Did you know that neurons have resonant properties? I didn't know that until I read this paper from the Johnston lab down in Austin, TX. Usually I think of synaptic transmission in terms of a single action potential or other event that releases neurotransmitter, so I don't end up in the frequency domain. But, of course, the pattern of release is just as important as the magnitude. Neurons can fire in rhythmic patterns that have additive effects downstream because of the principle of temporal summation. Neurons are continuously deciding whether to fire or not. In a simple view, they do this by summing the 'aye' and 'nay' votes across all inputs. Aye votes move the voltage potential across the neuronal membrane in a positive direction by allowing positive ions to flow into the cell. There is a time window in which an input can cast its vote and still be counted based on how long it takes the receiver neuron to respond to inputs and how long it takes to return to a clean slate after responding. This allows the same input to vote several times if it does so at just the right frequency. If it tries to go too fast it will run up against the membrane capacitance. You see, the membrane potential isn't exactly a count of the number of positive and negative charges inside and outside the cell. Rather, the ions have to be lined up right next to the membrane producing a capacitive current for the period of time it takes to push positive charges off the outside of the cell and line other ions up on the inside. People familiar with basic principles of electrical circuits know that charging a capacitor adds a time dimension. So if an input votes too fast, say, at a frequency of more that 10 to 20 Hz (I'm not sure of the exact number) the effect will be attenuated as the receiver simply can't add that fast.

On the other hand, firing too slow simply won't do either. Neurons have components that react to any change in membrane potential and push back toward baseline. Of particular interest is the H-conductance. The H-conductance is an ion channel in the membrane that allows positive ions to flow into the neuron (as long as the membrane potential is more negative than -30ish mV). Curses. I didn't want to, but I think I have to explain reversal vs. activation potential. Reversal: Ions want to flow to places where there is a lower concentration. They want room to spread out and have a nice big yard with a swingset for the kids and all that, but they also want to get away from other ions with the same type of charge. So if there were a bunch of K+ ions sitting inside a cell and only a few outside and a channel opened up they would want to go out, but if there are already a bunch of Na+ ions outside then they might think twice because there is too much positivity out there already. In other words, potential and concentration gradients are taken into consideration. SInce there is a lot of Na+ outside the cell, it wants to flow into the cell and drive the potential up. There is a lot of K+ inside a neuron, so it wants to flow out of the cell and drive the potential down. If you open a channel that allows both ions through they will arrive at a consensus (equilibrium, reversal) potential that accommodates everyone's needs. The H-conductance does just that and arrives at around -30 mV. If the membrane was at -30 mV, no net ion flow would occur. If it strays, the driving forces will push back towards -30. Normally, due to other considerations, a neuron sits near -55 to -60 mV, so activating the H-conductance will push the neuron toward a more positive potential. Now Activation: Ion channels respond to changes in membrane potential by opening and closing. The H-conductance is closed at positive potentials and becomes activated when the potential moves more negative than about -60 mV. This has the effect of opposing hyperpolarization, the increased difference in voltage across the membrane, because when the membrane tries to move more negative the channel opens and pushes positive. Since the activation potential for H-conductance is near the resting membrane potential, it can also oppose depolarization by shutting portion of channels and reducing the push towards positive. Finally, channel activation and inactivation takes time. It takes more time than charging the membrane capacitance. If an input wants to make a difference, it has to get its votes in before the H-conductance comes into play and brings everybody back to baseline. In this way, the H-conductance acts as a high-pass filter, only allowing speedy inputs to have a say.

