underlying pharmacological electroencephalography (EEG) biomarkers. (A) Identification of an EEG biomarker based on differences in brain signals between conditions. An example is an auditory event-related potential (ERP) that is reduced in the post-treatment condition (red) relative to the pre-treatment condition (blue). (B) Biophysical modeling enables testing of mechanistic hypotheses explaining how EEG biomarkers arise and change with pharmacological intervention. Hypotheses are formulated regarding drug-induced changes in neural activity, and corresponding model parameters are identified. (C) The default Human Neocortical Neurosolver (HNN) model is used as a starting point to test hypotheses by manually modifying model parameters or applying automated optimization and inference algorithms. Differences in parameter values between pre-treatment and post-treatment conditions correspond to model-based predictions. <a href="https://www.jove.com/files/ftp_upload/70618/70618fig1large.jpg" target="_blank">Please click here to view a larger version of this figure.</a>
This protocol uses the Human Neocortical Neurosolver (HNN), an open-source biophysical modeling framework, to link ERP biomarkers of treatment-related effects to their underlying cell- and circuit-level mechanisms33 (Figure 2). HNN is based on the principle that synchronous intracellular current flow in aligned pyramidal neuron dendrites generates the primary current dipoles underlying EEG signals6,35,36,37. The model represents a canonical neocortical column consisting of excitatory pyramidal neurons and inhibitory interneurons distributed across cortical layers 2/3 and 5. The default HNN network includes 100 pyramidal neurons and 33 inhibitory neurons per layer, forming a reduced yet biologically grounded representation of cortical circuitry. Pyramidal neurons are modeled with multicompartment dendritic structures to capture key morphological features38, while inhibitory neurons are represented as single compartments due to their limited contribution to extracellular currents33. Synaptic interactions include excitatory α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) and N-methyl-D-aspartate (NMDA) receptors, and inhibitory gamma-aminobutyric acid type A and gamma-aminobutyric acid type B (GABAB) receptors, with all neurons incorporating active ionic conductances governed by Hodgkin–Huxley dynamics.
<img alt="Neural network diagram; local/proximal/distal drives in cortical layers; neural model; brain connectivity." class="xfigimg" src="/files/ftp_upload/70618/70618fig2.jpg" data-altupdatedat="1788282455000" data-alt="Figure 2"> 
Figure 2. Schematic of the HNN model. Visualization of the major components of the HNN model, including local network connections between excitatory and inhibitory neurons, and exogenous input pathways termed “proximal drive” and “distal drive.” <a href="https://www.jove.com/files/ftp_upload/70618/70618fig2large.jpg" target="_blank">Please click here to view a larger version of this figure.</a>
Neural activity in HNN is driven by exogenous inputs representing feedforward and feedback pathways. Feedforward “proximal” drives correspond to inputs from lemniscal thalamus which target proximal dendrites, while feedback “distal” drives represent cortico-cortical and non-lemniscal thalamic inputs targeting distal dendrites. These inputs are modeled as trains of action potentials that evoke synaptic currents and generate intracellular current flow along pyramidal neuron dendrites. The resulting population-level current dipole is expressed in nanoampere-meters, enabling direct comparison with orientation-constrained source-localized EEG or magnetoencephalography (MEG) data. The default parameterization of HNN is informed by empirical data from somatosensory cortex studies39,40,41 and has been successfully applied to auditory42,43,44, visual45, and frontal cortical signals46, with model-derived predictions validated in subsequent experimental studies7,41,47.
HNN simulations can be applied at multiple stages of pharmaceutical research and development, including target validation, comparison of drug mechanisms of action, dose optimization, and hypothesis generation for follow-up experiments14,48,49,50. This enables users to incorporate mechanistic modeling into practical research workflows, supporting the generation and testing of hypotheses about how neurotherapeutics influence neural circuits. In this protocol, we focus on the early P1, N1, and P2 components of auditory ERPs, as these features are well characterized and provide constraints for hypothesis-driven modeling51. While the focus is on drug-induced changes, the approach can be extended to other neurotherapeutic interventions, such as brain stimulation or behavioral training, as well as to studies of CNS disorders.
The use of HNN follows an iterative modeling framework in which model structure and parameters are initially constrained by existing data and then refined through comparison with empirical observations. Large-scale neural models contain many parameters, but only a subset—referred to as parameters of interest—are adjusted to test specific hypotheses. These parameters are not selected arbitrarily; rather, they are chosen based on prior experimental evidence and literature describing potential mechanisms of action of the neurotherapeutic. In this protocol, parameters related to exogenous input timing and strength, local inhibitory connectivity, and dendritic ion channel conductances are selected as examples of biologically interpretable variables that may be influenced by neurotherapeutics.
Beginning with a default model, users first fit parameters to pre-treatment ERP data using a combination of manual tuning and automated optimization. Manual tuning adjusts global scaling and input parameters to approximate the empirical waveform, providing an intuitive understanding of how parameter changes affect model output. Automated methods such as covariance matrix adaptation evolution strategy (CMA-ES), Bayesian optimization, and constrained optimization by linear approximation are then used to refine parameter values and improve the fit. Once a pre-treatment model is established, parameters hypothesized to account for post-treatment changes are adjusted to fit post-treatment ERP data.
To address uncertainty in parameter estimation, simulation-based inference (SBI) is used to estimate distributions of parameter values that reproduce the observed data52,53. SBI accounts for the possibility that multiple parameter combinations can produce similar outputs and enables quantification of parameter uncertainty. Differences between pre-treatment and post-treatment parameter distributions can be assessed using an overlap index (OVL)54,55, providing insight into potential mechanisms of action.
A key advantage of this approach is that fitting the model to a specific data modality generates predictions across multiple scales of neural activity, including cell spiking, layer-specific local field potentials (LFPs), and current source density (CSD). These predictions provide targets for experimental validation using complementary techniques. If predictions are not supported by empirical data, the model can be updated by incorporating new constraints, forming an iterative cycle of hypothesis generation, testing, and refinement (Figure 3).
<img alt="EEG ERP workflow diagram; treatment-induced biomarker identification; model fit quantification." class="xfigimg" src="/files/ftp_upload/70618/70618fig3v4.jpg" data-altupdatedat="1788282455000" data-alt="Figure 3"> 
Figure 3. Iterative workflow for developing and testing ERP biomarker predictions with HNN. The workflow corresponds to the protocol steps. Identification of an EEG biomarker and initialization of the default HNN model are shown in red (Steps 1–2). Manual tuning and optimization are used to fit model parameters to pre-treatment and post-treatment ERP signals (purple; Steps 3–5). Uncertainty quantification using simulation-based inference (SBI) is shown in green (Step 6). Model predictions are then examined and compared with experimental data to validate or further constrain the model (orange; Step 7). <a href="https://www.jove.com/files/ftp_upload/70618/70618fig3largev2.jpg" target="_blank">Please click here to view a larger version of this figure.</a>
This protocol is designed for use with orientation-constrained, source-localized EEG or MEG data collected during evoked response paradigms. Standard preprocessing and source localization methods (e.g., minimum norm estimation [MNE]-Python56) can be used to generate the required input data. Source-level signals expressed in nanoampere-meters are directly comparable to HNN outputs. For fast sensory responses, source- and sensor-level signals are often highly similar, allowing insights from source-localized modeling to inform interpretation of sensor-level EEG data57,58.

