On model ensemble analyses of nonmonotonic data.

Mammalian ribonucleotide reductase (RNR) activity has been reported to be nonmonotonic in ATP. If many nonlinear models are to be fitted to such data automatically as part of a model search process, use of the same initial parameter values across all models can lead to too many poor fitting, monotonic least squares fits, i.e., false model rejections. We propose that such fits can be rescued by using as initial parameter estimates the final estimates of neighboring models that do have nonmonotonic fits; here models are neighbors if complexes that they represent differ by at most one ligand. We use this approach to show that troughs in RNR activity versus ATP can be fitted similarly well by models that do or do not demand a third ATP binding site.