Mechanisms mediating dynamic changes in neural responses during deep brain stimulation

Abstract

Background:

Deep brain stimulation (DBS) of the subthalamic nucleus (STN) to treat Parkinson’s disease (PD) generates local evoked potentials (DLEPs). DLEPs reflect neural activation by DBS, but the mechanisms underlying the dynamic changes in DLEPs during continuous DBS are not understood, and such knowledge is critical to using these signals for DBS programming and closed-loop control.

Methods:

We first systematically incorporated short-term synaptic depletion into 4 different synapse types in a biophysically realistic computational model of DLEPs. We then recorded DLEPs in naïve and a virally transfected rat model. Lastly, we conducted a retrospective analysis correlating DLEP dynamics to motor symptom severity in 13 patients with PD.

Results:

A bespoke biophysical model suggests that synaptic depression, through synaptic vesicle depletion of activated motor cortical synapses onto STN neurons, causes the dynamic changes in DLEP amplitude and latency during continuous DBS. Virally-driven overexpression of endophilin A1 or alpha-synuclein in rat motor cortex reduces the time constant of DLEP amplitude and latency, supporting a critical role of these synapses in mediating DLEP dynamics. Further, a correlation between the change in UPDRS-III scores with antiparkinsonian medications and the time constant of DLEP amplitude is observed in persons with PD, suggesting DLEP dynamics as a biomarker of PD progression.

Conclusions:

Collectively, these results suggest short-term synaptic depression of cortical synapses onto STN neurons mediate DLEP dynamics in STN DBS.

1. Introduction

Parkinson’s disease (PD) is a neurodegenerative disease characterized by early prominent death of dopaminergic (DA) neurons in the substantia nigra pars compacta (SNc), resulting in classic parkinsonian motor symptoms of tremor, rigidity, bradykinesia, and impaired balance/gait [1]. PD treatment often initially includes dopamine-based therapies to treat the motor signs and symptoms, but no disease-modifying treatments are currently available to slow disease progression [1,2]. As PD progresses, patients may experience complications including medication-resistant tremor, rapid motor fluctuations over minutes to hours, and levodopa-induced dyskinesia, all of which typically improve with deep brain stimulation (DBS) of subcortical motor targets. DBS is an effective surgical intervention for movement disorders including essential tremor and PD, often in conjunction with medications. DBS for PD entails implanting electrodes into the subthalamic nucleus (STN) or globus pallidus pars internus (GPi) to deliver electrical stimulation. In addition to treating the motor symptoms of PD, STN DBS improves overall quality of life [3]. While DBS is effective in treating many motor symptoms of PD, there are several refractory symptoms and side effects [1,4,5]. Moreover, there are still many competing – and sometimes contradictory – theories on the mechanism of action of DBS for PD, including suppression of neuronal firing, information lesion, action potential collision and axonal invasion (i.e., antidromic backpropagation to collateral axon projections) [2,6]. As a result, there is substantial interest in improving understanding of DBS mechanisms of action to advance DBS clinical implementation [7].

Evoked potentials (EP) represent the summation of neural activity evoked by DBS and may contribute to understanding neural mechanisms of DBS and clinical efficacy [4,8–12]. DBS local evoked potentials (DLEPs) reflect activated neural activity recorded locally at the site of stimulation and are under investigation as potential biomarkers for DBS lead localization [9,10,12–14], parameter selection [4,12], and adaptive DBS [9,10,12,15]. The potential utility of DLEPs relies on the assumption that DLEPs serve as a surrogate for motor symptom response, and a growing body of work supports this relationship. Notably, DLEP features change over time during continuous DBS with dynamics that align with changes in motor symptoms [8,15–17].

The process(es) that mediate the dynamic, time-dependent changes in DLEPs that occur during the first few minutes of continuous DBS remain unknown. The general supposition is that synaptic depletion at stimulated STN glutamatergic terminals is principally responsible for the observed dynamics, given the supraphysiological stimulation frequency of DBS [8,18]. Sermon et al. [19], through a biophysically-motivated model using Kuramoto oscillators, suggested that depletion of glutamatergic synapses from STN onto globus pallidus pars externus (GPe) during STN DBS mediates the dynamic changes in DLEPs. However, as noted by Wiest et al. [18], the synaptic depletion hypothesis remains speculative, particularly which synapses may be involved.

Here, we combined biophysically-realistic computational modeling and in vivo experiments to investigate the mechanisms underlying dynamic changes of DLEPs during STN DBS. We developed a rat model of DLEPs to validate the predictions of our computational model. Subsequently, we used viral transduction to modify synaptic dynamics and test model predictions in the newly developed rat model of DLEPs. We then translated these findings to underscore the utility of DLEPs as a biomarker and gathered retrospective human Unified Parkinson’s Disease Scale Rating Scale (UPDRS) data to provide preliminary clinical insight.

2. Methods

2.1. Computational model

The modeling builds upon a previously described model built in NEURON [8]. We converted the NEURON HOC implementation into Python, using NEURON as an extension module for Python [20], while keeping the general structure of the original model. In short, the model contains biophysically-realistic STN (Accession: 151460) and GPe (Accession: 114685) neurons from ModelDB (http://modeldb.science), and afferent cortical axons representing the hyperdirect pathway. The STN and GPe neurons are reciprocally connected and the cortical afferents project to the STN neurons. Simulations are composed of multiple simulation units each containing 10 of each cell type, allowing for both convergent and divergent projections. Unless specified otherwise, each simulation is composed of 50 simulation units for a total of 500 cells of each type and STN neurons randomly and uniformly populate a 6 × 2 × 8 mm prism representing the human STN. The Python implementation of the model also employs network parallelization [20] with Message Passing Interface (MPI) using OpenMPI [21] to reduce the elapsed time required to simulate 120 s of stimulation in the computational model. Each simulation unit is run on 10 CPUs and each CPU runs the calculation for 3 cells (1 of each type). Simulations were run on the Duke Compute Cluster. All simulations are done with a time step of 0.025 ms, unless otherwise specified. The elapsed time needed to simulate 120 s of stimulation varies from approximately 10 h–24 h for each simulation unit at a time step of 0.025 ms.

Triggered synaptic currents were simulated at model axon terminals similar to the original model. Briefly, synapses were modeled as simple exponentials triggered by presynaptic action potentials with 0.5 ms delays and other parameters matching empirical data. Excitatory cortical axons project to proximal and distal STN dendrites, excitatory STN synapses project to the distal dendrites of GPe neurons, and inhibitory GPe synapses project to the somas and dendrites of STN neurons. The model also includes intrapallidal inhibition in the GPe through inhibitory GPe synapses on GPe somas.

The model used a representation of the Medtronic 3387 lead for simulation of the stimulation electrode and flanking electrodes for recording. The electrodes have a 3 mm spacing. Point electrodes represented the electrodes [22], allowing for simplified electric potential calculations done natively in Python. The dorsal recording electrode and central stimulation electrode were positioned in the STN and the ventral recording electrode was positioned out of the STN. 100 μs/phase symmetric, biphasic current pulses of 3 mA intensity were applied assuming a homogeneous, infinite, isotropic medium with 0.3 S/m conductivity at a frequency of 130 Hz unless otherwise specified. Potentials at all neural elements were calculated using the point source equation for the potential at a distance r:

$$\phi (r)=\frac{I}{4\pi \sigma r}$$

where I is the source current, σ is the tissue conductivity, and r is the distance from the source.

Differential recordings were made with the dorsal and ventral electrodes. Neurons with neural elements less than 500 μm from any point electrode contacts were discarded to represent the physical presence of the DBS lead. Potentials observed at each recording electrode were calculated as:

$$\phi =\sum _k\frac{i_k}{4\pi \sigma r_k}$$

with i_k and r_k representing transmembrane current and distance from the electrode for each STN segment k, respectively, and σ is the tissue conductivity. Potentials from the ventral contact were subtracted from those of the dorsal contact to simulate differential recording.