<table><tbody><tr><td>Anaconda Python</td><td>Anaconda, Inc.</td><td>N.A.</td><td>Python distribution; Python version &amp;ge;3.9 and &amp;lt;3.14</td></tr><tr><td>Computer workstation</td><td>N.A.</td><td>N.A.</td><td>Operating system: Windows &amp;ge;10, Linux, or macOS. Minimum recommended hardware: &amp;ge;16 GB RAM, &amp;ge;8 CPU cores</td></tr><tr><td>EEGLAB</td><td>EEGLAB Developers</td><td>N.A.</td><td>Optional MATLAB-based toolbox for EEG preprocessing and ERP analysis</td></tr><tr><td>FieldTrip</td><td>Donders Institute for Brain, Cognition and Behaviour, Radboud University</td><td>N.A.</td><td>Optional MATLAB-based toolbox for EEG/MEG analysis</td></tr><tr><td>Human Neocortical Neurosolver (HNN-core)</td><td>HNN Developers</td><td>N.A.</td><td>Biophysical neural modeling software; version &amp;ge;0.6.0 used in this study</td></tr><tr><td>MATLAB</td><td>MathWorks</td><td>N.A.</td><td>Required to run EEGLAB and FieldTrip (if used)</td></tr><tr><td>MNE-Python</td><td>MNE Developers</td><td>N.A.</td><td>Used for EEG preprocessing and source localization</td></tr><tr><td>NumPy</td><td>NumPy Developers</td><td>N.A.</td><td>Used for numerical computations and random number generation</td></tr><tr><td>Pixi (package/environment manager)</td><td>Prefix.dev</td><td>N.A.</td><td>Used for managing dependencies in the associated code repository</td></tr><tr><td>PyTorch</td><td>PyTorch Developers</td><td>N.A.</td><td>Used for training SBI neural networks and setting random seeds</td></tr><tr><td>SBI (Simulation-Based Inference) package</td><td>SBI Developers</td><td>N.A.</td><td>Python package for parameter inference and uncertainty quantification</td></tr><tr><td>Windows Subsystem for Linux (WSL2)</td><td>Microsoft Corporation</td><td>N.A.</td><td>Required only for Windows-based installations</td></tr></tbody></table>

