Bayesian inference for biophysical neuron models enables stimulus optimization for retinal neuroprosthetics

  1. Jonathan Oesterle
  2. Christian Behrens
  3. Cornelius Schröder
  4. Thoralf Hermann
  5. Thomas Euler
  6. Katrin Franke
  7. Robert G Smith
  8. Günther Zeck
  9. Philipp Berens  Is a corresponding author
  1. Institute for Ophthalmic Research, University of Tübingen, Germany
  2. Naturwissenschaftliches und Medizinisches Institut an der Universität Tübingen, Germany
  3. Center for Integrative Neuroscience, University of Tübingen, Germany
  4. Bernstein Center for Computational Neuroscience, University of Tübingen, Germany
  5. Department of Neuroscience, University of Pennsylvania, United States
  6. Institute for Bioinformatics and Medical Informatics, University of Tübingen, Germany
10 figures, 7 tables and 1 additional file

Figures

From serial block-face electron microscopy (EM) data of retinal BCs to multicompartment models.

(A) Raw morphology extracted from EM data of an ON-BC of type 5o. (Bi) Processed morphology connected to three presynaptic cones (red) and several postsynaptic dummy compartments that are generated to create the synapses in the model (yellow). The cone and BC morphologies are divided into color-coded regions with a legend shown on the right. (Bii) Same as (Bi) but for an OFF-BC of type 3. (C) Three cylindrical compartments of a multicompartment model. Every compartment (blue) n consists of a membrane capacitance cmn, a membrane resistance rmn, a leak conductance voltage source Vrn, an extracellular voltage source Vexn and at least one axial resistor rin that is connected to a neighboring compartment. Vexn is only used to simulate electrical stimulation and is otherwise replaced by a shortcut. Compartments may have one or more further voltage- or ligand-dependent resistances ren with respective voltage sources ren to simulate ion channels (indicated in gray).

Discrepancy measures based on Equation 13 for the cone (red dashed line), the OFF- (blue solid line) and ON-BC (green dotted line).

The parameters defining the discrepancy measures are listed in Table 3. All discrepancy measures are between zero and one per definition.

Model for the external electrical stimulation of the retina.

(A) Schematic figure of the experimental setup for subretinal stimulation of ex vivo retina combined with epiretinal recording of retinal ganglion cells. Schematic modified from Corna et al., 2018. (B,C) Model for simulating the electrical field potential in the retina in 3D and 2D, respectively. The retina (darker blue) and the electrolyte above (lighter blue) are modeled as cylinders. The shown 3D model is radially symmetric with respect to the central axis (red dashed line). Therefore, the 3D and 2D implementations are equivalent, except that the computational costs for the 2D model are much lower. The 2D implementation is annotated with parameters that were either taken from the literature or inferred from experimental data. (D) Electrical field potential in the retina for a constant stimulation current of 0.5 µA for a single stimulation electrode with a diameter of 30 µm. Additionally, the compartments (black circles with white filling) of the ON-BC model are shown. The stimulation is subretinal meaning that the dendrites are facing the electrode (horizontal black line on bottom).

Discrepancies of samples from the cone and the BC models during and after parameter estimation.

(A, B, C) Sampled discrepancies for the cone (A), the OFF- (B), and ON-BC (C) respectively. For every model, the total discrepancy δtot (left) and the discrepancy between the simulated and target iGluSnFR trace δiGluSnFR (right) are shown. For every model, four optimization rounds with 2000 samples each were drawn (indicated by gray vertical lines). After the last round (indicated by dashed vertical lines, ‘p.’), 200 more samples were drawn from the posteriors. For the BCs, the number of compartments was increased in this last step to 139 and 152 for the OFF- and ON-BC, respectively. Additionally, 200 samples were drawn from assuming independent posterior marginals for comparison (indicated by blue vertical lines, ‘m.’). For every round, the discrepancy distribution (horizontal histograms), the median discrepancies (red vertical lines), the 25th to 75th percentile (blue shaded area) and the 5th to 95th percentile (gray-shaded area) are shown. (D) Discrepancies between different iGluSnFR traces of BCs to demonstrate the high precision of the model fit. The pairwise discrepancy computed with equation Equation 12 between eight iGluSnFR traces is depicted in a heat map. The column and row labels indicate which 𝝂t and 𝝂m were used in equation Equation 12 respectively. The traces consists of the optimized BC models (‘Model’), the targets used during optimization (‘Target’), experimental data from the same cell type without the application of any drug (‘No drug’) and experimental data from another retinal CBC type with the application of strychnine (‘BC4’ and ‘BC7’). Note that strychnine was also applied to record the targets.

