Philosophy of Computational Neuroscience

Just like experimental neuroscience, computational neuroscience can be done well or poorly.

computational models look beautiful (source)

This post was motivated by Janet Stemwedel's recent post in Adventures in Ethics and Science about the philosophy of computational neuroscience. There seem to be three views of the use of computational models in biology and neuroscience:

1. All models are bullshit.
2. Models rely on MATH, so of course they are right.
3. Some models are good and some are bad.

Obviously the first two are extremes and usually posited by people who don't know anything about computational neuroscience, and I am clearly advocating the third view. The only problem is that it is hard to tell if a model is good or bad unless you know a lot about it.

So here are some general principles that can help you divide the good and the bad in computational neuroscience.

1. The authors use the correct level of detail.

devil's in the details (source)

If you are trying to test how brain regions interact with each other, you don't need to model every single cell in each region, but you need to have enough detail to differentiate the brain regions from one another. Similarly, if you are trying to test how molecules diffuse within a dendrite, you don't need to model a whole cell, but you need to have enough detail to differentiate one molecule type from another. If you are trying to test how a cell processes information, you need to have a cell, as you may have learned in how to build a neuron.  Basically a model can be bad simply because it is applied to the wrong question.

2. The authors tune and validate their model using separate data.

When you are making a model you tune it to fit data. For example, in a computational model of a neuron you want to make sure your particular composition of channels produces the right spiking pattern. However, you also want to validate it against data. So how is tuning different from validating? Tuning is when you change the parameters of the model to make it match data. Validating is when you check the tuned model to see if it matches data. Good practice in computational neuroscience is to tune your model to one set of data, but to validate it against a different set of data.
For example, if a cell does X and Y, you can tune your model to effect X, but then check to see that the parameters that make it do X also make it do Y. Sometimes this is not possible. Maybe there is not enough experimental data out there. But if it is not possible, you should at least test the robustness of your model (see point 3).

3. The authors test the robustness of their model.

A robust computational model can be delicious (source)

One problem with computational models is that the specific set of parameters you've found by tuning the model might not be the 'right ones.' In fact they probably aren't the right ones. There are many different sets of parameters that can make a neuron spike slowly, for example.  And the chance that you hit on exactly the correct combination of things is very low. But that doesn't mean the model is not useful. You can still use the model to test effects that are not strongly altered by small changes in these parameters. So you need to test whether the specific effect you are testing is robust to parameter variation. If you are testing effect Q, you can increase the sodium channels by 10%, or the network size by 20% and see if you still get effect Q. In other words is 'effect Q' robust to changes in sodium channels or network size? If it is, then great! Your effect is not some weird fluke due to the exact combination of parameters that you have used.

These are the main things I try to pay attention to, but I am sure there are other important things to keep in mind when making models and reading about them. What are your thoughts?

© TheCellularScale

How to Build a Neuron: The ultimate shortcut

We've gone through all the main steps for Building a Neuron, and  compiled them here. But there is one last shortcut that I absolutely have to tell you about. It's pretty much the ULTIMATE shortcut.

Full of Computational Models (source)

If you really want a computational model of a neuron, but just don't want to do the work to build one, even using the shortcuts I've already provided... you can use someone else's.

Model DB is a repository where scientists can upload their computational models for all to see and use. A README file is attached to each model and explains the basics of  the model. In addition, there is always a citation to the paper in which the model was first published, so you can read about how it was used. With these tools (and possible emails to the authors) you should be able to download and run a complete model and replicate the figures that someone uses in a paper. The model I mention is step 5 is there, and is an example of a complete and usable neuron with synaptic and intrinsic channels.

Model DB 

Model DB is helping advance the field of computational neuroscience in two ways.

First it allows for post-publication review, where people can make sure that the model does what the authors say it does. Having access to the actual code used to run the simulations is better than just reading the 'methods section' of a paper. As with all methods sections, you just can't explain everything well enough for it to be replicated exactly as performed.