a protocol for uncovering neural mechanisms of neurotherapeutic effects on electroencephalography using the human neocortical neurosolver

A Protocol For Uncovering Neural Mechanisms Of Neurotherapeutic Effects On Electroencephalography Using The Human Neocortical Neurosolver

Electroencephalography (EEG) and electrophysiological methods provide millisecond-resolution biomarkers for central nervous system disorders and are widely used to assess treatment-related effects. However, limited understanding of the neural mechanisms generating these biomarkers impedes the development of diagnostics and therapeutics based on these signals. The Human Neocortical Neurosolver (HNN) is an open-source biophysical modeling software that links localized EEG biomarkers to their multiscale neural generators. This protocol demonstrates a hypothesis-driven workflow using HNN to test neural mechanisms of neurotherapeutic-induced EEG biomarkers by optimizing model parameters to achieve a fit between simulated and empirical current source waveforms. Corresponding multiscale cell- and circuit-level activity can then be visualized and quantified, providing validation targets for model predictions in follow-up empirical studies. An example is provided demonstrating how to examine the neural mechanisms underlying early event-related potential components of an auditory evoked response (P1, N1, and P2), and to assess changes following neurotherapeutic-induced modifications in neural circuit activity. This protocol enables the design of simulation experiments to generate testable predictions linking EEG biomarkers to underlying neural circuit mechanisms. A similar workflow can be applied to study disease mechanisms or other therapeutic interventions.

