Hacker News new | past | comments | ask | show | jobs | submit login
Ask HN: How close are we to replace animal models with software?
40 points by JPLeRouzic 41 days ago | hide | past | favorite | 48 comments
In pre-clinical trials for some diseases, animal models are notoriously nearly useless. The story of the C9orf72 mice model (a familial ALS/MND model) who lived healthily in a Broad lab, while their counterparts at Harvard died quickly [0, 1] has been widely discussed.

In general, for nervous system diseases, mice models are not translatable to humans. One reason might be that their nervous system is quite different from those of primates/humans, for example, great primates have direct motor neuron connections (no interneurons) for complex manipulations.

So scientists often try to develop animal models closer to primates, see for example [1]. But their methodology for validating animal models intrigues me.

Basically, they seek to disrupt to some extent some biological function, and if the behavior/phenotype of the animal model looks like that of human patients, then it is assumed the animal model is correct.

I wonder if it's not possible to use a software model of a human being with the same level of effectiveness. For example, there are very complex software models of human beings (see Biogears engine and similar [3]).

If we apply the same methodology to human software models and if when we disrupt a biological function, and if then the software model displays the pathological behavior, isn't as valid as an animal model?

I would be interested to know your advice.

[0] https://www.alzforum.org/news/research-news/gut-microbes-dif...

[1] https://pubmed.ncbi.nlm.nih.gov/32483373/

[2] https://pubmed.ncbi.nlm.nih.gov/39343009/

[3] https://www.biogearsengine.com




Not even close. We don't yet have a computational model that accurately reproduces all the properties of water, the solvent where most life of interest occurs. And, if and when we have one, and are able to develop accurate models for the myriad proteins involved (an even worse problem,) and we're able to develop accurate models for all their interactions (O(huge)), and so forth, the computational cost to simulate systems that would be relevant at the scale of a whole organism would be... jaw-dropping.


Thanks for your answer, but reproducing an organism down to the cell force fields is not my point.

My point is that animal models are extremely imperfect, otherwise we would not spend billions on clinical trials.

On the other hand, now there are many software that are used in medicine to predict the behavior of an organ in some circumstances. Whole-body simulations of pharmacokinetics and toxicology have existed for two decades and are now quite accurate.

I feel that we are now on the verge of integrating those many approaches. For instance:

https://pubmed.ncbi.nlm.nih.gov/38480804/

So I ask, if it's possible to use soon, the integration of these software components to replace the pre-clinical usage of animal models.


We can't even simulate a single cell. At least not accurately. You'd need to be able to simulate ribosomes, gene expression, epigenetic factors. Still a lot we don't understand because even single cells are incredibly complex electrochemical machines.


I agree but we also don't know what gravity is, and yet we can compute it with some degree of error. There might be a world we can just approximate enough of this cell, that the details matter less.


This actually illustrates the crux of the problem. Gravity is a single (highly accurate) approximation when applied to 2 objects. But when you have 3 objects interacting with each other, the system is chaotic: https://en.m.wikipedia.org/wiki/Three-body_problem

Even if we could estimate all the molecules, cells, and processes of a biological entity to a similar precision as gravity, the errors would compound exponentially.


Thanks for your answer, but we either don't understand the biology of the animal models that we use.

I think that medicine and drug regulations do not use the same conceptual framework as biology. Medicine and drug regulations are pragmatic, hence the concept of clinical trial: They test if something works while minimizing the confounding factors. They do not try to understand the biology down to cellular force fields of the animal model.

So why not have the same approach with software? Anyway, drugs authorities have already approved drugs based on software simulations [0, 1].

[0] https://www.fda.gov/media/163156/download

[1] https://www.fda.gov/science-research/about-science-research-...


> Anyway, drugs authorities have already approved drugs based on software simulations

The simulations you have linked aren't on the same scale. The paper uses the phrase "are powerful tools that COMPLEMENT traditional methods for gathering evidence"

And I'd argue this is more like detailed analysis rather than a simulation. It's like hoping you've found an immersive computer world like The Matrix, but it's a 2D side-scolling video game like Mario Bros.

I'd also question "approved drugs based on..." and instead argue they "didn't reject drugs based on..."


Yes, but while animals are a black box to some extent, the outcomes of the experiments are not. And you don't have to worry about whether your results are an artifact of your simulation.

