Saturday, June 29, 2019

The Subtle Mechanics of Regulatory Enhancers

Synergy is an overused buzzword elsewhere. But in biology, it has real meaning.

Our genes are mostly off, yet each cell needs some select portion of the genome turned on to do its job. Gene expression is an enormous field of study, encompassing the full life cycles of both mRNA and proteins, and much else. But turning the transcription of a gene on is by far the most powerful and typical way to control its expression. This is done by regulatory proteins that bind the DNA near the gene, either nearby at segments called promoters, (which is where the RNA polymerase complex is assembled), or far away at modular segments called enhancers. Proteins bound at enhancers are thought to loop around by way of the flexibility of DNA to touch the proteins bound at the promoters, forming a somewhat disordered protein mush that is more active (firing off more RNA polymerases) the bigger it is, and the more of its components are activating vs repressing.

A very general flow of early drosophila embryonic development, from egg to gastrula and from rough-out to finer position specification.

The fly model system has been a wonderful place to study enhancers, since it has a lot of them, they are clearly modular for different developmental tasks, and they frequently have critical and very visible influences on the fate of tissues in the adult. A recent paper used one developmental gene, hunchback (hb) to study in detail how its expression pattern is developed from the proteins bound to its enhancers.

Experimentally labeled micrographs of hunchback expression (in mRNA form, bottom) as driven by bicoid protein (middle), expressed from bicoid mRNA messages (top). How does the sharp midline cutoff of hunchback expression develop from the hazy gradient of bicoid protein, its primary transcriptional activator? We can ignore the odd posterior hunchback band, which is driven by some non-bicoid inputs. At this early stage, the drosophila embryo is a bag with cells mostly on the surface, and has not yet begun the body segmentation process, though glimmers are beginning on the molecular level, as shown here.

Hunchback was named for the shape of the larva of the fly when this gene is mutated. They are missing most of their thoracic segments. Hunchback is a very early gene in the lengthy and complex chain of events that specify the developmental pattern of the fly. It encodes a DNA-binding regulator that acts on the next set of developmental genes, the gap genes. It is expressed in a simple pattern of - on in the front/anterior, and off in the posterior, of the early embryo. A good deal of it is supplied by the mother in the egg, so it has been difficult to tease apart the effects of that inherited pool of mRNA, vs those of new gene expression within the embryo. But this paper studies only the embryo-based expression pattern, specifically using broken up parts of the enhancers to figure out how they generate a very sharp half-embryo-on / half-embryo-off pattern.

Some upstream portions of the hunchback gene from drosophila. Modular enhancer cassettes lie upstream, and the one studied in the current paper is the P2 enhancer, right next to the coding area of the gene, which for these experiments has been excised and put upstream of an easily assayed reporter gene, LacZ. 

The researchers study one of the three hunchback enhancers- the one that drives this early anterior-only expression (P2, above). The regulator that binds to this enhancer (at 6 distinct sites) is bicoid, a fully maternal factor supplied in the anterior part of the egg. (Mutants of bicoid have no head- the problems of mutants get more dramatic the earlier you go into the developmental cascade.) Bicoid has a very gradually tapering / diffusing distribution, from high in the anterior to low in the posterior. The six sites that bind bicoid in this main hunchback enhancer are known to have cooperative effects, and thus could account for the (non-linear) sharpness of the hunchback expression pattern- high in the anterior, then dropping off sharply at the midline. In this way a much more gradual gradient of bicoid is recomputed into a finer dividing line between front and back. This is a dynamic that is employed over and over again as finer divisions are made throughout development. Yet the authors maintain that this DNA site binding cooperativity is, on its own, not quantitatively sufficient to account for the pattern, and go in search of other explanations for how the bicoid and other factors drive this high-definition pattern.

As bicoid binding sites are removed from the hunchback P2 enhancer (left), the expression of the test gene becomes shallower in its anterior-posterior gradient and migrates towards the anterior.

