I think the broader intuition depends on your context.

If you have a very strong reason to believe that the true model contains no significant negative weights, then it may be ok to restrict weights to be positive. If, however, the true model has large negative weights, restricting it to positive domain will probably underestimate (a lot) the contributions of the variables that are correlated with the variable whose true weight is negative. One might imagine mathematical methods to solve this problem (perhaps try orthogonalising the variables, but more elegant ideas might also be possible), but if the true model does have large negative weights, then why not just estimate them?

A quicker idea might be to try a softer method, for example to define a model that penalises negative weights (this way, if negative weights are not truly significant but just artefacts of fitting, they will be squashed), or to use something like Bayesian regression with a strong prior on the weights being positive.

I don’t know what your context is, but I think brain variables are pretty correlated, and positive as well as negative weights are not unexpected. But perhaps in your application this is not the case.

I forgot to say, I think, also, that if the true model has large negative weights and you restrict the weights to be positive, the estimated parameters will vary a lot across different samplings.