The value of theory in biology

[glaucousNoise]

Hi folks,

I'm a new graduate student in biology. I'm interested in doing theory (either simulations of biomolecules, physical models of the brain/consciousness, or models of cellular dynamics). I have had very intriguing discussions with fellow students along the following lines:

1. There are two approaches to theoretical/computational biology. One is computational statistics, and the other is called, by the statisticians (computer scientists) "rule based modelling", i.e. normal theoretical science.

2. There is no point to theoretical science in biology. Why use molecular dynamics to figure out exactly how a protein moves when you can just do linear regressions on carefully performed experiments to determine which drugs bind or not?

Needless to say I find these views disagreeable (to be clear, some students from a stats background are more extreme than others, and some of the extremeness may be been caused by the imbibing of alcoholic beverages) but I think taking the strong view, that mechanistic, ground up approaches to biological problems are inferior to computational statistics creates for a more interesting discussion.

I was curious to know what people make of these views, and, in particular the following:
1. What evidence is there that mechanistic modeling in biology and other fields has been successful? I like to trot out Maxwell and transistors as examples of this but I don't know how robust they are.

2. Which areas of mechanistic modeling in biology do you think could produce impressive results in the future?

 

[Bystander]

https://www.physicsforums.com/threads/differences-in-teaching-phys-vs-chem-bio.827659/

 

[glaucousNoise]

This is really interesting, although I can't decide if my thread is redundant with it or not.

 

[Bystander]

redundant with it or not

That's two of us --- thought I'd give you the link by way of confusing your "discussion group" further.

 

[Choppy]

I do a little bit of modelling work in radiation biology.

I think the first argument that I would put forth is that biological modelling is necessary is because you can't do all of the experiments you want to do. Consider issues of medical or biological ethics, financial constraints, or resource constraints.

Secondly, you need models to figure out how all of these smaller systems come together in greater complex system. How does a solid tumour evolve over time? How will it respond to different types of radiation delivered with different dose fractionation schemes?

Third, sometimes you need to convince people to give you the money do to do an experiment. To do that you need some kind of an argument than an idea is viable. Running a computer simulation is cheap and it can let you know if an experiment is worth hiring a post-doc for in the first place.

Fourth, models allow you to twiddle small parameters that may not be worth twiddling experimentally.

Fifth, and these are not in any particular order, is that theoretical models of biological systems can be used in designing new products or algorithms. Look at the field of neural networks, for example.

 

[Ryan_m_b]

There's plenty of reason to do theoretical biology. When it's done right it's a synergistic effort between computer and lab, you use the former to identify potential targets for the latter.

 

[atyy]

Is there any difference between the two supposedly opposing views in the OP? Here are some proposals (please critcize!):

1) Statistical approaches pick the "best" model within a model class.

2) But mechanistic models pick the model class (eg. the relevant variables on which to do regression, or the requirement that the models for high energy physics have the form of a quantum theory)

In both cases, we require that the selected model make successful predictions - is that statistical or mechanistic?

So I'm not sure there is a difference between the approaches.

Fifth, and these are not in any particular order, is that theoretical models of biological systems can be used in designing new products or algorithms. Look at the field of neural networks, for example.

This is a very interesting example, Neural networks are an example of the "statistical" non-mechanistic approach - they are simply models with many parameters, and backpropagation, early stopping and dropout are ways of fitting the model and regularizing the fit.

On the other hand, it is true that the form of the model is inspired by biology, and in artificial neural networks and reinforcement learning can be seen as the theory of the mammalian cortex (before and after learning, but not during learning).

Currently, it is not clear that the dynamics of biological and artificial neural networks is the same during learning (I would actually say it is clear they are not, but there are thought-provoking suggestions that I'm being too naive https://www.cs.toronto.edu/~hinton/backpropincortex2007.pdf). But keeping to the naive idea for the moment, it is clear that if you just want an engineering product that does speech recognition, then the learning dynamics or the fitting algorithm that you use need not be the same as the brain's fitting algorithm. Also, you don't need to know the actual fitting algorithm to teach a biological neural network something (as people do everyday when they learn new things without knowledge of LTP or whatever the brain uses). However, the actual learning dynamics is of great interest to biologists, just as a good model of the dynamics is usually needed by aircraft engineers for Kalman filtering or whatever they use to control the plane.

https://www.physicsforums.com/threads/differences-in-teaching-phys-vs-chem-bio.827659/

This is really interesting, although I can't decide if my thread is redundant with it or not.

I gave an example of "mechanistic" modelling that has been useful in post #28 of that thread.

 

[glaucousNoise]

Is there any difference between the two supposedly opposing views in the OP? Here are some proposals (please critcize!):

1) Statistical approaches pick the "best" model within a model class.

2) But mechanistic models pick the model class (eg. the relevant variables on which to do regression, or the requirement that the models for high energy physics have the form of a quantum theory)

In both cases, we require that the selected model make successful predictions - is that statistical or mechanistic?

So I'm not sure there is a difference between the approaches.

well, computational statistics rarely infers mechanistic dynamics. If you are attempting to predict the fold of a protein using computational statistics, you might do a good job, but you are unlikely to explain how the protein folded.

I suppose you could do something like use, say, a hidden Markov model in conjunction with a Markov state model to build a mechanistic model (as has been done for this very problem) but now you've blended the two approaches.