Later on they cop to the fact that their system is more complex than they portrayed it at first. The enhancer they are working with (isolated from the rest of the enhancers and driving a fluorescent reporter gene) may have a few binding sites for some other factors, including krüppel, tailless, zelda, and indeed hunchback itself, forming a small positive feedback loop. When they scrubbed out those extra sites, expression was quite a bit farther from the wild-type condition, pushed towards the anterior (below). As extra demonstrations, they individually knock out hunchback protein expression, which clearly accounts for some of this effect, pushing expression of the wild-type enhancer farther anterior. And they do a similar demonstration for zelda, which has a similar, though much smaller, effect.

Some further experiments with a purified enhancer, where all non-bicoid sites have been removed (top, red). It is clear that the other sites (present in the wild-type P2 enhancer, black) have a key role driving expression to a more posterior position, but have little roll in the steepness of the cutoff of hunchback expression. Specific regulators are knocked out of the embryo and assayed on the wild-type P2 enhancer in the lower panels, hunchback and zelda, respectively, to show their individual effects.

Next, in search of the additional cooperativity factors, they make deficiencies in several of the common transcriptional components that occupy, not the enhancer, but the promoter where the RNA polymerase is going to be assembled. This is a bit tricky, since these will have quite general deleterious effects, complicating interpretation. But one example is shown below- CBP (Creb Binding Protein). This is an enzyme that modifies histones and is a common part of the core transcription complex, and knocking down its activity shifts the hunchback expression curve not only anteriorly, but also to a more shallow profile, indicating that this protein helps the cooperative effects of multiple bicoid activator proteins bound at the enhancer to take effect. This leads to a clear model of the system whereby the bicoid proteins are only partially cooperative among themselves as they bind to DNA. But their cooperativity is enhanced by all of them binding in concert to their targets in the core transcription complex, one of which may be this CBP protein, or others attached to it (below).

A sample experiment with the full enhancer, in a fly where the expression of one of the core transcription components, in this case CBP, has been knocked down. Not only is expression reduced in the sense of moving anteriorly, but the slope of the expression vs the gradient of bicoid is also reduced, indicating that this defect reduces the effective cooperativity of the six bicoid proteins bound at this enhancer. 

Surprisingly, this work mostly recapitulates work done twenty years ago in the same system. That earlier work had demonstrated that bicoid binds to its DNA sites in a cooperative fashion, specifically in directionally-specific pairs clearly suggesting a single side-by-side cooperative interface on the protein. It also showed that interactions with the core transcriptional apparatus, in that case TAFii60 and TAFii110 in yeast, could account for remaining amounts of cooperativity among the six or more bicoid binding sites on the hunchback enhancer. Thus all this is hardly news, whatever the new mathematical machinery brought to bear by Park et al. in the current paper (and this from Harvard, no less). By this point, we would be expecting to see a full structural reconsitution and recapitulation of the transcriptional activation system, with titrations to demonstrate its accuracy with respect to the concentrations of bicoid found in vivo.

Whatever the pace of progress, however, it is of small bricks like these that knowledge of biological mechanisms is built. Even this system, which is so well defined and long-studied, has endless complexities that arise when one looks closer than the schematic models that were originally advanced to explain it. For example, a recent paper discussed how bicoid activates some genes in the very posterior of the embryo, despite occurring at vanishing (nano-molar) concentrations. They showed that aggregations arise in the posterior that, despite the low average concentration, provide high local concentrations, and thus, plausible transcription activation activity by bicoid. Then there is the complexity of the core transcriptional apparatus, which has clearly impeded efforts to fully delineate the cooperative structures that form between it and activating / repressing regulators bound at peripheral sites.

  • RNAs enter the mix at enhancers and promoters as well.
  • Loving the Dead.
  • Jared Diamond on the current situation.

Saturday, June 22, 2019

Battle For the Truth

The battle of Midway- success comes from dedicated engagement with reality. Not from fantasy.

