ACS Fall 2007—ACD/Labs' Activities Schedule

A new approach for the pre-selection of wavelength, which can be applied with Partial Least Squares or other multivariate regression techniques, will be presented. This variable selection method utilizes the purity function from the SIMPLe-to-use Interactive Self-modeling Mixture Analysis (SIMPLISMA) algorithm to help determine the most significant and influential regions. The selected intervals are then individually tested in practical modeling and prediction to obtain the best possible subset of variables. This algorithm is simple, intuitive, and more readily relates to spectroscopically and chemically understandable features than iterative variable searches based solely on statistical selection criteria. The method was tested on a set of infrared spectra for proteins. The region selection was used to improve the quantitative determination of the fractions of two secondary structure elements, the α helices and β sheets in the protein polypeptide chain. Comparison will be made to the results obtained through interval PLS (i-PLS), an iterative search-based algorithm.

While modern computers and the internet have significantly lowered activation barriers for using and validating computational models, they have also increased the risk of improper use. A scan of recent literature will inevitably produce new computational models that can be used for predicting anything from activity to toxicology. Each will provide some information about how the models were developed, and statistics indicating how well the data has been modeled. While existing predictive models are valuable, it is imperative that these models be properly validated before use. Conducting a fair trial of computational predictors is essential to understanding the results provided, and will enable users to maximize the value of the model. This presentation will provide a summary of pitfalls to avoid, and guidance for designing an effective validation.