Hacker News Reader

Comment by dghughes | Hacker News Reader

I wonder would it be possible to accelerate a synthetic cell's evolution via software?

If you already know exactly what's in the cell (you built it!) could a computer speed up its life via a software model and jump ahead to something more complex which you build as version 2?

View on HN

You'd be hard pressed to beat biological evolution...

A liter of e. coli can have upwards of 100s of billions of cells. Which can double every 90 minutes. Try to run that many instances of a full environmental simulation in software. That's 1L of bacteria. Scaling that up to 100L is pretty easy. Try to do that on your super computer...

A straightforward evolutionary experiment to find resistance to an antibiotic can literally run trillions of trials in a stupid simple overnight culture. Slowly increase the dose of an antibiotic in a few liters of bacteria over the course of a day or two and see what survives.

To take it a step further, I wonder if it is possible to engineer better catalysts to speed up cell division. Evolutionary pressure should bias towards maximum reproduction rates (given a the right environmental conditions). But I wonder how much variation there is in cell division times between different species of bacteria.

I think you're describing the field of [bioprocess engineering](https://en.wikipedia.org/wiki/Bioprocess_engineering). A lot of time is spent trying to balance cell growth with the possibility of mutations being introduced into the system. Even between different strains of E. coli the doubling rate can vary. Moreover, one also needs to contend with the fact that prokaryotes typically cannot produce very complex proteins. This leads some processes to need to use eukaryotes such as CHO cells, which are much slower.

Not to mention millions of years of evolution.

Perhaps it would be more practical to use biological components to accelerate software...

And then use this to train neural nets ;)

As someone who codes and does a lot of wet biology/chemistry... What calculations would you suggest be accelerated biologically?

The comment was playing on how much more efficient biological components are compared to digital ones.

An actual attempt to answer your question.

Currently it takes quite a lot of computation to detect outlines of objects in a raster image. This is an ability innate to many animals. Suppose we had the technology to keep individual brain cells of an insect alive, and ability to interface with biological neurons, we could pass visual data through, say, an insect's eyes and brain, and get a signal on the other end that corresponds to vectors and lines that describe the object in the visual data. The insect brain is able to process images orders of magnitudes more efficiently than modern computers.

Or we could make some sort of Dog helmet, hopefully without a physical connection, that somehow transmit the vector data inside the Dog's brain to our computers.

Anyways, all of this looks almost impossible. We simply don't have the technology to interface with the detailed and intricate data that is present our bodies, as our brain process the world around us.

I do think one day we will be overlaying data in our own eyes using some sort of neural implant. That device will probably rely on existing vision-related, biological, computational infrastructure in our minds.

I wouldn't be surprised if it's two hundreds years away.

P.S. Here's an article relating to the idea of putting insect brains in robots: http://www.kurzweilai.net/robots-with-insect-brains It's talking about a replicated insect brain on digital hardware though.

Sth paralelizable. Like pathfinding, or factorization.

Could we do a bacteria that reproduces with part of genome changed in predictable way - like

    ... 100 ... ---> ... 110 ... -> ... 111 ...
                \               \-> ... 101 ...
                 \
                  -> ... 010 ... -> ... 011 ...
                                \-> ... 001 ...

And then checks that number in some way and signalizes somehow when it arrived at solution ? For example by releasing a toxine that kill others and reproducing without changes from that point on.

If we could, we could beat exponential-time problems in linear time as long as there's enough space and growth medium :)

The interesting problem you hit is although you suddenly CSV do things in linear time, you now have an exponential space problem.

So suddenly your problem scales from test tube to swimming pool to ocean sizes.

Right, so in the end constants matter, and the question is - are bacteries smaller than computers are fast?

That would be awesome, but it is unlikely with current software and hardware. Even small scale simulations (nanoseconds to microseconds of a single protein molecule) require hours of sim time (on large clusters).

To elaborate, time is the problem.

The number of atoms in the smallest viable cell is actually reasonably small, less than a billion atoms, such that you could hope to store it all in memory on a single beefy computer.

Meanwhile, you need something like picosecond resolution in the simulation, while you need seconds of simulated time to see interesting macro-level things happen. If you want to see cell division, you're probably talking a minimum of thirty minutes of simulated time.

Not to mention that the molecular dynamics simulations you're describing can't actually simulate chemical reactions. (the covalent bonds can't change in the sim, unless you do a much more intensive QM calculation and even that's iffy).

If you want evolution, you would also have to simulate the cell's environment. Without an environment, there's no natural selection and thus no evolution.

Actually, I was wondering if a learning algorithm could be coupled with a builder that uses real synthetically created biological components in an automated system. This way, the builder comes up with new combinations, automates all the science around evaluating the combination as quickly and automatically as possible, and the software/learning algorithm controls the development towards some generalized goal.

Ok, I admit this sounds ridiculously sci-fi, but, why do I feel like this is actually possible (just really, really hard)?

EDIT: To clarify, I am actually suggesting a real living system coupled with software/hardware, not a simulation. The idea is, I realize real life can't be "sped up" like a simulation, but if it's truly automated and kept running/fed/operational, the system could just sit there and keep trying things as long as it runs - over a long period of time, humans could monitor its progress.

I don't see why it couldn't be done, that's kind of a goal of programmable microfluidics.

http://groups.csail.mit.edu/cag/biostream/

It is sci-fi though, a biolab that operates like aws for biology experiments is featured in 'Rainbows End'.

Somebody did do a software-style analysis of a simple cell. Fascinating article, amazing results, pity I don't recall it. Worth looking for, shouldn't take long to find.

You're probably thinking of this paper by Jon Karr et al.: "A Whole-Cell Computational Model Predicts Phenotype from Genotype"

http://www.cell.com/abstract/S0092-8674(12)00776-3

Note the hit rate on predicting lethal knockouts - not much better than flipping a coin

I've been a programmer at the Folding@home project for about 8 years now. It takes massive computing power and a lot of time to simulate even one reasonably sized fast folding protein using the state of the art methods. Simulating a whole cell ab initio is way beyond the reach of current technology.

It would be cool to play with this idea using the OpenWorm platform.

Cells make proteins. Proteins fold up to do their jobs. Protein folding is O(2^N) to simulate in software, so probably not.

If only it was O(N^2). Seconddary, tertiary and quaternary structure definitely is not the same complexity as bubble sort

You didn't read my post. I said O(2^N) which means basically computationally intractable for large N and even if Moore's law were to continue indefinitely it would still be computationally intractable for large N. O(N^2) problems, on the other hand, are far more approachable by software.

Reply on news.ycombinator.com