In a quasi-experimental study (N = 60), grade 7/8 teachers students were taught to write arguments in content-area subjects. After instruction, students drew on document portfolios to write on a new topic: "Do the continents drift?" In a MANCOVA, students who participated in argument instruction scored significantly higher than a control class on the combination of dependent variables. A stepwise discriminant analysis indicated that instruction most strongly affected argument genre knowledge, which in turn accounted for variance in the other dependent variables. The features of argument texts that were most strongly associated with science learning were: the number of argument moves, the number of science propositions taken up from source documents, text length, and text coherence. These results support a constructivist model of writing to learn in which students use genre knowledge to select information from source documents and construct genre-specific relationships among ideas.