ADDED: It's a natural problem in Lie theory to find classification-free proofs of results which are first observed case-by-case. But I'm not aware of any substantial progress beyond Borel's careful commentary (in the 1983 volume), where he notes the equivalence of five statements about cohomology and the equivalence of three statements about orimes related to root systems and Weyl groups (with reference to Demazure and Steinberg). As he observes, the first five imply the last three, but only case-by-case study shows the converse as well.
Along the way Borel also points out an improvement over his original Nagoya formulation: while the "bad" primes (possibly 2, 3, 5) are those dividing a coefficient of the highest root, the "torsion" primes are the more limited ones dividing a coefficient of the coroot of this highest root. For instance in type
$G_2$, with respective short and long simple roots
$\alpha, \beta$, the highest root is
$3\alpha + 2\beta$ whereas its coroot is
$\alpha^\vee + 2\beta^\vee$.
There is a long history of study of the topology of a semisimple Lie group (equivalent to the topology of a maximal compact subgroup), in which a reduction is made to study of the root system and its Weyl group: for example, the determination of Betti numbers in terms of exponents or degrees for the Weyl group (Chevalley, ICM 1950). But this kind of transition is rather subtle. In the study of torsion primes, a key role is played by subgroups of maximal rank in a compact Lie group, these being correlated with certain subdiagrams of the extended Dynkin diagram. Much of the technology recurs in the study of p-compact groups, as Jesper observes. But not everything is well understood conceptually.