Now we have a window of input frequencies that can really strongly affect the cell. If they are too fast, they are filtered out by the membrane capacitance. If they are too slow, they are filtered out by the H-conductance. Really the H-conductance is just one type of conductance that might do the job. Any conductance that is activated near resting membrane potential and opposes change would work fine. You can measure how this plays out in a real cell using something called an impedance amplitude profile (ZAP). You measure the voltage change in a neuron as you inject current at different frequencies but constant amplitude. In practice, this is done really quickly as a sweep across the frequencies. The result is a peak voltage change that corresponds to the resonant frequency of the cell. Like so:



Narayanan and Johnston already knew that you could measure resonant properties of neurons. What we didn't know was that these properties varied in space and time. They measured resonance in CA1 pyramidal neurons. These are the major excitatory cells in an important region of the hippocampus, a brain structure responsible for memory encoding and spatial navigation. CA1 neurons are some of the best characterized neurons available because the CA1 region is highly accesible for in vivo recording and easily delineated for slice electrophysiology, and much is known about its specific inputs and outputs. Imagine an Egyptian pyramid. Now imagine a giant tree growing up through the center of it to about 10 times its height. That is what a pyramidal neuron looks like. The roots of the tree are basilar dendrites and the branches are apical dendrites. Dendrites are specialized structures for receiving input. One giant root will run out of the bottom of the pyramid and send output to some downstream cell. This is the axon. One of the first things that Narayanan and Johnston showed was that the resonant frequency of a CA1 neuron varies along the apical extent of the dendritic tree. The frequency increases as you get further toward the top of the tree, from 3 Hz to 8 Hz at the top. This correlates with the quantity of the channels responsible for H-conductance which also increases toward the tippy-tops of the tree. Input into CA1 neurons is spatially organized such that the entorhinal cortex inputs at the very tip of the apical extent while the CA3 region of the hippocampus inputs at sites more proximal to the cell body. One hypothesis that the authors put forward is that the resonant properties may be tuned to the specific types of inputs. Unfortunately, I can't tell you whether entorhinal neurons fire at 8 Hz vs 3 Hz. I think this is an interesting avenue, but I wonder why you would need to filter the inputs by frequency if you already have them filtered by space. I suppose if the entorhinal cortex naturally fires around 8 Hz and you want the maximal downstream effect then its not so much a matter of filtering out bad frequencies as enhancing the good ones.

The most interesting thing to me though was that certain excitation patterns could alter the resonant frequency. If, by direct stimulation, they caused the neuron to fire in bursts separated by about 100 ms, they could later observe a upward shift in the resonant frequency. They used several stimulation protocols. Of highest interest was the effect of inducing LTP. LTP (Long-Term Potentiation) is a cellular model for learning in memory. It involves the seleective strengthening of synapses between two coincidentally active neurons. There are various LTP inducing stimulation protocols. The one these folks used requires stimulating axons headed for the apical dendrites of the CA1 neuron while depolarizing the CA1 neuron's cell body to cause it to fire. Thus input activity is paired with downstream firing and that particular input is strengthened. Coincident firing is detected by a special receptor (the NMDA receptor) that is activated only when post-synaptic (dendritic, receiving end) membrane depolarization is paired with neurotransmitter release (from an axon of another neuron). After LTP induction, that input now has a bigger say in the overall activation election of the downstream neuron. The analogy between LTP and learning has been argued for decades now and some good evidence exists that this is a legitimate model. Early attempts to show this involved blocking the NMDA receptor in LTP and in learning and showing that both were impaired. Here is the key interesting thing for me about Narayanan and Johnston's paper. Blocking the NMDA receptor not only blocked LTP, but also blocked a global, non-input specific, upward shift in the resonant frequency of CA1 neurons. Now we have two physiological phenomena that you are manipulating when you inject an animal with an NMDA receptor antagonist. Is learning disrupted because of failed LTP or failed resonance shifts?