Memorial Day brings up the Greatest Generation and the battles it fought to keep the World Wars away from the US, and save other countries from tyranny along the way. One of the greatest of those battles was at Midway, about half a year after Pearl Habor. It carries some object lessons in why this generation was so successful, and what makes America great. I am watching Battlefield 360, a History Channel program that profiles the aircraft carrier Enterprise. The production is absurdly over-the-top and padded with cheap filler, yet also full of compelling history.

The story of this battle, as for all others, is a search for truth, which then leads to success. A mere 33 years after the Wright brother's first flight, the US christened aircraft carriers like the USS Enterprise, which carried 90 airplanes. Mastering flight was the first step to a new form of long-range mobile warfare. The US had broken the Japanese naval code, enabling us to know the truth of what Japan planned for its invasion of Midway. The US ran an experiment to nail down the meaning of one word in the code, which clearly denoted a place, but which place? The truth came out when the Japanese took the bait and relayed our (fake) news that Midway was short of water. The word was their code name for Midway.

The US had learned quite a bit about the dangers of fire aboard aircraft carriers, which are awash with fuel, bombs, and artillery rounds, which led to a variety of novel equipment and training, such as CO2 purging of fuel lines before facing attacks, fire-fighting foams, and dedicated, pre-positioned fire control crews. This led to the Enterprise being able to take three direct bomb hits and not sink, while in the battle of Midway, the US sank four enemy carriers with only a couple of bombs each. The Japanese had not learned the value and truth of protective design and effective fire suppression.


The US had radar, a new way to find out the truth of enemy positions, so critical to both defense and offense. As in the old board game of battleship, naval warfare is a game of cat and mouse. The more you know, and the more you can blind your opponent, the more successful you will be. The search for truth has been so integral as to be almost unconscious in our military (not to mention "intelligence") culture. And in the post-war era, it led to a broader cultural commitment to education and research which has formed US preeminence in physics, chemistry, and biology, among many other fields, including notably climate science.

Which all leads one to wonder why lying is now one of our leading national characteristics. Who does our president regard as the enemy, and who the friend? Why is, for him, truth in journalism so dangerous? What is the morality of a whole party lying habitually about fundamental economics, about public interest regulation, about democratic values, and about the future of the planet? What has happened to us?

Saturday, June 15, 2019

Can Machines Read Yet?

Sort of, and not very well.

Reading- such a pleasure, but never time enough to read all that one would like, especially in technical fields. Scholars, even scientists, still write out their findings in prose- which is the richest form of communication, but only if someone else has the time and interest to read it. The medical literature is, at the flagship NCBI Pubmed resource, at about 30 million articles in abstract and lightly annotated form. Its partner, PMC, has 5.5 million articles in full text. This represents a vast trove of data which no one can read through, yet which tantalizes with its potential to generate novel insights, connections, and comprehensive and useful models, were we only able to harvest it in some computable form.

That is one of the motivations for natural language processing, or NLP, one of many subfields of artificial intelligence. What we learn with minimal effort as young children, machines have so far been unable to truly master, despite decades of effort and vast computational power. Recent advances in "deep learning" have made great progress in pattern parsing, and learning from large sets of known texts, resulting in the ability to translate one language to another. But does Google Translate understand what it is saying? Not at all. Understanding has taken strides in constricted areas, such as phone menu interactions, and Siri-like services. As long as the structure is simple, and has key words that tip off meaning, machines have started to get the hang of verbal communication.

But dealing with extremely complex texts is another matter entirely. NLP projects directed against the medical literature have been going on for decades, with relatively little to show, since the complexity of the corpus far outstrips the heuristics used to analyze it. These papers are, indeed, often very difficult for humans to read. They are frequently written by non-English speakers, or just bad writers. And the ideas being communicated are also complex, not just the language. The machines need to have a conceptual apparatus ready to accommodate, or better yet, learn within such a space. Recall how perception likewise needs an ever-expanding database / model of reality. Language processing is obviously a subfield of such perception. These issues raises a core question of AI- is general intelligence needed to fully achieve NLP?


