> Many software tools and databases support SBML directly. You don’t write in SBML yourself—your software does it for you.
So its likely that you aren't supposed to parse/write the SBML directly. The BioModels website also provides PNGs so you can look at the state machines visually.
These models are not exactly state machines, they’re usually expressed as systems of differential equations. They’re sometimes expressed as stochastic equations, which are like random state machines, I think.
But yeah, SBML is the standard and there are a number of tools that let you express reactions between bio molecules in easy to read formats. These then output SBML, which is easier to share.
Kind of impressive that academics were able to settle on some real standardization for this field
Perhaps they should be expressed as (analog) electrical circuits. Then you get differential equations in a format that is common and relatively easy to read.
Interesting to see Biomodels on HN. This kind of computational modeling is still fairly niche in biology but I think it’s going to play a much larger role in the future in understanding biology. If you look at a graph of all the interactions between proteins in a cell, it’s pretty clear that everything influences everything else, but we still just don’t understand the details of those interactions. It’s also clear that those atomistic level details can matter quite a lot. Tools like AlphaFold are going to be really important for that, but model repositories like biomodels will help us integrate those details into the networks that actually give rise to the larger scale behavior of cells.
Happy to answer any questions about computational modeling in biology. I’m just a phd student but I’m writing a review about this at the moment so I’ve been thinking about it quite a bit
Be very very good at differential equations. Systems/synthetic biology primarily rely on dynamics to model behavior and side effects. AlphaFold is actually pretty different from the biomodel type of modeling.
This is a good intro to (non deep learning based) computational/synthetic biology. ML in biology is somewhat orthogonal but it is rapidly growing.
Any tips for breaking into the field? Been doing ml/signal processing type stuff towards the end of grad school, but did some diffeq/dynamical systems stuff at the beginning and was considering pivoting towards biological applications of ML. Interviewed with pumas.ai last summer but they eventually turned me down.
Go into industry not academia. Academia won’t pay well and they do not appreciated good quality software engineering. For most labs, ML is considered a tool, not a primary driver of innovation. Look towards places like Isomorphic labs and similar companies that are cash rich and has a strong engineering heritage.
Regarding being very good at differential equations: if I remember correctly, one way to approach solving them is to convert them to difference equations and solve them numerically (and not analytically). Is this true? As you can probably tell, I am not good at differential equations at all, so I was thinking about using difference equations as a crutch.
The systems of differential equations produced by these types of models tend to be analytically intractable, so "numerical methods" are absolutely essential here. The phrase you want to look for is "initial value problem" (IVP): you know the initial values of the quantities described by the equations and you want to find out how those quantities evolve over time.
Full disclosure, I haven’t read it, but I like his research and he’s a good communicator.
It’s tough because it’s a newish field and not super well defined. The frustrating answer is that the best way to understand the field is maybe to read a book about biology, then a book about computational modeling, then a few papers where they put those together.
Yeah they haven’t made much of an impact yet but I’m almost certain they’ll help improve biological models in the future.
Check out INDRA (http://www.indra.bio/). It uses standard NLP tech to convert natural language references to computational models. But, that old school NLP is pretty limited in “understanding” context and details and so the models it makes aren't really comparable to human built models. LLM’s could change that, building really high quality, correct models. That opens up the possibility that we could take all of the biological knowledge in papers and integrate it into models we could use to make much more detailed predictions about cellular behavior.
That said, there’s a ton of work to be done to make it reliable and validate the results.
Interesting, indeed. I wrote my PhD in mathematical modeling in biology nearly 15 years ago, when systems biology was all the hype - I left the field ages ago, but it seems like there has not been significant progress into using mathematical models to guide hypothesis generation and experimental validation.
So while I agree with you in principle, it seems like fusion: always twenty years in the future.
Historically much of computational biology was driven by people with no wet lab experience and no access to a wet lab so hypotheses went untested because nobody wanted to bet their grant on someone else's work they didn't understand or trust. Now that maths and computation are everywhere in biology there are more hybrid dry/wet labs or dry and wet labs working in long term collaborations. So I think there's been some progress but maybe not as much as one could have hoped.
Exactly. Thinking like a physicist is still pretty uncommon in biology, especially biological education. Hence the strong influence physicists had on the field. It also seems like there‘s still a mathematical framework missing to handle biological questions appropriately. Maybe we should revisit Rosen‘s Life Itself.
20 comments
[ 2.8 ms ] story [ 52.9 ms ] threadThe seem to be using a markup format called [SBML](https://synonym.caltech.edu/).
> Many software tools and databases support SBML directly. You don’t write in SBML yourself—your software does it for you.
So its likely that you aren't supposed to parse/write the SBML directly. The BioModels website also provides PNGs so you can look at the state machines visually.
But yeah, SBML is the standard and there are a number of tools that let you express reactions between bio molecules in easy to read formats. These then output SBML, which is easier to share.
Kind of impressive that academics were able to settle on some real standardization for this field
http://seti.harvard.edu/unusual_stuff/misc/lorenz.htm
Happy to answer any questions about computational modeling in biology. I’m just a phd student but I’m writing a review about this at the moment so I’ve been thinking about it quite a bit
Do folks mostly use MatLab or is Python more popular? (IIRC, MatLab used to have a nice GUI for things like system dynamics.)
This is a good intro to (non deep learning based) computational/synthetic biology. ML in biology is somewhat orthogonal but it is rapidly growing.
http://be150.caltech.edu/2020/content/index.html
Traditional bioinformatics is mostly text mining and string processing. Modern bioinformatics is heavily dominated by ML.
Regarding being very good at differential equations: if I remember correctly, one way to approach solving them is to convert them to difference equations and solve them numerically (and not analytically). Is this true? As you can probably tell, I am not good at differential equations at all, so I was thinking about using difference equations as a crutch.
https://en.wikipedia.org/wiki/Numerical_methods_for_ordinary...
As far as textbooks go, I guess I’d recommend Uri Alon’s book: https://www.routledge.com/An-Introduction-to-Systems-Biology...
Full disclosure, I haven’t read it, but I like his research and he’s a good communicator.
It’s tough because it’s a newish field and not super well defined. The frustrating answer is that the best way to understand the field is maybe to read a book about biology, then a book about computational modeling, then a few papers where they put those together.
Check out INDRA (http://www.indra.bio/). It uses standard NLP tech to convert natural language references to computational models. But, that old school NLP is pretty limited in “understanding” context and details and so the models it makes aren't really comparable to human built models. LLM’s could change that, building really high quality, correct models. That opens up the possibility that we could take all of the biological knowledge in papers and integrate it into models we could use to make much more detailed predictions about cellular behavior.
That said, there’s a ton of work to be done to make it reliable and validate the results.
So while I agree with you in principle, it seems like fusion: always twenty years in the future.