Demo of bmtool chemical synaptic tuner using only Neuron¶

By Gregory Glickert

First we must define some general settings and the settings for the connection we would like to tune. Below is an example of what this could look like for excitatory and inhibitory connections. Currently all of these settings, but the ones in the spec_syn_param are needed in order to use the tuner. If you dont include the general settting when initalizing the tuner then by default it will use these settings. The spec_settings are going to depend on your exact use case and connection type.

If you are using the tuner for a BMTK network you can look at this notebook, but make sure to also check out the bmtk_tuner notebook in this same directory.

In [1]:

general_settings = {
    'vclamp': True, # if vclamp should start on or off used mostly for singleEventv
    'rise_interval': (0.1, 0.9), #10-90%
    'tstart': 500., # when the singleEvent should start
    'tdur': 100.,    # Dur of sim after single synaptic event has occured
    'threshold': -15., #threshold for spike in mV
    'delay': 1.3, # netcon delay
    'weight': 1., # netcon weight
    'dt': 0.025, # simulation dt
    'celsius': 20 # temp of sim
}

conn_type_settings = {
    'Fac2FSI': { # facilitating synapse 
        'spec_settings': {
            'post_cell': 'FSI_Cell', 
            'vclamp_amp' : -70., # voltage clamp amps
            'sec_x': 0.5, # location of synapse
            'sec_id': 1,  # location of synapse 
            "level_of_detail": "AMPA_NMDA_STP", # name of mechanism from the modfile
        },
        'spec_syn_param': { # synaptic parameters from modfile 
            'initW': 0.76,
            'tau_r_AMPA': 0.45,
            'tau_d_AMPA': 7.5,
            'Use': 0.13,
            'Dep': 0.,
            'Fac': 200.
        },
    },
    'Dep2FSI': { # depressing synapse
        'spec_settings': {
            'post_cell': 'FSI_Cell',
            'vclamp_amp': -55,
            'sec_x': 0.5,
            'sec_id':0,
            "level_of_detail": "GABA_A_STP",
        },
        'spec_syn_param': {
            'initW': 20,
            'tau_r_GABAA': 0.9,
            'tau_d_GABAA': 15,
            'e_GABAA':-75,
            'Use': 0.4,
            'Dep': 190.,
            'Fac': 0.
        },
    },

}

general_settings = { 'vclamp': True, # if vclamp should start on or off used mostly for singleEventv 'rise_interval': (0.1, 0.9), #10-90% 'tstart': 500., # when the singleEvent should start 'tdur': 100., # Dur of sim after single synaptic event has occured 'threshold': -15., #threshold for spike in mV 'delay': 1.3, # netcon delay 'weight': 1., # netcon weight 'dt': 0.025, # simulation dt 'celsius': 20 # temp of sim } conn_type_settings = { 'Fac2FSI': { # facilitating synapse 'spec_settings': { 'post_cell': 'FSI_Cell', 'vclamp_amp' : -70., # voltage clamp amps 'sec_x': 0.5, # location of synapse 'sec_id': 1, # location of synapse "level_of_detail": "AMPA_NMDA_STP", # name of mechanism from the modfile }, 'spec_syn_param': { # synaptic parameters from modfile 'initW': 0.76, 'tau_r_AMPA': 0.45, 'tau_d_AMPA': 7.5, 'Use': 0.13, 'Dep': 0., 'Fac': 200. }, }, 'Dep2FSI': { # depressing synapse 'spec_settings': { 'post_cell': 'FSI_Cell', 'vclamp_amp': -55, 'sec_x': 0.5, 'sec_id':0, "level_of_detail": "GABA_A_STP", }, 'spec_syn_param': { 'initW': 20, 'tau_r_GABAA': 0.9, 'tau_d_GABAA': 15, 'e_GABAA':-75, 'Use': 0.4, 'Dep': 190., 'Fac': 0. }, }, }