Second, Model DB reduces the amount of overlapping work that scientists have to do. You've seen how tedious it can be to extract the Boltzmann curves of each intrinsic channel. Model DB allows you to see if someone has already done it for your cell type. You can use it just as it is, or you can follow the citations and re-extract it. In either case you have saved yourself valuable time.

One of the advantages of computational models is that they can be used to answer TONS of questions. Most of the time a huge amount of work is put into building the model, and then the model is used to answer a few questions, but the possibilities are endless. Having the models freely accessible allows for faster advancement in both computational neuroscience and the field of neuroscience in general. 

"ModelDB provides a resource for the computational neuroscience community that enables investigators to increase their understanding of published models by enabling them o run the models as published and build on them for further research. Its use can aid the field of computational neuroscience to enter a new era of expedited numerical experimentation." Migliore et al., 2003 

© TheCellularScale

Model DB link: http://senselab.med.yale.edu/modeldb/default.asp

Migliore M, Morse TM, Davison AP, Marenco L, Shepherd GM, & Hines ML (2003). ModelDB: making models publicly accessible to support computational neuroscience. Neuroinformatics, 1 (1), 135-9 PMID: 15055399


Hines ML, Morse T, Migliore M, Carnevale NT, & Shepherd GM (2004). ModelDB: A Database to Support Computational Neuroscience. Journal of computational neuroscience, 17 (1), 7-11 PMID: 15218350

How to Build a Neuron: Step 5

And now, the final step in how to build your computational model of a neuron: Add Synaptic Channels. All the steps in this series can be found here.

Synapses connect neurons (source)

So you already have a neuron, and you've added intrinsic channels to it. The next thing you want to do is add synaptic channels so you can hook this neuron up to other cells.

The main synaptic channels you want to add are the excitatory channels: NMDA and AMPA and the inhibitory channel GABA. These channels don't have the same kind of activation and inactivation curves and the intrinsic channels do because they aren't activated by voltage, they are activated by a neurotransmitter.

AMPA and NMDA receptors are activated primarily by glutamate, and cause an influx of sodium and calcium ions. Since both sodium and calcium ions are positively charged, this depolarizes the cell membrane and brings it closer to firing an action potential.

AMPA receptors (source)

GABA receptors, on the other hand are primarily activated by GABA, and cause and influx of chloride ions into the cell. Because chloride ions are negatively charged, this hyperpolarizes the cell membrane and brings it further away from firing an action potential.

So if you want to have a realistic model of a neuron, you need to add an approximation of these channels. This is easier than adding intrinsic channels, because it is an on/off style (binary) rather than an analogue activation. So basically you just put in the parameters you want like how fast does the channel open and close, how much current does it allow through when activated, and where are they on the neuron.

Of course deciding these parameters is not always easy. A paper out this year in PLoS Computational Biology describes 4 different ways the NMDA receptor can be configured and analyzes the consequences during different stimulation patterns. 

Evans et al., (2012) Figure 3

The 4 NMDA configurations (based on the 4 different GluN2 subunits) vary in their sensitivity to a magnesium block, how fast they decay, and their maximal current. Above are their responses to the same stimulation patterns (an STDP protocol). Even though they were all receiving the same input pattern, they each show a very different response.

So when considering adding synaptic channels to your model neuron, take the time to find out what the configuration of the receptors should actually be in the type of neuron you are building.


© TheCellularScale

If you are good at following clues, you will realize that I am very, very familiar with this paper.


Evans RC, Morera-Herreras T, Cui Y, Du K, Sheehan T, Kotaleski JH, Venance L, & Blackwell KT (2012). The effects of NMDA subunit composition on calcium influx and spike timing-dependent plasticity in striatal medium spiny neurons. PLoS computational biology, 8 (4) PMID: 22536151

How to Build a Neuron: step 4

And now, the next step in neuron building! You can see all the previous steps and shortcuts here. Step 4 is adding intrinsic channels to your neuron.