Why would a resonance shift matter? Here's one reason. Neurons don't work alone. Thousands of neurons have to send input to one downstream neuron to get any reaction. They need to do this in a temporally coordinated fashion. One of the best methods for temporal coordination is oscillatory firing. Watch a tug of war sometime and note the effectiveness of "1,2,3 PULL!" compared to everybody struggling on their own time. This coordinated group of neurons is referred to as an ensemble. If members of the ensemble are connected to each other, they can settle on a frequency of oscillation that best excites everyone at once. If anyone gets out of line and starts going faster or slower, the big PULL will drag them back in, unless they are so far out that they simply can't get down with a certain tempo. If a CA1 neurons is part of some larger ensemble that really loves to fire at 3 Hz and then its resonant frequency jumps up to 5 Hz, it will be that much less responsive to its former buddies. Instead, some other more enticing fast-paced ensemble might recruit that dude into their little 5 Hz cult. The implication is that NMDA receptor-dependent learning could be caused by changes in ensemble size and strength rather than or in addition to strengthening or weakening of specific synapses. A good place to begin on testing this implication would be to record from CA1 neurons in live, behaving animals (this is routine) during a learning task and note whether their preferential firing pattern shifts in frequency and whether this is coordinated across multiple neurons.

The H-conductance has more effects than just determining the resonant frequency. The Narayanan and Johnston paper was published alongside an article describing their effect on active properties of dendrites (dendritic calcium spikes) and third paper featuring the H-conductance prominently.

Labels: CA1, LTP, Memory, Resonance

The first genome-wide association study on human episodic memory back in 2006 showed an association between the T allele of a gene called KIBRA and better performance on certain list-learning tasks. That study contained two replications in different populations, and the outcome was independently replicated in healthy, elderly folks. Next, another group showed an association between the T allele and very late-onset Alzheimer's. There are some issues with interpreting this study that I certainly didn't think of the first time I read it. Almeida et al. point out that this could be due to 'survivorship bias' wherein the C allele carriers that were gonna get AD got it a lot earlier and left the T allele folks to provide the 'very late-onset' crowd (or at least that's how I interpret survivorship bias).

Two studies have come out in the past few months. One replicates the effect of the T allele on memory with a little smaller effect size than before. The second fails to find any effect at all. One experiment in this latter report was an exact replication of the 2006 memory study with a population of European origin (German vs. Swiss. That shouldn't matter should it?).I don't know how to explain the failure to replicate, but it is duly noted. Perhaps it really really matters how well you vet your cohort. For instance (from Almeida and co again):

We did not find evidence in support of our original hypothesis that CC carriers would be at greater risk of MCI (ed: Mild Cognitive Impairment) (although we did observe a trend in that direction), nor were we able to show any evidence of an effect of the gene on the memory scores of older people with MCI. These results suggest that the effect size of the T→C polymorphism decreases with increasing impairment of episodic memory, and that the KIBRA gene plays all but a limited role after scoresfall below a certain threshold, as is the case in MCI.

I don't think there is any evidence that the cohort that failed to replicate had especially bad memory, but I'm not an expert in human memory assessment. A few more molecular details below:

Kibra had an especially good tie-in to memory because in yeast two-hybrid studies it binds to PKM zeta which is established as a key player in maintenance of several types of memory and synaptic plasticity in a completely separate literature. The molecular situation is foggy as well though. We don't have any published assessment of the function or localization of endogenous Kibra protein in neurons. In fact, most of the molecular work has been done with an overexpressed GFP fusion protein. The group that discovered Kibra reports that it is a 125-kDa protein with specialized "WW" protein interaction domains at one end, while the group that reported the Kibra-memory association used a custom antibody to detect human Kibra protein and identified a 100-kDa truncated protein. One final issue is that the memory-SNP (and all SNPs in linkage disequilibrium) in human Kibra is intronic, which means we have no straightforward prediction as to how it might alter protein function. Papassotiropoulos et al.(2006) could not find a difference in the total amount of Kibra protein in human brain tissue with different alleles. Either we have to predict that the SNP produces an expression change that they couldn't detect or that the SNP alters splicing such that the protein sequence changes but the size doesn't.

Labels: Kibra, Memory

Do two points make a line?

Better Memory and Neural Efficiency in Young Apolipoprotein E epsilon4 Carriers

The apolipoprotein E (APOE) epsilon4 allele is the major genetic risk factor for Alzheimer's disease, but an APOE effect on memory performance and memory-related neurophysiology in young, healthy subjects is unknown. We found an association of APOE epsilon4 with better episodic memory compared with APOE epsilon2 and epsilon3 in 340 young, healthy persons.