I think the answer is yes- the ability to read human text with full understanding assumes a knowledge of human metaphors, general world conditions, and specific facts and relations from all areas of life which amounts to general intelligence. The whole point of NLP, as portrayed above, is not to spew audio books from written texts, (which is already accomplished, in a quite advanced way), but to understand what it is reading fully enough to elaborate conceptual models of the meaning of what those texts are about. And to do so in a way that can be communicated back to us humans in some form, perhaps diagrams, maps, and formulas, if not language.

The intensive study of NLP processing over the Pubmed corpus reached a fever pitch in the late 2000's, but has been quiescent for the last few years, generally for this reason. The techniques that were being used- language models, grammar, semantics, stemming, vocabulary databases, etc. had fully exploited the current technology, but still hit a roadblock. Precision could be pushed to ~ %80 levels for specific tasks, like picking out the interactions of known molecules, or linking diseases with genes mentioned in the texts. But general understanding was and remains well out of reach of these rather mechanical techniques. This is not to suggest any kind of vitalism in cognition, but only that we have another technical plateau to reach, characterized by the unification of learning, rich ontologies (world models), and language processing.

The new neural network methods (tensorflow, etc.) promise to provide the latter part of the equation, sensitive language parsing. But from what I can see, the kind of model we have of the world, with infinite learnability, depth, spontaneous classification capability, and related-ness, remains foreign to these methods, despite the several decades of work lavished on databases in all their fascinating iterations. That seems to be where more work is needed, to get to machine-based language understanding.


  • What to do about media pollution?
  • Maybe ideas will matter eventually in this campaign.
  • Treason? Yes.
  • Stalinist confessions weren't the only bad ones.
  • Everything over-the-air ... the future of TV.

Saturday, June 8, 2019

Whom do You Trust?

We have placed our lives in the hands of Silicon Valley companies. Do they earn and keep that trust?

It used to be that banks made a big show of trustworthiness and stability. They would build classical edifices of stone to signify their solidity, and use names like "trust", "fidelity", "savings", "citadel", etc. This dates back to the 1800's, when there was no regulation, and banks could collapse from one day to the next, taking all their depositor's money with them. We got a brief taste of that in the recent banking / credit crises- the savings and loan debacle, and the 2008 subprime catastrophe. But generally, banks these days are rather boring from a depositor perspective, more concerned about appearing friendly and neighborly than awesomely immovable. Deposit insurance and other regulations have removed virtually all the risk of retail banking, and computers have simplified and automated its accounting mechanics. On the other hand, bigger investors (and borrowers) would have been wise to pay more attention to the trustworthiness of such institutions and their products through the subprime, securitization, and housing bubble periods, when so many were sold a bill of goods.

In this new computerized world, our faith turns out to be more at risk elsewhere, among the custodians not so much of our money, but of our selves in all exposed dimensions - messages, emails, pictures, documents, conversations, backups - our data. Our financial data is still of highest concern, but now that communications have migrated to myriad "platforms", we have so much more to worry about. What used to be securely private is now much less so. Electronic communication used to be confined to ATT, which came under significant regulation. Now it is a wild west of whoever can convince us to try a new service sure to enhance our lives or reputations, and all for free. Google led the way with incredible search capabilities, followed by Amazon, Myspace, Facebook, Paypal, Twitter, iTunes, Instagram, Roku, Pinterest, Linkedin, Netflix, Reddit, and countless other purveyors and services. Every one requires an account, with lock and key, every one collects our data, and most monetize it for ads, spam, and who knows what else.


Do they merit our faith? This becomes an increasingly urgent question as more of our lives migrate to digital form, and the companies we deal with gain increasing power by virtue of their custody over those forms. Are they responsible fiduciaries? Facebook and Google offer an instructive contrast. Google lives mostly by search, and while using ads, has carefully kept the search space clean enough to facilitate use. Its YouTube subsidiary is perhaps its most social media-y, pushing suggestions drawn from the user's viewing history. Since stochastically, this will ramify outwards into new areas, it can facilitate those looking for more extreme content to head in that direction. But different companies clearly carry different ecosystems and ideas of where to draw the line.

