A novel set of algorithms and computational tools has been developed within PROMETHEUSTM a comprehensive computational environment for molecular modelling, design and simulation. These have been developed for the design of peptide analogue sets from a single lead and have been applied to optimise the potency of a 15-residue peptide. The chosen lead peptide, CAMELO, was designed in the laboratory of Boman.1h5 It is a hybrid of cecropin A, a 37-residue antimicrobial peptide produced by silk moth larvae, and melittin, a 26-residue peptide extracted from bee venom, and combines to some extent the cytolytic activity of melittin while retaining the selectivity of cecropin A for prokaryotic cells. Our objective was to test analogues of CAMELO and to Ðnd alternative sequences with increased potency against a panel of bacterial strains.The physicochemical properties of the molecules were described by the residue-based parameters of Hellberg and co-workers (Z scales)6 and Norinder (ID Scales)7 * To whom correspondence should be addressed. and covariance functions derived from these (Table 1). A third group of descriptors, referred to as the "Design parametersÏ, represent particular molecular properties which have been suggested as relevant to structureÈ activity relationships for this class.A D-optimal design8 in the Principal Components sub-space of these descriptors produced a QSAR training set of 60 molecules well distributed in the possible design space for 15mer peptides.The antibacterial potencies of the peptides were assayed and shown to be well spread in activity space. Partial Least Squares (PLS) analysis was used to develop a variety of QSAR models. The best model identiÐed had a cross-validated R2 value \ 0.65. Descriptors based on covariances of the Z-scales appeared to provide the most predictive model of the training set data, and with this data set, they performed better than the Norinder scales. The covariance scales were calculated using similar methods to Wold et al.9Like them, we found that the covariance parameters gave an improved model of the data, which may be