Your instincts are good - if we could simulate people in silico we could basically understand and cure every disease - but the scale of such a simulation is literally (not figuratively!) astronomical. Biological systems are way, way, way more complex than they appear and our computers are (currently) hopelessly inadequate.


I can imagine a world where you can simulate a simple organism like a worm without knowing how to simulate the cells that make it. This is because emergent behaviour of complex systems can be much simpler than the constituent parts of that system (temperature vs. knowing particle velocities).


I can write a program today that can simulate a worms "emergent behavior" without simulating the cells. Maybe a little 3D model that wriggles around. I can even do a human!

But I think what we're looking for here is something closer to emulation. In the same way that video game console emulators seek to reproduce the exact bugs that the original hardware produced - the value here does not come from abstracting away the lowest levels of behavior.


> the value here does not come from abstracting away the lowest levels of behavior

I totally agree.


How are you going to simulate the administering of a medicinal molecule? Or the knocking out of a gene?


My modelling the relevant parts of the worm that are involved in these processes.

You don't need an atomistic DFT simulator to get a good enough simulation of fluid dynamics, enough to design jets or rockets that go into space.


> modelling the relevant parts of the worm that are involved in these processes.

You mean cell chemistry?


It seems to me that if you can fully simulate a worm you're probably only an order of magnitude of difficulty away from a person, but to get there you need to have crossed dozens across many domains.


But none of that means you should take a drug bc someone’s elaborately technologized guess is that it’ll be okay


there's https://openworm.org/ for simulating a c. elegans, but think you're underestimating the complexities


Modeling a creature is so damned hard. Take a single celled organism and break it down to the most basic one you can imagine: I'm willing to bet a model of its interaction with drugs would take decades to be useful in any manner. We're only just getting __ok__ models of how protiens fold and even then within specific temperature boundaries. Even then the error rate is high enough that they'd be a lot of issues. You cant just have 5% of your protiens being wrong. Even if we assume perfect protien folding models it's another problem to model correctly how they'll interact wit X drug or eachother. If you want to start somewhere, try modeling hemoglobin. It's got a simple job: Move oxygen through blood, dump it, and pick up CO2. Yet it's fussy as hell. if the protien is slightly off it'll become useless. And you can't cheat because plenty of diseases that are all about homoglobin being "slightly off" so you can't just hack it so all your homoglobins are perfect magic carriers, you'd have to model it perfectly and watch how drugs may interact with it.

Every step of this modeling would be hellish. Not to mention just how much stuff you'd need to model and have simulated in parallel for even the most basic of creature simulation. Parallelization is trivial for nature. Once it figured out how to create one cell the the next 2 or a trillion was easy. In computer simulations the first model just as resource intensive to simulate as the next... simulating a trillion cells? Ooofff. You'd be lucky to get a second of simulated time after a months run.


Biology is an organic computation that runs at the speed of existence. There may be no way to simulate this computation in silicon at the same scale, insofar as biology is already massively parallel and may very well involve quantum computation.

Put another way, if we accept for a moment that the universe is a simulation, it may be fundamentally impossible for an in-simulation simulator to ever reach the computational power of its parent simulator.


Lots of biological processes are indeed parallel but that doesn't mean they are all different. Once the most important ones are understood we can reasonably simulate how biology works. Just like you can simulate a much more complex CPU on a lower end CPU - just at a reduced clock speed.


So what you're saying is, the most efficient AI is not a data center of GPUs, it's... a giant tank of lab-grown brains.


There's an entire field called Biological Computing which studies precisely that (not AI, more around general purpose computing).

https://en.m.wikipedia.org/wiki/Biological_computing


As a counter to my original thought, perhaps simulating the computation of a skull-sized organic system is possible by building an earth-sized silicon system.


>So what you're saying is, the most efficient AI is not a data center of GPUs, it's... a giant tank of lab-grown brains.

That's literally what a think tank is.

Or a research university, for that matter. The labs are where our brains grow.


They didn’t say that at all. It seems like you wanted to say it.


Biologist here. We're definitely not close to replacing animal models with software. However, many aspects of animal models can be studied using cell culture systems. I work on ovarian organoids, which can be used to develop fertility therapies.

Still, organoids will merely reduce the number of animals used in drug development rather than eliminating them entirely. Before giving a drug to humans, testing whether it's safe to give to animals is a step that can't be skipped for the foreseeable future.