Potassium Channel (source)

Pretty much all neurons need sodium and potassium channels so they can fire action potentials, but other channels such as calcium channels are also commonly seen in computational models.

To add these channels you have to extract the parameters from known data. This means extracting Boltzmann curves and time constant information so you can tell the channel which voltages activate it and inactivate it and how fast to open and close.

Activation (Boltzmann) curve for fast sodium channel

This step is tricky and can take a long time, but there is some software that can help. The Enguage Digitizer is one tool I could not live without.

Enguage is basically a tool that allows you to manually trace curves from published figures to get the curve data as an excel or .csv file. First you add axis points using the button at the top that has red plus signs on it. You tell the software what values each of the 3 corners of the graph are. Then you click the blue plus signs button and start to trace your graph, like so:

using Enguage digitizer to extract channel data

Then you export the data as whichever type of file you want. Pretty nice!
I like to have the data this way because then I can overlay this figure trace with any other trace I want and can manually fit an equation to it.

Channels are a hugely important part of a computational model. A recent paper from Eve Marder's lab shows that even with a very simple morphological model (just a soma), interesting electrical characteristics can be seen simply by manipulating the channels.

Kispersky et al., 2012 from Figure 1

Kispersky et al., (2012) introduce an interesting paradox. They show that when you increase the sodium channel conductance you see more action potentials with low current injections (like 200pA). This is expected because the sodium channel is what causes the upswing of the action potential and more sodium is thought to mean more excitability. However, the authors find that when a high current injection is given (like 10nA), the increased sodium channel conductance actually decreases the firing rate. This is counter-intuitive because it goes against the more sodium=more excitability rule.

This is a pretty cool finding published in the Journal of Neuroscience using only a simple one-compartment model. The finding is based entirely on channel manipulation, and demonstrates how important these intrinsic channels are to any computational model.


© TheCellularScale


Kispersky TJ, Caplan JS, & Marder E (2012). Increase in sodium conductance decreases firing rate and gain in model neurons. The Journal of neuroscience : the official journal of the Society for Neuroscience, 32 (32), 10995-1004 PMID: 22875933

How to Build a Neuron: step 3

Steps 1 and 2 of neuron-building, as well as an important set of shortcuts can be found in the How to Build a Neuron index. Step 3 is deciding which simulation software or programming language you want to use.

Simulated Neuron in Genesis (source)

The big two are Genesis and Neuron. They are pretty similar in a lot of ways, but Genesis runs in Linux and Neuron runs in Windows. However, you can run Genesis in Windows if you install the Linux environment Cygwin.

Both programs can read in morphological data, but they use different syntax and coding procedures. There are other types of neural simulators as well, and an ongoing problem in the field of computational neuroscience is compatibility between programs. If someone has done the work to make a beautiful Purkinje cell in Genesis like the one above, it will take a lot of time and effort to translate that neuron into a different simulator such as Neuron.

Gleeson et al., (2010) explains this problem and presents a possible solution in the form of the "Neuron Open Markup Language" or NeuroML.

"Computer modeling is becoming an increasingly valuable tool in the study of the complex interactions underlying the behavior of the brain. Software applications have been developed which make it easier to create models of neural networks as well as detailed models which replicate the electrical activity of individual neurons. The code formats used by each of these applications are generally incompatible however, making it difficult to exchange models and ideas between researchers....Creating a common, accessible model description format will expose more of the model details to the wider neuroscience community, thus increasing their quality and reliability, as for other Open Source software. NeuroML will also allow a greater “ecosystem” of tools to be developed for building, simulating and analyzing these complex neuronal systems." -Gleeson et al (2010) Author Summary

NeuroML is basically a "simulator-independent" neuronal description language. A neuron built with or converted to NeuroML should be able to run on Neuron, Genesis, and plenty of other platforms. Gleeson et al. validated NeuroML by using a simulated pyramidal neuron converted to NeuroML format and run with several different simulators.

