A novel one-class SVM based negative data sampling method for reconstructing proteome-wide HTLV-human protein interaction networks

Abstract

Protein-protein interaction (PPI) prediction is generally treated as a problem of binary classification wherein negative data sampling is still an open problem to be addressed. The commonly used random sampling is prone to yield less representative negative data with considerable false negatives. Meanwhile rational constraints are seldom exerted on model selection to reduce the risk of false positive predictions for most of the existing computational methods. In this work, we propose a novel negative data sampling method based on one-class SVM (support vector machine, SVM) to predict proteome-wide protein interactions between HTLV retrovirus and Homo sapiens, wherein one-class SVM is used to choose reliable and representative negative data, and two-class SVM is used to yield proteome-wide outcomes as predictive feedback for rational model selection. Computational results suggest that one-class SVM is more suited to be used as negative data sampling method than two-class PPI predictor, and the predictive feedback constrained model selection helps to yield a rational predictive model that reduces the risk of false positive predictions. Some predictions have been validated by the recent literature. Lastly, gene ontology based clustering of the predicted PPI networks is conducted to provide valuable cues for the pathogenesis of HTLV retrovirus.

Figure 1. One-class SVM OCSVM_{(ν, γ)} parameters tuning.

Eight representative parameter pairs (ν, γ) are chosen from the parameter space according to OCSVM_{(ν, γ)} LOOCV performance. The blue bars illustrate the recognition rate of the training positive data and the brown bars illustrate the predicted positive rate of proteome-wide predictions by OCSVM_{(ν, γ)}.

Figure 2. Two-class SVM TCSVM_{(ν, γ)} LOOCV ROC curves.

For each representative parameter pair (ν, γ), one negative dataset is sampled from the predicted outcomes of the correponding OCSVM_{(ν, γ)}. The sampled negative data are merged with the positive data to train two-class SVM TCSVM_{(ν, γ)}. The ROC curves and corresponding AUC scores are used to estimate the quality of the negative data sampled by OCSVM_{(ν, γ)}.

Figure 3. Two-class SVM TCSVM_{(ν, γ)} LOOCV performance on dataset S, S1, S2 and S3.

For each representative parameter pair (ν, γ), negative datasets are sampled from the predicted outcomes of the correponding OCSVM_{(ν, γ)} to be the negative data of dataset S, S1, S2 and S3, and then train four two-class SVM TCSVM_{(ν, γ)}. The blue bars denote TCSVM_{(ν, γ)} LOOCV Accuracy and the brown bars denote TCSVM_{(ν, γ)} LOOCV MCC.

Figure 4. Two-class SVM TCSVM_{(ν, γ)} proteome-wide predicted positive rates.

From the predicted posisitve rates, K values derived are derived to be used as constraint on OCSVM_{(ν, γ)} model selection. Lower bar signifies higher K value.

Figure 5. Percentage of HTLV targeted human proteins predicted by two-class SVM TCSVM_{(ν, γ)}.

Lower bars are supposed to signify lower risk of false positive predictions. The metric together with K value is used as constraint on model selection of one-class SVM OCSVM_{(ν, γ)}.

Figure 6. Details of percentage of HTLV targeted human proteins predicted by TCSVM_{(ν, γ)}.

From the metric, K value can be derived for each HTLV protein to conduct fine-grained model selection of one-class SVM OCSVM_{(ν, γ)}. The parameter pair (ν, γ) with more lower bars are preferred.

Figure 7. Comparative ROC curves between the final model TCSVM_{(ν = 1, γ = 3)} and random sampling model TCSVM_random.

From the points of view of AUC scores, TCSVM_{(ν = 1, γ = 3)} outperforms TCSVM_random.

Figure 8. Gene ontology based clustering of predicted PPI subnetworks - biological processes.

Three human signaling pathways predicted to be targeted by HTLV proteins are illustrated as examples: GO:0007219 - Notch signaling pathway. GO:0050852 - T cell receptor signaling pathway. GO:0046426 - negative regulation of JAK-STAT cascade. The diamond denotes HTLV proteins and the ecllipse circle denotes human proteins.

Figure 9. Gene ontology based clustering of predicted PPI subnetworks - molecular functions.

Three human Molecular functional modules predicted to be targeted by HTLV proteins are illustrated as examples: GO:0017124 - SH3 domain binding. GO:0002039 - p53 binding. GO:0004553 - hydrolase activity. The diamond denotes HTLV proteins and the ecllipse circle denotes human proteins.