Data were analyzed in MATLAB 2022a (MathWorks, Natick, MA) on the Duke Compute Cluster using bespoke functions and scripts. The differential recording of all included STN neurons were summed to produce simulated DLEP waveforms. DLEPs were then low pass filtered at 1 kHz (3rd order Butterworth). Next, DLEPs were zeroed from 20 μs before to 500 μs after stimulation, high pass filtered at 0.1 Hz, and low pass filtered at 10 kHz (both 3rd order Butterworth filter) to mirror the blanking and filtering done with previous intraoperative recordings [8,23,24]. Lastly, DLEPs were averaged across stimulation responses over 1 s bins. The peak-to-peak amplitude of each 1 s averaged DLEP was calculated as the difference between the maximum and minimum values between 2.5 ms and 6.25 ms after stimulation, and latency was the time at which the maximum value occurred.

2.1.1. Tsodyks-Markram model of short-term synaptic plasticity

We used the Tsodyks-Markram (TM) model to simulate short-term synaptic plasticity [25]. The TM model was implemented in NEURON and found in ModelDB (Accession:3815). The model is a phenomenological description of synapses that, in this implementation, regulates the degree of conductance change at an activated synapse (Figs. 2A and 1F). The model divides synaptic resources into three fractions: recovered synaptic resources x, active synaptic resources y, and inactive synaptic resources z. These resources are governed according to the following system of equations:

$$\frac{dx}{dt}=\frac{x}{{\tau }_{rec}}-\mathit{ux}\delta \left(t-t_s\right)$$

$$\frac{dy}{dt}=-\frac{y}{{\tau }_1}+\mathit{ux}\delta \left(t-t_s\right)$$

$$z=1-y-x$$

Where τ_rec is the recovery time of inactive synaptic resources, u is the fraction of recovered resources x used by each presynaptic spike, t_s is the presynaptic spike time, and τ₁ is the decay constant of active synaptic resources y. Note that the synaptic conductance used in the model follows the same dynamics as the active synaptic resources y (i.e., synapse conductance decays with time constant τ₁). With this model, high frequency stimulation depletes x as synaptic resources build up as z.

Fig. 2. A biophysically-realistic model predicts short-term synaptic depletion (STSD) of hyper-direct pathway mediates the DBS local evoked potential (DLEP) time dynamics observed empirically.

A. STSD as described by the Tsodyks-Markram model includes the fraction of synaptic resources readily available for release (x), the resources actively in the synaptic cleft (y), and the inactive resources (z) in the process of being recycled back into the readily available resources. This figure (left) was created with BioRender.com. B. Sobol sensitivity analysis suggests hyperdirect pathway STSD is the most critical synapse for recapitulating DLEP time dynamics. The error bars represent the 95 % confidence intervals. C. Simulation-based inference suggests that there is a tight combination of STSD parameters that recapitulate empirical DLEP dynamics. D. The inclusion of STSD at the hyper-direct pathway synapses recapitulate the increasing latency and decreasing amplitude of DLEPs observed empirically. E-G. The addition of STSD at the STN projection to GPE (E), GPE projection to STN (F), and intrapallidal projection (G) do not reproduce the DLEP time dynamics. Line and dot colors for E-G follow the colorbar at the bottom of the figure.

Fig. 1. Manipulation of hyperdirect pathway synaptic weights affects DLEP latency and amplitude dynamics.

A. DLEPs decrease in amplitude and increase in latency over the course of minutes during continuous 130 Hz DBS in participants with PD. B. The model includes excitatory subthalamic nucleus (STN), inhibitory globus pallidus pars externus (GPe), and excitatory cortical axons (CTX) representing the hyper-direct pathway. Reprinted from Schmidt et al., 2020. C. The original model simulated over 120 s of 130 Hz DBS yields DLEPs (above) that do not change over time (below). Each DLEP waveform line (above) is a 10-s average, and each dot on the scatter plot (below) represents the DLEP peak-to-peak amplitude and latency for each 1-s bin over 120 s. D. When hyperdirect pathway synapse weights are incrementally decreased each second over 120 s of DBS, DLEP parameters change similar to empirical DLEPs. Line and dot colors for A, C, and D follow the colorbar at the bottom of the figure.

2.1.2. Sobol sensitivity analysis

Sobol SA is a variance-based analysis method, meaning it evaluates how the variance of an input contributes to the variance of an output [26]. When given a number of model inputs and outputs, Sobol SA estimates the total Sobol index, the contribution of each input parameter (i.e., main effects), and their interactions with other inputs (i.e., second order interactions and higher) to the overall variance of the model output. For example, total Sobol index of 0.7 for an input parameter indicates 70 % of the output variance is due to the variance of the given parameter. Generally, total Sobol indices greater than 0.8 are considered very important parameters, important parameters are 0.5–0.8, unimportant parameters are 0.3–0.5, and below 0.3 are considered irrelevant [26,27].

We used Sobol SA on the TM model inputs to identify the most important synapses for DLEP time dynamics. We added the TM mechanism to all synapses and kept the TM input parameters the same for each synapse type, leaving four different TM input sets: inputs for all synapses from STN→GPe, GPe→STN, hyperdirect pathway CTX→STN, and intrapallidal GPe→GPe. All other relevant synapse parameters were unchanged (e.g., synapse weights were unchanged, etc.) and τ₁ was given the same values as the original simple exponential decays of synapse conductances. This left only two input parameters τ_rec and u to analyze for each TM mechanism set, resulting in a total of 8 input parameters per model evaluation.

To perform Sobol SA, we used Saltelli’s sampling method [28] with a total of 1280 model evaluations for 8 input parameters. We ran the model 1280 times to get DLEP outputs for each input parameter set with a time step of 0.075 ms. We determined the amplitude and latency of the DLEP peak for each second of the 120 s simulation. Empirical data from one patient’s DLEPs over 120 s of 130 Hz stimulation were used to calculate root mean squared errors (RMSE) as a singular concise model output metric. The RMSEs were the model outputs for the total Sobol index calculations. The sample input sets generation (i.e., Saltelli’s sampling method) and total Sobol index calculations were done in Python’s SAlib version 1.4.6.1 [29,30].

2.1.3. Sequential neural posterior estimation

After determining that the inclusion of the TM mechanism in cortical axon synapses was critical for DLEP time dynamics, we needed a means of selecting appropriate τ_rec and u parameters. Simulation-based inference [31] enabled us to examine efficiently the full range of possible τ_rec-u parameter space. We used 512 model simulations with a time step of 0.075 ms, drawing parameters using the previously mentioned Saltelli method. The parameters τ_rec and u were sampled from the ranges 0–100, 000 ms and 0–0.001, respectively. We calculated the normalized amplitude and latency of every 10th second of each simulation to train the neural network in the sequential neural posterior estimation (SNPE) via the sbi module (version 0.19.2) in Python version 3.9.7 [32].

Specifically, we trained a mixture density network (MDN) to learn the association between the input parameters θ (i.e., τ_rec and u) and the output observations y (i.e., DLEP amplitude and latency at 10 s intervals across 120 s of stimulation). This density estimator was then used to build the posterior distribution p(θ|y) (i.e., the distribution of input parameters θ given observation y) according to the SNPE-C algorithm. With the newly trained posterior distribution p(θ|y), the full posterior of parameters can be estimated for new observations. Thus, we can provide a target observation (e.g., DLEP data from intraoperative patient recordings) and have a distribution of the likely parameter combinations that will yield a similar simulation output to the empirical observation. We sampled p(θ|y) 10,000 times, given DLEP amplitude and latency data from a patient over 120 s as the target observation. This allowed us to visualize the most likely parameter sets θ in the DLEP model that would yield observations similar to the patient DLEPs.

