Dear lord, this is exactly why everyone thinks neural networks are AGI about to take over the world. Neural networks are not magic.

> I'm pretty sure that if we threw a lot of genomes, healthy and defective, to a neural network, and then showed it a genome and told "make it healthy", it would.

It's not quite that simple.

This is an active research area that one of my former professors is involved in. Things are not as simple as you hope they would be.

You're optimizing in a space of millions of discrete dimensions (one for each base pair), with little knowledge about independencies. This is in contrast to tasks like image recognition, where we can make use of the spacial structure of the pixels to build effective models like convnets.

Additionally, medical datasets, especially ones with genetic data, rarely contain more than a few hundred datapoints, which is not where you want to be for deep learning.

Machine learning is still useful in this area, especially to find genes or gene combinations with high correlation to certain diseases. But we are very far away from having a model that maps genome -> healthiness.

That's an interesting idea, but with my limited knowledge of neural networks and deep learning, don't these things usually need to know when they are "successful?" (i.e. the distance gone right in a mario game or something). How do they not only suss out various options for genomes, then test the effects?

Sorry, but this is kind of laughable and shows a lack of familiarity. The data and existing knowledge base in biology is nowhere near sufficient to train a model like the one you described. Our genomes and the molecular biology it powers is astoundingly complex and small, it will take decades still to begin to understand how all the pieces work together.

I don't think you are wrong in principle but in practice this simply wouldn't work. We don't have enough sequenced genomes in the world to train such a network and even if we did there is no guarantee that a network-produced genome would stay in the manifold of healthy humans when trying to change an existing defective genome to a healthy one. Either way it would be to expensive to find out if a new genome actually works. Finally with a sequenced genome going for a couple gigabytes I don't see us having a large enough usable data set of these any time soon.

I'm sure that machine learning will come into play at some point, but we aren't able to model complex organic molecules in order to understand or evaluate "healthy" on silicon.

In order to use ML on actual organisms, you'd need a way to try a lot of different modifications and then evaluate them after actual growth.

Doing that kind of experimentation with human embryos, fetuses, and beyond has huge ethical minefields as well as many biological engineering challenges.