Figure 4—source data 1

Sample discrepancies of all samples shown in (A-C).

https://cdn.elifesciences.org/articles/54997/elife-54997-fig4-data1-v2.zip
Figure 4—source data 2

Discrepancies shown in (D), and the respective (mean) iGluSnFR traces.

https://cdn.elifesciences.org/articles/54997/elife-54997-fig4-data2-v2.zip
Figure 5 with 2 supplements
Optimized cone model.

(A) Normalized light stimulus. (B) Somatic membrane potential relative to the resting potential for the best parameters (blue line) and for 200 samples from the posterior shown as the median (gray dashed line) and 10th to 90th percentile (gray shaded area). A histogram over all resting potentials is shown on the right. (C) Release rate relative to the resting rate. Otherwise as in (B). (D) Simulated iGluSnFR trace (as in (B)) compared to target trace (orange). Three regions (indicated by gray dashed boxes) are shown in more detail below without samples from the posterior. Estimates of positive and negative peaks are highlighted (up- and downward facing triangles respectively) in the target (brown) and in the simulated trace (blue and cyan respectively). Pairwise time differences between target and simulated peaks (indicated by triangle pairs connected by a black line) are shown as histograms for positive (blue) and negative (cyan) peaks on the right. The median over all peak time differences is shown as a black vertical line.

Figure 5—source data 1

iGluSnFR traces used for constraining the cone and BC models.

https://cdn.elifesciences.org/articles/54997/elife-54997-fig5-data1-v2.zip
Figure 5—source data 2

Stimulus, target, and cell responses, including responses with removed ion channels.

https://cdn.elifesciences.org/articles/54997/elife-54997-fig5-data2-v2.zip
Figure 5—figure supplement 1
iGluSnFR traces used for constraining the cone and BC models.

iGluSnFR traces used for constraining the cone and BC models. (A,C,E) All single ROI traces used for the cone, OFF- and ON-BC, respectively. For the cone, the traces are shown in different colors for the two recorded axon terminals. (B) Aligned mean traces of both cone axon terminals (same colors as in (A)), and the mean of both means (orange), which was used as the target during parameter inference. (D,F) Mean traces (orange) plus-minus one standard deviation (gray-shaded area) across all single ROI traces for the OFF- and ON-BC, respectively.

Figure 5—figure supplement 2
Effect of active ion conductances on the optimized cone model.

Voltage and iGluSnFR traces for the optimized cone model with different ion channels removed. In every row, a different (set of) ion channels has been removed from the model by setting the respective conductances to zero. The simulated somatic membrane voltages (first column) and iGluSnFR traces (third column) and the differences with respect to the simulation with all ion channels (second and fourth column, respectively) are shown. The fifth column shows how removing the ion channel changed the total discrepancy δtot and distance to the experimental target δiGluSnFR. The row titles refer to the channel that has been removed, for example in the first row, the HCN channel has been removed. In the last row (passive), all but the calcium channels in the axon terminals were removed.

Figure 6 with 6 supplements
Optimized BC models.

(A) Normalized light stimulus. Responses of the OFF- and ON-BC are shown in (B–D) and (E–G), respectively. (B, E) Somatic membrane potential relative to the resting potential for the best parameters (blue line) and for 200 samples from the posterior shown as the median (gray dashed line) and 10th to 90th percentile (gray-shaded area). A histogram over all resting potentials is shown on the right. (C, F) Mean release rate over all synapses relative to the mean resting rate. Otherwise as in (B). (D, G) Simulated iGluSnFR trace (as in (B)) compared to respective target trace (orange). Three regions (indicated by gray dashed boxes) are shown in more detail below without samples from the posterior. Estimates of positive and negative peaks are highlighted (up- and downward facing triangles, respectively) in the target (brown) and in the simulated trace (blue and cyan, respectively). Pairwise time differences between target and simulated peaks (indicated by triangle pairs connected by a black line) are shown as histograms for positive (blue) and negative (cyan) peaks on the right. The median over all peak time differences is shown as a black vertical line.