Then the modfiles must be compiled in order for the tuner to work properly

In [2]:

import os

# if already compiled then lets delete the folder and force a recompile
if os.path.isdir('modfiles/x86_64'):
    os.system("rm -rf modfiles/x86_64 ")
# compile the mod files
if not os.path.isdir("modfiles/x86_64"):
    os.chdir('modfiles')
    os.system("nrnivmodl")
    os.chdir("..")

import os # if already compiled then lets delete the folder and force a recompile if os.path.isdir('modfiles/x86_64'): os.system("rm -rf modfiles/x86_64 ") # compile the mod files if not os.path.isdir("modfiles/x86_64"): os.chdir('modfiles') os.system("nrnivmodl") os.chdir("..")

/Users/gregglickert/miniconda3/envs/bmtk/bin/nrnivmodl:10: DeprecationWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html
  from pkg_resources import working_set

/usr/bin/xcrun
/Users/gregglickert/Documents/GitHub/bmtool/docs/examples/notebooks/synapses/synaptic_tuner/modfiles
Mod files: "./AMPA_NMDA_STP.mod" "./GABA_A_STP.mod" "./Gfluct.mod" "./cadad.mod" "./cal2.mod" "./can_mig.mod" "./exp2syn_stp.mod" "./gap.mod" "./h_kole.mod" "./imCA3.mod" "./kBK.mod" "./kap_BS.mod" "./kdmc_BS.mod" "./kdrCA3.mod" "./kdr_BS.mod" "./kdrinter.mod" "./leak.mod" "./nainter.mod" "./napCA3.mod" "./natCA3.mod" "./nax_BS.mod" "./vecevent_coreneuron.mod"

 -> Compiling mod_func.cpp
 => LINKING shared library ./libnrnmech.dylib
Successfully created arm64/special

initialize tuner¶

Now we can initialize the synaptic tuner. When initializing you will have to change a few arguments depending on your use case. other_vars_to_record can be any variable in your synaptic mechanism, while slider_vars can be any range variable you wish to tune in the synapse. If the variable is not defined in the spec_syn_param than the tuner will get the value from the mechanism and try to set up some sliders to tune it.

In [3]:

mechanisms_dir = 'modfiles'
templates_file = 'templates.hoc'

# Initialize our tuner
from bmtool.synapses import SynapseTuner

tuner = SynapseTuner(mechanisms_dir=mechanisms_dir, # where x86_64 is located
                    templates_dir=templates_file, # where the neuron templates are located
                    conn_type_settings=conn_type_settings, # dict of connection settings
                    general_settings = general_settings, # dict of general settings
                    connection = 'Dep2FSI', # key in connection settings for which connection you want to tune
                    current_name = 'i', # name of current variable in synapase
                    other_vars_to_record = ['record_Pr', 'record_use'], # Other synaptic variables you wish to record besides the normal ones
                    slider_vars=['initW','Dep','Fac','Use']) # Range variables you want to tune to adjust synaptic response.

mechanisms_dir = 'modfiles' templates_file = 'templates.hoc' # Initialize our tuner from bmtool.synapses import SynapseTuner tuner = SynapseTuner(mechanisms_dir=mechanisms_dir, # where x86_64 is located templates_dir=templates_file, # where the neuron templates are located conn_type_settings=conn_type_settings, # dict of connection settings general_settings = general_settings, # dict of general settings connection = 'Dep2FSI', # key in connection settings for which connection you want to tune current_name = 'i', # name of current variable in synapase other_vars_to_record = ['record_Pr', 'record_use'], # Other synaptic variables you wish to record besides the normal ones slider_vars=['initW','Dep','Fac','Use']) # Range variables you want to tune to adjust synaptic response.

Warning: no DISPLAY environment variable.
--No graphics will be displayed.

SingleEvent¶

The SingleEvent method will run a short pulse and then print out the synaptic properties for the synapse.

