It is typical to use log normalized data, so if you already have it in that format that would be a good place to start.
A dataset will typically support a certain *number* of predictors rather than a fixed percentage. That, in turn, will depend on the size of your dataset (number of observations), the nature of the target variable and the method you will be using. So unfortunately no easy answers. If your dataset is sufficiently large, take advantage of dividing your data into three portions so you can try various numbers of genes without compromising your ability to assess the final performance of the models.