 

[atyy]

well, computational statistics rarely infers mechanistic dynamics. If you are attempting to predict the fold of a protein using computational statistics, you might do a good job, but you are unlikely to explain how the protein folded.

I suppose you could do something like use, say, a hidden Markov model in conjunction with a Markov state model to build a mechanistic model (as has been done for this very problem) but now you've blended the two approaches.

Yes, that's pretty much what could also be taken from Choppy's and my comments about neural networks - hidden markov models used to be state of the art in speech recognition, but then it became neural networks plus hidden markov models. I'm not sure what the current state of the art is.

I presume you say that using HMMs is a blend since (1) statistical methods are used to infer the HMM parameters, which may not be unique given the data (2) the choice of HMMs to guess mechanism is based on assumptions in chemical kinetics and the need to do experiments to pick among the various HMMs consistent with the extant data?

 

[Murdstone]

Some tangential reading - Life : Its Nature Origin and Development - A. E. Oparin What is Life - Schrodinger The Double Helix James D. Watson. If you have not read the following you have not done science. The definitive works.

It takes an exceptional mind to do theoretical Biology. Most only wish there was more.

There is very little theoretical science in Biology.

Biology, in particular, Evolutionary Biology lacks rigor.

There is no definition of Evolutionary Biology. None. The most you get is something like Evolution is fact. Yes, things have changed but how?

The precise definition of Evolution and how these changes come about are vague.

It is like you state - some round peas and some wrinkled peas - and based on this you explain 5 billion years of evolution. It does not work.

 

[Andy Resnick]

<snip>
2. Which areas of mechanistic modeling in biology do you think could produce impressive results in the future?

From my perspective, modeling has been more-or-less successfully applied at the two extreme levels of biological organization: at the cell/molecular level and at the population dynamics level

There's not much in between. While models of the cardiovascular system are fairly mature, compared to systems biology, there's hardly any results in the large space spanned by tissue-, organ-, organ system- and organism-levels of biology: plenty of room for new approaches and new discoveries.

 

[Silicon Waffle]

[...]
1. What evidence is there that mechanistic modeling in biology and other fields has been successful? I like to trot out Maxwell and transistors as examples of this but I don't know how robust they are.

I agree with all ideas about statistical data analysis and mathematical modeling. I personally prefer math models myself in pharmaceutical and disease or virus related research. Without models, I can't predict e.g the possible growth of viruses in some phase the patient has to enter.
How about modeling cancerous tumor growth or more specialized agents of our immune system such as T cells, B cells, macrophages or even about what protein is on and off during the time any of them gets activated and targeted by specific chemical species?

 

[glaucousNoise]

I was going to say that computational modeling of proteins hasn't really been that successful (we can't access the relevant timescales, the forcefields can't handle disorder very well etc) but there was a very interesting paper where enhanced sampling methods managed to produce "in sillico", as they say, a plausible transition for a transporter.

I find statisticians/mathematicians/computer scientists to be very baffling sometimes; theoretical computer scientists recently 'proved' that edit distance was the best possible algorithm for comparing genomes. It turns out that such a bizarre conclusion arises from restricting oneself to not incorporating new knowledge about genomes into one's algorithms, but rather purely viewing the problem of comparing two genomes as a matter of comparing two strings with no initial guess of the underlying structure.

 

[Ryan_m_b]

There is very little theoretical science in Biology.

Biology, in particular, Evolutionary Biology lacks rigor.

There is no definition of Evolutionary Biology. None. The most you get is something like Evolution is fact. Yes, things have changed but how?

The precise definition of Evolution and how these changes come about are vague.

It is like you state - some round peas and some wrinkled peas - and based on this you explain 5 billion years of evolution. It does not work.

Little of this is true. There are various definitions of evolutionary biology, the very subject is a body of work constituting a theory and the field is hugely rigourous. I have no idea what your basis is for assuming that Mendel's observations on inheritance (which he described using mathematics) are the sole basis for extrapolating the entire history of evolution on Earth.

 

[Ygggdrasil]

Rule-based modeling and simulation can be helpful for interpreting data from experiments where the readout can be difficult to analyze. For example, there are many techniques where the data analysis runs into the "inverse problem." The classical example here is x-ray crystallography. Given a protein structure, it is easy to generate its diffraction pattern. It is much harder, however, to generate a protein structure from a diffraction pattern. Refining crystal structures requires iterative rounds of building a structural model, generating the diffraction pattern of the model, crosschecking the model's diffraction pattern against the real diffraction pattern, then refining the model.

More recent examples includes interpretation of data from a technique called Hi-C, which reports on the three-dimensional organization of DNA in the nucleus. Given a polymer model for how DNA is organized in the nucleus, it's possible to generate a Hi-C map, but there are no methods to take a Hi-C dataset and generate a polymer model from that data. In two recent studies, two independent groups tested various rule-based models for how DNA is organized in the nucleus, and both came up with some very interesting hypotheses that will likely guide future research in this field. These studies, in my opinion, show a great example of how how experimental data can guide theorists to develop new models and how these new theoretical models provide hypotheses for experimentalists to test in the future.

Fudenberg et al. 2015 Formation of Chromosomal Domains by Loop Extrusion. http://biorxiv.org/content/early/2015/08/14/024620
Sanborn et al. 2015 Chromatin extrusion explains key features of loop and domain formation in wild-type and engineered genomes. http://www.pnas.org/content/early/2015/10/22/1518552112