The T allele of KIBRA was found associated with better memory in a genome-wide association study last year. And now this:

Age-dependent association of KIBRA genetic variation and Alzheimer's disease risk.

An association between memory performance in healthy young, middle aged an elderly subjects and variability in the KIBRA gene (rs17070145) has been recently described. We analyzed this polymorphism in 391 sporadic Alzheimer's disease (AD) patients and 428 cognitively normal control subjects. The current study reveals that KIBRA (rs17070145) T allele (CT and TT genotypes) is associated with an increased risk (OR 2.89; p=0.03) for very-late-onset (after the age of 86 years) AD.

I haven't done an exhaustive search, but there appears to be about four more human memory-associated genes. I wonder if they'll form a pattern. By the way, it looks like these cats Papassotiropoulos and de Quervain have the memory genome-wide association game on lock. Can someone with the right kind of knowledge take a look and see if they are doing things in a robust way? I don't really know what alternative hypotheses you have to get rid of before you can claim an association.

Labels: ApoE4, Kibra, Memory

Rapid Erasure of Long-Term Memory Associations in the Cortex by an Inhibitor of PKM{zeta}
Reut Shema, Todd Charlton Sacktor, Yadin Dudai

Little is known about the neuronal mechanisms that subserve long-term memory persistence in the brain. The components of the remodeled synaptic machinery, and how they sustain the new synaptic or cellwide configuration over time, are yet to be elucidated. In the rat cortex, long-term associative memories vanished rapidly after local application of an inhibitor of the protein kinase C isoform, protein kinase M zeta (PKM{zeta}). The effect was observed for at least several weeks after encoding and may be irreversible. In the neocortex, which is assumed to be the repository of multiple types of long-term memory, persistence of memory is thus dependent on ongoing activity of a protein kinase long after that memory is considered to have consolidated into a long-term stable form.

The authors used conditioned taste aversion (you may be familiar with this learning paradigm if you've ever made yourself sick off tequila). Injection of a peptide inhibitor of this enzyme (PKM zeta) completely removed the aversive association. This isn't a paper about how to erase memories for any clinical application because, for instance, injecting the drug erases multiple olfactory associations (i.e. we don't have a clue how to achieve specificity). It is a paper about how memory works, and it is pretty remarkable that a simple mechanism like persistent kinase activation may be central to this neural function.

Labels: Memory, PKM Zeta

Epigenetics. What is it? Let's ask some reputable sources.

From the glossary of a review in the current Nature Reviews Genetics:

Epigenetic - Refers to mitotically or meiotically heritable changes in gene expression that do not involve a change in DNA sequence.

From the Cell epigenetics review issue a month and a half ago:

Epigenetics, in a broad sense, is a bridge between genotype and phenotype-a phenomenon that changes the final outcome of a locus or chromosome without changing the underlying DNA sequence. ... More specifically, epigenetics may be defined as the study of any potentially stable and, ideally, heritable change in gene expression or cellular phenotype that occurs without changes in Watson-Crick base-pairing of DNA.

Both definitions suggest that, to be epigenetic, something must 1) be heritable and 2) not be attributable to DNA sequence. Imagine if you took that "ideally" very liberally and dropped the heritability requirement. Now a change in gene function must be attributable to something besides DNA sequence. At that point epigenetics becomes gene regulation. Living cells are constantly performing gene regulation in myriad ways. If we loosen the definition enough, epigenetic could mean signal transduction, regulation of protein synthesis and degradation, and any and all transcriptional modulation. Definitions can and do change. One thing I'd like to do in this post is generate discussion as to the most useful way to use the term epigenetics, so that we can communicate most efficiently about the various biological topics now gathering under its umbrella.