2.2. Rat model

2.2.1. Animals and Surgery

Animal care and experimental procedures were approved by Duke University Institutional Animal Care and Use Committee (IACUC). 70 female Sprague Dawley rats (250–350 g weight) were used in this study, housed in a temperature-controlled environment on a 12 h light-dark cycle with ad-libitum water and food. Rats were given one of five different treatments: no injections (i.e., naïve, n = 4), 6-hydroxydopamine (6-OHDA; n = 4), AAV-CaMKIIα-EGFP (n = 7), AAV-CaMKIIα-EndoA1-P2A-EGFP (n = 7), AAV-CaMKIIα-αSyn-P2A-EGFP (n = 7). Details on 6-OHDA injections and viral injections can be found below (see 6-OHDA injection and viral injections).

Animals were placed under sevoflurane anesthesia (induction: 7 %; maintenance: 3.5 %) and fixed in a stereotaxic frame. Body temperature was maintained at 37C with a heated water blanket while heart rate and oxygen saturation were monitored. Meloxicam (2.0 mg/kg, s.q.) and Bupivacaine (0.2 ml of 0.25 %, s.q.) were administered for pain management. Dexamethasone (0.5 mg/kg, s.q.) was administered as an anti-inflammatory. Craniotomies were performed over the STN (AP: 3.6 mm, ML: ±2.6 mm) and intraoperative recordings were conducted with a 0.5 MΩ tungsten microelectrode to confirm STN stereotaxic location before placement of a microelectrode array (MEA) in the STN (from brain surface, DV: 6.8 mm–7.2 mm). The MEA is a 2 × 2 75 μm 10 kΩ platinum-iridium (Pt-Ir) electrode with 300 μm interelectrode spacing (Microprobes for Life Science, Gaithersburg MD). Bone screws were placed in the skull before STN MEA implantation. A bone screw over the cerebellum served as a ground for recordings. Dental acrylic was used to affix implants to the skull and bone screws. Recording/stimulating experiments were conducted the day after MEA implantation.

2.2.2. Electrical stimulation and recording

We recorded and stimulated from the 2 × 2 MEAs implanted in the STN – 2 microelectrodes for differential recording and the remaining 2 electrodes for bipolar stimulation. After passing through 0.2 μF capacitors in series (C315C224K1R5TA7303, KEMET, Ft. Lauderdale, FL), the signal entered the first stage amplifier (SR560, Stanford Research Systems, Sunnyvale, CA). The first stage amplifier was DC-coupled and set for differential recording. The output of the first stage amplifier led into a second stage amplifier with a low pass filter at 10 kHz and high pass filter at 0.1 kHz. The second stage output was then digitized by a NIDAQ 6216 (National Instruments, Austin, TX) with 16-bit resolution. Stimulation was done with a Caputron LCI 1208 HP current isolator through the remaining two microelectrodes with a 0.03Hz high pass filter. Electrodes delivered bipolar, symmetric biphasic current stimulation at a maximum amplitude of 120 μA. We alternated stimulation pulse polarity (i.e., cathodic phase first, followed by anodic phase first) to retain neural response to stimulation while reducing stimulation artifact when averaging multiple stimulus pulses.

2.2.3. 6-OHDA injection and Viral injections

Anesthesia, stereotaxic frame, and medications were applied as described above in Animals and Surgery. Craniotomies were done over the medial forebrain bundle (MFB; AP: 2.00 mm, ML: ±2.00 mm) for 6-OHDA injection. 6-OHDA injections were applied to one hemisphere to induce unilateral dopaminergic lesions in the substantia nigra pars compacta (SNc), commonly used to create a rodent model of Parkinson’s disease. Ten μl of 6-OHDA (2.5 mg/mL in 0.2 % ascorbic acid in saline) was delivered to the MFB (DV: 7.5 mm) at a rate of 2 μL/min using a 30-gauge Hamilton Syringe. Pargyline (50 mg/kg) and desipramine (5 mg/kg) were administered intraperitoneally at least 30 min before 6-OHDA injection to inhibit monoamine oxidase and prevent noradrenergic neuron loss, respectively. The animal then recovered for 7–14 days before MEA implantation.

Virus was delivered in four locations (AP: +0, +2, +4, and +4 mm; ML: ±1.5, ±2, ±1.5, and ±3.5 mm; DV: 1 mm) to transduce a large portion of the excitatory Layer V pyramidal neurons of the motor cortex projecting to the STN. Transfecting at the MC as opposed to the STN serves two purposes: (1) MC injection avoids also affecting STN neurons since both STN and hyperdirect pathway would be transfected by the excitatory neuron specific CaMKIIα promoter and (2) requiring both viral injection and MEA implant in the STN, a small and deep nucleus, would result in more surgical difficulty and more animals excluded for incorrect injection site and MEA placement. Each injection delivered 0.5 μL at a rate of 0.1 μL/min using a Hamilton 32-gauge syringe. Control AAV expressing EGFP (AAV2:CaMKIIα:EGFP:WPRE, Vector ID: VB231220–1683fgc), AAV expressing rat endophilin A1 and EGFP (AAV2:CaMKIIα:rSh3gl2:P2A:EGFP:WGPRE, Vector ID: VB231220–1681 dgs), and AAV expressing human α-Synuclein and EGFP (AAV2: CaMKIIα:hSNCA:P2A:EGFP:WPRE, Vector ID: VB240313–1260fex) were all sourced from VectorBuilder (Chicago, IL). All viruses were ultra-purified with >10¹³ gc/mL. Animals recovered for 7–14 days after viral injections before MEA implantation.

2.2.4. Analysis of DLEPs

Data were analyzed in MATLAB (MathWorks) using bespoke functions and scripts. The signals were averaged across stimulation responses over 1 s bins and low pass filtered at 1 kHz (3rd order Butterworth filter). The peak-to-peak amplitude V_pp for every 1-s averaged DLEP was calculated using MATLAB’s findpeaks to identify the peak and trough that occurred at least 2 ms after stimulation. The latency was defined as the time after stimulation at which the peak amplitude occurred. The 1-s V_pp values over 2 min of stimulation were normalized and then fit to:

$$V_{\mathit{pp},\mathit{norm}}=a{e}^{\frac{-t}{{\tau }_v}}+b$$

where t is stimulation time, and a, b and τ_V are fitted constants. The constant τ_V is the V_pp time constant and thus in units of seconds. The 1-s latency values were similarly fit to:

$$\mathit{L}\mathit{a}\mathit{t}\mathit{e}\mathit{n}\mathit{c}\mathit{y}=a-b{e}^{\frac{-t}{{\tau }_{\mathit{Lat}}}}$$

where t is stimulation time, and a, b and τ_Lat are fitted constants. Again, the constant τ_Lat is the latency time constant and thus in units of seconds.

2.2.5. Histology

Animals were sacrificed via transcardial perfusion with saline followed by 4 % paraformaldehyde. Then, brains were removed and postfixed for 24 h in 4 % paraformaldehyde and left to equilibrate in a cryoprotectant solution of 30 % sucrose. After samples equilibrated, they were frozen in optimal cutting temperature (OCT) compound (Fisher Scientific, 23–730-571) at −80C until sectioning. 40 μm thick coronal sections were cut on a cryostat and stored in a phosphate buffered saline (PBS) for immediate use or PBS with azide (ChemCruz, sc-296028) for short term storage.

