A variety of clinical observations suggest that certain forms of epilepsy are due to long-term, progressive changes in neural networks that eventually provoke spontaneous and recurring seizures. This process of network transformation, known as epileptogenesis, is a potentially important therapeutic target and also serves as an extremely interesting model of central nervous system plasticity. This article reviews some of the significant, recent advances in our understanding of mechanisms underlying epileptogenesis in different forms of epilepsy. The most substantial progress has been made in work related to temporal lobe epilepsy (TLE), where the biochemical, electrophysiological and anatomical changes in the hippocampus have been intensively studied. This has led to a number of cogent and testable hypotheses, including the concept that dentate granule cell hyperexcitability in TLE is due to a selective loss of hilar neurons that renders inhibitory cells 'dormant.' Studies of other forms of focal epilepsy suggest that a seizure focus may develop as a result of axonal reorganization or immune-mediated effects on membrane channels. Epileptogenesis in generalized epilepsies remains poorly understood, although recent work using models of absence epilepsy point to the critical role of GABAB or T-type calcium channels in the thalamus. Also, new transgenic mouse lines with epilepsy phenotypes have introduced candidate genes, such as those encoding the serotonin 5-HT2C receptor or the alpha subunit of calcium/calmodulin kinase II, that may be responsible for epileptogenesis. Finally, a large amount of investigation has focused on seizure-induced gene expression and it is now clear that seizures can cause a cascade of changes in the expression of gene products that are likely to play a role in network plasticity. Progress in developing 'anti-epileptogenic' therapies will require further advances in understanding the mechanistic roles of these various biochemical and anatomical changes in the transformation of normal to hyperexcitable neural networks.