I took a look at the Biogears website that you linked, and it looks like a physiology simulator, i.e. more of a model of a plumbing system than a full organism. Something that can model heart rate and blood pressure won't be able to say if a cancer drug will work (or if it will have a toxic side effect).


> Something that can model heart rate and blood pressure won't be able to say if a cancer drug will work (or if it will have a toxic side effect)

Thanks for having a look at Biogears.

(edit) Please have a look at: "Constraint-based modelling predicts metabolic signatures of low and high-grade serous ovarian cancer"

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11344801/

https://github.com/katemeeson/repository_to_accompany_paper_...

Most successful pre-clinical trials *fail* at predicting phase III success. The ratio is enormous. It's time to put some effort into alternative approaches.


Biology is insanely complicated. Simulating that to any degree of fidelity is really, really, hard - at least in medically interesting domains.

(E.g. we don't even have a simple model that given some parameterized human as input and a simulated macro diet can predict their body mass...)


There's a huge difference between when we know how an organ system behaves in a particular circumstance vs. when we don't. Take, for example, the hepatic function in biogears: https://github.com/BioGearsEngine/core/blob/trunk/projects/b... If you know all the constants and how the liver reacts to a given stimulus, then sure, you don't need to test anything out in vivo. But the reason we do experiments is that we don't know how the liver will react in a given circumstance, so we can't rely on known equations to tell us the answer.

I also think that while there are circumstances where animal models are not helpful, those tend to make the news because they are the exception rather than the rule. There are many, many diseases where animal models were critical for figuring out at least where to look for human disease processes. In addition, a lot of the issues with mouse models are not due to the fact that mice are inherently a poor substitute for humans, but that the models (the specific genes mutated) were a poor mimic of human diseases. For example: "Measurements of gait and grip strength showed that their muscle deficits were in fact mild, and post-mortem examination found that the animals died not of progressive muscle atrophy, but of acute bowel obstruction caused by deterioration of smooth muscles in the gut." [https://www.nature.com/articles/507423a]


Biology is a borderline infinite system depending on how deep you go.

We need to be able to model molecular forcefields to be able to model DNA expression to be able to model protein expression to be able to model, layers and layers of higher order molecules just to represent a cell.

Then...you have combining cells to make an organism and the interaction of the organism with its environment which affects all of the above.

I think a monkey will be able to understand how to use a cell phone before humans understand how biology works.


Do you actually need to simulate at that minuscule level of detail?

Or is it possible for a system to be built that can approximate biology similar to how LLMs approximate cognition without true understanding and reasoning?


Maybe it’s possible to build something analogous to LLMs, but it’s going to need to be based on completely different technology, due to: 1) the lack of abundant training data (nothing analogous to a trillion-word internet to scrape) 2) the unacceptability of hallucinations (can’t just dose people; need some other way to validate)


I don't know.

I guess it depends on how accurate you want to get and when testing therapies for humans you probably want to be pretty accurate.

There may be some abstraction layer that provides 'good enough' accuracy though that I'm not aware of.


Not close. This might be the best whole-animal sim available, and it is a for a worm: https://openworm.org/


And not just any worm, but a C. elegan, which always grows the exact same way every time. The whole reason those worms are so well understood was due to ease of analysis. Hell, it always has 302 neurons which always wire up in the same way. It couldn't be easier to simulate, and it's really hard to simulate!


To be clear, open worm is not even close to serving as a viable animal sim for bio research on worms even


Unknown Unknowns man. We may be able to iterate faster with a computer model, but the next step will be animal models perhaps followed by human test subjects.


The body of animal is computationally irreducible.

It is impossible to create a shortcut using software that can "skip ahead" and accurately predict what the body will do given arbitrary initial conditions.

This is possible only for stuff like eclipses because they are reducible.

To simulate an animal model, you will have to replicate it's environment, all it's proteins all it's, all it's hormones, all it's cellular structures and all it's physical & psychological behaviour.

Not only that, we also have know know 100% how they all work.

An animal is a complex system and complex systems often can't be reduced in to simple constituents used to make accurate predictions.