Gleeson et al., (2010) Figure 7

Zooming in:

Neuron, Genesis, Moose, Psics comparison

All the simulators overlay so tightly that you can barely tell that they are separate lines.

So when building you neuron, take care to follow the NeuroML format and then you and others can use it with any simulator you want.

© TheCellularScale


Gleeson P, Crook S, Cannon RC, Hines ML, Billings GO, Farinella M, Morse TM, Davison AP, Ray S, Bhalla US, Barnes SR, Dimitrova YD, & Silver RA (2010). NeuroML: a language for describing data driven models of neurons and networks with a high degree of biological detail. PLoS computational biology, 6 (6) PMID: 20585541

Cut your brain some SLACK

Action potentials are the main means of communication between neurons, and their exact timing can be really important. But the specific timing of action potentials is really important in the auditory system, because the auditory system encodes (among other things) information about sound wave frequency.

Sound waves (source)

I've previously written about auditory processing with regards to the wonder that is the chicken brain, but today we will focus on timing-specificity in the mammalian brainstem. Specifically, some weird channels in the Medial Nucleus of the Trapezoid Body (the MNTB).


Mammalian Auditory Brainstem (source)

At the Society for Neuroscience meeting, I learned about the sodium-activated potassium channels which help the electric fish fire super-fast super-large action potentials. I was suprised to learn that sodium-activated potassium channels are located in many parts of the mammalian brain.

A paper from the Kaczmarek lab at Yale explains that these sodium-activated potassium channel (SLICK and SLACK) are present in the mouse auditory brainstem and contribute to the 'temporal accuracy' of the MNTB neurons. Yang et al. (2007) record the action potentials from these neurons at a range of frequencies and show that the neuron can 'keep' up with the frequencies better when more sodium is present.

Yang et al., 2007 Figure 9B

In the figure above, the 'flatter' the line, the better the 'temporal accuracy.' They also made a computational model of this neuron and ran simulations altering the sodium values and reversal potential.

Yang et al., 2007 Figure 9D

Their model simulations are similar to their experimental recordings, in that more sodium results in more temporal accuary of the action potential. They confirmed that this was dues to a sodium-activated potassium channel by directly activating SLACK and seeing a similar improvement in temporal accuracy.

The SLACK channel still blows my mind, but its role in helping the auditory system fire with the utmost precision actually makes a lot of sense.


© TheCellularScale



Yang B, Desai R, & Kaczmarek LK (2007). Slack and Slick K(Na) channels regulate the accuracy of timing of auditory neurons. The Journal of neuroscience : the official journal of the Society for Neuroscience, 27 (10), 2617-27 PMID: 17344399

How to Build a Neuron: Shortcuts

So you want to build a neuron, but don't have the time to fill and stain it, digitally reconstruct it, or even to knit one.

Knitting Neuroscience from Knit a Neuron

Well you are in luck because a lot of scientists have collected a lot of data already and some of them are even willing to openly share their work. 

While it is great that people are willing to share their data, that willingness alone is not enough to actually make the data widely accessible (or searchable for that matter). To bridge the chasm, other scientists have developed databases and repositories.  These databases and repositories store large datasets and organize them in a searchable way. 

The first shortcut to building a neuron I will discuss is the Cell Centered Database (CCDB).

Sounds a little like "self-centered" but represents just the opposite: scientists willing to share their data with everyone

In 2003, Martone and colleagues created the CCDB as a repository for 2D, 3D, and 4D images of cells that could be downloaded and used by researchers around the globe. There is a ton of data here, protein stains, electron microscopy, and fluorescent confocal images just to name a few.  While you could do a lot with this kind of information, I am just going to give you one example of how it can be used as a major short cut in the process of building a neuron.

