It seems that backpropagation might be a good abstraction for biological learing rules after all: "local predictive coding converges asymptotically (and in practice rapidly) to exact backprop gradients on arbitrary computation graphs using only local learning rules" from https://arxiv.org/abs/2006.04182