Link — https://en.wikipedia.org/wiki/Computational_irreducibility


For additional insight, we are not actually able to fully model a single bacterial culture, gene expression is random and we are barely getting to know one gene at a time when considering dynamics and not just steady state (we also have some idea how to represent an average bacterium's gene expression at steady state, if they were all doing the same thing, but they aren't)

As far as I know, we are actually closer to (less far from) in vitro models where you culture human cells into organs very close to real ones and apply drugs to those organs. I think they already do that with skin for cosmetics but everything else is far away (key word: I think, not sure).


Hi! My professional job is to write emulators, my wife is a molecular biologist, and as a kid I had dreams of writing the exact software you are referencing, so I think I have a decent handle on this question.

We are no where close. We are ludicrously far away.

Let's define the exact scenario: we want to replace clinical trials with a software simulation of the human body. If the simulation shows ill side-effects, we can deny approval of a treatment.

1. We can barely emulate other computers. It's tempting to look at something like a Nintendo emulator and think "oh this isn't that hard" but it is. Most video game emulators get about 90% of the emulation right and it's good enough for most games. But a common practice is to carry patches for all the software to patch the software. Hilariously, this is sometimes because the software is working around a hardware quirk or bug, but then it turns out difficult to emulate that quirk or bug, so we patch out the hack. If you want a perfect emulator it's really hard [0] If you're testing for bad drug interactions in a human simulation, it's exactly these quirks/bugs you want to accurately simulate!

2. The software of cells is DNA and the genes contained within. And genes encode for proteins which are amazing at doing a huge amount of varied tasks. But these are the basic building blocks, and we've only begun to scratch the surface. We made huge progress but we barely understand. [1] Imagine trying to work on an emulator of a microchip, but we don't quite understand how transistors work.

3. There's mind-body feedback loop with the endocrine system [2]. On top of everything else, we need to simulate the brain. Sure we can use a simplified model of that, but animal models are also simplification. The whole point was to try to get more accurate, and how accurate do you need to guarantee results? I know this argument is a bit absurd but it's to point out there's no finish-line, only more and more difficultly as higher accuracy is demanded.

4. How would we develop this simulator. Let's suppose I have my initial prototype. I've simulated various known drugs and got results, and I've tuned my parameters. But this is a massive complex system. Once I run a new novel drug, the point is that it's doing something new! So, if I have a bad reaction, is it a bad drug, or a simulation bug? Each scenario is new and poses to surface incorrect modeling between complex subsystems. You can argue that we'd build our confidence over time, but that means we'll see the long path to simulator development. There have been some attempts but they have appeared to not provide predictive results [3]

5. When asked to debate whether or not we could simulate the human body, the pro-simulation side invoked fantasy: "Exascale or quantum computing will enable algorithms that we are yet to conceive of" suggesting that we are very far away if it is possible. [4]

[0] https://arstechnica.com/gaming/2011/08/accuracy-takes-power-...

[1] https://www.quantamagazine.org/how-ai-revolutionized-protein...

[2] https://www.psychologytoday.com/us/blog/the-brain-body-conne...

[3] https://en.wikipedia.org/wiki/Virtual_Physiological_Human

[4] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8638236/


Thanks for your answer, but pharmaceutical companies have used software for decades to describe how a drug behaves in a body and to do toxicology studies.

And there are software approaches for humans like this one:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7285886/


On my research, I've done it with a plant (to a limited extent, obviously). The reason why I chose a plant is because they're less-complicated versions of animals, at least as far as my simulation was concerned.

We are "far" but it's that kind of thing that an unexpected breakthrough gets you 80% of the way. My bet would be, doable by 2040.


It depends on what you mean by "replace". If you are willing to limit the term to pharmacokinetics then companies like https://www.verisimlife.com are having success there. Full disclosure - I used to work there.


Centuries, if not millennia (at least).

It’s possible future AI advances speed that up, but isn’t imminent regardless.


I was under the impression that one of the holy grails of biological modeling was protein folding. I know we made progress but I don’t think we are even close to that milestone yet.


It would be replaced by organoids, but in certain cases it can't be given biology might even have quantum physics in the cards


Not going to happen. A model cannot be "valid" because by definition it's not ground truth (aka all models a wrong, some aren't too wrong). It can be good for some cases where it is known to match the reality good enough, but the cases any scientist care about are exactly those where existing models does not match the reality.


Wait until we cracked protein folding, then add 20 years.




Consider applying for YC's W25 batch! Applications are open till Nov 12.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: