23
$\begingroup$

My Background:

I'm a mathematics graduate student with a physics background. I have a very little biology knowledge and a little knowledge of machine learning and statistics.

Topic:

I recently found out about the following computational biology breakthrough: A computer managed to independently develop an explanatory theory for a 120 year old problem in biology. It is detailed in the popular media here and here and in great detail in this paper.

Questions:

I lack domain expertise but I'm very curious as to how this was possible and why it happened now. For example:

  • Were these methods novel (from a machine learning perspective), or does the novelty lie in using old tools/techniques and adapting them to a workable problem?
  • Mathematically speaking, what kind of models were produced by the computer? Are these standard in this field of genetics/biology? Is this area particularly mathematically tractable?
  • What about this particular biological problem made it amenable to this machine learning approach?
  • What other kinds of low hanging fruit (in biology or science in general) could we expect to see tackled with these techniques?
$\endgroup$
  • 2
    $\begingroup$ Popscience blogs have a flair for exaggeration. This is not the first application of GA and machine learning has been used to solve a lot of problems in biology. I am sure that those who have written the blogs have not even read the article. $\endgroup$ – WYSIWYG Jun 7 '15 at 6:15
  • $\begingroup$ Agreed. However, my first exposure to the topic was through the popsci articles via social media so I figured I'd throw those in to give a picture of how I came to understand the topic. $\endgroup$ – Christian Bueno Jun 7 '15 at 19:58
  • 1
    $\begingroup$ I find this question to be of low quality. Details are hidden in links, and premise is speculative at best. $\endgroup$ – Oct18 is day of silence on SE Jun 8 '15 at 23:43
20
$\begingroup$

The fruit, sadly, does not hang so low.

Short version

Lobo et al (the work you refer to) is a nice and not especially novel application of basic Systems Biology modeling approaches to the wound healing system in flat worms. The main barrier to the wider application of such work is the lack of the necessary experimental data. Lobo et al themselves don't really have enough data to justify their models.

Long version

Lobo et al can be thought of as consisting of two separate parts: the models themselves, and their genetic algorithm based search of modelspace.

The Models

The models aren't novel. Their basic form is extremely common in Systems Biology work. Take a look at the equations in the methods section of their paper and in their Supporting Information. For example:

enter image description here

This is a very standard way of representing a set of coupled biological "species". Some of those species correspond directly to the concentration of some specific protein (ie βcat), whereas other species are more vaguely constructed (ie head) and presumably correspond to some generalized set of biological elements (eg proteins, DNA, lipids, etc).

The first term in each of the equations is a production term. It represents how the given species is increasing with time. Each of the fractions takes the form of something called a Hill Function. These come up all over the place in biological modeling. The important thing to know about Hill Functions is that they vary either from 0->1 (for activators) or from 1->0 (for inhibitors) as the concentration of the relevant species varies from 0->infinity, with a smooth transition region in between.
The min and max operators are meant to stand in for the Boolean operators AND and OR, respectively. In other words, if two of the Hill Functions are related by a min operator, they both have to be in an ON state in order for the overall term to be in an ON state. If the two Hill Functions were instead related by a max operator, only one of them would need to be in the ON state for the term to be in the ON state overall.

The second term is a degradation term. It represents how the given species is decreasing with time. The particular form of the degradation term (directly proportional to species concentration) derives from the fact that the main sink for generic biomolecules is dilution via cellular growth and division.

The third term (the thing with ∇² in it) is a diffusion term. This describes how the species moves around the in the worm, assuming that it is free to do so (not all proteins/species can exit the cell). Not all of the species have this diffusion term.

The Search of Modelspace

Their modelspace search isn't novel either, although some very specific aspects of it may be.

the basics of modelspace search

The basic problem of modelspace search works like this: imagine a generalized form of your model in which every species interacts with every other species in every possible way. Each of the equations will have a (large) set of free parameters associated with it. You then take a set of experimental observations and-using any one of a number of methods-find a set of parameter values such that your model will reproduce those observations. This is called parameter fitting. In order to find the unique set of parameters that describes the "real" system that you're trying to model, in general you need at least as many independent experimental observations as you have independent parameters in your model. If you have more parameters than you have data, then you will be able to find infinitely many different parameter sets that reproduce your experimental data.