In [4]:

tuner.SingleEvent()

tuner.SingleEvent()

('baseline', 0.03107559632675816)
('sign', 1.0)
('latency', 1.35)
('amp', 0.16000855570656114)
('rise_time', 1.35)
('decay_time', 15.182347895738655)
('half_width', 14.850000000000001)
Current Integral in pA*ms: -15777.35

InteractiveTuner¶

The InteractiveTuner will deliver an input to the cell at a desired weight and frequency. The frequency by default will be 8 spikes then a 250ms delay and then 4 more spikes.

Paired-pulse ratio is (Avg 2nd pulse - Avg 1st pulse) ÷ 90th percentile amplitude.

Induction is (Avg (6th, 7th, 8th pulses) - Avg 1st pulse) ÷ 90th percentile amplitude.

Recovery is (Avg (9th, 10th, 11th, 12th pulses) - Avg (1st, 2nd, 3rd, 4th pulses)) ÷ 90th percentile amplitude

In [5]:

tuner.InteractiveTuner()

tuner.InteractiveTuner()

VBox(children=(HBox(children=(Dropdown(description='Connection:', options=('Dep2FSI', 'Fac2FSI'), style=Descri…

Output()

Frequency reponse¶

We can also see how the STP parameters vary with different train frequencies

In [6]:

results = tuner.stp_frequency_response(log_plot=False)

results = tuner.stp_frequency_response(log_plot=False)

Analyzing frequencies:   0%|          | 0/16 [00:00<?, ?it/s]

SynapseOptimizer¶

If we don't feel like tuner by hand we can also try to optimize an output of our model. In this example we will optimize and find the best STP parameters that give the induction and paired pulse response we want. Something to note is that the optimizer does not know what the trace should look like and only knows the features. So it might get some wild trace that happens to work. Also if you are using the random init_guess and don't like the voltage trace then run it again. The seed is different each time so the optimization will be different and can result in a better fit.

In [12]:

from bmtool.synapses import SynapseOptimizer

# Create the optimizer
optimizer = SynapseOptimizer(tuner)

# Define parameter bounds these can be any range variable you wish to tune
param_bounds = {
    'Dep': (0, 800.0),
    'Fac': (0, 100.0),
    'Use': (0.1, 1.0),
}

# Define target metrics these are the metrics that the tuner will try to automatic get the synapse to respond with
# max amps is an absolute value
target_metrics = {
    'induction': -0.75,
    'simple_ppr': 0.8,
    'recovery': 0.0,
}

# currently the metrics in the SynapseOptimizer are 
#            - induction: measure of synaptic facilitation/depression
#            - ppr: paired-pulse ratio from allen
#            - simple_ppr ration of first to second amplitude 
#            - recovery: recovery from facilitation/depression
#            - max_amplitude: maximum synaptic response amplitude
#            - rise_time: time for synaptic response to rise from 20% to 80% of peak
#            - decay_time: time constant of synaptic response decay
#            - latency: synaptic response latency
#            - half_width: synaptic response half-width
#            - baseline: baseline current
#            - amp: peak amplitude from syn_props
def custom_cost(metrics, targets):
    # equal zero unless using train input
    induction_error = (metrics['induction'] - targets['induction']) ** 2
    ppr_error = (metrics['simple_ppr'] - targets['simple_ppr']) ** 2
    recovery_error = (metrics['recovery'] - targets['recovery']) ** 2

    return induction_error + 3 * ppr_error + recovery_error 

# Run optimization with custom cost function
result = optimizer.optimize_parameters(
    target_metrics=target_metrics,
    param_bounds=param_bounds,
    run_single_event=True,  # Run and use parameters from SingleEvent
    run_train_input=True,   # Run and use parameters from train input 
    train_frequency=50,     # Freq in Hz of train input
    train_delay=250,        # delay in ms of second train
    init_guess='middle_guess',    # either random or middle_guess. Random will start the synapse witha random value in the param_bound. Middle guess will pick the middle value in the param_bounds
    cost_function=custom_cost,
    method='SLSQP')          # check out https://docs.scipy.org/doc/scipy-1.15.0/reference/generated/scipy.optimize.minimize.html
                               

