Applying one-class learning algorithms to predict phage-bacteria interactions

Abstract

The need to predict phage-bacteria interactions is a nowadays concern to overcome bacterial resistance issue; public genome databases contain highly imbalanced datasets which have hindered this task. Throughout this paper we will investigate, implement and evaluate One-Class Learning algorithms in order to predict phage-bacteria interactions using only positive samples. We will use the programming language Python aided by Scikit-Learn, Tensorflow and keras to develop the machine learning models and test them with real phagebacteria interactions datasets. We trained the models using cross validation technique generating a gridsearch with all the datasets to find several combinations of hyperparameters available. Furthermore, we optimized those hyperparameters by using Pareto fronts based on seven different performance metrics, improving the efficiency of each algorithm for a given dataset. To refine each algorithm's performance separately we used the ensemble learning technique with an odd number of algorithms by simple voting. Finally, we managed to achieve an overall performance of 80% in predicting phage-bacteria interactions trained only with positive classes, this percentage in practice means that when a patient has an infection resistant to antibiotics, we have 80% of saving the life rather than maybe a 0% while finding the correct phage for the pathogenic host.

Learning Peptide Properties with Positive Examples Only

“…The positive examples are significantly outnumbered by a large number of protein pairs for which no interactions have been identified. Similar situations can be found in drug-drug interaction identification [33], small non-coding RNA detection [34], gene function [35,36] and phage-bacteria interaction [37] prediction, and biological sequence classification [38,39].…”

Learning peptide properties with positive examples only

“…(2) The positive examples are significantly outnumbered by a large number of protein pairs for which no interactions have been identified. Similar situations can be found in drug–drug interaction identification, 33 small non-coding RNA detection, 34 gene function 35,36 and phage–bacteria interaction 37 prediction, and biological sequence classification. 38,39…”

“…(2) The positive examples are signicantly outnumbered by a large number of protein pairs for which no interactions have been identied. Similar situations can be found in drug-drug interaction identication, 33 small non-coding RNA detection, 34 gene function 35,36 and phage-bacteria interaction 37 prediction, and biological sequence classication. 38,39 To address the challenges above, we demonstrate on a positive-unlabeled (PU) learning framework to infer peptide sequence-function relationships, by solely exploiting the limited known positive examples in a semi-supervised setting.…”

Prediction of virus-host associations using protein language models and multiple instance learning

“…As we know, there is no information on negative associations in the public datasets. Most phages are likely to infect a range of hosts belonging to the same species [18,19], therefore, in order to mitigate possible errors, some researchers [20,21] construct negative virus-host associations by selecting those viruses from the remaining viruses that do not infect the given host, instead of relating to hosts with different taxa from the given host. Inspired by this method, we construct putative negative virus-host associations by collecting viruses from a specific range.…”