Cost-Sensitive Classifier Evaluation Using Cost Curves

Holte, Robert C.; Drummond, Chris

doi:10.1007/978-3-540-68125-0_4

Cited by 8 publications

(3 citation statements)

References 25 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In order to determine whether a scaling factor ν different from 1 would improve the discrimination of ID segments from structured proteins, we constructed ROC curves at different scaling factors. We determined an optimal cutoff value at each scaling factor by selecting the threshold with the highest Matthew's correlation coefficient (MCC) [37], [38]. Out of all the scaling factors that we tested, ν of 1 results in a classifier that has the highest AUC (Table S1).…”

Section: Resultsmentioning

confidence: 99%

On the Importance of Polar Interactions for Complexes Containing Intrinsically Disordered Proteins

Wong

Gsponer

2013

PLoS Comput Biol

View full text Add to dashboard Cite

There is a growing recognition for the importance of proteins with large intrinsically disordered (ID) segments in cell signaling and regulation. ID segments in these proteins often harbor regions that mediate molecular recognition. Coupled folding and binding of the recognition regions has been proposed to confer high specificity to interactions involving ID segments. However, researchers recently questioned the origin of the interaction specificity of ID proteins because of the overrepresentation of hydrophobic residues in their interaction interfaces. Here, we focused on the role of polar and charged residues in interactions mediated by ID segments. Making use of the extended nature of most ID segments when in complex with globular proteins, we first identified large numbers of complexes between globular proteins and ID segments by using radius-of-gyration-based selection criteria. Consistent with previous studies, we found the interfaces of these complexes to be enriched in hydrophobic residues, and that these residues contribute significantly to the stability of the interaction interface. However, our analyses also show that polar interactions play a larger role in these complexes than in structured protein complexes. Computational alanine scanning and salt-bridge analysis indicate that interfaces in ID complexes are highly complementary with respect to electrostatics, more so than interfaces of globular proteins. Follow-up calculations of the electrostatic contributions to the free energy of binding uncovered significantly stronger Coulombic interactions in complexes harbouring ID segments than in structured protein complexes. However, they are counter-balanced by even higher polar-desolvation penalties. We propose that polar interactions are a key contributing factor to the observed high specificity of ID segment-mediated interactions.

show abstract

Section: Resultsmentioning

confidence: 99%

On the Importance of Polar Interactions for Complexes Containing Intrinsically Disordered Proteins

Wong

Gsponer

2013

PLoS Comput Biol

View full text Add to dashboard Cite

show abstract

“…Use of well-known discretization methods can cause information loss and hypotheses missing in the version space construction of a numeric data set. These discretization methods include the entropy-based discretization method [8], [9], [10], Gini-Index based method [11], Chisquare based method [12], [13], 1-rule method [14], [15], and unsupervised equal-length or equal-density methods. In fact, use of these discretization methods makes almost every version space empty as investigated by this work.…”

Section: Examplementioning

confidence: 99%

Version Space Completeness for Novel Hypothesis Induction in Biomedical Applications

2018

2018 International Joint Conference on Neural Networks (IJCNN)

View full text Add to dashboard Cite

Use of traditional discretization methods caused a heavy loss of hypotheses in the induction of version spaces. We present a new discretization method, named two-point discretization, to construct an interval covering all the positive data points of a variable as purely as possible. We prove that the two-point discretization is a necessary and sufficient condition to guarantee the completeness of version spaces (i.e., no loss of hypothesis). A linear complexity algorithm is proposed to implement these theories. The algorithm is also applied to real-world bioinformatics problems to induce significant biomedical hypotheses which have been never discovered by the traditional approaches.

show abstract

“…In other words, both classifiers perform equally well if the cost of classifying benign and malignant tumors is kept the same. However, if we would like to change the cost of classifying benign and malignant tumors, for example, we decided to give more cost for missing malignant tumors than missing benign tumors then both classifiers perform differently (see Holte & Drummond (2011)). The later observation explains why the SVM and Parzen classifier have an overlapping performance which is easy to explain from the ROC curves.…”

Section: Analyzing the Classifiers Performancementioning

confidence: 99%