Here is a third definition from David Sweatt and Jonathan Levenson:

A third definition posits that epigenetics is the mechanism for the stable maintenance of gene expression that involves physically 'marking' DNA or its associated proteins. This allows genotypically identical cells (such as all cells in an individual human) to be phenotypically distinct (for example, a neuron is phenotypically distinct from a liver cell). The molecular and physical basis for this type of change in DNA or chromatin structure is the focus of this review. By this definition, the regulation of chromatin structure is equivalent to epigenetics.

In this definition, they are equating a mechanism for epigenetics with epigenetics. Now we have the same term at two levels of analysis. If you're a Hofstadter fan you can probably think of a lot of mischief to stir up when you run into this situation, but that's not the real problem. The problem stems from the fact that we already know that epigenetics isn't restricted to the regulation of chromatin structure, so in simplifying the term to one set of mechanisms we are discarding information and excluding mechanisms that we may not have even discovered yet. A lot of the time, Sweatt and Levenson use the term 'epigenetic mechanisms' to mean regulation of histone modification and DNA methylation. If we mean chromatin structure, why not say that? Why do Sweatt and company trade in 'chromatin modification' for 'epigenetics'? It's not an entirely arbitrary thing to do. The best understood mechanisms for epigenetic phenomena are DNA methylation and histone modification, hands down, and if you search Pubmed for epigenetics you will have a very difficult time finding an article that doesn't involve some combination of these chromatin 'marks'.

The most-studied instance of epigenetics is probably imprinting. It has something for everyone, from an exciting battle-of-the-sexes evolutionary explanation to a cellular mechanism veiled in mystery. Very briefly, imprinting occurs when, for instance, you get the exact same DNA sequence for Gene A from mom and dad (you get two of each gene in case you forgot), but only the Dad copy of Gene A is active. The Mom copy is imprinted and thus silenced. Some percentage of human genes are imprinted. I think it's less than ten percent, can't remember. How is the gene expressed differently if the DNA sequence is the same? Well, during egg-ogenesis Mom added a methyl group (CH3, not a very reactive group, just makes the molecule bulkier) to a bunch of the cytosines in and around Gene A, thus changing them to 5-methyl-cytosines. This is called DNA methylation. In this example the gene is maternally imprinted, but don't get it twisted. Dads can imprint too. DNA methylation can affect gene expression directly by hiding sequences that would normally be recognized by regulatory proteins or indirectly as there are proteins that recognize methylated DNA and can recruit chromatin modification complexes to that locus. By the way, if any of this is terribly interesting to you please go check out one or two of the billion reviews on this stuff as I am not an expert in this area and I'd hate to lead you astray.

DNA is often packed up by wrapping it around proteins called histones like you wrap the string around your yo-yo before you put it in the drawer. Unwrapping DNA from the histones allows other proteins (like RNA polymerase) access to the DNA and generally means an upregulation of that gene's activity. You can attach or detach acetyl groups or methyl groups (or phosphate groups or ubiquitin groups or SUMO groups, histones must look like Katamari Damacy) to the histones to make them more or less attracted to the DNA. Histone acetylation (by HATs) and deacetylation (HDACs, histone deacetylases) are core mechanisms for switching a section of chromosome between transcriptionally active and inactive states. There is a complex relationship between DNA methylation and histone modifications. As a basic scheme, DNA methylation is more often associated with silent genes wrapped tightly around deacetylated histones. The two occur independently though. Many transcriptional activators play their part by dragging some histone modification machinery to the gene they are interested in, so histones are dynamically modified in many cases of transcriptional regulation, whereas DNA methylation has been thought to be a more stable, long-lasting modification.

So, yes. Very often 'epigenetic mechanism' refers to DNA or histone modification, but there are other ways of passing information between generations of organisms or cells without touching the DNA sequence. Yeast can transmit information across generations in the form of prions, infectious protein conformations. Think of Ice-9 in Cat's Cradle. When prions touch other normally folded proteins of the same type, they can transfom the normal protein. So a protein with the exact same amino acid sequence can fold up two different ways, one of which (the prion form) can be passed on to the offspring of the prion-infected yeast. Thus, a phenotypic determinant is passed across generations with no help from the DNA sequence. Also, last year a French group reported an instance of epigenetics in which accumulation of some abnormal RNA molecules in mouse sperm is responsible for inheritance of a specific coat pattern. All that is to say is that there are lots of ways to transmit information across generations and cell divisions.

