Motivation: Large-scale biological analyses produce huge amounts of data. As a consequence, automation in the data analysis process is needed. Sample screening problems in NMR high-throughput protein structure analysis are the typical examples. Especially, screening by protein (1)H-(15)N heteronuclear single quantum coherence (HSQC) spectra must be done quantitatively by a human expert. One popular solution for this problem is data mining. Machine learning methods can automatically extract rules and achieve high accuracy in prediction when a good quality training dataset is prepared. However, they tend to be a black box and the learned machines suffer the risk of overfitting to the dataset.
Results: We propose a model which evaluates HSQC spectra for feature construction. The model calculates similarity between the measured chemical shifts and those of a random coil peak model. We applied our feature construction method for the machine learning discrimination of folded protein HSQC spectra from unfolded ones, and compared our model-based features with those of conventional sequence-based features and image recognition features. The results revealed that our method has sufficient discrimination power and less overfits on training data, as compared to the other methods. In addition, our method succeeded reduction of input data complexity towards further investigation.
Supplementary information: Supplementary data are available at Bioinformatics online.