So say you want to make a model of a cerebellum purkinje cell, but you don't have the time or lab facilities to fill and stain your own neuron.  You could go to CCDB, type in 'purkinje neuron' in the search box and download whichever 3D image stack suits your fancy. 

example Purkinje neuron that I just got from CCDB

With this data you could go straight to step 2: reconstructing the neuron. 

But what if you don't have the time to digitally reconstruct the neuron?  We have already discussed how much time reconstructing a neuron can take, so it's pretty easy to see why you would want to bypass that step too. And in fact, there is a database for that!

Halavi et al (2008) developed Neuromorpho.org as a repository for neural reconstructions. Neuromorpho.org has almost 8,000 downloadable digital reconstructions of neurons, which as they say on the website represents over 200,000 hours of manual reconstruction time. 

NeuroMorpho.org, for all your neural needs.

Similar to CCDB, Neuromorpho offers much more than just a shortcut for lazy computational modelers. It has such detailed information about each neuron that a whole project could be done simply by comparing neural characteristics of different cell classes or different species. 

But my job here is to tell you how you can use it as a shortcut to building a neuron.

Say you want to build a computational model of a CA1 Hippocampal Pyramidal Cell, but you don't want to stain it and you don't want to reconstruct it.  Well, just go to Neuromorpho.org and click 'browse by brain region' and then on 'hippocampus'. Then look through the 1,000 hippocampal cells (organized by class) that have already been reconstructed for you...

Pyramidal cell in the Hippocampus from Neuromorpho.org
 

...and pick your favorite. 

Then you can jump right on through to step 3. (coming soon)


Halavi M, Polavaram S, Donohue DE, Hamilton G, Hoyt J, Smith KP, & Ascoli GA (2008). NeuroMorpho.Org implementation of digital neuroscience: dense coverage and integration with the NIF. Neuroinformatics, 6 (3), 241-52 PMID: 18949582

Martone ME, Tran J, Wong WW, Sargis J, Fong L, Larson S, Lamont SP, Gupta A, & Ellisman MH (2008). The cell centered database project: an update on building community resources for managing and sharing 3D imaging data. Journal of structural biology, 161 (3), 220-31 PMID: 18054501

How to Build a Neuron: Step 2

Recently we've discussed the first step in how to build a neuron. Today we will discuss step 2: reconstructing that stained cell.

Hippocampus CA1 Pyramidal neuron (from Neuromorpho.org)

There are a couple of ways that you turn an image (or image stack) of a neuron into a digital neuron file like the one pictured above.  Basically there is an easy way and a hard way.  The hard way is to reconstruct the neuron manually, where you literally trace the neuron by hand.  The easy way is to auto-trace the neuron.

In a recent Frontier's in Neuroinformatics article, Myatt et al. (2012) explain the hard to easy gradient in reconstruction methods.

• "Manual (Camera lucida). Prisms are employed to visually overlay the microscope image onto a piece of paper, and the neuron is then traced by hand. Although primarily used for 2D tracings, 3D reconstructions can be derived from these with time consuming post-processing (Ropireddy et al., 2011).
• Semi-manual (e.g., Neuron_Morpho, Neurolucida). Digital segments are added by hand through a software interface, typically sequentially, beginning at the soma, and working down the dendritic tree.
• Semi-automatic [e.g., NeuronJ (Meijering et al., 2004; 2D reconstruction only) and Imaris (3D reconstruction)]. User interaction defines the basic morphology, such as identifying the tree root and terminations, but branch paths are traced by the computer
• Fully automatic (e.g., Imaris, NeuronStudio; Rodriguez et al., 2003, AutoNeuron add-on for Neurolucida). The entire morphology is extracted with minimal user-input. " (Myatt et al., 2012)

You may ask: "Why not just do it the easy way?" Good question.  It is actually surprisingly difficult to make a versatile program that can accurately reconstruct neurons.  So difficult in fact that in 2010 an open challenge was issued with a monetary prize for the best automatic reconstruction algorithm. Five teams competed in this DIADEM challenge and the results and process are explained in detail in a special issue of Neuroinformatics. (And in less detail in this HHMI press release)

automatic reconstructions of neurons (source)

Advances in automatic reconstruction are being made at an astounding pace, but most neural reconstructions are still being done in a semi-manual or semi-automatic way. 