Free floating sections were washed thrice for 10-min intervals with PBS and then incubated in a blocking solution (10 % goat serum/0.3 % Triton X-100, PBS) for 1 h at room temperature, then overnight with the appropriate antibodies at 4C in blocking solution with 1 % goat serum. Primary antibodies include Tyrosine Hydroxylase (Sigma-Aldrich, AB152), endophilin A1 (Synaptic Systems, 159002), and alpha-synuclein (Thermo Fisher, AHB0261). For fluorescent imaging, we incubated sections in a 1 % blocking solution with the respective secondary antibody conjugated to Alexa 488 (Abcam, ab150113/ab150077). Tissue sections were washed thrice for 10-min intervals prior and post fluorescent incubation and mounted onto a slide with a cover slip after adding a DAPI mounting medium (SouthernBiotech, 0100–20).

2.3. Human participant information

2.3.1. Experimental setup

The Duke University Health System Institutional Review Board (IRB) approved all protocols and participants all provided written informed consent. Participants were all volunteers with PD undergoing DBS lead implantation in the STN. 13 participants were included in this study with 2 participants tested both in Stage I and II (Table S1, Supporting Information). Participants received sedation as part of routine medical care.

The intraoperative research began after microelectrode recordings, DBS lead placement, and clinical assessment of therapeutic DBS by the attending neurologist were completed for patients undergoing DBS lead placement (i.e., Stage I). In cases where patients were having their implantable pulse generator (IPG) placed in a separate surgery (i.e., Stage II), the intraoperative research began after the neurosurgeon tunneled the DBS lead extension to the chest. A counter-electrode was placed on the chest to allow monopolar stimulation. Ground was either the metal guide-cannula for the DBS lead during stage I or a metal ribbon retractor placed in the chest pocket in stage II.

We recorded DLEPs from the initial participants using previously developed instrumentation to record evoked potentials during DBS [8, 23,24]. Briefly, the signal was passed through the SR560 first-stage amplifier, then anti series diodes (E101, SEMITEC USA, Torrance, CA) and series capacitors (0.22 μF, R76MN32205050J, KEMET, Fort Lauderdale FL). The second stage SR560 was blanked from 20 μs before to 100–500 μs after each stimulation pulse, low pass filtered at 10 kHz, and high pass filtered at 0.1 Hz. For the initial cohort, we used the contacts above and below the monopolar stimulating contact for differential recordings. Later DLEPs were recorded with equipment that allowed for programmatic changes in recording/stimulating contacts using electrical switches (DG403, Analog Devices, Wilmington, MA) to direct appropriate contacts into recording or stimulating equipment. The DAQ directed each contact’s switch into either recording or stimulating mode. The recording arm was similar to the previous setup: the signal passed through diodes and capacitors and was high pass filtered (0.1Hz). The signal was then digitized by a g.HIAMP biosignal amplifier (g.tec medical engineering GmbH, Austria) with 24-bit resolution. DBS pulses were controlled by the DAQ and isolated using a bp isolator (FHC, Bowdoin, ME).

2.3.2. DLEP and UPDRS analysis

DLEPs were analyzed similarly to the rat DLEPs: low pass filtered at 1 kHz, with V_pp and latency calculated using MATLAB’s findpeaks. UPDRS scores were retrospectively gathered from patient records to relate DLEP dynamics to motor symptoms. UPDRS is a tool for evaluating various motor and non-motor aspects of Parkinson’s Disease divided into four parts, with part III as the “motor evaluation” [33]. Neurologists record UPDRS part III scores off and on anti-parkinsonian medication in the same clinic visit. We gathered the most recent UPDRS part III scores obtained before DBS surgery, ranging from 1 to 7 months before surgery. The τ values were linearly fit to UPDRS part III scores using MATLAB’s fitlm function.

2.4. Statistical analysis

Statistical analyses were conducted in MATLAB and all tests were assessed at an alpha-level of 0.05. We used two-sample t-tests to compare τ values between naïve rats and 6-OHDA treated rats. We used one-way ANOVA with multiple comparisons test to compare τ values in the viral experiments. After statistically significant ANOVAs, specific pairs were tested with Tukey’s honest significant difference test. ANOVA was also used to determine whether the slopes of τ-UPDRS part III linear regressions were non-zero.

3. Results

3.1. Computational model recapitulates DLEP dynamics from humans with PD

We sought to understand the mechanisms underlying the dynamic changes in both the amplitude and latency of local evoked potentials (DLEPs) that occur over the initial few minutes of continuous therapeutic STN DBS in participants with PD (Fig. 1A). We began with a biophysically based computational model to determine whether synaptic plasticity, and at which specific synapses, could reproduce the dynamic changes of DLEPs. The model [8] reproduces DLEPs using 3 neural elements: reciprocally connected STN and GPe neurons and cortical axon projections synapsing onto STN neurons, representing the hyperdirect pathway (Fig. 1B). The original model produced stable steady-state DLEPs with little to no change over time and did not reflect the dynamic changes in amplitude and latency observed in persons with PD during STN DBS (Fig. 1C).

As an initial test of the hypothesis that DLEP dynamics are mediated through synaptic plasticity, we conducted direct manual manipulation of synaptic weights. We classified synapses by the pre- and post-synaptic cell types (e.g., CTX→STN synapses, STN→GPe synapses, etc.) and either increased or decreased the weight of all synapses of a given type by 0.0002 μS every second over 120-s simulations. A decrease in the CTX→STN synapse weights alone was sufficient to change both the DLEP amplitude and latency in a manner consistent with empirical data (Fig. 1D). No other single specific synapse manipulation produced both the decrease in amplitude and increase in latency seen experimentally (Fig. S1, Supporting Information).

Progressively decreasing the synaptic weight of hyperdirect excitatory pathway axonal terminals reproduced DLEP dynamics. In light of the supraphysiological stimulation frequency used in DBS, we considered synaptic depression highly likely and incorporated the Tsodyks-Markram (TM) model of short-term synaptic plasticity (STSP) into the computational model [25]. The TM model of STSP represents the recycling and recovery of a pool of readily available synaptic vesicles (SV) during synaptic activation (Fig. 2A). At high frequencies, the readily releasable pool of SVs is depleted as vesicles are used more quickly than the pool can be replenished.

Excluding possible effects of facilitation to focus solely on synaptic depression resulted in two free parameters: the time constant of synaptic recovery τ_rec and the release probability u (Fig. 2A). We conducted a Sobol sensitivity analysis of STSP mechanisms added to all synapse types (Fig. 2B) to assess the importance of all synapse types present in the model on DLEP dynamics. The total Sobol indices indicated that the most important parameters for the DLEP dynamics were those of the hyperdirect pathway glutamatergic synapses.

We employed a simulation-based inference technique called sequential neural posterior estimation (SNPE) to determine efficiently the most likely combination of τ_rec and u to generate the appropriate dynamics. This analysis yielded a highly constrained region in parameter space (Fig. 2C). The model with STSP included in hyperdirect pathway synapses, with τ_rec = 50,000 ms and u = 0.00009 (Fig. 2D), recapitulated well DLEP dynamics. To confirm that STSP of the hyperdirect pathway is uniquely critical to reproduce DLEP dynamics, we also ran simulations with the TM model incorporated independently in the STN→GPe, GPe→STN, and GPe→GPe synapses (Fig. 2E–G) with the same TM model parameters. While DLEPs did evolve over time with the inclusion of STSP in the STN→GPe and GPe→STN synapses, neither yielded decreasing amplitudes and increasing latency (Fig. 2E–F), and the inclusion of STSP to the intrapallidal synapses GPe→GPe showed minimal effect (Fig. 2G). These model results largely align with the Sobol sensitivity analysis (Fig. 2B). Running the computational model with τ_rec = 1000 ms and u = 0.5, i.e., parameters associated with classic STSD, did not replicate observed DLEP dynamics (Fig. S2, Supporting Information). Further, the DLEP dynamics were preserved after revising the ratio of the number of STN:GPe neurons to 1:2 (Fig. S3, Supporting Information).

