[iguy]

For complex traits, I believe people think that linear effects dominate. Which is another way of saying that they see no advantage to adding a hidden layer, which encodes interactions, compared to adding up individual effects.

The caveat here "for complex traits" means things like height, for which we know there are hundreds to thousands of common variants which matter. Some things aren't like that, e.g. simple recessive gene effects are interactions! And, with more data, this may change.

[gbrown]

Yeah, you see a lot of penalized regression techniques applied in that field. With that said, identifying useful non-linearity might just require moar samples.