I think the thing about chromatin modification that was initially appealing to Sweatt and colleagues in terms of memory was its apparent stability. Research into cell differentiation, the process by which your basic stem cell produces the diverse array of cell-types required, has implicated chromatin modification as a strong influence. The idea is that to make, for instance, a neuron, you have to permanently set up a certain transcriptional program to make things like neurotransmitter receptors, so you go ahead and permanently turn on those genes. In some sense you could extend cell differentiation to a highly refined level so that the cell types aren't broad classes like neuron and skin cell, but rather 'neurons storing memory A' and 'neurons storing memory B'. Following this train of thought or something like it, they began to look for signs of chromatin modification in response to learning. They discovered that after a hippocampus-dependent form of fear conditioning, histones in the hippocampus are modified. Acetylation and phosphorylation are high one hour after training, but return to baseline by 24 hours. Given that we know that many genes are upregulated in response to fear conditioning and that transcription factors that activate genes have a history of modifying histones it would have been quite a surprise if we hadn't seen any changes. At this point, I began to get a little unhappy with the papers coming out of the Sweatt lab because I felt like the histone-modification story wasn't telling us much that we didn't know.

So far, the epigenetics in memory story has been presented in a theoretical framework emphasizing the stability of chromatin modifications and the persistence of long-term memories. Just so I don't mischaracterize, here is a quote from 2006:

One question that has eluded neuroscientists is, How can the brain store information over the lifetime of an organism in the face of molecular turnover? This question is especially relevant for understanding a complex process such as cognition, which relies heavily on the ability to store and recall information for periods longer than the half-lives of most of the molecules utilized in these processes. Chromatin is the one structure that remains relatively constant in almost every cell of a metazoan. It is not surprising that many recent studies in the nervous system indicate that from invertebrates to mammals, chromatin is a dynamic structure that integrates potentially hundreds of signals from the cell surface and effects a coordinated and appropriate transcriptional response. More important, chromatin is perhaps the only structure in a neuron capable of such higher-level signal integration and information storage that is not continually turned over.

To rephrase: Memories are persistent. Therefore the molecular mechanisms for memory are persistent. Most molecules don't last, so they can't do the job. Chromatin structure lasts, so it is a good candidate for long-term memory storage.

Here are my various criticisms of this line of thinking:

1) You can't have your cake and eat it too. Is chromatin a dynamic structure changing in the face of new stimuli or is it a solid, stable basis for long-term information storage? Once a cell's epigenome has been modified to store a memory does it somehow get write-protected? It seems to me that there are some chromatin modifications that are highly stable (i.e. the ones that control cell differentiation) that should be very unresponsive to the day-to-day vagaries of cell signaling, otherwise you would have neurons de-differentiating all willy-nilly. Another separate set of chromatin modifications must respond to cellular stimuli and allow dynamic gene expression to fit the needs of the cell in the current situation.

2) A cell only has one epigenome, but encodes more than one memory. That's the wonderful thing about synapses; they allow multiple excitability configurations. An average neuron has thousands of synapses. Memory 1 can light up synapses X, Y, and Z and cause the cell to fire a burst of action potentials whereas Memory 2 lights up a different set of synapses causing the cell to fire at a slower, more consistent rate. Maybe when synapse X loses its ability to excite the cell you forget a little bit and the burst of action potentials is weaker. That's one scenario. The point is illustrated empirically in studies like Vazdarjanova and Guzowski, (2004), which shows that 35% of cells in the CA1 region of the hippocampus light up in response to a single experience, and there is some 15% overlap between cell populations responding to two different contexts. The point of that report is actually to show that representations are relatively non-overlapping, but they aren't so sparse that you can afford a cell per experience. Is there some way to propose chromatin modification as a long-term memory mechanism that doesn't make this sparsity requirement?

