The Ferocious Complexity of the Cell | Hacker News Reader

The Ferocious Complexity of the Cell

rbharath.github.io

One of the major misunderstandings even many biologists have about how a cell works is a complete misconception of how jam-packed and dynamic a living cell actually is. See the following images for a pretty simulation of what a cell actually "looks" like on the inside:

http://mgl.scripps.edu/people/goodsell/books/MoL2figures/Fig...

Webpage with more info:

http://occamstypewriter.org/scurry/2010/08/30/the_crowded_ce...

Physicists/chemists often make fun of biology as "blob-logy" because we tend to draw biological systems as Blob-1 interacting with Blob-2 which leads to a change in Blob-3. It's a very useful abstraction, but it's also limited. In reality all these blobs are jammed together and moving around, creating a lot of random noise in the system. This is because every component in a cell is working at nano-scale (proteins are to a 1st approximation about 5nm in diameter). The rules are different at that scale, we don't experience anything like them in our daily lives.

In other words, the components of the cell aren't doing chemistry at a bench, passing the results of one reaction to the next reaction, it's more like doing chemistry in a mosh pit.

Biology has developed coping mechanisms for this insanity, but analogies drawn with things like computer processors and macro-world items are going to be somewhat wrong because of the inherent differences in scale.

Here's a list of "Molecule of the Month" images from David S. Goodsell. Some great looking images.

http://pdb101.rcsb.org/motm/motm-image-download

I also recommend the book described at the link, with more pics like that. It's short and very approachable.

Does this misunderstanding translate to medicine and medical research ?

This article seems to equate "exhaustively understand" with "can be simulated down to an atomic level". I think that's not a good way to go about it. A cube of silicon with 1cm side length has about 10^20 atoms. Storing one bit for each atom would use exabytes of memory. Luckily, we don't have to simulate the cube at that level of precision to get a pretty good idea about how it behaves.

The "we could just simulate the whole brain down to the quarks" argument is an argument for the existence of AI. Nobody seriously suggests that it's the practical path forward to achieving it.

I agree with your general premise, that abstraction is essential and sufficient for us to understand a lot of the functions and mechanisms of the cell.

I do, however, think that the author is right in calling the cell ferociously complex, and that a lot of the abstractions we use hide and gloss over a chaotic system whose particulars can explain a lot of unexplained or misunderstood processes. That's fine to help us get an initial understanding, but you always want to drill down to a lower level abstraction until you (don't) reach the bottom of the stack of turtles.

Biology undergoes a major paradigm shift couple of decades, and minor paradigm shifts on an even more regular basis, and I attribute that largely to the increasing resolution with which we're able to observe things.

Ten years ago, microscopes were shittier, we were using microarrays instead of RNA/DNA-seq, we weren't using biotinylated-fusion proteins... but that stuff was necessary for us to really prove in a persuasive way ground-breaking shifts like that the DNA->Protein->Function paradigm isn't sufficient to describe a lot of what goes on in a cell.

PREEMPTIVE EDIT: I know that DNA->Protein->Function paradigm was questioned way earlier than 2007, but it's taken a lot of persuasive evidence since then using modern techniques to get things like lncRNAs implicated in cancer

"a lot of the abstractions we use hide and gloss over a chaotic system whose particulars can explain a lot of unexplained or misunderstood processes."

Is not that always the case with abstractions? In fact, is not their point?

The article says:

"How far is human science from understanding a human brain? [..] Assuming that each neuron contains 175 trillion atoms, the total brain contains roughly 2 * 10^25 atoms."

Well, by the same logic, we could argue that we don't understand cars.

>Well, by the same logic, we could argue that we don't understand cars

Or, by similar logic, that we don't understand how the planets of the solar system interact. It seems that in the history of science, initial views of complex things as chaotic got gradually replaced with better abstractions. As in, the word 'planet' means 'wanderer', then there were epicycles, perfect circles, elipses... I'm not a biologist, but maybe the current abstractions and models of cells are waiting for their Newtons and Mendeleevs to figure out the right models.

The epicycle theory was actually pretty good and explained planet motion to the precision that measurements of that time provided.