STN spike times were also recorded to quantify further STN activity underlying DLEPs for simulations from Fig. 1I–K (Fig. S4, Supporting Information). As the DLEPs increased in latency and decreased in amplitude during continuous stimulation, STN neuron firing rates (FRs) decreased and firing occurred later in time, although the FR peak occurred later in time than the corresponding DLEPs (e.g., Fig. 2D vs Fig. S4A, Supporting Information). None of the other synaptic manipulations resulted in the orderly transition in STN FR (Fig. S4A, Supporting Information) consistent with the DLEP dynamics.

3.2. DLEPs are present in rat STN

The computational model predicted that synaptic depression of the activated hyperdirect pathway glutamatergic synaptic terminals onto STN neurons could produce the dynamic changes in DLEP amplitude and latency observed experimentally. To test this prediction empirically, we first confirmed the presence of DLEPs in a rat model. Apart from a recent study using an ovine model [34], DLEPs have been studied only in human and nonhuman primates and have not been reported in small animal models. We conducted intraoperative recordings while implanting 2×2 microelectrode arrays (MEAs) into the STN (Fig. 3A) and observed that DLEPs abruptly emerged with submillimeter movement of the electrode (Fig. 3C). Further, post-mortem histology showed that animals in which we observed DLEPs had electrodes positioned in the STN, while animals where we did not detect DLEPs had electrodes placed outside of the STN (Fig. S5, Supporting Information). DLEPs recorded in rats occurred ~4 ms after stimulation (Fig. 3D), similar to human recordings.

Fig. 3. Rat model of DLEPs.

A. Microelectrode arrays (MEAs) were inserted into the subthalamic nucleus (STN) of rats as shown through X-ray imaging. B. Representative image showing post-mortem confirmation of correct electrode placement in STN. Inset: dotted outline marks the STN. C. DLEPs were observed in the STN only. Ten second stimulation trials were performed intraoperatively as the MEA approached the STN. The distance between the red dot and blue dot representing the MEA location around the STN during the two stimulation trials is 0.2 mm. We see DLEPs in blue, but not in red. The grey lines show 10 randomly selected single-pulse DLEPs used to make the 10 s averages in red and blue. D. Representative plot of DLEPs over time. Each line represents DLEPs averaged over 10 s intervals during a 120 s stimulation trial. This figure also graphically defines the amplitude (Vpp) and latency. E-F. Representative figures showing how τ values are calculated. Each dot represents the calculated DLEP amplitude and latency for each second during the trial and the red lines are the fitted exponential curves that yield the latency (E) and amplitude (F) τ values. The grey lines show the latency and amplitude changes of the other 7 animals. For amplitude, the grey lines are normalized to fit on a single scale (grey left y-axis). G-H. The τ values were calculated for both naïve rats and rats treated with 6-OHDA to produce hemi-parkinsonian rats. There was no statistically significant difference between naïve and 6-OHDA treated rats for latency τ (G; p = 0.137, n = 4 rats per group) and amplitude τ (H; p = 0.989, n = 4 rats per group).

We recorded DLEPs in both awake, freely moving naïve rats and in rats rendered hemi-parkinsonian by unilateral 6-OHDA injection into the medial forebrain bundle. DLEPs were present in both naïve and hemi-parkinsonian animals (Fig. 3E and F), and we did not detect differences in the dynamics of DLEPs upon onset of stimulation in naïve and hemi-parkinsonian animals (Fig. 3G and H; two-sample t-test, latency τ: p = 0.137, amplitude τ: p = 0.989; n = 4 animals per group).

3.3. Viral transfection to manipulate synaptic dynamics

After confirming that DLEPs can be reliably observed in rats, we employed viral transfection to alter excitatory pyramidal neurons of the motor cortex (MC) known to project to the STN [35–38]. We injected adeno-associated virus (AAV) with a CaMKIIα promoter into MC and implanted stimulating/recording electrodes in STN to manipulate selectively the synaptic terminals of the hyperdirect pathway within STN (Fig. 4A).

Fig. 4. Viral experiments confirm changes in STSD of hyperdirect pathway impact DLEP time dynamics.

A. AAV was injected throughout the motor cortex (MC) of rats and MEAs were implanted at the STN. By injecting AAV at the MC and stimulating/recording at the STN, we can specifically investigate the hyper-direct pathway projections from MC to STN. This figure was created with BioRender.com. B. All viruses included an EGFP reporter. Left: AAV was injected throughout the MC. The dotted line identifies the MC. Right: MC projections reach the distant ipsilateral STN. Inset: dotted lines identify the STN. C-E. Representative figures showing the DLEP waveforms (C), latency (D), and amplitude (E) over time for a rat virally transduced with EGFP. F-H. Representative DLEP waveforms, latency plot, and amplitude plot for a rat transduced with endophilin A1. I-K. Representative DLEP waveforms, latency plot, and amplitude plot for a rat transduced with alpha-synuclein. L-M. The τ values calculated for EGFP, endophilin A1, and alpha-synuclein transduced animals. The latency τ (L) is significantly different between EGFP and endophilin A1 (one-way ANOVA: F_2,18 = 7.17, p < 0.01, Cohen’s F = 0.89; Tukey’s HSD post-hoc test; EGFP vs. endophilin A1: p < 0.01, Cohen’s D = 1.72; EGFP vs. alpha-synuclein: p = 0.103, Cohen’s D = 1.07; n = 7 rats per group). The amplitude τ (M) is significantly different between EGFP and endophilin A1 (one-way ANOVA: F_2,18 = 16.51, p < 0.0001, Cohen’s F = 1.35; Tukey’s HSD post-hoc test; EGFP vs. endophilin A1: p < 0.0001, Cohen’s D = 2.65; n = 7 rats per group) and between EGFP and alpha-synuclein (EGFP vs. alpha-synuclein: p < 0.01, Cohen’s D = 2.20;, n = 7 rats per group). **p ≤ 0.01, ****p ≤ 0.0001.

We transfected rats individually with either endophilin A1 with EGFP or alpha-synuclein with EGFP, as well as EGFP alone as a control. Endophilin A1 positively regulates release probability and short-term plasticity [39], while alpha-synuclein attenuates SV recycling through clustering of SVs [40–43]. Post-mortem histology confirmed that injection in MC yielded distant expression of EGFP in hyperdirect axon terminals within the STN (Fig. 4B). We quantified the dynamics of DLEPs in all three groups: EGFP-only (Fig. 4C–E), endophilin A1 overexpressing (Fig. 4F–H), and alpha-synuclein overexpressing rats (Fig. 4I–K). The time constants of DLEP latency changes were significantly different between EGFP and endophilin A1, but not EGFP and alpha-synuclein (one-way ANOVA: F_2,18 = 7.17, p < 0.01, Cohen’s F = 0.89; Tukey’s HSD post-hoc test; EGFP vs. endophilin A1: p < 0.01, Cohen’s D = 1.72; EGFP vs. alpha-synuclein: p = 0.103, Cohen’s D = 1.07). The time constants of DLEP amplitude changes of endophilin A1 and alpha-synuclein overexpressing rats were significantly different from that of EGFP (one-way ANOVA: F_2,18 = 16.51, p < 0.0001, Cohen’s F = 1.35; Tukey’s HSD post-hoc test; EGFP vs. endophilin A1: p < 0.0001, Cohen’s D = 2.65; EGFP vs. alpha-synuclein: p < 0.01, Cohen’s D = 2.20; n = 7 animals per group). Immunohistochemistry for rat endophilin A1 and human alpha-synuclein at 7 days and 2 months post-transduction directly confirmed protein overexpression (Fig. S6, Supporting Information).