3) The molecular turnover problem was never a problem. Sure molecules turnover, so the memory can't be housed in a single individual molecule. Dendritic spines (structures forming excitatory synapses), on the other hand, can remain stable for as long as we can record them. All of the molecular constituents of the spine turn over, but the structure remains. All of your skin cells die and are replaced, but your body remains intact. It's not an issue. The only way that molecular turnover would be an issue is some critical mass of the molecules in a structure turned over at once. I can imagine some mechanisms involving the post-synaptic scaffolding and structurally induced biochemical states, but the mechanism doesn't matter for putting the molecular turnover issue to bed. We have observed neuronal structures remaining intact for long periods of time. Apparently there are parts of the cell that are stable besides the chromatin.

I can maybe think of a way out of #2, but #3 seems very clear and very much in conflict with the theoretical framework for this work. So when the new Miller and Sweatt came out I was thinking, "This is just the histone papers again with DNA methylation." I skimmed it, saw some of the experiments I expected, and gave it a general thumbs down. Now that I've given it a few reads I think the paper has some relevant and really novel findings, a major one being that DNA methylation is regulatable in nondividing cells.

The enzymes that methylate DNA are called DNA methyltransferases (DNMTs). They most commonly add the methyl group to the C in the sequence, CG. Remember that DNA is double-stranded so you get a CG - GC quartet of nucleotides at these preferential methylation sites. There are two major classes of DNMT in mammals: de novo and maintenance. Maintenance DNMTs make sure that if a portion of the genome was methylated in a dividing cell, say a neuronal precursor cell, then that portion will be methylated in the daughter cells produced by the division. During DNA replication just prior to cell division you have double stranded DNA containing one of the original strands (with methylated CGs) and one newly synthesized strand. A maintenance DNMT recognizes the dsDNA with only a methylated CG facing an unmethylated GC and makes them match. De novo DNMTs can make new methylation sites where there was no indication of methylation to begin with. Their major role is during the 'reset' of methylation states that occurs during gametogenesis and in the preimplantation embryo, and they have been considered relatively quiet in the fully developed adult.

The first thing Miller and Sweatt did was to show that the level of de novo DNMT mRNA is increased after fear conditioning. They then injected DNMT inhibitors into the hippocampus after fear conditioning and blocked long-term memory consolidation. I initially thought this experiment was suspect because I didn't understand that DNA methylation is not expected to be part of a cell's normal function. I associated DNA methylation with histone deacetylation and my understanding is that histone acetylation states are very commonly altered to tweak this or that gene's level up or down. In reality, people have hardly tested the effects of DNMT inhibitors in non-dividing cells. The literature is full of attempts to selectively target cancer cells with DNMT inhibitors because they have a habit of dividing too much. You would still like a more specific manipulation, a way of inhibiting DNMTs at specific genomic loci would be nice, but the behavioral control experiments in this paper go a long way toward showing that the drugs are not simply killing off neurons. The inhibitors are only effective at a time point shortly after training, defining a maximum window of some 6 hours in which relevant genes are demethylated and remethylated or vice versa in response to behavioral training.

They next showed regulation of methylation at genes that are positively and negatively associated with memory formation. Protein phosphatase 1, a 'molecular constraint on learning and memory', is transiently methylated leading to reduced levels, whereas reelin, which has been associated with synaptic plasticity shows reduced methylation levels and increased mRNA expression. This second finding is pretty important. The identity of the mammalian demethylating enzyme is unclear leading some to question if an active demethylase exists at all. Demethylation in dividing cells can be achieved by simply being lazy about methylating new DNA. According to Hermann et al. in 2004, the 'best evidence so far for active demethylation in vivo" was the rapidity of sperm chromatin demethylation after fertilization. I'd say reduced methylation at the reelin locus in non-dividing cells within an hour after behavioral training is quite strong evidence. In fact, this window of very rapid demethylation at a specific locus suggests a strategy for investigating the biochemistry of mammalian demethylation. It should be possible to trap any proteins that are associated with the reelin promoter in the hour or so after training using techniques like chromatin immunoprecipitation.