Here's where things start to get complicated, and a bit fuzzy. What separates modelspace search from standard parameter fitting is that you usually employ some version of the Principle of Parsimony in modelspace search. The Principle of Parsimony can be thought of as equivalent to Occam's Razor: the simplest model is usually best. In practice, what this means is that when you don't have enough experimental data to uniquely parameterize your model, you should favor versions of your model with as few species-species interactions as possible.

The Principle of Parsimony is one specific example of a broad category of modelspace constraints known as "prior knowledge". This refers to everything you know in advance about what your final model "should" look like. Prior knowledge tends to be specific to the particular modelsearch at hand, and it can be incorporated into the search algorithms in many different ways.

With enough prior knowledge the modelsearch can be constrained to an extent such that, in theory, it should become possible to identify the one true parameter set, even when you only have a relatively small set of experimental observations. Good luck constructing a formal proof for that, though. In fact, for a complex system it's usually impossible to quantify the extent to which a given piece of prior knowledge reduces the number of experimental observations required to uniquely fit the parameter set.

the modelspace search in Lobo

For the type of model used in Lobo, a rough calculation says there's going to be

$3n^2 + 4n$

independent parameters (where n is the number of species). Further, there's going to be a set of parameters that describes whether each species-species interaction is activating or inhibiting, and another set of parameters that describes the Boolean logic of those interactions.

The "comprehensive" model in Lobo has 14 species, so there are at least

$3\cdot14^2 + 4\cdot14 = 644$

644 free parameters to fit during the modelspace search. Against those 644 parameters they have exactly 16 independent experimental observations.

$\endgroup$
21
$\begingroup$

The paper by Lobo and Levin is an attempt to learn a model that represents the inner workings of a biological system by fitting parameters to data. This is a common topic in "systems biology", a model-based approach to studying biology that is popular in some fields.

Even for small systems, this is a phenomenally hard problem. Unlike most machine learning problems, achieving good accuracy (on independent test sets) is not enough; the model structure itself must be correct, meaning that it should reflect the actual mechanisms operating in the cell or organism. It's like trying to identify the actual structure of an integrated circuit by only observing inputs and outputs --- not only identifying the logical rule that the system exhibits, but the actual wiring diagram that implements that rule. It is of course quite possible that alternative models exist that perfectly captures the behavior, and yet is completely different "on the inside". In other words, the system may not be identifiable from the data --- two different models may produce exactly the same behavior. So there's no way to know beforehand if these problems are even theoretically solvable (let alone practically feasible).

Lobo and Lewin claim to have accomplished this task for a small system of gene products. As far as I can tell, the central piece of evidence is that the model they come up with does capture some mechanisms previously known to exist. Whether this is just plain luck, overfitting, or a special case where (parts of) the system is actually identifiable from data is hard to tell. But it's notable that Lobo and Levin fall back on demonstrating predictive accuracy in the end, which is not enough to support their thesis. The critical test is to take a new prediction from their model and actually test it experimentally, to prove their (strong!) claim that their method can actually infer new mechanisms. Doing so is expensive and time-consuming, and much more difficult to get right than tweaking models to fit mechanisms you already knew beforehand --- which is exactly why I want to see that evidence. If the authors really believe that their method is a breaktrough, those experiments should have been done.

Finally, these type of algorithms are applied to models containing only a handful of components, and typically to systems that have been studied for decades and are very well-characterized to begin with. In most of biology, the problems are orders of magnitude more complex, and often we don't even know what the components are. So I thik there are a limited number of cases where this approach is feasible.

$\endgroup$
3
$\begingroup$

Christian, great idea to ask this question here before taking important decisions.

Are those media articles a hype?

Yes. Over the last 10 years I constantly see those hype stories in media about "revolutionary" large-scale-study/big data projects with mind-blowing numbers (gigabases, teraflops, terabytes, thousands of papers and hundreds of genes).

Journalists are competing for readers with pompous titles as well as scientists need to promote their results. I recall that ~7 years ago Russian national TV came up with nothing less than "Scientists unravel the mystery of Russian soul", when one of our scientific institutions, close to government, finally managed to actually launch (not buy and put on hold as usual) a next-generation sequencing machine and sequence a full genome. Curiously "Russian soul" belongs to half-jewish, half-armenian guy. :)

Though it's nice that in media papers journalists and scientists managed to explicitly articulate the problem we can't solve: how to predict morphology using molecular biology data. This work is not even close to doing that. It essentially predicts a binary trait, "heads or tails".

