Over the last seven years there has been a productive collaboration between the groups of Dr Sternberg (ICRF) and Dr Muggleton (Turing Institute, Glasgow and then Oxford) in applying machine learning to structural biology. The early work used the program GOLEM, e.g. rules governing the location of a-helices in the all-a were identified that yielded a prediction of 80 Following a pioneering study on ILP and drug design (King et al., 1992), recent work employed PROGOL and a study probing structure-activity relationship for mutagenic nitro-aromatic compounds identified rules that were both of high predictive accuracy and provided intelligible chemical insights (King et al., 1995). Importantly PROGOL was able to find a predictive relationship within a subset of compounds that was not previously identified by experts. This encourages us that the approach will be productive when applied to protein topology. This proposal extends stems from an earlier study using GOLEM to derive topological relationships governing the a/b class of proteins (King et al., 1994). Data mining was performed on a topological database of protein structures established by Dr Rawling's group at the ICRF (Rawlings et al., 1985; Clark et al., 1991). The rules obtained had predictive accuracy and provided new insights.