# Plot results
optimizer.plot_optimization_results(result)

from bmtool.synapses import SynapseOptimizer # Create the optimizer optimizer = SynapseOptimizer(tuner) # Define parameter bounds these can be any range variable you wish to tune param_bounds = { 'Dep': (0, 800.0), 'Fac': (0, 100.0), 'Use': (0.1, 1.0), } # Define target metrics these are the metrics that the tuner will try to automatic get the synapse to respond with # max amps is an absolute value target_metrics = { 'induction': -0.75, 'simple_ppr': 0.8, 'recovery': 0.0, } # currently the metrics in the SynapseOptimizer are # - induction: measure of synaptic facilitation/depression # - ppr: paired-pulse ratio from allen # - simple_ppr ration of first to second amplitude # - recovery: recovery from facilitation/depression # - max_amplitude: maximum synaptic response amplitude # - rise_time: time for synaptic response to rise from 20% to 80% of peak # - decay_time: time constant of synaptic response decay # - latency: synaptic response latency # - half_width: synaptic response half-width # - baseline: baseline current # - amp: peak amplitude from syn_props def custom_cost(metrics, targets): # equal zero unless using train input induction_error = (metrics['induction'] - targets['induction']) ** 2 ppr_error = (metrics['simple_ppr'] - targets['simple_ppr']) ** 2 recovery_error = (metrics['recovery'] - targets['recovery']) ** 2 return induction_error + 3 * ppr_error + recovery_error # Run optimization with custom cost function result = optimizer.optimize_parameters( target_metrics=target_metrics, param_bounds=param_bounds, run_single_event=True, # Run and use parameters from SingleEvent run_train_input=True, # Run and use parameters from train input train_frequency=50, # Freq in Hz of train input train_delay=250, # delay in ms of second train init_guess='middle_guess', # either random or middle_guess. Random will start the synapse witha random value in the param_bound. Middle guess will pick the middle value in the param_bounds cost_function=custom_cost, method='SLSQP') # check out https://docs.scipy.org/doc/scipy-1.15.0/reference/generated/scipy.optimize.minimize.html # Plot results optimizer.plot_optimization_results(result)

Optimization Results:
Final Error: 2.91e-07

Target Metrics:
induction: -0.750 (target: -0.750)
simple_ppr: 0.800 (target: 0.800)
recovery: 0.000 (target: 0.000)

Optimal Parameters:
Dep: 144.855
Fac: 99.077
Use: 0.680

========================================
Short Term Plasticity Results for 50Hz with 250 Delay
========================================
Simple PPR: Above 1 is facilitating, below 1 is depressing
PPR:        Above 0 is facilitating, below 0 is depressing.
Induction:  Above 0 is facilitating, below 0 is depressing.
Recovery:   A measure of how fast STP decays.

Simple Paired Pulse Ratio (PPR)
    Calculation: Avg 2nd pulse / Avg 1st pulse
    Values: 54.576 / 68.223 = 0.800

Paired Pulse Response (PPR)
    Calculation: (Avg 2nd pulse - Avg 1st pulse) / 90th percentile amplitude
    Values: (54.576 - 68.223) / 66.858 = -0.204

Induction
    Calculation: (Avg(6th, 7th, 8th pulses) - Avg 1st pulse) / 90th percentile amplitude
    Values: 18.053 - 68.223 / 66.858 = -0.750

Recovery
    Calculation: (Avg(9th, 10th, 11th, 12th pulses) - Avg(1st to 4th pulses)) / 90th percentile amplitude
    Values: 44.122 - 44.122 / 66.858 = 0.000

========================================

('baseline', -0.0002407732493736603)
('sign', 1.0)
('latency', 1.35)
('amp', 0.06798172722533025)
('rise_time', 1.35)
('decay_time', 15.182348785123999)
('half_width', 14.850000000000001)
Current Integral in pA*ms: 1362.98