Figure 6—source data 1

Stimulus, target, and cell responses, including responses with removed ion channels.

https://cdn.elifesciences.org/articles/54997/elife-54997-fig6-data1-v2.zip
Figure 6—figure supplement 1
Heatmaps of the OFF-BC.

Heatmaps of the OFF-BC in response to the OFF-step of the light stimulus. A video for the full stimulus is available in the supplementary material. (A) Membrane voltage (color-coded) of all compartments mapped on fully detailed morphology. (B) Calcium concentration (color-coded) in BC for all compartments for which calcium is simulated. Other compartments are shown in gray. (C) Release sites with color-coded release rates (green dots). Compartments are shown in gray. (D) Light stimulus. The time for which the heatmaps are shown is indicated by the black vertical line.

Figure 6—figure supplement 2
Heatmaps of the ON-BC.

Heatmaps of the ON-BC in response to the ON-step of the light stimulus; otherwise as in Figure 6—figure supplement 1.

Figure 6—figure supplement 3
Effect of active ion conductances on the optimized OFF-BC model.

Voltage and iGluSnFR traces for the OFF-BC with different ion channels removed. Models were simulated for the five posterior samples (out of 200) with the lowest total discrepancy δtot under different conditions. The samples are color-coded across all panels. In every row, a different (set of) ion channels has been removed from the posterior samples by setting the respective conductances to zero. The simulated somatic membrane voltages (first column) and iGluSnFR traces (third column) and the differences with respect to the simulation with all ion channels (second and fourth column, respectively) are shown. The fifth column shows how removing the ion channel changed the total discrepancy δtot and distance to the experimental target δiGluSnFR for all samples individually (small vertical lines) and as a mean over all samples (red vertical line). The row titles refer to the channel that has been removed, for example in the first row, the HCN channel has been removed. Ca @ S means, that all somatic calcium channels were removed. In the last row (passive), all but the calcium channels in the axon terminals were removed.

Figure 6—figure supplement 4
Effect of active ion conductances on the optimized ON-BC model.

Voltage and iGluSnFR traces for the ON-BC with different ion channels removed; otherwise as in Figure 6—figure supplement 3.

Figure 6—video 1
Animated heatmaps of the OFF-BC.
Figure 6—video 2
Animated heatmaps of the ON-BC.
Figure 7 with 2 supplements
Parameter distributions of the BC models.

1D-marginal prior (dashed gray line) and posterior distributions (solid lines) are shown for the OFF- (blue) and ON-BC (green). The parameters of the posterior samples with the lowest total discrepancy are shown as dashed vertical lines in the respective color. XY@Z refers to the channel density of channel XY at location Z. Locations are abbreviated; S: soma, A: axon, D: dendrite and AT: axon terminals (see Figure 1 and main text for details). Note that although these 1D-marginal distributions seem relatively wide in some cases, the full high-dimensional posterior has much more structure than a posterior distribution obtained from assuming independent marginals (see Figure 4). Not all parameter distributions are shown.

Figure 7—source data 1

Prior and posterior parameters for the OFF- and ON-BC.

https://cdn.elifesciences.org/articles/54997/elife-54997-fig7-data1-v2.zip
Figure 7—figure supplement 1
2D-marginals for the OFF-BC.

Parameter distributions of the OFF-BC model. 2D-marginal prior (gray) and posterior distributions (red). Distributions are shown as contour lines at 30, 60 and 90% of the maximum of the respective probability density functions. Not all parameter distributions are shown.

Figure 7—figure supplement 2
2D-marginals for the ON-BC.

Parameter distributions of the ON-BC model. 2D-marginal prior (gray) and posterior distributions (red). Distributions are shown as contour lines at 30, 60, and 90% of the maximum of the respective probability density functions. Not all parameter distributions are shown.

Estimation of the conductivity σretina and the relative permittivity εretina of the retina for simulating electrical stimulation.

