Róbert Busa-Fekete

Bacterial evolution of antibiotic hypersensitivity

The evolution of resistance to a single antibiotic is frequently accompanied by increased resistance to multiple other antimicrobial agents. In sharp contrast, very little is known about the frequency and mechanisms underlying collateral sensitivity. In this case, genetic adaptation under antibiotic stress yields enhanced sensitivity to other antibiotics. Using large‐scale laboratory evolutionary experiments with Escherichia coli, we demonstrate that collateral sensitivity occurs frequently during the evolution of antibiotic resistance. Specifically, populations adapted to aminoglycosides have an especially low fitness in the presence of several other antibiotics. Whole‐genome sequencing of laboratory‐evolved strains revealed multiple mechanisms underlying aminoglycoside resistance, including a reduction in the proton‐motive force (PMF) across the inner membrane. We propose that as a side effect, these mutations diminish the activity of PMF‐dependent major efflux pumps (including the AcrAB transporter), leading to hypersensitivity to several other antibiotics. More generally, our work offers an insight into the mechanisms that drive the evolution of negative trade‐offs under antibiotic selection.

Boosting products of base classifiers

In this paper we show how to boost products of simple base learners. Similarly to trees, we call the base learner as a subroutine but in an iterative rather than recursive fashion. The main advantage of the proposed method is its simplicity and computational efficiency. On benchmark datasets, our boosted products of decision stumps clearly outperform boosted trees, and on the MNIST dataset the algorithm achieves the second best result among no-domain-knowledge algorithms after deep belief nets. As a second contribution, we present an improved base learner for nominal features and show that boosting the product of two of these new subset indicator base learners solves the maximum margin matrix factorization problem used to formalize the collaborative filtering task. On a small benchmark dataset, we get experimental results comparable to the semi-definite-programming-based solution but at a much lower computational cost.

Genome-wide analysis captures the determinants of the antibiotic cross-resistance interaction network

Understanding how evolution of antimicrobial resistance increases resistance to other drugs is a challenge of profound importance. By combining experimental evolution and genome sequencing of 63 laboratory-evolved lines, we charted a map of cross-resistance interactions between antibiotics in Escherichia coli, and explored the driving evolutionary principles. Here, we show that (1) convergent molecular evolution is prevalent across antibiotic treatments, (2) resistance conferring mutations simultaneously enhance sensitivity to many other drugs and (3) 27% of the accumulated mutations generate proteins with compromised activities, suggesting that antibiotic adaptation can partly be achieved without gain of novel function. By using knowledge on antibiotic properties, we examined the determinants of cross-resistance and identified chemogenomic profile similarity between antibiotics as the strongest predictor. In contrast, cross-resistance between two antibiotics is independent of whether they show synergistic effects in combination. These results have important implications on the development of novel antimicrobial strategies.

Antagonism between Bacteriostatic and Bactericidal Antibiotics Is Prevalent

ABSTRACT Combination therapy is rarely used to counter the evolution of resistance in bacterial infections. Expansion of the use of combination therapy requires knowledge of how drugs interact at inhibitory concentrations. More than 50 years ago, it was noted that, if bactericidal drugs are most potent with actively dividing cells, then the inhibition of growth induced by a bacteriostatic drug should result in an overall reduction of efficacy when the drug is used in combination with a bactericidal drug. Our goal here was to investigate this hypothesis systematically. We first constructed time-kill curves using five different antibiotics at clinically relevant concentrations, and we observed antagonism between bactericidal and bacteriostatic drugs. We extended our investigation by performing a screen of pairwise combinations of 21 different antibiotics at subinhibitory concentrations, and we found that strong antagonistic interactions were enriched significantly among combinations of bacteriostatic and bactericidal drugs. Finally, since our hypothesis relies on phenotypic effects produced by different drug classes, we recreated these experiments in a microfluidic device and performed time-lapse microscopy to directly observe and quantify the growth and division of individual cells with controlled antibiotic concentrations. While our single-cell observations supported the antagonism between bacteriostatic and bactericidal drugs, they revealed an unexpected variety of cellular responses to antagonistic drug combinations, suggesting that multiple mechanisms underlie the interactions.

Counter-Example Generation-Based One-Class Classification

For One-Class Classification problems several methods have been proposed in the literature. These methods all have the common feature that the decision boundary is learnt by just using a set of the positive examples. Here we propose a method that extends the training set with a counter-example set, which is generated directly using the set of positive examples. Using the extended training set, a binary classifier (here i¾?-SVM) is applied to separate the positive and the negative points. The results of this novel technique are compared with those of One-Class SVM and the Gaussian Mixture Model on several One-Class Classification tasks.

