This is an interesting idea. I wonder if we could then use something like [0] to trace back "gay-ness" as well as other "personality features" (I'm not sure what the correct term is here) to the exact genes.
One potential issue I see is that the DNA of any person has an arbitrary length, which poses some challenges in the design of a neural network. Traditionally, this has been solved with LSTMs or RNNs, but as far I know these are designed for data with a temporal dimension (such as text or speech, which progress with time). I'm not sure if that's true for DNA.
Or.. there is no such thing as a gayness gene. Genetics are generally not a strong predictor for human behavior. Gestational environment, what sort of environment you grew up in, whether or not you've suffered any head injuries or have another developmental disorder, internal hormonal environment, the culture you live in... these are all much better predictors.
You could of course argue that environment in the womb is largely genetic (and has a 50% correlation with the kid's genes, which is damn influential).
But once we open that door we're going down an extended phenotype rabbithole where you can start to talk about social interactions amongst non-immigrants being partially genetic because of correlations within regions, etc, and hilarity (with a healthy dash of idiocy) ensues.
The greater point being that saying something is either genetic or due to environment is always a pointless exercise because there is no separation, there are merely genes that spread without ever deciding how to allocate blame or credit for their successes and failures.
One potential issue I see is that the DNA of any person has an arbitrary length, which poses some challenges in the design of a neural network. Traditionally, this has been solved with LSTMs or RNNs, but as far I know these are designed for data with a temporal dimension (such as text or speech, which progress with time). I'm not sure if that's true for DNA.
[0] https://arxiv.org/abs/1509.06321