(A) Electrical circuits used during parameter estimation. (B) Stimulation voltages vstim at 25 (left) and 40 Hz (right). From the six experimentally applied amplitudes, only the amplitudes used for parameter inference are shown. (C) Experimentally measured currents iames through electrolyte (Ames’ solution) without retinal tissue for the stimulus voltages vstim in (B). The mean traces over all (but the first two) repetitions are shown (black dashed lines). Sine waves were fitted to the mean traces (solid lines) with colors referring to the voltages in (B). (D) Simulated voltages over the electrolyte vames using the fitted currents in (C) as stimuli applied to the circuit in (Ai). (Aii) Electrical circuit used to model the electrode plus interface. (E) Stimulus frequency and amplitude dependent estimates of Re (i) and Ce (ii) based on the electrical circuit shown in (Aii) for 25 (black) and 40 Hz (gray). Note that the values were derived analytically (see main text). The values corresponding to the stimulus voltages shown in (B) are highlighted with respective color. (Aiii) Electrical circuit used to estimate σretina and εretina. The respective values for Re and Ce are shown in (E) and are dependent on vstim. The current iretina through the model is measured for a given stimulus voltage vstim. (F) Sampled parameters of εretina and σretina and the respective sample losses. First, samples were drawn in a wide logarithmic space (gray dots) and then in a narrower linear parameter space. The best sample (lowest discrepancy) is highlighted in red. (G) Simulated currents iretina (solid lines) through the circuit in (Aiii) with optimized parameters (red dot in (F)) and respective experimentally measured currents (broken lines). Here, results for all six stimulus amplitudes are shown for both frequencies.

Figure 8—source data 1

Currents, voltages, and samples from optimization.

https://cdn.elifesciences.org/articles/54997/elife-54997-fig8-data1-v2.zip
Threshold of electrical stimulation for experimentally measured RGCs and simulated BCs of photoreceptor degenerated mouse retina.

(A) Stimulation currents measured experimentally and used as stimuli in the simulations. (B) xy-positions of BCs (crosses) and electrodes (red dots) for 1×1, 2×2, 4×4, 10×10 stimulation electrodes, respectively. Every electrode is modeled as a disc with a 30 µm diameter. Except for the electrode configuration, the models were as in Figure 8. (C, D) Mean synaptic glutamate release relative to the size of the readily releasable pool vRRPmax of simulated OFF- and ON-BCs, respectively. Values can be greater than 1, because the pool is replenished during simulation. Errorbars indicate the standard deviation between simulations of BCs with different parametrizations; for both BCs, the five best posterior samples were simulated. Glutamate release is shown for different charge densities (x-axis) and cell positions (colors correspond to xy-positions in (A); for example the darkest blue corresponds to the central BC). Experimentally measured RGC thresholds (gray dashed lines) plus-minus one standard deviation (gray-shaded ares) are shown in the same plots.

Figure 9—source data 1

Current traces and vesicle release for all simulations.

https://cdn.elifesciences.org/articles/54997/elife-54997-fig9-data1-v2.zip
Figure 10 with 1 supplement
Electrical currents optimized for selective ON- and OFF-BC stimulation.

(A) Illustration of the stimulus parametrization. A stimulation current was computed by fitting a cubic spline through points defined by p1p4 and p* (black dots) that were placed equidistantly in time. p1p4 were free parameters and p* was chosen such that the stimulus was charge neutral, that is the integral over the current (gray-shaded area) was zero. Currents were normalized such that the maximum amplitude was always 0.5 µA. (B) Illustration of the electrical stimulation of an OFF- (i) and ON-BC (ii) multicompartment model. (C, E) Stimuli optimized for selective OFF- and ON-BC stimulation, respectively. The two best stimuli observed during optimization are shown. (D, F) Cumulative vesicle release (as a mean over all synapses) relative to the size of the readily releasable pool vRRPmax in response to the electrical stimuli shown in (C) and (E), respectively. Stimuli and responses are shown in matching colors. Release is shown for the five best posterior cell parametrizations for the OFF- (i) and ON-BC (ii) as a mean (line) plus-minus one standard deviation (shaded area).

Figure 10—source data 1

Optimized stimuli and BC responses, including responses with removed ion channels.

https://cdn.elifesciences.org/articles/54997/elife-54997-fig10-data1-v2.zip
Figure 10—figure supplement 1
Electrical stimulation with optimized stimuli and removed ion channels.

