Multi-body dynamical (MBD) systems, such as bearings, gearboxes, and suspension systems, are vital in many industries. Developing digital twins for these systems is crucial for early fault detection, health monitoring, and efficient maintenance. To this end, real-time prediction of dynamics is essential. Physics-based prediction models are robust but depend on precise measurements of material and physics parameters, like stiffness and damping, which may not always be accurately measurable. Additionally, simulating systems with complex forces, e.g., elastic-hydrodynamic interactions in lubricated bearings, can be computationally intensive. This has led to the adoption of data-driven methods, but they often fail to generalize beyond their training data. Physics-informed machine learning, incorporating system-specific partial differential equations (PDEs), improves generalization but faces challenges in MBD systems due to coupled PDEs, and dynamic boundary conditions driven by time-varying interactions. This study presents a novel graph neural network (GNN) architecture with physics-informed inductive biases for modeling dynamical systems \cite{RIB}. GNNs offer a modular and scalable framework for learning multi-body dynamics by representing system components as nodes and learning pairwise interactions over connecting edges \cite{GNS}. Our model incorporates physics-informed biases into the GNN's message-passing framework, enabling it to learn dynamics by observing the trajectory data alone. This is achieved by encoding pairwise messages as vectors and interpreting each message-passing step as a single-step forward Euler approximation of Newtonian dynamics applied to nodes. Additionally, by leveraging the graph structure and differential node predictions, we introduce a novel method for inferring parameters of pairwise interactions. This aspect enables our model to work in both prediction and inference modes, leading to potential application in system identification when trained on historical trajectory data. Our model was evaluated on diverse spring-mass simulations with distinct force functions, stiffness, rest lengths, and damping parameters. The model demonstrated strong generalization, accurately predicting extended dynamics trajectory roll-out on untrained systems and inferring correct interaction parameters like stiffness and damping coefficients.