We introduce a graph-theoretic formalism suitable for modeling biochemical networks marked by combinatorial complexity, such as signal-transduction systems, in which protein-protein interactions play a prominent role. This development extends earlier work by allowing for explicit representation of the connectivity of a protein complex. Within the formalism, typed attributed graphs are used to represent proteins and their functional components, complexes, conformations, and states of post-translational covalent modification. Graph transformation rules are used to represent protein-protein interactions and their effects. Each rule defines a generalized reaction, i.e., a class of potential reactions that are logically consistent with knowledge or assumptions about the represented biomolecular interaction. A model is specified by defining 1) molecular-entity graphs, which delimit the molecular entities and material components of a system and their possible states, 2) graph transformation rules, and 3) a seed set of graphs representing chemical species, such as the initial species present before introduction of a signal. A reaction network is generated iteratively through application of the graph transformation rules. The rules are first applied to the seed graphs and then to any and all new graphs that subsequently arise as a result of graph transformation. This procedure continues until no new graphs are generated or a specified termination condition is satisfied. The formalism supports the generation of a list of reactions in a system, which can be used to derive different types of physicochemical models, which can be simulated and analyzed in different ways. The processes of generating and simulating the network may be combined so that species are generated only as needed.