3.4. DLEP dynamics correlate to motor symptoms in persons with PD

Considering that alpha-synuclein overexpression influenced DLEP dynamics and alpha-synuclein’s known link to PD through protein aggregation and misfolding [44,45], we investigated whether DLEP dynamics in persons with PD may reflect disease severity (Fig. 5A). We gathered UPDRS part III scores from 13 patients in whom we conducted intraoperative measurements of DLEPs, with DLEP measurements in 2 patients conducted during both Stage I and II surgery (Table S1, Supporting Information) totaling 15 intraoperative DLEP observations, and quantified the correlation between UPDRS III scores and DLEP τ values. One observation was excluded for having a negative τ and another observation was excluded as an extreme outlier to avoid skewing the linear regression results.

Fig. 5. DLEP amplitude time dynamics correlate with UPDRS III.

A. Graphical representation of the intraoperative experimental setup. Adapted from Swan et al., 2014. B. We correlated the amplitude τ values calculated from patients intraoperatively to the on-med and off-med UPDRS part III scores gathered ≤6 months before DBS implantation. The amplitude τ is positively correlated to the off-med UPDRS part III (above; ANOVA, p = 0.027, n = 13), and on-med UPDRS part III (below; p < 0.01, n = 13). C. The correlation between amplitude τ and percent improvement in UPDRS part III (p = 0.018, n = 13). Dotted red lines represent the 95 % confidence intervals.

UPDRS part III scores were positively correlated with amplitude τ both on and off medications (Fig. 5B; off-med UPDRS part III vs. amplitude τ, ANOVA: p = 0.027, R² = 0.371; on-med UPDRS part III vs. amplitude τ, ANOVA: p < 0.01, R² = 0.624). We calculated the percent improvement in pre-operative UPDRS scores from off-meds to on-meds to use as a patient-normalized metric, with the idea that greater improvement in symptoms serves as a surrogate for greater disease burden [46–48]. We correlated τ values to percent improvement in pre-operative UPDRS part III scores from off-meds to on-meds (Fig. 5C), and this yielded a negative linear relationship for DLEP amplitude τ (ANOVA: p = 0.018, R² = 0.411). We then tested the relationship between DLEP amplitude τ values and percent improvement in pre-operative UPDRS part III scores after accounting for time since diagnosis and levodopa equivalent daily dose (LEDD [49]) using multiple linear regression compared to simple linear regression. The p-value of the amplitude τ coefficient when accounting for time since diagnosis and LEDD was 0.011 (Tables S2 and S3; Supporting Information). DLEP latency τ values were not correlated with UPDRS III score changes (Fig. S7, Supporting Information).

4. Discussion

We investigated the mechanisms underlying the dynamics of evoked potentials recorded in STN during DBS using both computational modeling and in vivo experiments, and we propose a potential new clinical application of DLEP dynamics. Our results suggest that short-term synaptic plasticity of the hyperdirect pathway synaptic terminals onto STN neurons, activated locally within the STN by DBS, mediates the dynamic changes in DLEP amplitude and latency observed during minutes of continuous stimulation. This study developed both a biophysically-realistic computational model and small animal model that provide a new level of insight into the mechanisms of deep brain stimulation local evoked potentials.

Our computational model replicated the biophysical details of neurons including ion channels, complex morphologies, and extracellular stimulation/recording. Researchers have attempted to create biophysically-motivated DLEP models by employing Kuramoto oscillators as simplified point neurons [19,50,51]. Sermon et al. [19], in contrast to our findings, relied on the same logic as outlined by Wiest et al. [18] to suggest that synaptic depletion of STN to GPe synapses underlies the dynamic changes in evoked potentials. However, this was not a direct result of their model because their model only included STN point neurons, making it difficult to determine which synapses are responsible for the DLEP dynamics or to make predictions on how changes to specific synapse types would affect DLEPs. Our direct synaptic manipulations (Fig. S1, Supporting Information), Sobol sensitivity analysis (SA), and TM synapse investigation (Fig. 1) demonstrate that synaptic depletion of STN to GPe synapses alone do not reproduce empirical DLEP dynamics, contrary to Sermon’s and Wiest’s conclusion.

We employed a mechanistically agnostic approach with our biophysically-realistic model, and after extensive direct manipulations of synaptic properties (Fig. S1, Supporting Information), we arrived at a hypothesis of synaptic depression and the specific neural elements involved. Both unbiased sensitivity analysis and simulation-based inference techniques independently supported that local activation of hyperdirect pathway terminals leads to synaptic depression and mediates DLEP dynamics. This agnostic modeling approach and our biophysically-realistic model with explicit neural elements allowed us to come to this conclusion despite the well-reasoned analysis provided by Sermon et al. [19] and Wiest et al. [18]. As noted by Wiest et al. [18], their synaptic depletion hypothesis remained speculative without further validation. With our animal experiments, we further validated and supported the conclusion from our model.

While our agnostic modeling approach suggested that synaptic depletion of the hyperdirect pathway, rather than STN to GPe projections, is critical to reproduce DLEP dynamics, our work points to similar order of magnitude for τ_rec and u as used by Sermon et al. for the resting or reserve pool (RtP) [19]. Sermon et al. used a three-vesicle pool model of synaptic transmission. In short, STSD can be conceptualized with 3 vesicle pools that vary in size and ability to be mobilized: the readily-releasable pool (RRP), the recycling pool (RP), and the resting pool (RtP). The RRP is a small but easily-mobilized pool that is depleted within a second and replenished within seconds. On the other extreme is the RtP, which holds most vesicles, is least easily mobilized, and is depleted or replenished over minutes. As explained by Sermon et al., RRP is the primary pool for synaptic transmission during physiological conditions, but RtP becomes the primary pool during prolonged supraphysiological stimulation. This can be seen explicitly by comparing our computational model outputs using parameters aligned with RtP (Fig. S2A and C; Supporting Information) compared to classic STSD reflecting RRP (Fig. S2B and D; Supporting Information). As seen in Fig. S2B and D, the RRP is quickly depleted and consequently does not contribute greatly to the slow dynamics observed in DLEP recordings. Thus, our model independently concluded that the RtP is responsible for the DLEP dynamics observed over the course of minutes, in agreement with Sermon et al.

Our computational model represented equal numbers (500) of STN and GPe neurons, but anatomical work shows GPe neurons to be more abundant than STN neurons [52]. While the number and ratio of neurons were not reflective of these data, we did ensure the sparsity and density of projections were accurate [8]. For example, 30 % of GPe projections synapsed onto STN somas, 40 % onto STN proximal dendrites, and 30 % onto STN distal dendrites [53,54]. Additionally, we ensured that there were both divergent and convergent connections in the network. Neural population dynamics, like DLEPs, emerge from the structure of connectivity (e.g., connection density, convergent and divergent connections) [55]. In reciprocal networks like the computational model, connectivity features dictate how signals spread, how oscillations form, and the stability of the oscillation. This was demonstrated in simulations using an STN:GPe ratio of 1:2 (Fig. S3, Supporting Information). The network produced marginally different DLEP dynamics compared to those in Fig. 2D but retained the reduction in amplitude and increase in latency. The differences in the DLEPs and DLEP dynamics between Fig. 2D and Fig. S3 are principally a feature of the network connectivity rather than the STN:GPe ratio. This can be seen in the variability in DLEP dynamics between Fig. S3A–C, in which the connection ratios are the same, but they have different random connections.

We introduced and validated a small animal model of DLEPs, which allowed us to evaluate further our hypothesis and provides a valuable resource to the research community. Prior research on DLEPs was conducted exclusively on human and non-human primates [8–10,14,15,18, 19,56,57], except for a single ovine study [34], and this greatly limits the range of potential experimental manipulations. The rat model of DLEPs enables application of a host of tools, such as optogenetics and chemogenetics, that are accessible and well-suited to answering mechanistic questions.

