And various dedicated AI models all made poor predictions of the results of that experiment, which casts doubts about the value of the AlphaGenome predictive map.
A virus is much simpler than a human, but even for that simple virus the effects of most of the mutations could not be predicted. A half of the mutations had harmful effects, and for a half of those it is unknown for now why they were harmful.
For a human the uncertainty about the effects of a mutation will be far greater than for one of the simplest viruses.
> A virus is much simpler than a human, but even for that simple virus the effects of most of the mutations could not be predicted.
Is this the correct logic? I (very naively) would think that a small mutation of a virus, with the percent change that is being made, would be more of a fundamental change in the nature of the thing, so maybe less predictable. Where poking genes of a human being more of a (possibly strong!) statistical nudge, since most things seem to be spread across many genes, what's being expressed for sometimes environmental reasons, and even influenced by "junk DNA".
I think you have it completely backwards. A smaller genome makes small tweaks much more likely to produce measurable effects, which means your data quality is much better. A larger genome with more gene interaction is more likely to produce non-measurable or non-subtle effects with perturbations of a single parameter. If a positive effect can only be measured by simultaneously mutating multiple inter-dependent genes, and you don't have a data point with those mutations, the effect can never be predicted. The model may even predict the exact opposite: if you have a bunch of data points showing that varying one gene at a time causes the organism to die, the model will predict that the sum of those mutations also causes death. With larger genomes you need exponentially more data points to discover subtle effects.
The Deus Ex game series is known for relatively high-effort writing you can find while exploring, and I think this fictional e-mail from Deux Ex: Human Revolution has some renewed relevance after 15 years:
> We've been working with VersaLife’s new Chaos Model Genomatrix, which is trying to apply fractal mathematics to the study of genetic mutation. In essence, we've been trying to predict the future of human evolution using past mutations stored in the database of our mitochondria (which is passed uninterrupted from ancestral Eve through our mothers; it gives us an accurate and universal roadmap for our common past).
> The thing is… the sample you sent fits one of our evolutionary models of the future. We’re talking about someone who is ahead of the genetic curve by one step, maybe two. Do you know what this could mean for medicine alone? You have a living, breathing Nobel Prize on your hands!
this is literally not possible. while I'm excited for these kinds of things, alphafold is already imperfect, and this basically side-steps actual structural predictions. it sounds good, but frankly biology is not there yet. in practice not even alpha fold can tell me if my gfp fusion will work...
Meanwhile someone went and did the experiment on the equivalent problem in a virus: https://www.science.org/content/blog-post/mutate-em-all-and-...
And various dedicated AI models all made poor predictions of the results of that experiment, which casts doubts about the value of the AlphaGenome predictive map.
A virus is much simpler than a human, but even for that simple virus the effects of most of the mutations could not be predicted. A half of the mutations had harmful effects, and for a half of those it is unknown for now why they were harmful.
For a human the uncertainty about the effects of a mutation will be far greater than for one of the simplest viruses.
> A virus is much simpler than a human, but even for that simple virus the effects of most of the mutations could not be predicted.
Is this the correct logic? I (very naively) would think that a small mutation of a virus, with the percent change that is being made, would be more of a fundamental change in the nature of the thing, so maybe less predictable. Where poking genes of a human being more of a (possibly strong!) statistical nudge, since most things seem to be spread across many genes, what's being expressed for sometimes environmental reasons, and even influenced by "junk DNA".
I think you have it completely backwards. A smaller genome makes small tweaks much more likely to produce measurable effects, which means your data quality is much better. A larger genome with more gene interaction is more likely to produce non-measurable or non-subtle effects with perturbations of a single parameter. If a positive effect can only be measured by simultaneously mutating multiple inter-dependent genes, and you don't have a data point with those mutations, the effect can never be predicted. The model may even predict the exact opposite: if you have a bunch of data points showing that varying one gene at a time causes the organism to die, the model will predict that the sum of those mutations also causes death. With larger genomes you need exponentially more data points to discover subtle effects.
The Deus Ex game series is known for relatively high-effort writing you can find while exploring, and I think this fictional e-mail from Deux Ex: Human Revolution has some renewed relevance after 15 years:
> We've been working with VersaLife’s new Chaos Model Genomatrix, which is trying to apply fractal mathematics to the study of genetic mutation. In essence, we've been trying to predict the future of human evolution using past mutations stored in the database of our mitochondria (which is passed uninterrupted from ancestral Eve through our mothers; it gives us an accurate and universal roadmap for our common past).
> The thing is… the sample you sent fits one of our evolutionary models of the future. We’re talking about someone who is ahead of the genetic curve by one step, maybe two. Do you know what this could mean for medicine alone? You have a living, breathing Nobel Prize on your hands!
(From Megan Reed's computer, in the prologue.)
If you like that abstract geometric headline image, you might also like the source artist their model “learned” it from: https://leegriggs.com/xgen-rendered-with-arnold-for-maya
the catch with a genome-wide map is the benchmark labels came from conservation, so you can't tell prediction from re-reading the prior.
> benchmark labels came from conservation
And from human disease...?
this is literally not possible. while I'm excited for these kinds of things, alphafold is already imperfect, and this basically side-steps actual structural predictions. it sounds good, but frankly biology is not there yet. in practice not even alpha fold can tell me if my gfp fusion will work...
gfp --- green fluorescent protein
glow in the dark jelly fish bits that biologist stick in other critters to see if, when and where it shows up (under a blacklight).
Some more discussion: https://news.ycombinator.com/item?id=49611251