I agree that better methods of observation tend to yield radical new insights and I'm pretty excited about all the cool things we can do, for example by adding fluorescent markers to biomolecules. But I don't believe that the simplest model that can perfectly explain a behavior of a cell is even in the same ballpark as an atom-scale simulation.

> The "we could just simulate the whole brain down to the quarks" argument is an argument for the existence of AI. Nobody seriously suggests that it's the practical path forward to achieving it.

Yes, but you have to consider that funding will normally go to research programs that can show steady progress - even if it'll be 10 million times more work (of the same kind) to have a potentially useful result.

Here's the discoverer of the electron, J. J. Thomson, explaining it far better than I can:

“If you pay a man a salary for doing research, he and you will want to have something to point to at the end of the year to show that the money has not been wasted. In promising work of the highest class, however, results do not come in this regular fashion, in fact years may pass without any tangible results being obtained, and the position of the paid worker would be very embarrassing and he would naturally take to work on a lower, or at any rate, different plane where he could be sure of getting year by year tangible results which would justify his salary. The position is this: you want this kind of research, but if you pay a man to do it, it will drive him to research of a different kind. The only thing to do is to pay him for doing something else and give him enough leisure to do research for the love of it.”

Luckily at the current stage any work that improves our atom-scale brain simulations is wildly beneficial to all kinds of fields, so the funding is not wasted.

But the information contained in the two examples (silicon cube vs cell) are several orders of magnitude apart in complexity. I don't think that's really a great counter to the thesis of the OP.

Though I agree that simulating down to an atomic level is probably not necessary, understanding will come from analysis closer to that level of granularity of analysis than it would from [given so many units of volume of a cell, we can conclude that X and thus achieve understanding].

Even higher level formal approaches are ferociously complex:

Within each cell is very complex gene regulatory network (GRN). GRN behavior can be modeled as complex stochastic recurrent neural network. Genes are on/off and they can put other genes on and off with gradients and interaction of the proteins they produce.

Then there is the functioning of neurons on top of that.

There are even higher level abstractions, like cortical columns and other structures. It's conceivable that there are 4-7 levels of complex functional interaction that produces human behaviour and understanding all of those levels is important.

The brain of C. elegans (roundworm) has been mapped exactly (connectome is known) and we know it's 302 neurons and 8000 synapses well (it has little over 1000 cells total) but we still can't understand how its primitive brain works. It don't even have spiking neurons and it's still a mystery.

It's probably hopeless to try to simulate a cell at the atomic level. Instead, it seems to me that computational biologists try to build models of cells[1] where if I understand correctly all occuring molecules are tracked as concentrations, not as particles. Presumably, possible reactions are established not by solving the Schrödinger equation but with empiric models and whatnot.

I don't know how closely such a model can match an actual cell behavior, but I don't see any reason it can not do so pretty well, while also being computationally efficient thus doable.

1. https://en.wikipedia.org/wiki/Cellular_model

The challenges involved are still insanely tremendous. For ex, correctly simulating folding a lot of the larger human proteins de novo is still beyond our capabilities.

It's not exactly the same problem. A cell model will hopefully not have to know the exact geometric configuration of a protein, as long as it contains a rich enough collection of possible reactions with other proteins. Determining those reactions is difficult indeed, but it only has to be done once, and one can hope it will be done soon with quantum computers.

The complexity of cellular biochemistry is why the most effective near-term approaches to medical biotechnology involve sidestepping the need to learn very much more about how it all works in detail. They are engineering rather than scientific efforts.

This is most apparent in aging research, though the principle applies to most medicine. The idea is to identify the root cause differences between normal and diseased metabolism, a task that is a lot easier than figuring our how exactly those root causes propagate and expand into the disease state over time. Root causes are those that have no other cause beyond the normal operation of cells. You then intervene to revert the root cause differences and see what that does to metrics of disease state and progression. The goal is to return metabolism to its normal state, a state that we don't fully understand, but that we know works.

