We can't experiment directly on humans - we can only look for correlations, which always causes trouble with determining causality. Biological questions are inherently complicated, so any hypothesis is necessarily little more than a stab in the dark, even if we had the ability to compensate for genetic variation et cetera, which we most certainly don't. Large-scale studies are expensive, time-consuming, difficult to organize, and would inevitably end up politicized from 200 different directions, and then driven into the ground in the next campaign season.
In short, any and all dietary studies are doomed to be 99% failures.
But that 1% is important. I don't think there's much question that we have a better understanding today, of what constitutes healthy food, than we did 50 years ago. We know, in particular, that foods rich in certain substances (I don't want to try to classify them all in one or two words, because I don't study this stuff and would get it wrong - I'm talking about tradition "junk foods" and sweets, like cupcakes and potato chips.), are strongly correlated with a wide variety of health issues, the incidence of which has increased over the last several decades. Furthermore, these foods are inherently attractive to children, and (perhaps more controversially?) display limited addictive properties.
Yeah, our knowledge is limited, and yeah, we'll have missteps. But I don't think this is "premature optimization" - the program's running really slowly, and everyone's complaining about it. We don't understand a lot of why it takes so long to finish, but there is something we can do that would probably make it at least 5% faster. Why not give it a shot, and then go back to profiling?
Oh, and have an upvote, for contributing to the conversation. Because that's the etiquette around here, although some seem to have forgotten.