Electrical stimulation with optimized stimuli and selectively removed ion channels. (A) Stimulation currents optimized to selectively stimulate the ON- (left and middle) and OFF-BC (right). (B) Evoked extracellular voltage at the presynaptic BC compartments (as a mean over all synapses) for the OFF- (blue) and ON-BC (green), respectively. (C) Change of the membrane voltage during stimulation at the presynaptic BC compartments for the five best posterior BC parametrizations. Otherwise as in (B). (D) Cumulative vesicle release (as a mean over all synapses) relative to the size of the readily releasable pool vRRPmax for the same BC parametrizations. (E) As in (D), but the CaT channels at the axon terminals of the OFF-BC were removed. (F) As in (D), but the CaL channels at the axon terminals of the OFF-BC were removed.

Tables

Key resources table
Reagent type (species)
or resource
DesignationSource or referenceIdentifiersAdditional information
Genetic reagent (mouse)B6;129S6-Chattm2(cre)LowlJJackson laboratory
JAX 006410
RRID:IMSR_JAX:006410
Genetic reagent (mouse)Gt(ROSA)26Sor tm9(CAG-tdTomato)HzeJackson laboratory
JAX 007905
RRID:IMSR_JAX:007905
Strain (mouse, female)B6.CXB1-Pde6brd10Jackson laboratory
JAX 004297
RRID:IMSR_JAX:004297
Strain
(Adeno-associated virus)
AAV2.7m8.hSyn.iGluSnFRVirus facility,
Institute de la Vision, Paris
Software, algorithmNeuronChttps://retina.anatomy.upenn.
edu/rob/neuronc.html
RRID:SCR_014148Version 6.3.14
Software, algorithmSNPE
https://github.com/mackelab/delfi
See Inference algorithm
Software, algorithmCOMSOL MultiphysicsCOMSOL MultiphysicsRRID:SCR_014767
Table 1
Ion channels of biophysical models.
  1. Regions of ion channels and the respective abbreviations as in Figure 1.

    D refers to the combination of DD and PD. All refers to the combination of IS/S, A, and AT.