The rat model was used to test the hypothesis that emerged from our computational model simulations and provided valuable insight into DLEP mechanisms. The presence of DLEPs in naïve animals supports that DLEPs are not specific to the Parkinsonian state, but rather intrinsic to basal ganglia circuits, as suggested previously [9,34]. This finding aligns with the computational model, as the model suggests DLEPs arise from neuroanatomical connections not unique to PD. Moreover, we found no difference in the DLEPs between naïve and 6-OHDA treated rats (Fig. 3). The computational model predicts there to be no difference in DLEPs between naïve and 6-OHDA treated rats as 6-OHDA interferes with dopaminergic circuit elements not considered relevant in the computational model (i.e., dopamine producing cells in the nigrostriatal pathway). The computational model suggests DLEP dynamics arise due to synaptic depletion of the hyperdirect pathway projection terminals within STN, rather than DA neuron projections, aligning with the rat experiments. 6-OHDA treated rats emulate PD by destroying DA neurons of the SNc, which does not affect hyperdirect pathway projections. Thus, the computational model predicts no changes in DLEP dynamics in 6-OHDA treated rats, as observed experimentally. The increasing DLEP latency and decreasing DLEP amplitude in the animal model occurred on a time scale similar to human DLEPs (i.e., during the first few minutes of continuous DBS), suggesting similar underlying mechanisms mediating DLEP dynamics in both species. The results in the rat model supported the hypothesis that synaptic depression of hyperdirect pathway projection terminals within STN mediates the dynamic changes in DLEP amplitude and latency, contrary to the synaptic depletion of STN to GPe synapses suggested by Sermon et al. [19] and Wiest et al. [18].

We used viral transfection to manipulate synaptic dynamics as a causal test of the hypothesis that STSD at the hyperdirect synapse was responsible for the dynamic changes in the amplitude and latency of DLEPs during DBS. According to this hypothesis, if synapses depress more quickly, then the DLEP dynamics would occur faster. We chose to decrease rather than increase τ to avoid possible ceiling effects, attributed either to using an insufficiently low stimulation frequency to depress synapses or to the presence of other rate limiting mechanisms that resist increased synaptic recycling efficiency.

Our first candidate protein was endophilin A1, which positively regulates release probability and short-term plasticity [39] and is implicated in SV recycling and endocytosis [39,58–61]. To test further our hypothesis, we used a second protein with a different mechanism that we expected to yield the same result of decreasing τ. Alpha-synuclein is an attenuator of SV recycling [40–43], with overexpression of alpha-synuclein clustering SVs and thus disrupting recycling [41,43,62]. Alpha-synuclein is also of interest, as it is commonly implicated in Parkinson’s disease pathophysiology [63–65]. As expected, overexpression of endophilin A1 and alpha-synuclein resulted in decreased DLEP time constants (Fig. 4L–M), suggesting short-term synaptic depression of hyperdirect pathway glutamatergic synaptic terminals mediates DLEP dynamics.

The combination of viral delivery to MC and MEA implant in STN allowed us to manipulate selectively the synaptic dynamics of the hyperdirect pathway without affecting the synapses between STN and GP. However, the MC is a highly connected brain region, and we see, for example, EGFP expressed in other brain regions including the thalamus, which is to be expected. It is unlikely that regions outside of the STN, GP, and hyperdirect pathway contributed to the altered DLEP dynamics considering the time scale at which DLEPs occur – more distant structures and polysynaptic pathways would take longer than 4 ms to propagate to STN. Further, the Sobol SA suggests outside regions are unlikely to be involved in DLEP dynamics considering that even the local inputs from STN and GP do not produce the observed DLEP dynamics. As well, synaptic depression multiple synapses away is expected to have an increasingly small effect on downstream processes.

A potential limitation of our viral experiments was the comparatively short transduction time. Our electrophysiological measurements were conducted 8–15 days after viral injection. Because the proteins were chosen with the expectation that they would accelerate the time-dependent decline in DLEP amplitude, we were concerned that exceedingly robust expression would lead to DLEPs that diminished too quickly to be observed. To address concern regarding the short transduction time, we used immunohistochemistry to confirm alpha-synuclein and endophilin A1 protein expression 7 days post transduction. At 7 days post transduction, there is no visible difference in endophilin A1 expression at the STN between transduced and non-transduced hemispheres (Fig. S6A, Supporting Information). This is likely due to using a viral vector with rat endophilin A1 which is endogenously expressed throughout the brain, making confirmation of transduction more challenging among the background of endogenous endophilin A1. Even at 2 months after transduction there was little difference in endophilin A1 expression (Fig. S6B, Supporting Information), highlighting the difficulty of observing the overexpression of a ubiquitously expressed protein. However, the expression of human alpha-synuclein shows a subtle but clear difference in alpha-synuclein at 7 days post transduction (Fig. S6A, Supporting Information), and a very clear difference at 2 months (Fig. S6B, Supporting Information). The notable presence of human alpha-synuclein expression at the STN at 7 days (Fig. S6A, Supporting Information) and the clear extension of viral transduction from MC to STN (Figure S6B, Figure S6B, Supporting Information) makes it likely that both endophilin and alpha-synuclein overexpression is present 7 days after transduction.

The rodent experiments are also limited by the technical difficulty of placing electrodes precisely in the rodent STN – a nucleus requiring submillimeter accuracy – and recording sub millisecond signals evoked by local stimulation in freely moving animals. These results provide the first instances of DLEPs in rodent and required substantial troubleshooting and technique development, which partially hindered our ability to gather rodent data efficiently for this study. As the sample size of the rodent experiments is relatively small, the results should be considered preliminary pending replication and larger sample sizes made possible by this newly established protocol.

The results of alpha-synuclein over-expression on DLEP dynamics inspired additional exploratory investigation into the potential clinical utility of DLEP dynamics. The Braak hypothesis suggests that PD is caused by a pathogen that initiates alpha-synuclein aggregation (Lewy Body pathology) in the nose and digestive tract, spreading through the lower brainstem up through the midbrain and finally to the cortex [44,65]. This accumulation of alpha-synuclein is associated with progressive neuronal death through mitochondrial, lysosomal, and calcium homeostasis dysfunction [45]. Yet before neuronal cell death, there are early alpha-synuclein dependent synaptic changes seeming to link neurodegenerative inflammation and synaptic dysfunction. Altogether this raises the possibility that changes in DLEP dynamics, reflecting synaptic dysfunction, may have utility as an indirect measure of disease progression. Thus, we compared retrospective UPDRS scores to patient-specific DLEP dynamics, under the assumption UPDRS scores are an adequate surrogate for disease severity.

Based on the changes in DLEP dynamics following alpha-synuclein expression, we expected the DLEP time dynamics to be anti-correlated to UPDRS scores. However, we found the opposite with on and off med UPDRS scores (Fig. 5B). Similarly, Wiest et al. [15] reported that UPDRS III scores of PD subjects in off medication state were positively correlated with time to DLEP steady state. If alpha-synuclein in the hyperdirect pathway hastens DLEP dynamics (i.e., lower τ), Fig. 5B and the findings reported by Wiest et al. would suggest that patients with less alpha-synucleinopathy oddly have more severe PD symptoms (i.e., higher UPDRS scores). This may be reconciled by patient heterogeneity in alpha-synuclein propagation and differential vulnerability to alpha-synuclein across brain regions. The DA neurons of the SNc are known to be especially susceptible to alpha-synuclein accumulation [45]. Thus, it may be that patients with particularly vulnerable SNc lose the integrity of their dopaminergic system before significant spread and impact on the motor cortex and hyperdirect pathway. Conversely, those with more resilient SNc would have a less dramatic difference in alpha-synuclein accumulation between SNc and hyperdirect pathway. If true, patients with more susceptible SNc would not only have a greater motor UPDRS before hyperdirect pathway is impacted but also be less responsive to dopaminergic medications due to the extensive disruption of the dopaminergic system. In fact, we find the DLEP dynamics are anti-correlated with percent improvement in UPDRS scores (Fig. 5C) which aligns with the theory. Amplitude tau significantly correlates with percent improvement in UPDRS in the simple linear regression (p = 0.018, Fig. 5C; Table S2, Supporting Information) and in a multiple linear regression analysis including both years since diagnosis and LEDD (p = 0.011; Table S3, Supporting Information), suggesting that neither of these latter variables account for the variance in the statistical model. Yet, time to diagnosis as a measure of disease duration is limited by factors such as patient access to healthcare, large variance in rate of disease progression, and even ambiguity and variability in how long a patient may have had PD before diagnosis. Similarly, LEDD is affected by factors such as patient tolerance to medication and provider decision-making. Nonetheless, as a preliminary retrospective result, further study is needed to validate this clinical correlation.