This section presents a scenario in which a neurotherapeutic with an unknown mechanism of action is investigated using the HNN modeling software. The goal is to use pre-treatment and post-treatment EEG signals to generate predictions on how the neurotherapeutic alters neural circuits. The results are presented for demonstration purposes to illustrate how HNN modeling can be applied to investigate neurotherapeutic mechanisms.
Developing mechanistic hypotheses underlying EEG ERP biomarkers (Step 1)
In this example, a hypothetical sensory ERP paradigm is used to examine how the neurotherapeutic alters the signal (Step 1). Figure 1A shows a pre-treatment auditory ERP (blue) alongside a hypothetical post-treatment ERP (red; see also Figure 9). The pre-treatment auditory ERP is experimentally recorded source-localized data from Kohl et al.43, and the hypothetical post-treatment ERP is generated by scaling the pre-treatment waveform with a Gaussian-tapered window. As shown, the hypothetical neurotherapeutic produces a large decrease in the magnitude of the P1, N1, and P2 components relative to the pre-treatment ERP.
Note that in Kohl et al.43, from which the pre-treatment ERP data was obtained, the HNN simulations used a model in which pyramidal neurons were enhanced with more realistic calcium channel dynamics than in the default HNN model. As a result, the simulation results in Kohl et al.43, differ slightly from those shown here. The Kohl et al. 2020 model (and other updated HNN models) can be accessed through the Python API (https://jonescompneurolab.github.io/hnn-core/stable/generated/hnn_core.calcium_model.html#hnn_core.calcium_model). Access to such expanded models through the GUI is currently under development.
Next, identify model parameters representing treatment-related effects (i.e., parameters of interest) that are hypothesized to explain how the neurotherapeutic reduces the P1, N1, and P2 magnitudes (Steps 1.6–1.8). Broad categories of candidate neural mechanisms (and corresponding model parameters) include timing of exogenous synaptic inputs, local neuronal ion channel conductances, local synaptic connectivity, and exogenous synaptic connectivity (Figure 1B). In this example, candidate mechanisms from each category are evaluated using HNN to assess how changes in these parameters impact the simulated ERP.
Parameters of interest
<ol>
	<li>Standard deviation of the first (thalamocortical) proximal drive (i.e., thalamocortical synchrony), representing variability in synchronization of initial feedforward sensory inputs.</li>
	<li>Muscarinic potassium (Km) channel conductance in layer 5 (L5) pyramidal neurons, controlling neuronal excitability, such that excitability decreases as conductance increases.</li>
	<li>Local GABAB receptor strength, corresponding to a slow inhibitory synapse delivered by interneurons to all cells in the local network.</li>
	<li>Conductance strength of the feedback (corticocortical) distal drive, representing the strength of the ~100 ms sensory-evoked feedback input to AMPA and NMDA synapses in supragranular layers.</li>
</ol>
Establishing the pre-treatment ERP model fit (Steps 3–4)
Simulate the pre-treatment ERP by following Steps 3–4 (final pre-treatment simulation shown in Figure 8C). A successful result is indicated by a close match between simulated and empirical waveforms, as quantified by a high correlation coefficient and low RMSE.
Establishing the post-treatment ERP model fit (Step 5)
Use the pre-treatment ERP model as a starting point and apply manual tuning and parameter optimization to determine whether the parameters of interest can reproduce the empirical post-treatment ERP. A successful fit indicates that the hypothesized parameters are sufficient to explain treatment-related changes in the ERP waveform.
Uncertainty quantification with SBI (Step 6)
Due to parameter degeneracy inherent in biophysical models, uncertainty quantification using SBI (Step 6) is essential for making predictions about pre- to post-treatment parameter changes. A critical prerequisite for SBI is achieving accurate fits to pre-treatment and post-treatment ERPs (Steps 3–5). If accurate fits are not achieved, posterior samples generated by SBI may not reproduce the empirical waveforms, leading to unreliable predictions.
If a successful fit cannot be achieved in Steps 3–5, revise the selection of parameters of interest and their prior ranges before applying SBI.
In this example, SBI is applied only to the four post-treatment parameters of interest, while all other parameters are held fixed. Although applying SBI to a larger parameter set can improve robustness, it substantially increases computational cost (see Discussion).
SBI is used to estimate full parameter distributions that generate simulated ERPs closely matching target waveforms. Briefly, SBI is a Bayesian inference approach that trains a neural network to map model outputs to distributions of model parameters52,53,55. The trained network is then applied to empirical waveforms to infer parameter distributions consistent with the data. This requires prior hypotheses on parameter ranges.
In this example, a uniform prior distribution is defined over the four parameters of interest: thalamocortical synchrony, pyramidal neuron dendritic Km conductance, local GABAB conductance, and corticocortical feedback strength. Prior bounds are defined as scalar multiples of default values: 0–5× for thalamocortical synchrony and 10−1–101× for the remaining parameters.
Figure 10A shows the resulting parameter distributions for pre-treatment and post-treatment ERPs, visualized using a pairplot. Diagonal panels display univariate distributions, while off-diagonal panels display bivariate relationships. Mechanistic predictions correspond to parameters with strongly separated distributions between conditions.
