High-quality Training Set Research Articles

A hybrid ensemble learning approach to star–galaxy classification

There exist a variety of star-galaxy classification techniques, each with their own strengths and weaknesses. In this paper, we present a novel meta-classification framework that combines and fully exploits different techniques to produce a more robust star-galaxy classification. To demonstrate this hybrid, ensemble approach, we combine a purely morphological classifier, a supervised machine learning method based on random forest, an unsupervised machine learning method based on self-organizing maps, and a hierarchical Bayesian template fitting method. Using data from the CFHTLenS survey, we consider different scenarios: when a high-quality training set is available with spectroscopic labels from DEEP2, SDSS, VIPERS, and VVDS, and when the demographics of sources in a low-quality training set do not match the demographics of objects in the test data set. We demonstrate that our Bayesian combination technique improves the overall performance over any individual classification method in these scenarios. Thus, strategies that combine the predictions of different classifiers may prove to be optimal in currently ongoing and forthcoming photometric surveys, such as the Dark Energy Survey and the Large Synoptic Survey Telescope.

A novel empirical free energy function that explains and predicts protein–protein binding affinities

A free energy function can be defined as a mathematical expression that relates macroscopic free energy changes to microscopic or molecular properties. Free energy functions can be used to explain and predict the affinity of a ligand for a protein and to score and discriminate between native and non-native binding modes. However, there is a natural tension between developing a function fast enough to solve the scoring problem but rigorous enough to explain and predict binding affinities. Here, we present a novel, physics-based free energy function that is computationally inexpensive, yet explanatory and predictive. The function results from a derivation that assumes the cost of polar desolvation can be ignored and that includes a unique and implicit treatment of interfacial water-bridged interactions. The function was parameterized on an internally consistent, high quality training set giving R 2 = 0.97 and Q 2 = 0.91. We used the function to blindly and successfully predict binding affinities for a diverse test set of 31 wild-type protein–protein and protein–peptide complexes ( R 2 = 0.79, rmsd = 1.2 kcal mol − 1 ). The function performed very well in direct comparison with a recently described knowledge-based potential and the function appears to be transferable. Our results indicate that our function is well suited for solving a wide range of protein/peptide design and discovery problems.

A computationally based identification algorithm for estrogen receptor ligands: part 2. Evaluation of a hERalpha binding affinity model.

The objective of this study was to evaluate the capability of an expert system described in the previous paper (S. Bradbury et al., Toxicol. Sci. 58, 253-269) to identify the potential for chemicals to act as ligands of mammalian estrogen receptors (ERs). The basis of the expert system was a structure activity relationship (SAR) model, based on relative binding affinity (RBA) values for steroidal and nonsteroidal chemicals derived from human ERalpha (hERalpha) competitive binding assays. The expert system enables categorization of chemicals into (RBA ranges of < 0.1, 0.1 to 1, 1 to 10, 10 to 100, and >150% relative to 17ss-estradiol. In the current analysis, the algorithm was evaluated with respect to predicting RBAs of chemicals assayed with ERs from MCF7 cells, and mouse and rat uterine preparations. The best correspondence between predicted and observed RBA ranges was obtained with MCF7 cells. The agreement between predictions from the expert system and data from binding assays with mouse and rat ER(s) were less reliable, especially for chemicals with RBAs less than 10%. Prediction errors often were false positives, i.e., predictions of greater than observed RBA values. While discrepancies were likely due, in part, to species-specific variations in ER structure and ligand binding affinity, a systematic bias in structural characteristics of chemicals in the hERalpha training set, compared to the rodent evaluation data sets, also contributed to prediction errors. False-positive predictions were typically associated with ligands that had shielded electronegative sites. Ligands with these structural characteristics were not well represented in the training set used to derive the expert system. Inclusion of a shielding criterion into the original expert system significantly increased the accuracy of RBA predictions. With this additional structural requirement, 38 of 46 compounds with measured RBA values greater than 10% in hERalpha, MCF7, and rodent uterine preparations were correctly categorized. Of the remaining 129 compounds in the combined data sets, RBA values for 65 compounds were correctly predicted, with 47 of the incorrect predictions being false positives. Based upon this exploratory analysis, the modeling approach, combined with a high-quality training set of RBA values derived from a diverse set of chemical structures, could provide a credible tool for prioritizing chemicals with moderate to high ER binding affinity for subsequent in vitro or in vivo assessments.