Steel-framed structures are extensively used in high occupancy buildings, where fire safety is a critical concern. While steel offers high strength and ductility under normal conditions, its mechanical properties degrade significantly at elevated temperatures, increasing the risk of collapse during fire incidents. To mitigate this risk, passive fire protection measures are widely applied to steel components. This helps delay the temperature rise of steel members during fire exposure, maintaining their load-bearing capacity for a longer duration. This paper presents a progressive framework for performance-based structural fire design that balances practicality with engineering rigour. The first stage introduces a streamlined performance-based structural fire engineering (PBSFE) approach that preserves the essential modelling of structural response while avoiding the computational demand of fully nonlinear finite element analyses with coupled heat transfer. This enables efficient, physics-informed determination of fire protection thickness for real building frames and makes performance-based design more accessible for routine engineering practice. Building on this foundation, we further examine whether fire protection requirements can be predicted without any structural simulation. To this end, we propose a Graph Neural Network (GNN) model that learns to map structural topology and member properties directly to fire protection thickness. GNNs operate on graph-structured data, which matches the representation of building frames: columns can be treated as nodes and beams as edges linking them.