If you are interested in reconstructing some neurons, you can download Neuromantic for free or Neurolucida for money. There is other reconstruction software available, summarized nicely in Myatt et al. 2012, but these are the two I am most familiar with. 

In the next edition of "How to Build a Neuron" I will tell you how you can completely skip step 1 (the staining of the neuron) and step 2 (the reconstruction of the neuron). 

For ease of access, the whole "How to Build a Neuron" series is archived.


© TheCellularScale


Myatt DR, Hadlington T, Ascoli GA, & Nasuto SJ (2012). Neuromantic - from semi-manual to semi-automatic reconstruction of neuron morphology. Frontiers in neuroinformatics, 6 PMID: 22438842

How to Build a Neuron: Step 1

There are many reasons to try to build a neuron, but fully building a model neuron is an extensive process with many steps.  Today we will discuss the very first step in the neuron-building process: determining the activity and  shape of the neuron.

Biocytin filled cortical neurons (source)

To determine the shape of neuron, you have to stain it somehow.  There are several ways to do this, but we will focus on the biocytin filling method.
To determine the activity of a neuron, you have to use electrophysiology to record its electrical activity. The biocytin filling method makes use of the same patch clamp electrode to record the electrical activity of the neuron and to fill it with the biocytin molecule that can be later dyed.  So this method is perfect for building a neuron because with it you can correlate the shape of the neuron directly with its activity patterns. 

Neural activity correlated with neural morphology (source)

A recent Nature Protocols paper by Marx et al. (2012) provides step by step details for how to fill and dye a neuron using the biocytin method. 

The basic biocytin staining protocol is as follows:

1. make brain slices
2. fill the neuron with biocytin while recording its electrical activity
3. fix the brain slice in paraformaldehyde
4. quench the endogenous peroxidase
5. connect the biocytin to avidin (using the vectastain ABC kit)
6. colorize the avidin (using DAB and nickel)
7. mount the slices on gelatin subbed slides
8. dehydrate the slices SLOWLY through very small steps of ethanol concentration
9. clear with xylene and coverslip

Marx et al. provide some excellent specifics in the paper that make the whole process understandable and more importantly, doable. They even have a troubleshooting section which explains what might have gone wrong under several conditions.

Marx et al., 2012 Figure 2

One of their best tips in the paper is to dehydrate the slices very slowly.  They show that when you dehydrate the tissue quickly, you get a cork-screw artifact (A) that is not physiologically meaningful, but when you dehydrate slowly, you get a more accurate morphology. 

So there you have it, Step 1 of neuron building.  Step 2 will be coming soon, can be found here.
And all the "build a neuron" steps will be indexed here.

© TheCellularScale


Marx M, Günter RH, Hucko W, Radnikow G, & Feldmeyer D (2012). Improved biocytin labeling and neuronal 3D reconstruction. Nature protocols, 7 (2), 394-407 PMID: 22301777

Neurons are like Fireworks

Neuron Firework (source)

In honor of American Independence, today here are some beautiful pictures of neurons that look sort of like fireworks.

(source)

And some actual neuroscience studies using light-up cells.  One example, I've already covered is the imaging of neurons as they fire. Since neurons fire really fast, you can see them light up under the microscope.



Computational neuroscience can lead to fireworks as well.  When neurons fire in an artificial network, they light up in beautiful patterns:  (It really sparkles at 2:00)




 © TheCellularScale

Neuron-controlled robots: reverse-cyborgs

Last post we discussed robotically controlled biology.  In this post we will talk about biologically controlled robots.