A practical example is the presence of senescent cells. They cause problems, they are not present in large numbers in young, healthy people, so remove them. Removing them sidesteps the whole involved process of trying to understand exactly how they cause problems, in detail, at the level of cellular mechanisms. Researchers are working on that goal of full knowledge, but it will be years until that slowly starts to produce useful therapeutic options, and there are probably decades yet in any research program or combination of research programs that seeks to alter the behavior of senescent cells to make them less harmful. Meanwhile, removing senescent cells works reliably to improve health and turn back disease, and you don't need to know why that happens in order to obtain the benefits, beyond the high-level sketch of "less harmful signaling".

Unfortunately most medical research takes an entirely different approach, which is to start with analysis of the disease state (complex, expensive, slow) and then work back through proximate causes of disarray. At each point at which a therapy might be attempted, try to adjust the proximate cause. Since this doesn't touch on root causes, it is usually marginal, and even the successes just postpone the inevitable.

For comparison, here are the opinions from some of the leaders in pathology informatics, coincidentally out today:

Andy Beck (winner of the Camelyon16): http://www.archivesofpathology.org/doi/full/10.5858/arpa.201...

Alexis Carter (Emory, works mainly on wrangling clinical databases and generally keeping tech from killing patients): http://www.archivesofpathology.org/doi/full/10.5858/arpa.201...

Here's a thought experiment for all Singularitarians.

If a modern-day Music Man appeared in your town with virtual everlasting life in a box, and all the miraculous brain-uploading equipment attached, then started selling tickets and "uploading" your neighbors... how would you be able to tell whether he's actually migrating their consciousnesses, or just using a deep network to train simulations of their minds inside the virtual world while concurrently euthanizing them?

Not a singularitarian myself but personally, I don't care as long as the simulation is sufficiently good (waves hands vigorously). Though I would prefer in the case of a duplicate that I wasn't euthanized.

What does it matter to you that a copy of you is uploaded, but you are left to wither away in meatspace?

That's nice for the copy, I guess, but how will it benefit you?

I think you could at least make the argument your copy has a similar sort of relationship to you that a child does, so you might reasonably be a little more invested in its happiness than some other random AI (not that I am optimistic about the prospects for uploading consciousness).

It matters because the original and the copy belong to the same reference class[1]

1. https://en.wikipedia.org/wiki/Self-sampling_assumption

Similar to how I want people to remember me, and how I would like children of mine to live on and prosper. We are all, on some level, replicating machines who want to organize the world for more of ourselves.

Edit: Even our bodies are completely renewed over time, and any sense of self similarity that relies on the particular arrangement of matter seems fraught with problems to me.

Philosophical zombie problem.

Why would you need to kill the original person in the first place? Couldn't there just be a copy of me running around in their fancy version of the Sims?

Also, what is the difference between a simulation of a conscious person vs. actual consciousness?

> he's actually migrating their consciousnesses,

What exactly do you mean by that?

Isn't that top 5 greatest questions? I don't think anyone has an answer - I'd say for the purpose of the exercise, you have to just suspend disbelief.

Forgive me, quoting the first few sentences of the article:

> Fifty years ago, the first molecular dynamics papers allowed scientists to exhaustively simulate systems with a few dozen atoms for picoseconds. Today, due to tremendous gains in computational capability from Moore’s law, and due to significant gains in algorithmic sophisticiation from fifty years of research, modern scientists can simulate systems with hundreds of thousands of atoms for milliseconds at a time. Put another way, scientists today can study systems tens of thousands of times larger, for billion of times longer than they could fifty years go. The effective reach of physical simulation techniques has expanded handleable computational complexity ten-trillion fold.

Doesn't setting the initial bar to the first academic paper set it so low that you'd have massive gains regardless after a few decades?

I'm not sure what your point is.

They don't try to tease out whether it was raw computing power or better algorithms, just that it was tens of trillions of times more computed.

Also, minor nitpicking -- they said papers not paper, likely meaning the first several years to decade of it.

To be fair, they don't try to tease out if it is algorithmic or computational because they state it was both.

To make a simulation of many atoms, you must do a lot of approximations. For example that the electrons are affected only by the nearby atoms, and use an approximate model for each atom. It has a lot of assumptions, but it provides good enough result and the runtime is O(n).