Finally, they emphasize that the methylation changes are transient. Everything returns to baseline by 24 hours after conditioning. This very obviously contradicts the idea of DNA methylation as a memory storage tool. For some reason, the discussion section still won't let it go though. Look at this:

This attribute is, however, consistent with the role of the hippocampus as a structure contributing to memory consolidation but not memory storage. It will be interesting to determine if the Crick/Holliday mechanism plays a role in perpetuating long-term changes in adult neurons in the cortex, at known sites of long-term memory storage.

The hippocampus' role in various types of memory may or may not be temporary. This is a broadly debated issue and in my humble opinion, the folks making the case for a long-term role for the hippocampus are doing a fine job. Regardless, no one suggests that the hippocampus is only involved for less than a day. So there must at the very least be a mechanism besides DNA methylation that helps maintain the memory for a week or so.

I spent a lot of time looking into ways to specifically affect DNA methylation at certain loci. I found a paper where they made chimeric DNMTs with a custom DNA-binding domain and inserted the recognition sequence for that domain into the area to be methylated. To achieve this in vivo with any sort of temporal specificity would require the combination of at least three transgenes either through traditional methods or injection of viral vectors. The temporal specificity (within 6 hours) would still not really be on point, so I guess this isn't the solution.

So now we know that some genes are regulated by methylation state in response to learning. The obvious next question is which genes. How does the cell decide which genes to turn up and which ones to turn down? Is there a signature pattern of CGs that indicates memory repressor vs memory enhancer genes? Do transcription factors that are already implicated in memory-related gene expression recruit DNA demethylase activity?

A broader question remains. Why should mRNAs get expressed in response to learning anyway? Why involve the nucleus when the action is out at the synapses? I'm convinced that its important, but for my money, local protein synthesis from pre-existing RNAs seems a more specific response to changes in input patterns. Does anyone reading have an explanation for the need to produce new RNAs so far from the locus of plasticity?

Labels: Epigenetics, Memory

There's a new article by Miller and Sweatt in Neuron claiming that DNA methylation is a step in memory formation. They show methylation and demethylation of particular genes (reelin and PP1) following fear conditioning, and they show that inhibition of DNA methyl transferases (the name says it all) during the memory consolidation can disrupt memory. I hope now that they've started naming genes at which the methylation matters they'll study something more specific. The idea of inhibiting all DNA methylation in a cell for any length of time seems too blunt an instrument. The cells responsible for that memory could've just keeled over or radicaly changed because some cell cycle regulator got turned up too high for instance. They do show that animals can form a new, strong memory a couple days after administration of the drug, but in some cases, animals with a hippocampal lesion can perform these tasks. The nervous system can achieve learning (especially learning as important as conditioned fear) through many means, so seemingly normal behavior after an insult isn't that strong a control.

The scope of the processes disrupted by a DNMT inhibitor is indicated by this sentence from Miller and Sweatt:

Many developmentally important processes utilize this "prima donna" of epigenetics (Scarano et al., 2005 and Santos et al., 2005), including gene imprinting, cell differentiation, X chromosome inactivation, and long-term transcriptional regulation (Bestor et al., 1988 and Okano et al., 1998).

Sorry to be such a naysayer. It is an interesting hypothesis. It is likely the case that DNA methylation is regulated during memory formation. If we just think that memory will require transcriptional regulation then probably some DNA modifications will have to be done and undone. Cis-regulatory sequences that control a gene's level of activation can act through recruitment of histone acetyl transferases and other chromatin modifiers, so a precedent exists. But I don't think there is any special connection between epigenetics and memory. Memory requires cells to be cells and work properly, so it will require transcriptional regulation and DNA/chromatin modification. For any further connection, I think we have to start naming names.

Labels: Epigenetics, Memory