Neural networks Flashcards
To build a model - identify cellular components
-does stimulating the neuron influence network output? <>
-is this neuron active when the network is active?
<>
So errors of omission and comission
To build a model - identify synaptic components
-Ideal is paired intracellular recordings
<>
-can use extracellular recordings
<>
-monosynaptic or polysynaptic connection?
<>
Wilson 2008 said that synapses with a short latency are ‘unequivocally’ monosynaptic, but there are plenty of exceptions
Wickersham et al 2007 - removed the glycoprotein coat of a virus used to label neurons, which meant it couldn’t do reverse transcription, so could show us only monosynaptic connections. But only 9 of the 11 assumed monosynaptic connections actually were. Can we assume this is experimental error, or could this mean many of the connections we think are monosynaptic actually aren’t?
To build a model - identify cellular properties
Lots (143 families) of different voltage gated ion channels, and they differ cell-to-cell (splicing, post-translational modification, different subunits)
These result in differen cell-intrinsic firing properties.
Think Tim O’Leary - even between animals, cells can have different intrinsic properties! Conductances vary 3-5 fold.!
To build a model - identify synaptic properties
Even the same class of synapse will have different properties - quantal parameters will vary, and STP will vary
~200 transmitters
Activity-dependent and synapse-specific (even divergence within the same cell)
To build a model - look for motifs
Koch 2012 - He used Bell’s number to suggest that it would take 2000 years to examine a network of 1000 components using all Selverston’s 1980 criteria. BUT his networks had everything connected to everything - he then said we don’t need to assume this, but we should look at motifs instead, such as the one shared by the mammalian hippocampus and the aplysia
Arshavsky et al 1993 - /Clione/ mollusk and /Xenopus/ have different neural architecture but both have halfcentres, with mutual inhibition providing alteration, and post-inhibitory rebound contributing to rhythmicity, and descending inhibition to turn the whole thing off.
Types of plasticity
Activity-dependent (i.e. STP and LTP)
Neuromodulation (effects via GPCRs, may have no effect on their own but alter response to an existing input. Can make networks multi-functional. STG has 30 neurons, 60 neuromodulators)
Homeostatic (to maintain network output in the face of an external challenge. E.g. Younger et al 2013 - at the Drosophila NMJ, reduced postsynaptic NT receptor sensitivity is precisely counterbalanced by increased NT release, which is not the result of an increase in active zone number, but in probability of release. This is achieved through altered channel expression - combined mRNA of Degenerin and ENaC subunits was increased during homeostatic plasticity, and blocking them abolished the plasticity without altering baseline NT release.)
Intrinsic vs extrinsic plasticity
Extrinsic is from a neuron or other influence not part of the network itself - e.g. neuromodulation was thought to be totally extrinsic before (CoG neurons as an example)
Intrinsic is when the plasticity derives from a cell within the network, so most of the STP mechanisms, but also some neuromodulation - e.g. Dorsal Swim Interneuron in Tritonia releases 5-HT rhythmically. If you block 5-HT receptors, you no longer get rhythmic network output. There’s a grey area between intrinsic plasticity and basic network/cellular properties.
BCM Model - necessity
1982 - to avoid a positive feedback loop that causes false salience, excitotoxicity or silent synapses, there must be a threshold for plasticity that shifts when the synapse strength shifts. They made a mathematical model for V1, but never used the word ‘metaplasticity’ (that was Abraham and Bear, 1996)
my own point - there must be two thresholds, because there’s a range that results in no LTP/LTD at all
BCM model - evidence
5Hz stimulation blocks LTP in CA1 neurons for 60-90 minutes. This block is NMDA-dependent, because it’s abolished by AP-5 (so NMDA not only mediates LTP/LTD, but also monitors postsynaptic activity to regulate plasticity). A stronger stimulus was still able to cause LTP, so it’s a threshold shift.
10Hz stimulation primed LTP in CA1 neurons
600 x 1Hz stimulation primed LTD. Effect abolished by AP-5
Hulme et al 2014- Priming that inhibited LTP was blocked by carbenoxelone (blocks astrocyte gap junctions) or by adenosine receptor antagonist. When the priming stimulation is applied in the strata oriens, calcium waves are seen in astrocytes in the strata radiata, which are blocked
Combinatorial plasticity by different neuromodulators
Brezina et al - Aplysia feeding circuit has two neurons synapsing onto tongue, both releasing ACh, one releasing Small Cardioactive Peptide and one releasing MyoModulator. SCP causes increase in amplitude and narrows AP, but not as much at high concentrations. MM increases amplitude at low concs, decreases it at high. Both increase cAMP (which opens Ca channels, increasing both contraction and relaxation rate) but also activate K+ channels (which act as shunts). SCP is more potent at the former, MM at the latter.
The combined effects mean more flexibility than either in isolation, e.g. affect relaxation rate without affecting contraction rate.
Problems for model building - off-target effects
Ochi et al - stimulated the motor cortex with optogenetics, and paw movement was lost. Lesioned the motor cortex, and paw movement was retained. So somehow the basal ganglia and cerebellum must have been affected by motor cortex activity. Diaschisis is about downstream effects, and feedback is essentially ubiquitous, so diaschisis is unavoidable (i.e. should be considered the norm!)
EEGs pick up electrical fields, which we assume are irrelevant and don’t affect neuronal activity, but if they do then picking a network apart into individual components is useless, and we can’t study quiescent networks because they’d normally be influenced by activity in distant neurons.
Also vicariation - when one region is lesioned/inactivated, another can compensate.
New techniques (4)
Imaging (esp Calcium or voltage dependent) - temporal precision has increased, we can image the nervous system almost in real-time
Connectomics - like a Google maps of neural networks
Genetic manipulations - can look at effects of perturbing development on a system
Computer modelling - more sophisticated, and easier to run many simulations on (e.g. Prinz et al 2004 database approach to finding equivalent networks in STG)
Optogenetics
Allows us to stimulate or inhibit specific cells, with greater temporal precision than drugs, and bidirectionality.
BUT how ‘specific’ can we really be? We just express it in all cells that have a common marker, so may get false inclusion.
Analyses are complicated by different levels of expression of the probes in different neurons
And activating the laser can cause the probes to heat up, directly affecting function and potentially damaging tissue.
Fictive locomotion - problems
Evidence that networks may be experimentally useful rather than functionally isolated
Isolated CPG gives a slower locomotive pattern than normal, e.g. Wilson’s locust flight
Roberts et al showed the brainstem is needed to maintain spinal CPG output. Brainstem activity is not tonic, but patterned cycle-to-cycle interaction.
Is constant drug application (required to maintain fictive activity) physiological?
Perhaps fictive activity is just a network component?
Adding intracellular recording to ventral root recording often yields less regular traces - maybe because the intracellular recording is affecting network output, but more likely that this is more accurate and there are fewer studies so fewer examples of regular activity (and more of the actually more common irregular activity)
Regular behaviour isn’t even physiological - may represent ‘trapping’, as experienced in Parkinson’s disease. Real locomotion is irregular, a critical system on the border between regular and chaotic activity, to allow flexibility.
CPGs without sensory input
Using UAS-Gal4 system to target TTX to peripheral NS in Drosophila slows larval crawling.
De-afferented locust thoracic ganglion has decreased interval stability in wingbeats
Mutations in Erg3 gene specifically disrupt muscle spindle activation, and when expressed in mice impairs locomotor coordination
If fictive activity were truly representative of normal activity, then these effects would not happen.
