Multitask learning for host–pathogen protein interactions

Kshirsagar, Meghana; Carbonell, Jaime; Klein-Seetharaman, Judith

We followed an identical procedure for all algorithms. For the 10-fold CV experiments we train on 8-folds, use 1-fold as heldout and another as test. The optimal parameters (i.e. the best model) were obtained by parameter tuning on the held-out fold. The test fold was used to evaluate this best model-these results are reported in Section 7. The range of values we tried during the tuning of the regularization parameter () were 150-10 À4 . For , the parameter controlling overfitting in multitask pathwaybased learning (MTPL), we used a fixed value of ¼ 1. For Mean MTL, we tune both and . To handle the high-class imbalance in our data, we used a weight-parameter W pos to increase the weight of the positive examples in the logistic loss terms of our function. We tried three values and found W pos ¼ 100 performed the best on training data. Table 3 reports for each bacterial species, the average F1 along with the standard deviation for the 10-fold CV experiments. The performance of all baselines is similar, and our method outperforms the best of the baselines by a margin of 4 points for B.anthracis, 3.4 points for F.tularensis and 3.2 points for Y.pestis and 3.3 for S.typhi. The overall performance of all methods on this dataset is twice as good as that on the others. We believe that the difference in the nature of the datasets might explain the above observations. While the S.typhi dataset comprises small-scale interaction studies, the other datasets come from high-throughput experiments. Owing to its smaller size, it has less variance making it an easier task. This dataset is also likely to be a biased sample of interactions, as it comes from focussed studies targeting select proteins. The coupled learner (Coupled) performs slightly worse than Indep. This is explained by the fact that Indep. has more flexibility in setting the regularization parameter for each task separately, which is not the case in Coupled. It is interesting to note that the independent models that use the pathway matrices P s and P t as features (i.e. Indep-Path) show a slightly worse performance than the Indep. models that do not use them. This seems to suggest that the cross-task pathway similarity structure that we enforce using our regularizer has more information than simply the pathway membership of proteins used as features. Precision-Recall curves: We also plot the P-R curves for MTPL. Please see the Supplementary Section 3.