If you make less approximation, you get a accurate results that consider more of the quantum details of the problem, but the runtime is something like O(n^6) or O(n^9). If I can abuse the notation, it's more "accurate" to calculate the run time as O((5n)^9) or O((10n)^9), i.e. the invisible multiplicative constant is big and you can only calculate molecules with a few atoms. With some additional approximations you can use them for bigger molecules, but not too big, something like up to 100 atoms.

For an accurate simulation, the problem is exponential, the run time is something like O(comb(10n,5n)^3) where comb is the combinatorial number, something roughly like O((10n)^(15n)). This can be used for very small molecules with one or two Carbons or Oxygens and a few Hidrogens (because they are cheaper).

And last year I went to a research talk about the relativistic correction in the H2 molecule. They couldn't calculate all the relativistic correction for a molecule with more than 2 atoms last year.

Lets imagine you can simulate a sperm a cell over it's cell cycle. What do you want to learn about the sperm after you've got your simulation? After you've simulated the sperm you have not yet gained any insight towards how the sperm.

Analogously, you don't need to do molecular dynamics simulations of rocket ships in order to understand how it works. In fact that would be the wrong abstraction. Similarly an atom by atom simulation of the human body would be the wrong abstraction to understand physiology.

Computational models need to be useful. I am not convinced this approach to simulate the cell would be useful.

I think computational efforts could be greatly reduced if we focus only on protein-protein interactions for any pair of proteins (or a small group of them). From this, we obtain a functional description of every protein, and this allows us to perform simulations at a much higher level.

Of course, the number of proteins that can be produced from a given genome is still very large, and the number of binary/ternary/... interactions is even larger.

The trouble is a protein's function is tied to to its solvent, and the small molecules it interacts with on timescales orders of magnitude smaller than a protein itself, AND protein complexes up to microns in size can contain strucutres critical for functionality at time scales orders of magnitude larger.

There just is no way to accurately simulate across so many orders of magnitude when the effects both above and below it efficiently impinge on that function.

> Following the historical example from the previous paragraph, and assuming that Moore’s law continues in some form, we are at least 50 years of hard research from achieving thorough understanding of the simplest of human cells.

Things fall apart here. Moore's law has been dead for a while.

I'm pretty sure none of the futurists who believe in uploading themselves into a computer expect to do it by completely simulating a brain on the quantum level. Indeed, that would in some sense miss the point since that would accurately simulate all the degeneration a human biological brain would experience, in which case why bother?

Secondly, I think there's an implicit definition that a "full simulation" is equal to "fully understanding something", but that's not the case. For any non-trivial process, we say we "understand" it if we have a cognitively-tractable simplified model that works well enough for our use case. "Understanding" almost never implies that we have the ability to fully simulate it. I "understand" quite a lot of computer technologies without the ability to simply sit down and spew out an implementation of them with no further thought but my already-existing understanding; nevertheless, I "understand" them well as evidenced by my ability to harness them thoroughly and effectively.

And for the other direction, the mere ability to produce vast quantities of data does not mean that we have "understanding", either. You can give me the source code for a large simulation [1], say a weather prediction program, and you can give me the incoming data, and in principle I could manually compute that entire program one step at a time, but at the end I still might not "understand" anything I just produced. On its own, even if we were handed a computer system that did have a full simulation of a neuron in it that still would not grant us "understanding" of a neuron; it would grant us a tool that we might be able to use to obtain understanding, but it would not itself be "understanding".

So the literal facts in the essay are interesting, the historical perspective is interesting, the observation that full simulation of the physical world at certain significant scales is still dozens of orders of magnitude away is interesting and relevant, but the conclusion that therefore brain simulation is impossible and the reasonably-inferred conclusion that we can never understand these other large structures is logically invalid, due to a strawman view of what "brain simulation is" and due to a very peculiar and not-widely-shared implicit definition of "understanding".

[1]: Note I'm not saying anything about the simulation corresponding to reality today, I'm strictly referring to the terabytes of data the simulation itself can produce on its own terms.

Reply on news.ycombinator.com