I can't naively hope that very high performance (exascale) computing helps finding the right theories here. But I do hope that people don't waste time digging in the wrong direction. Facing complexity is better than not being aware of it.

ps: I hope though, in the case of non mental disorders, that finer and cheaper technology allow for preemptive discovery, continuous monitoring and non invasive therapies.

Much appreciated.

The problem as always with machine learning methods is figuring out how to construct your machine correctly and what data to feed it. Mostly this is about good data collection, not about what machine you use. I am somewhat agnostic on whether you train using a linear model, SVM, neural network, etc. - all of these are likely to be analytically useful, provided I can give them an interesting and well-contsructed input set. But if the data I'm feeding into such a system is too noisy, doesn't sample deeply enough to catalog the extent of variation, or we don't have good descriptors for correctly labeling the dataset, it's useless to try to train a machine, we'll just have GIGO.