The inference or reconstruction of various biological networks, including regulatory, signalling or metabolic pathways, from large-scale heterogeneous data is currently an active research subject with several important applications in systems biology. While several approaches proposed so far cast this problem as inferring a graph de novo from genomic data, I will argue in this talk that the network of interest is often partially known and that the reconstruction process should use this partial knowledge to guide the inference of the missing edges. I will then review how this paradigm leads naturally to various supervised machine learning algorithms for graph inference, and illustrate the relevance of the approach through several examples of successful prediction of missing enzymes in metabolic networks.