Facebook has been notorious for its obscure, ever-shifting intefaces, its constant foisting of new content and tools, and its devil-take-the-hindmost attitude to user data and privacy. Everything is open, except its own operations. All data is ripe for pushing to advertisers, and whatever it takes to get more clicks goes. Where Google remains dominant and comfortable in its search sphere and ancillary businesses, Facebook had made a scientific project of developing the most addictive tools to get people uncomfortable in their social networks, forced to like and be liked in an endless and downward hedonic treadmill. As an introvert, I am largely unaffected, but others seem to be hopelessly ensnared in the depressing exercise of social comparison.

Then there is the fake news. The new platforms act as publishers with vast powers of propagation, to viral degress unheard of in past ages of humdrum paper publishing. But at the same time they eschew the responsibility of publishers to vet media they purvey and provide a gatekeeping function that has been a critical, if unacknowledged heart of the rights of the free press. We have yet to get used to this world where power and reach are unconnected with curated cogency and minimal economic marketability.

Reputation is coming to the fore, as it once did for banks. Apple is making a great deal of its security and login operations, that they as a philosophy and business do not sell user data, being in the hardware business instead. Facebook has taken a big reputational hit through its bad behavior, particularly its release of data to the Republican-affiliated trolls in England, but also for its many other practices and attitudes. Governance is another issue. Facebook is extremely unusual in its monarchical shareholder model, where the founder has all the voting power, and the public none. How was this allowed as a "public" company?

Regulation is needed, on many levels. That has been the time-tested way to address market failures in the face of new technologies and market practices. Reputation alone is a poor way to police companies that have grown too big for many, if not most people, to do without. Antitrust, corporate governance, user data protection and use restrictions, transparency of data custody, and responsible free speech curation are all areas that need work. We should have a government that is willing to do that work, instead of one that lurches from one tweet to the next.


  • Stiglitz on the next chapter of capitalism.
  • What does socialism mean, today?
  • What on earth are people thinking, supporting Biden?
  • There is more to say.
  • On lies.
  • The cold war is back, and trilateral.
  • Some arguments against a job guarantee, which actually sound more like arguments for it.
  • Winter is coming. (Press "max")

Sunday, June 2, 2019

Backward and Forward... Steps to Perception

Perception takes a database, learning, and attention.

We all know by now that perception is more than simply being a camera, getting visual input from the world. Cameras see everything, but they recognize nothing, conceptualize nothing. Perception implies categorization of that input into an ontology that makes hierarchical sense of the world, full of inter-relationships that establish context and meaning. In short, a database is needed- one that is dynamically adaptive to allow learning to slice its model of reality into ever finer and more accurate categories.

How does the brain do that? The work of Karl Friston has been revolutionary in this field, though probably not well-enough appreciated and admittedly hard for me and others not practiced in mathematical statistics to understand. A landmark paper is "A theory of cortical responses", from 2005. This argues that the method of "Empirical Bayes" is the key to unlock the nature of our mental learning and processing. Bayesian statistics seems like mere common sense. The basic proposition is that the likelihood of some thing is related to our naive model (hypothesis) of its likelihood arrived at prior to any evidence or experience, combined with evidence expressed in a way that can weight or alter that model. Iterate as needed, and the model should improve with time. What makes this a statistical procedure, rather than simple common sense? If one can express the hypothesis mathematically, and the evidence likewise, in a way that relates to the hypothesis, then the evaluation and the updating from evidence can be done in a mechanical way.

Friston postulates that the brain is such a machine, which studiously models the world, engaging in what statisticians call "expectation maximization", which is to say, progressive improvements in the in detail and accuracy of its model, driven by inputs from sensory and other information. An interesting point is that sensory input functions really as feedback to the model, rather than the model functioning as an evaluator of the inputs. We live in the model, not in our senses. The overall mechanism works assiduously to reduce surprise, which is a measure of how inputs differ from the model. Surprise drives both attention and learning.