A Survey of Preference-Based Online Learning with Bandit Algorithms

In machine learning, the notion of multi-armed bandits refers to a class of online learning problems, in which an agent is supposed to simultaneously explore and exploit a given set of choice alternatives in the course of a sequential decision process. In the standard setting, the agent learns from stochastic feedback in the form of real-valued rewards. In many applications, however, numerical reward signals are not readily available—instead, only weaker information is provided, in particular relative preferences in the form of qualitative comparisons between pairs of alternatives. This observation has motivated the study of variants of the multi-armed bandit problem, in which more general representations are used both for the type of feedback to learn from and the target of prediction. The aim of this paper is to provide a survey of the state-of-the-art in this field, that we refer to as preference-based multi-armed bandits. To this end, we provide an overview of problems that have been considered in the literature as well as methods for tackling them. Our systematization is mainly based on the assumptions made by these methods about the data-generating process and, related to this, the properties of the preference-based feedback.

Tune and mix: learning to rank using ensembles of calibrated multi-class classifiers

In subset ranking, the goal is to learn a ranking function that approximates a gold standard partial ordering of a set of objects (in our case, a set of documents retrieved for the same query). The partial ordering is given by relevance labels representing the relevance of documents with respect to the query on an absolute scale. Our approach consists of three simple steps. First, we train standard multi-class classifiers (AdaBoost.MH and multi-class SVM) to discriminate between the relevance labels. Second, the posteriors of multi-class classifiers are calibrated using probabilistic and regression losses in order to estimate the Bayes-scoring function which optimizes the Normalized Discounted Cumulative Gain (NDCG). In the third step, instead of selecting the best multi-class hyperparameters and the best calibration, we mix all the learned models in a simple ensemble scheme.Our extensive experimental study is itself a substantial contribution. We compare most of the existing learning-to-rank techniques on all of the available large-scale benchmark data sets using a standardized implementation of the NDCG score. We show that our approach is competitive with conceptually more complex listwise and pairwise methods, and clearly outperforms them as the data size grows. As a technical contribution, we clarify some of the confusing results related to the ambiguities of the evaluation tools, and propose guidelines for future studies.

GraphClus, a MATLAB program for cluster analysis using graph theory

Cluster analysis is used in numerous scientific disciplines. A method of cluster analysis based on graph theory is discussed and a MATLAB(TM) code for its implementation is presented. The algorithm is based on the number of variables that are similar between samples. By changing the similarity criterion in a stepwise fashion, a hierarchical group structure develops, and can be displayed by a dendrogram. Three indexes describe the homogeneity of a given variable in a group, the heterogeneity of that variable between two groups, and the usefulness of that variable in distinguishing two groups. The algorithm is applied to both a synthetic dataset and a set of trace element analyses of lavas from Mount Etna in order to compare GraphClus to other cluster analysis algorithms.

Tree-Based Algorithms for Protein Classification

The problem of protein sequence classification is one of the crucial tasks in the interpretation of genomic data. Many high-throughput systems were developed with the aim of categorizing the proteins based only on their sequences. However, modelling how the proteins have evolved can also help in the classification task of sequenced data. Hence the phylo-genetic analysis has gained importance in the field of protein classification. This approach does not just rely on the similarities in sequences, but it also considers the phylogenetic information stored in a tree (e.g. in a phylogenetic tree). Eisen used firstly phylogenetic trees in protein classification, and his work has revived the discipline of phylogenomics. In this chapter we provide an overview about this area, and in addition we propose two algorithms that well suited to this scope. We present two algorithms that are based on a weighted binary tree representation of protein similarity data. TreeInsert assigns the class label to the query by determining a minimum cost necessary to insert the query in the (precomputed) trees representing the various classes. Then TreNN assigns the label to the query based on an analysis of the query’s neighborhood within a binary tree containing members of the known classes. The algorithms were tested in combination with various sequence similarity scoring methods (BLAST, Smith-Waterman, Local Alignment Kernel as well as various compression-based distance scores) using a large number of classification tasks representing various degrees of difficulty. At the expense of a small computational overhead, both TreeNN and TreeInsert exceed the performance of simple similarity search (1NN) as determined by ROC analysis, at the expense of a modest computational overhead. Combined with a fast tree-building method, both algorithms are suitable for web-based server applications.

Preference-based reinforcement learning: evolutionary direct policy search using a preference-based racing algorithm

We introduce a novel approach to preference-based reinforcement learning, namely a preference-based variant of a direct policy search method based on evolutionary optimization. The core of our approach is a preference-based racing algorithm that selects the best among a given set of candidate policies with high probability. To this end, the algorithm operates on a suitable ordinal preference structure and only uses pairwise comparisons between sample rollouts of the policies. Embedding the racing algorithm in a rank-based evolutionary search procedure, we show that approximations of the so-called Smith set of optimal policies can be produced with certain theoretical guarantees. Apart from a formal performance and complexity analysis, we present first experimental studies showing that our approach performs well in practice.