Knowledge-based leader-following synchronization problem of heterogeneous nonlinear multi-agent systems is challenging since the leader's dynamic information is unknown to all follower nodes. This paper proposes a learning-based fully distributed observer for a class of nonlinear leader systems, which can simultaneously learn the leader's dynamics and states. The class of leader dynamics considered here does not require a bounded Jacobian matrix. Based on this learning-based distributed observer, we further synthesize an adaptive distributed control law for solving the leader-following synchronization problem of multiple Euler-Lagrange systems subject to an uncertain nonlinear leader system. The results are illustrated by a simulation example.