The Hybrot: a rat neuron controlled robot

In 2001, S. Potter published a paper on the "Animat". A set of cultured neurons on a multi-electrode array (MEA, purple circle in above image) interfaced with a simulated robot.  That is, not a physical moving around robot as pictured above, but a computer program simulating what a robot/animal could do. 

They made a virtual room for the animat to 'explore'. (If you can make a virtual environment for a worm, I suppose you can make one for a petri dish of cultured neurons) The signal from the cultured neurons determined where the animat went. If one group of neurons fired, the animat moved left, if another group fired it moved forward, and so forth. (The actual equations translating neuronal activity to animat movement were more complex than this, but you get the idea.) 

So here's the really cool thing: When the animat 'hit a wall', a set of neurons were stimulated with an electric pulse. They also gave the cultured neurons a sort of vestibular system, stimulating a different area depending on which direction the Animat was traveling.

Although this Animat study was using a simulated environment and a simulated robot, using cultured neurons to control an actual robot was only a matter of time. 





Neurons are somehow even cooler when they are combined with robots, no?

So what I think is really exciting about this reverse-cyborg system is that you can study the formation of neuronal networks in response to realistic experience. The feedback system used in the Animat could reveal how natural synaptic plasticity and other network-forming processes could organize a set of neurons. I am particularly interested in the effects of neuromodulation on these neurons.  If they form a certain kind of network under normal conditions, how would that change if they were bathed in dopamine during the 'experience' or serotonin, or whatever. (Pick your favorite neurotransmitter).

It is easy to think that this robot has a 'brain' but really the cultured neurons are not organized like the brain at all.  Watching a network form in a dish is fascinating and can yield information about how neural networks form in general, but don't assume that this will tell us how networks actually form in an actual brain. 

Robots sure are cute (source)

These methods can be used to discover really interesting things about neurons and networks, but other kinds of study (such as ones using real, intact brains) are need to find out what actually happens.

© TheCellularScale


Demarse TB, Wagenaar DA, Blau AW, & Potter SM (2001). The Neurally Controlled Animat: Biological Brains Acting with Simulated Bodies. Autonomous robots, 11 (3), 305-310 PMID: 18584059

A Tale of Two Huxleys

Andrew Huxley is one of the founders of both modern electrophysiology and  computational neuroscience, and is consequently a personal hero of mine. His recent (May 30, 2012) death inspired me to learn more about his life.

Andrew Huxley (1917-2012)

Andrew Huxley along with Alan Hodgkin discovered the mechanisms which governed the action potential in nerve cells. They inserted micro-electrodes into the squid giant axon and recorded the sodium and potassium currents which generated and propagated the action potential. They shared the Nobel prize for physiology and medicine (with John Eccles) in 1963.

(squid giant axon)

Andrew Huxley is a hero of neuroscience because he (along with Alan Hodgkin) was not only able to develop the equipment and techniques necessary for the complex electrophysiological recordings of the squid axon, but he was also able to understand and mathematically interpret the results of their experiments. Hodgkin and Huxley's mathematical interpretation of their experimental results is basically the beginning of modern computational neuroscience. Their equations describing the flow of ions based on voltage and on concentration are still used in computational models of neurons today. Their famous series of papers (1952) in the Journal of Physiology culminates in their mathematical model of the action potential.

Time constants and steady state curves for activation and inactivation of sodium (Na) and potassium (K) channels (source)

This paper is fascinating to read because of the meticulous thought process that can be traced through it, and because of how much was not known about neurons at the time. The simple composition of the cell membrane was not clear and the fact that sodium and potassium ions actually flow in and out of channels formed by proteins was unknown.

