Multi-agent trajectory prediction is crucial for many real-world applications. This task faces challenges in effectively capturing individual temporal patterns and complex interactions between intelligent agents. Existing models either solely focus on single agent dynamics or neglect multi-agent collective interactions. To address these difficulties, we propose a graph neural network-based trajectory forecasting approach, named ForceGNN. Specifically, it combines Transfomer to extract temporal patterns for each agent and utilizes a multi-scale hypergraph neural network to simulate complex crowd interactions. Meanwhile, we incorporate a social force model with strong inductive biases, abstracting pedestrian interactions as social forces, thereby achieving more efficient trajectory prediction. We compare ForceGNN with state-of-the-art deep learning methods on the UCY/ETH and SDD datasets, achieving optimal performance on both.