Is this a new breakthrough approach in science, connecting areas that noone could imagine to connect?

No. This approach is called Systems Biology. It's been getting popular, I'd say, since late 90s-2000s. This paper is not like "look, I've invented a DNA computer/quantum computer" etc. It's an ordinary systems biology study, its quality is a matter of discussion.

Is this a good systems biology paper?

I don't claim that it's bad. But some things sound strange to me:

  1. I don't think that de-novo predicting the network of interactions should be done by systems biology, when some facts are known from literature and could be enriched by Omics. If you're modelling a molbiol/biochem network, too much freedom might render your network totally different from reality.

  2. Robustness test is unconvincing to me. Ok, "We dropped 3 crucial papers" and still network holds. Ok, network predicted some results from papers, not taken into account upon its construction. But this is not a systematic test. Split the data into two parts, optimize the network on one part and test it on another one.

  3. Small number of genes involved in the described process and binary trait in consideration make the authors do overambitious claims, IMHO.

Can a smart and diligent person as yourself make a better career, doing systems biology than something else? Will your higher IQ and/or mathematical background give you an edge?

I think, no. I think, this area downplays the difference in intellect between people, while you might prefer a more discriminating area, even if you allow the assumption that you're not the single brightest person to ever live on Earth. :)

You should have worked out (or have from birth) some mental habits (e.g. habit to consider simple examples with small numbers of elements; use modular thinking; not to get stuck with a single approach - switch tricks; etc. etc.) that less smart people lack. Systems biology often works with too complex objects to imagine, thus your edge might be lost.

Complexity:

In reality cascades of biological pathways include huge numbers of players involved. Moreover often a single protein has several seemingly unrelated functions. I was a bit shocked at first, when aconitase, a Crebbs-cycle protein turned out to be at the same time an iron-responsive element and regulator of mRNA turnover. What a mad engineer could've created this?! Later I heard that many Crebbs-cycle proteins have seemingly unrelated "part-time jobs".

Stability of solution:

Look at any pathway in KEGG. Or this one: a famous cascade of cell signal transfer: . Although, it's not that complex, when you know all those proteins by heart, it might be hard to work with it quantitatively. Solutions might be unstable. Lack of experimental data may lead your simulation in a wrong direction. For instance, famous p53 protein (in this cascade it has 3 arrows or so) in reality has more than a hundred counter-agents, loss of interaction with one of them might lead your simulation to a state, critically different from reality.


Publish-ability:

Another problem for systems biology papers is lack of comprehensible results. People are usually interested in a paper, when they can build something on top of it or at least when it's an interesting special case to know about. To be honest, I can't make myself read them. What's the scientific takeaway of this paper for instance? It doesn't exist. I can't use it as a scientist.

Systems biology guys often speak of holism approach as opposed to reductionism. To me it's an empty word - the only comprehensible (and thus publishable) way to approach complex entities for human being is modular thinking - you just replace a complex cascade with a black box with a set of inputs/outputs and rates and treat it as a single object, reducing the complexity that way.

I feel that systems biology approach is better suited for commercial use in industry than for small-lab science. You need tools to mine data, to visualize results etc. Companies can afford building them for themselves. As for the results, in industry you don't need them to be publishable. In agricultural industry you can optimize producing organism's metabolism with systems biology means to make it output more product of interest e.g. in cows or chemicals-producing organisms (though same goal can often be attained with pure genetic engineering easier). May be pharmacology can be interested in this to optimize their drugs to find best way to modify particular coefficients in particular reactions.

Practical achievements based on fundamental biological science still make use of reductionist works. They stem from little facts that cervical carcinoma is caused by a specific virus and vaccination against it seems (silver bullet warning) to prevent it. Or the fact that CCR5 mutants are seemingly immune to AIDS or similar results for causative agents for stomach cancer or hepatitis. Or useful discoveries of tools, such as siRNAs, Crispr/CAS9 or Tal effectors.


If you still decide to give biology a try, it won't take you too long to grasp the molecular biology. You can mostly work with it without knowledge of biology or chemistry. In a month you'll grasp the basic concepts such as replication-transcription-translation, in a year will be familiar with most key processes, in ~3 years will be a good specialist in some area.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.