  2. If multiple regions are stated for a neuron, the ion channel density differs between them.

Table 2
Ion channel implementation details and optimized channel parameters.
ChannelNeuronCTypeStatesParametersChannel remarks and references
Kainate rec.AMPA1MS7STC, τγBased on Jonas et al., 1993.
mGluR6mGluRSTCSee NeuronC documenation.
CaLCA0HH(4)ΔVα, ταBased on Karschin and Lipton, 1989.
CaTCA7MS12ΔVα, ταModification of Lee et al., 2003.
CaPCaPKSee NeuronC documenation.
HCN1/2/4K4MS10Based on Altomare et al., 2001.
KVK0HH(5)ΔVα, ταBased on Hodgkin and Huxley, 1952.
KirK5MS3ΔVαModification of Dong and Werblin, 1995.
ClCaCLCA1MS12Modification of Hirschberg et al., 1998.
NaVNA5MS9ΔVα, ΔVγ, τallBased on Clancy and Kass, 2004.
Table 3
Parameters of discrepancy measures.
δConeBC (3a | 5o)References
pltltupupltltupu
δRateRest05080100003 | 47Choi et al., 2005; Sheng et al., 2007; Berntson and Taylor, 2003; Singer and Diamond, 2006
δVRest−80−55−40−20−80−65−45−30Cangiano et al., 2012; Ichinose et al., 2014; Ichinose and Hellmer, 2016; Ma et al., 2005
δRateΔ0506510005
δVΔ0510200515 | 2540Ichinose et al., 2014; Ichinose and Hellmer, 2016
δVmin−7560−100−80Cangiano et al., 2012; Ma et al., 2005; Saszik and DeVries, 2012
δVmax--−35−20--−100Cangiano et al., 2012; Ma et al., 2005; Saszik and DeVries, 2012
Appendix 1—table 1
Cone model parameter priors.
ParameterUnitaiµibiReferences: direct and additional
CmμFcm20.911.3Oltedal et al., 2009
RmkΩcm215100Oltedal et al., 2009
VrmV-90-70-50
CaLmScm20.1210Morgans et al., 2005; Mansergh et al., 2005; Cui et al., 2012
KVmScm200.13Van Hook et al., 2019
HCN@SmScm20310Knop et al., 2008; Hellmer et al., 2016; Van Hook et al., 2019
HCN@AmScm20110Knop et al., 2008; Hellmer et al., 2016; Van Hook et al., 2019
HCN@ATmScm200.110Knop et al., 2008; Hellmer et al., 2016; Van Hook et al., 2019
ClCamScm200.110Yang et al., 2008; Caputo et al., 2015
CaPμScm20.110100Morgans et al., 1998
τα(CaL)0.7511.5
τα(KV)0.1110
ΔVα(CaL)mV−505
ΔVα(KV)mV−10010
CaPKμmol0.015100
vRRPmaxvesicles102030Thoreson et al., 2016; Bartoletti et al., 2010
gl0.313
Appendix 1—table 2
BC model parameter priors.
ParameterUnitai
(3a|5)
µi
(3a|5)
bi
(3a | 5)
References: direct and additional
CmμFcm20.91.181.3Oltedal et al., 2009
RmkΩ cm2126100Oltedal et al., 2009
VrmV−90−70−50
CaL@SmScm200.53Cui et al., 2012, Van Hook et al., 2019
CaL@ATmScm20.10.53Cui et al., 2012, Van Hook et al., 2019
CaT@SmScm20|n/a0.5|n/a3|n/aCui et al., 2012, Van Hook et al., 2019; Hu et al., 2009; Satoh et al., 1998
CaT@ATmScm20|n/a0.5|n/a3|n/aCui et al., 2012, Van Hook et al., 2019; Hu et al., 2009; Satoh et al., 1998
KV@DmScm200.42Ma et al., 2005
KV@PAmScm200.42Ma et al., 2005
KV@AmScm200.42Ma et al., 2005
Kir@SmScm2012Cui and Pan, 2008, Knop et al., 2008
HCN@DmScm200.22Hellmer et al., 2016, Knop et al., 2008
HCN@SmScm200.22Hellmer et al., 2016, Knop et al., 2008
HCN@ATmScm200.22Hellmer et al., 2016, Knop et al., 2008
NaV@DAmScm2020100Hellmer et al., 2016
CaP@SμScm20.110100Morgans et al., 1998
CaP@ATμScm20.110100Morgans et al., 1998
τγ(Kainate)1|n/a5|n/a20|n/a
τα(CaL)0.512
τα(CaT)0.5|n/a1|n/a2|n/a
τα(KV)0.1110
τall(NaV)0.512
ΔVα(CaL)mV−10010
ΔVα(CaT)mV−10|n/a0|n/a10|n/a
ΔVα(KV)mV−10010
ΔVα(Kir)mV−505
ΔVα(NaV)mV−505
ΔVγ(NaV)mV−505
CaPKμmol0.1|0.0120100
STCmmol0.05|10.5|1.51|3
vRRPmaxvesicles4815Singer and Diamond, 2006
gl0.513
Appendix 1—table 3
Other NeuronC parameters.
ParameterUnitValueRemarks
vnamV65Reversal potential sodium
vkmV−89Reversal potential potassium
vclmV−70Reversal potential chloride
dnaommol151.5Extracellular sodium concentration
dkommol2.5Extracellular potassium concentration
dclommol133.5Extracellular chloride concentration
dcaommol2Extracellular calcium concentration
dicafrac1Fraction of calcium pump current that is added to total current
use_ghki1Use Goldman-Hodgkin-Katz equation
cone_timec0.2Time constant of cone phototransduction
cone_loopg0.0Gain of calcium feedback loop in cones
cone_maxcondnS0.2Max. conductance of of outer segment
timincms0.001 | 0.01Simulation time step (Figure 9)
plotims0.01 | 1Recording time step (Figure 9 and Figure 10 | otherwise)
stimincms0.01 | 0.1Synaptic time step (Figure 9 and Figure 10 | otherwise)
srtimestepms0.001 | 0.01 | 0.1Stimulus update time step (Figure 9Figure 10 | otherwise)
predurs5 | ≥10Simulation time to reach equilibrium potential (Cone | BCs)

Additional files

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Jonathan Oesterle
  2. Christian Behrens
  3. Cornelius Schröder
  4. Thoralf Hermann
  5. Thomas Euler
  6. Katrin Franke
  7. Robert G Smith
  8. Günther Zeck
  9. Philipp Berens
(2020)
Bayesian inference for biophysical neuron models enables stimulus optimization for retinal neuroprosthetics
eLife 9:e54997.
https://doi.org/10.7554/eLife.54997