This document proposes GraphTrans, a framework that uses graph neural networks (GNNs) to learn local structure and a modified Transformer to learn global structure from graphs. It introduces challenges in capturing long-range dependencies with GNNs alone. GraphTrans leverages a GNN backbone to learn local structure and adds a Transformer to pool local embeddings and extract global structures. The document evaluates GraphTrans on biomolecular and computer programming benchmark datasets and analyzes the impact of adding a CLS token.