<p>Traditional subspace methods for direction-of-arrival (DOA) estimation face significant challenges in practical scenarios involving coherent sources, multiple sources, few-snapshots, and low signal-to-noise ratio (SNR) conditions. While deep learning approaches have emerged as alternatives, they often lack interpretability, unlike classical methods. This paper presents G-ESPRIT, a novel framework that integrates graph neural networks with differentiable subspace learning for robust DOA estimation. Our approach employs a Graphformer architecture to model spatial relationships between array elements and generate enhanced covariance representations. By incorporating differentiable ESPRIT into the training pipeline, G-ESPRIT enables end-to-end gradient-based optimization while preserving the rotational invariance properties essential for accurate angle estimation. The graph-based processing effectively captures complex inter-element dependencies, showing superior performance in these challenging scenarios. Experimental results demonstrate that G-ESPRIT achieves substantial improvements over both traditional subspace algorithms and recent deep learning methods. The proposed method maintains computational efficiency while providing interpretable outputs that facilitate source enumeration and performance analysis.</p>