I understand you have multiple predictors expressed as percentages and you use some kind of multivariate modelling.

Transformations (maybe except log transformation) make the interpretation of the results less easy. So I would try first without transformation and check if (how severely) the model assumptions are violated. What the model assumptions are depend on the method you chose, but

• one assumption is usually the error distribution, often normal;

• another usual assumption of multivariate models concerns linearity.

Transformations often help getting close to multiple model assumptions in the same time, but sometimes you need to find a compromise imperfectly getting near multiple assumptions.

When several variables need to be transformed I usually get an overview with the package 'acepack' of R.

If the interesting response is expressed as percentage, it can be modeled by beta-regression. See http://cran.r-project.org/web/packages/betareg/index.html