"The next question to consider is how changes in the distribution of a charged particle might affect the ease with which sodium ions cross the membrane. Here we can do little more than reject a suggestion which formed the original basis of our experiments (Hodgkin, Huxley & Katz, 1949). According to this view, sodium ions do not cross the membrane in ionic form, but in combination with a lipoid soluble carrier which bears a large negative charge and which can combine with one sodium ion but no more. Since both combined and uncombined carrier molecules bear a negative charge they are attracted to the outside of the membrane in the resting state. Depolarization allows the carrier molecules to move, so that the sodium current increases and membrane potential is reduced. The steady state relation between sodium current and voltage could be calculated for this system and was found to agree reasonable with the observed curve at 0.2msec after the onset of a sudden depolarization. This was encouraging, but the analogy breaks down if it is pursued further. In the model the first effect of depolarization is a movement of negatively charged molecules from the outside to the inside of the membrane. This gives an initial outward current, and an inward current does not occur until combined carriers lose sodium to the internal solution and return to the outside of the membrane. In our original treatment the initial outward current was reduced to vanishingly small proportions by assuming a low density of carriers and a high rate of movement and combination. Since we now know that sodium current takes an appreciable time to reach its maximum, it is necessary to suppose that there are more carriers and that they react or move more slowly. This means that any inward current should be preceded by a large outward current. Our experiments show no sign of a component large enough to be consistent with the model. This invalidates the detailed mechanism assumed for the permeability change but it does not exclude the more general possibility that sodium ions cross the membrane in combination with the lipoid soluble carrier. " (Hodgkin &Huxley 1952) (emphasis mine)

They describe the ions being bound on one side of the membrane, carried through and released on the other side. If you did not have any idea about membrane channels, this would make sense. What is so beautiful about this is that their experiments and model constrain the vague theory.  However the ions get across the membrane, it must be this fast, this strong, and this dependent on temperature.
They continue:

                "A different form of hypothesis is to suppose that sodium movement depends on the distribution of charged particles which do not act as carriers in the usual sense, but which allow sodium to pass through the membrane when they occupy particular sites on the membrane. On this view the rate of movement of the activating particles determines the rate at which the sodium conductance approaches its maximum but has little effect on the magnitude of conductance. It is therefore reasonable to find that temperature has a large effect on the rate of rise of sodium conductance but a relatively small effect on its maximum value. In terms of this hypothesis one might explain the transient nature of the rise in sodium conductance by supposing that the activating particles undergo a chemical change after moving from the position which they occupy when the membrane potential is high. An alternative is to attribute the decline of sodium conductance to the relatively slow movement of another particle which blocks the flow of sodium ions when it reaches a certain position in the membrane." (Hodgkin &Huxley 1952) (emphasis mine)

Without any structural or molecular analysis of the membrane, Hodgkin and Huxley speculate that there might be sodium channels. They also discuss whether potassium has an entirely separate mechanism of membrane-transport, or whether it is the same one as sodium, but switched in affinity and timecourse in response to membrane depolarization.  Rather than quoting the entire paper here, I urge you to read it as an example of a truly beautiful train of scientific thought.

Aldous Huxley (1894-1963)

Speaking of truly beautiful trains of thought, a different Huxley, half brother to Andrew and 23 years his senior, was a world famous novelist. Aldous Huxley is known best for writing Brave New World, a dystopian novel about a 'perfect' future in which everyone has a place and likes it.

"Till at last the child's mind is these suggestions, and the sum of the suggestions is the child's mind. And not the child's mind only. The adult's mind too-all his life long. The mind that judges and desire and decides-made up of these suggestions. But all these suggestions are our suggestions... Suggestions from the State."
- Aldous Huxley, Brave New World, Ch. 2

Aldous Huxley was on track to become a scientist or doctor, but was struck by an illness which rendered him functionally blind for 3 years, preventing him from maintaining this course of study.

I am not sure which delights me more, that Aldous Huxley is a novelist with a scientist brother, or that Andrew Huxley is a scientist with a novelist brother.

© TheCellularScale

HODGKIN AL, & HUXLEY AF (1952). A quantitative description of membrane current and its application to conduction and excitation in nerve. The Journal of physiology, 117 (4), 500-44 PMID: 12991237