Inspection of the univariate distributions shows that thalamocortical synchrony exhibits the greatest pre- to post-treatment separability (lowest OVL of 0.07) and increases following treatment (Figure 10A(iii), red). This indicates that the HNN framework predicts modulation of thalamocortical synchrony as a potential mechanism of action.
Posterior predictive validation
Validate inferred parameter distributions using a PPC. Generate independent parameter samples from the posterior distribution and simulate corresponding ERPs. A successful PPC is indicated when simulated waveforms closely match the empirical ERP.
As shown in Figure 10B and Figure 10C, both pre-treatment (Figure 10B, blue) and post-treatment (Figure 10C, red) waveforms closely match simulations generated from posterior samples (black), with correlation coefficients of 0.99 and 0.96, respectively (averaged over 10 independent samples). These results confirm that the inferred parameter distributions produce accurate waveform reconstructions.
An example of an unsuccessful PPC is provided in Supplementary Figure 1. The example follows the same structure as Figure 10 and uses the same trained SBI network; however, an alternate post-treatment waveform is used that is not well represented in the training set (e.g., ERP waveforms with a positive deflection at the N1 latency). The failed PPC is indicated in Supplementary Figure 1C, where the correlation coefficient is low (e.g., Corr &lt; 0.95). Notably, the posterior distribution in Supplementary Figure 1A shows highly separated parameter distributions. Without performing a PPC, these results could be misinterpreted as meaningful differences between pre-treatment and post-treatment conditions. This example highlights the importance of conducting a PPC alongside interpretation of posterior distributions, as results from a failed PPC are unreliable and should not be further analyzed.
Model examination and validation (Step 7)
Using the HNN model, it is possible to directly inspect and visualize cell- and circuit-level activity, such as spiking, underlying each ERP simulation (Step 7.2.2). Figure 11A and Figure 11B shows simulated ERPs sampled from pre-treatment and post-treatment parameter distributions, alongside corresponding cell-specific spiking activity (Figure 11C and Figure 11D).
<img alt="Neural activity comparison; graphs show dipole moment (nAm) pre and post-treatment; spike raster plots." class="xfigimg" src="/files/ftp_upload/70618/70618fig11.jpg" data-altupdatedat="1788282455000" data-alt="Figure 11"> 
Figure 11. Cell-level spiking activity underlying EEG biomarker generation. (A) Pre-treatment ERP (blue) with a single posterior predictive simulation (black). (B) Post-treatment ERP (red) with a corresponding posterior predictive simulation (black). (C) Simulated spiking activity underlying the pre-treatment ERP. (D) Simulated spiking activity underlying the post-treatment ERP. <a href="https://www.jove.com/files/ftp_upload/70618/70618fig11large.jpg" target="_blank">Please click here to view a larger version of this figure.</a>
The waveforms are visualized without smoothing to emphasize the contribution of spike timing to the current dipole. In experimental EEG signals, large neuronal populations produce spatially averaged signals that appear smoother. Because HNN simulates a smaller population (200 pyramidal neurons), smoothing is used to approximate larger-scale activity (&gt;100,000 neurons).
A notable difference between conditions is reduced spiking activity in L5 pyramidal neurons following treatment (Figure 11C and Figure 11D, red dot). Note that Figure 11 shows a single sample from the posterior distribution; multiple samples should be analyzed to generate robust predictions. These results demonstrate that the hypothetical neurotherapeutic alters multiscale circuit activity, resulting in decreased P1–N1–P2 amplitudes.
Predictions such as these can be directly tested through invasive electrophysiology (e.g., high-density laminar probe recordings) or other imaging modalities (Step 7.3). Newly acquired data can then be used to further constrain model predictions. While this protocol focuses on fitting macroscale EEG data to infer microcircuit activity, the framework can also be applied in reverse by fitting microcircuit data (e.g., spiking, LFP/CSD) to infer macroscale EEG signals.
Supplementary Figure 1. Example of a failed posterior predictive check in the SBI workflow. Plots are organized identically to Figure 10. The pre-treatment data (blue) is identical to Figure 10. The hypothetical post-treatment data was generated identically as before (waveform multiplied with a Gaussian-tapered window), but transformed to produce a positive peak that is not well represented in the training set of HNN simulations. (A) Pairplot visualization of parameter distributions estimated using SBI. Diagonal panels (i–iv) show univariate distributions for individual parameters, including (i) thalamocortical synchrony, (ii) dendritic Km conductance, (iii) GABAB conductance, and (iv) corticocortical feedback strength. Distributions for pre-treatment (blue) and post-treatment (red) conditions demonstrate high separability for all parameters (OVL &lt; 0.1). Off-diagonal panels show bivariate relationships between parameters. (B) Posterior predictive check (PPC) for pre-treatment ERP; simulated waveforms (black) closely match empirical data (blue). (C) PPC for post-treatment ERP; simulated waveforms (black) are highly dissimilar to the empirical data (red), with Corr &lt; 0.95 indicating a failed PPC.<a href="https://www.jove.com/files/ftp_upload/70618/JoVE70618_Supplementary%20Figure%201.pdf" target="_blank">Please click here to download this file.</a>

