Phillip Isola's fifth lecture in MIT's 6.7960 Deep Learning course covers graph neural networks (GNNs). He builds the topic from the ground up, showing how GNNs relate to multilayer perceptrons and convolutional networks through the shared idea of message passing across structured data. The lecture works through how information propagates between nodes and edges in a graph, and how this generalizes the sliding-window logic of CNNs to irregular structures like molecules, social networks, or citation graphs. Isola then turns to theory, laying out known limits on how expressive GNNs can be, including cases where distinct graph structures are indistinguishable to standard message-passing schemes, and what these limits mean for practical model design. Eighty-one minutes of blackboard and slide-based instruction aimed at students already comfortable with basic neural network architectures.