Driven by advancements in artificial intelligence, end-to-end learning has become a key method for system optimization in various fields, including communications. However, applying learning algorithms such as backpropagation directly to communication systems is challenging due to their non-differentiable nature. Existing methods typically require developing a precise differentiable digital model of the physical system, which is computationally complex and can cause significant performance loss after deployment. In response, we propose a novel end-to-end learning framework called physics-guided learning. This approach performs the forward pass through the actual transmission channel while simplifying the channel model for the backward pass to a simple white-box model. Despite the simplicity, both experimental and simulation results show that our method significantly outperforms other learning approaches for digital pre-distortion applications in coherent optical fiber systems. It enhances training speed and accuracy, reducing the number of training iterations by more than 80%. It improves transmission quality and noise resilience and offers superior generalization to varying transmission link conditions such as link losses, modulation formats, and scenarios with different transmission distances and optical amplification. Furthermore, our new end-to-end learning framework shows promise for broader applications in optimizing future communication systems, paving the way for more flexible and intelligent network designs.