Biological evolution favors genome structures that align mutations with environmental shifts. This mathematical framework of selectability functions similarly to kernel alignment in neural network training, utilizing G-matrices and neutral networks to optimize learning.