DataFramesMeta.jl was designed with intention of explicit passing of column names. If you want, you can pass arbitrarily complex expression to it also escaping it with $ (though it is useful only in complex cases - here it is just a noise and probably using combine is better):
@combine(groupby(df, :gr), $(names(df) .=> (x->sum(log.(x))) .=> x -> x * "_logsum" ))
(note that it also shows you that you can pass a function that renames your input columns).
As @pdeffebach commented column selection having a specified type is achieved using the names function.