This protocol demonstrates how physics-based neural simulations can be used to interpret electrophysiological biomarkers of neurotherapeutics and uncover their effect on neural circuits, providing a mechanistically grounded approach for neurotherapeutic development.

Electroencephalography (EEG) and electrophysiological methods provide millisecond-resolution biomarkers for central nervous system disorders and are ...

a-protocol-for-uncovering-neural-mechanisms-neurotherapeutic-effects

Research

JoVE Journal

Neuroscience

We present a method for collecting electrocorticographic signals for research purposes from humans who are undergoing invasive epilepsy monitoring.  We show how to use the BCI2000 software platform for data collection, signal processing and stimulus presentation.  Specifically, we demonstrate SIGFRIED, a BCI2000-based tool for real-time functional brain mapping.

Recording Human Electrocorticographic (ECoG) Signals for Neuroscientific Research and Real-time Functional Cortical Mapping

A procedure is described for manipulating the activity of cerebral cortical pyramidal neurons optogenetically while the electroencephalogram, electromyogram, and cerebral lactate concentration are monitored. Experimental recordings are performed on cable-tethered mice while they undergo spontaneous sleep/wake cycles. Optogenetic equipment is assembled in our laboratory; recording equipment is commercially available.

Simultaneous Electroencephalography, Real-time Measurement of Lactate Concentration and Optogenetic Manipulation of Neuronal Activity in the Rodent Cerebral Cortex

This video demonstrates the process of recording cortical and subcortical neuronal activity using electroencephalography (EEG) and deep brain stimulation (DBS) electrodes during an attention-based visual stimulation test.

Recording Cortical and Subcortical Neuronal Activity Using Electrode Systems

The present protocol aims at assessing cognitive-emotional functions in the basal ganglia by simultaneous neurophysiological recording of local field potentials and non-invasive brain cortical activity (EEG). The procedure is exemplified by the use of paradigms involving speech stimuli with emotional connotation or the Flanker task involving cognitive control.

Combined Invasive Subcortical and Non-invasive Surface Neurophysiological Recordings for the Assessment of Cognitive and Emotional Functions in Humans

Resting-state functional-connectivity MRI has identified abnormalities in patients with a wide range of neuropsychiatric disorders, including epilepsy due to malformations of cortical development. Transcranial Magnetic Stimulation in combination with EEG can demonstrate that patients with epilepsy have cortical hyperexcitability in regions with abnormal connectivity.

A Multimodal Imaging- and Stimulation-based Method of Evaluating Connectivity-related Brain Excitability in Patients with Epilepsy

This protocol describes key steps involved in assessing the sensitivity of the brain of one person to the stimulus processing of a close other by selecting pairs of partners, recording their electroencephalogram (EEG) simultaneously and computing their event-related brain potentials (ERPs).

 How to Find Effects of Stimulus Processing on Event Related Brain Potentials of Close Others when Hyperscanning Partners

We present two analytical protocols that can be used to analyze intracranial electroencephalography data using the Statistical Parametric Mapping (SPM) software: time-frequency statistical parametric mapping analysis for neural activity, and dynamic causal modeling of induced responses for intra- and inter-regional connectivity.

Analyzing Neural Activity and Connectivity Using Intracranial EEG Data with SPM Software

Standard EEG analysis techniques offer limited insight into nervous system function. Deriving statistical models of cortical connectivity offers far greater ability to investigate underlying network dynamics. Improved functional assessment opens new possibilities for diagnosis, prognostication, and outcome prediction in nervous system diseases.

Statistical Modelling of Cortical Connectivity Using Non-invasive Electroencephalograms

A novel protocol for personalized transcranial magnetic stimulation in combination with electroencephalography (TMS–EEG) mapping of cortical networks, informed by multiple magnetic resonance imaging (MRI) data modalities.

Neuroimaging-Guided TMS&#8211;EEG for Real-Time Cortical Network Mapping

This protocol describes a two-phase surgical method to create a large, resealable, dura-sparing cranial window in mice. This technique enables chronic, multi-modal electrophysiological recordings from distributed, deep-brain networks, such as the Default Mode Network, over several weeks.

Surgical Protocol for a Large, Resealable Cranial Window Enabling Longitudinal, Multi-Modal Electrophysiology Recordings of Mouse Default Mode Network

Learn to use the Human Neocortical Neurosolver to link EEG biomarkers to neural circuit mechanisms. This computational protocol enables hypothesis-driven modeling of neurotherapeutic effects on brain dynamics.

A Protocol For Uncovering Neural Mechanisms Of Neurotherapeutic Effects On Electroencephalography Using The Human Neocortical Neurosolver (Video) | JoVE Journal

A Protocol For Uncovering Neural Mechanisms Of Neurotherapeutic Effects On Electroencephalography Using The Human Neocortical Neurosolver

Explore More Videos