Another interesting point is the relationship between inference and learning. The model exists to perform inference- that is the bottom-up process of judging the reality and likely causes of some event based on the internal model, activated by the input-drive attention. We see a ball fall down, and are not surprised because our model is richly outfitted with calculations of gravitation, weight, etc. We infer that it has weight, and no learning is required. But suppose it is a balloon that floats up instead of falling- a novel experience? The cognitive dissonance represents surprise, which prompts higher-level processing and downward, top-down alterations to the model to allow for lighter-than-air weights. Our inferences about the causes may be incorrect. We may resort to superstition rather than physics for the higher-level inference or explanation. But in any case, the possibility of rising balls would be added to our model of reality, making us less surprised in the future.
The brain as a surprise-minimizing machine. Heading into old age, we are surprised by nothing, whether by great accumulated experience or by a closure against new experiences, and thus reach a stable / dead state. 

This brings up the physiology of what is going on in the brain, featuring specialization, integration, and recurrent networks with distinct mechanisms of bottom-up and top-down connection. Each sensory mode has its specialized processing system, with sub-modules, etc. But these only work by working together, both in parallel, cross-checking forms of integration, and by feeding into higher levels that integrate their mini-models (say for visual motion, or color assignment) into more general, global models.
"The cortical infrastructure supporting a single function may involve many specialized areas whose union is mediated by functional integration. Functional specialization and integration are not exclusive; they are complementary. Functional specialization is only meaningful in the context of functional integration and vice versa."

But the real magic happens thanks to the backward connections. Friston highlights a variety of distinctions between the forward and backward (recurrent) connections:

Forward connections serve inference, which is the primary job of the brain most of the time. They are regimented, sparsely connected, topographically organized, (like in the regular striations of the visual system). They are naturally fast, since speed counts most in making inferences. On the molecular level, forward connections use fast voltage-gated receptors, AMPA and GABA.

Backward connections, in contrast, serve learning and top-down modulation/attention. They are slow, since learning does not have to obey the rapid processing of forward signals. They tend to occupy and extend to complimentary layers of the cortex vs the forward connecting cells. They use NMDA receptors, which are roughly 1/10 as fast in response as the receptors use in forward synapses. They are diffuse and highly elaborated in their projections. And they extend widely, not as regimented as the forward connections. This allows lots of different later effects (i.e. error detection) to modulate the inference mechanism. And surprisingly, they far outnumber the forward connections:
"Furthermore, backward connections are more abundant. For example, the ratio of forward efferent connections to backward afferents in the lateral geniculate is about 1 : 10. Another distinction is that backward connections traverse a number of hierarchical levels whereas forward connections are more restricted."

Where does the backward signal come from, in principle? In the brain, error = surprise. Surprise expresses a violation of the expectation of the internal model, and is accommodated by a variety of responses. An emotional response may occur, such as motivation to investigate the problem more deeply. More simply, surprise would induce backward correction in the model that predicted wrongly, whether that is a high-level model of our social trust network, or something at a low level like reaching for a knob and missing it. Infants spend a great deal of time reaching, slowly optimizing their models of their own capabilities and the context of the surrounding world.
"Recognition is simply the process of solving an inverse problem by jointly minimizing prediction error at all levels of the cortical hierarchy. The main point of this article is that evoked cortical responses can be understood as transient expressions of prediction error, which index some recognition process. This perspective accommodates many physiological and behavioural phenomena, for example, extra classical RF [receptive field] effects and repetition suppression in unit recordings, the MMN [mismatch negativity] and P300 in ERPs [event-related potentials], priming and global precedence effects in psychophysics. Critically, many of these emerge from the same basic principles governing inference with hierarchical generative models."

This paper came up due to a citation from current work investigating this model specifically with non-invasive EEG methods. It is clear that the model cited and outlined above is very influential, if not the leading model now of general cognition and brain organization. It also has clear applications to AI, as we develop more sophisticated neural network programs that can categorize and learn, or more adventurously, develop neuromorphic chips that model neurons in a physical rather then abstract basis and show impressive computational and efficiency characteristics.