The hypothesis that STN DLEP dynamics are largely mediated through spread of alpha-synucleinopathy to cortex, causing synaptic dysfunction, is consistent with the 6-OHDA vs naïve animal experiments. The 6-OHDA model directly damages DA neurons, not the hyperdirect pathway or synaptic release. Like 6-OHDA, alpha-synuclein causes DA neuron degeneration when aggregating in the substantia nigra. However, alpha-synuclein can also spread to the motor cortex as described in the Braak hypothesis, where it can then have effects on synaptic release of the hyperdirect pathway. Thus, finding no difference in DLEP dynamics between the parkinsonian 6-OHDA rat model and naïve animals suggests DA neuron degeneration is not related to STN DLEP dynamics, as predicted by the computational model. The computational model predicts the synaptic depletion of hyperdirect pathway is critical for DLEP dynamics, which is not affected by the targeted DA neuron degeneration caused by 6-OHDA. Alpha synuclein, however, can cause both DA neuron degeneration in SNc and synaptic dysfunction in the hyperdirect pathway.

Notably, there seems to be a relationship between the time course of motor symptoms alleviation and DLEP dynamics. STN DBS in PD patients relieves tremor symptoms in seconds [66], bradykinesia and rigidity over minutes [67], and axial symptoms hours to days after DBS onset [68]. The empirical data and computational model follow well with the time course of bradykinesia improvement but not tremor or axial symptom improvement [8,15–17]. The computational model does not account directly for the time course of symptom reversal, and we do not claim that the changes in DLEPs are causal of symptom change. Nonetheless, dynamic changes in neural responsiveness, as reflected in the DLEP, may be a part of the mechanism explaining the time course of tremor vs. bradykinesia and rigidity vs. axial symptoms. The similarity between the DLEP dynamics and the time course of bradykinesia and rigidity symptom alleviation may suggest that bradykinesia and rigidity are mediated by short term plasticity and/or circuit elements more directly stimulated by DBS, while tremor and axial symptoms are related to other effects of DBS such as immediate neuromodulation effects in the case of tremor or longer term plasticity in the case of axial symptoms [69].

Our study is focused on STN DLEP dynamics rather than mechanisms of motor symptom alleviation through DBS. While STN DBS likely engages multiple mechanisms of action to alleviate PD symptoms, such as antidromic activation of motor cortex, the mechanisms behind DLEP dynamics specifically are likely a smaller subset of the myriad mechanisms driven by DBS [6]. The synaptic depression hypothesis is supported by the absence of DLEP dynamics during stimulation at 20 Hz [10], a frequency too low to deplete the reserve vesicle pool [15,70,71]. The short-term depletion of the reserve pool also aligns with the time scale of observed DLEP dynamics on the scale of seconds to minutes [70] and the partial recovery of DLEP amplitude and latency with a 5 s DBS pause reported by Wiest et al. [18]. Beyond being theoretically congruent, STSD of the hyperdirect pathway in STN DBS is supported by computational modeling and, for the first time, in vivo measurements in rodents.

DBS for PD is also often targeted to the GPi and DLEP dynamics can be observed in this nucleus as well [51,72]. The observation of DLEP dynamics in GPi DBS – where no hyperdirect pathway is present – seemingly refutes our hypothesis that STSD of hyperdirect pathway mediates DLEP dynamics in STN DBS. However, GPi DLEP dynamics support generalization of our hypothesis. The hyperdirect pathway is critical for DLEP dynamics in STN DBS because it projects to the STN, and we expect that activation of glutamatergic STN projection terminals within the GPi are similarly critical for DLEP dynamics in GPi DBS because STN neurons project to the GPi and can be locally stimulated within GPi. Further, the excitatory hyperdirect pathway mediates STN DLEP dynamics because the hyperdirect synapses depress during sustained high frequency DBS. Thus, the hyperdirect pathway is uniquely important in STN DBS because it is the excitatory projections to the site of stimulation. We expect then that local activation within GPi of STN axonal projections to be critical for GPi DLEPs for the same reason, but this remains to be evaluated in future studies.

As an exploratory retrospective analysis, the human subjects results of this study are limited by the small sample size of the study. The human subjects results are preliminary and require replication in a larger prospective cohort for robustness. Additionally, the UPDRS III scores in this study come from a specific subpopulation: subjects that are eligible for DBS. Of note, patients with significant cognitive decline, severe dementia, or those with primary symptoms unresponsive to DBS such as gait freezing or speech problems are not candidates for DBS and thus not included in this study. The medication responsiveness of a patient with PD as a surrogate for disease burden may not hold in sufficiently advanced PD because PD patients become less responsive to medications [46–48,73]. Advanced PD patients likely also have substantial cortical involvement, which may saturate DLEP dynamics as a viable surrogate marker of PD disease severity (i.e., DLEP tau values are limited to zero seconds in cases of extreme cortical involvement). While this specific subpopulation is a limitation of this study, it is conveniently a population that likely provides the greatest variation in both motor symptoms and cortical involvement and thus is ideal for a correlation analysis.

Furthermore, measuring motor symptoms alone overlooks many other factors in PD that are informative of disease burden such as cognitive symptoms. These limitations hold for UPDRS III scores but might not constrain DLEP dynamics. Having access to non-motor assessments of PD, such as other parts of the UPDRS, or other markers like alpha-synuclein levels on neurosurgical tools [74] could be particularly valuable in establishing the relationship between DLEP dynamics and PD disease burden. Alternatively, repeated longitudinal measures could more directly assess the relationship between disease burden and DLEP dynamics. Ultimately, persons with PD are a heterogenous population with various dominant motor and non-motor phenotypes, disease progressions, and genetic factors, thus requiring further study to understand DLEP dynamics in relation to PD disease state. The present results provide the justification to include more intentional, prospective study designs to investigate this possible new application of DLEPs.

5. Conclusion

We developed a biophysically-realistic computational model to investigate the mechanistic underpinnings of DLEP dynamics in STN DBS. Computational modeling identified STSD of the hyperdirect pathway as critical for DLEP dynamics. We then developed a rat model of DLEPs to test the computational model prediction. Viral transduction-mediated modification of hyperdirect pathway synaptic dynamics affected DLEP dynamics as predicted in the computational model. Linking changes in DLEP dynamics to cortical synaptic dysfunction points to the possible clinical utility of DLEP measures. Lastly, we correlated the DLEP dynamics of PD patients with UPDRS part III scores in an exploratory retrospective analysis. DLEP dynamics correlated with UPDRS scores, and this preliminary finding raises the possibility of using DLEPs as a biomarker of PD severity. Collectively, these results suggest STSD of the hyperdirect pathway mediates DLEP dynamics in STN DBS, points to a possible new clinical application of DLEPS, and encourages further investigation to build upon our results.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.brs.2025.103002.

Data availability

Code for the computational model will be available upon acceptance. Rat model data are available upon request. To preserve the anonymity of the participants, participant raw data are available upon request and a data use agreement only.