The minimum inhibitory concentration (MIC) is the lowest concentration of an antimicrobial agent (commonly known as an antibiotic) that inhibits bacterial growth. The interpretation of an MIC value for a bacterial organism, recovered from an infection guides the treatment of that infection by informing the physician if the treatment has a chance to be successful or not. Methods to generate the MIC values depend on bacterial growth and could take 24 hours to several days, delaying the implementation of appropriate therapy and increasing rates of morbidity and mortality. In this study, XGBoost and Neural Network (NN) models were used to predict the MIC values for 13 antimicrobial agents in parallel. The XGBoost model was then used for feature selection, and those features were used as input to the NN. The 29 selected features from the XGBoost model gave an F1 score of 0.808 when applied to a NN model. That NN model has a +-1 2-fold dilution accuracy of ׅ89 for all antibiotics. The results prove that the XGBoost model is capable of selecting important features from the dataset and those selected features can be used to generate a highly effective NN model that can predict MIC results.

Since the beginning of antimicrobial therapy, antimicrobial resistance from bacteria has been a threat. Now, multiple companies track antibiotic resistance annually. Monitoring helps watch for new resistant mechanisms to restrict spread. This monitoring also determines if a new antimicrobial agent must be created. Proper prescription of an antimicrobial agent also plays a key role in slowing resistance. Proper prescription to a patient requires the determination of a Minimum Inhibitory Concentration (MIC). These MICs are slow to determine, which raises morbidity and mortality rates in hospital settings. Physicians cannot predict the creation of new resistance mechanisms based on the prescription of one agent over another. It is also impossible for researchers to know what agents to create in the future. Because of this, there is a race against the clock when new resistant bacteria emerge, to create appropriate antimicrobial agents to counter the new mechanisms of resistance. In this study, we propose a new simulation algorithm named Bacterial Antibiotic Adversarial Genetic Algorithm (BAAGA) that simulates the generation of new mechanisms of resistance and new antimicrobial agents in a stochastic manner. BAAGA shows the ease at which bacteria can gain resistance mechanisms and how slow the process is to find new antimicrobial agents. We also demonstrate how BAAGA can produce possible new antimicrobial agents to counter mechanisms of resistance that currently do not exist.

Antimicrobial Susceptibility Testing (AST) is necessary to properly identify if a pathogen is susceptible or resistant to a given antimicrobial agent. However, current methods utilizing Minimum Inhibitory Concentration (MIC) testing are slow, leading to delay in implementing effective antimicrobial therapy. Machine Learning (ML) offers a faster alternative. This paper proposes a novel way to predict MIC values by fingerprinting isolates using three different K-Mer count sizes, rather than using a single, large K-mer, of three, four, and five. Results from our models achieved an average F1 score of 0.75 in comparison to other ML models that achieved 0.67 and 0.58 F1 scores by previous approaches.

Antibiotic resistance in bacteria is being recognized as a global health threat. Patients with infections caused by antibiotic resistant bacteria have higher mortality rates and use more hospital resources. The diagnostic of infections caused by antibiotic resistant bacteria is cumbersome and requires bacterial isolation, identification of the species, followed by antibiotic susceptibility testing. This process can take up to 72h. DNA sequencing methods are becoming more convenient and affordable, and the presence of different genes can help predict the susceptibility results of bacteria against antibiotics. We aim to use Machine Learning (ML) to predict the Minimum Inhibitory Concentration (MIC) for bacterial isolates from resistance genes present in their DNA. To do this, we will use information about the presence of beta-lactamase genes identified by whole genome sequencing and its correlation with the susceptibility results (MIC values) for beta-lactam antibiotics. The data was collected worldwide during the years 2016, 2017, and 2018. Species analyzed were Escherichia coli and Klebsiella pneumoniae. All models were trained on different datasets, but all predicted for 2018 data. The training data was split by separator of either year, species, or continent. The species and continent separators had 2016 and 2017 data combined while the year separator had the training datasets of 2016 data, 2017 data, or 2016 and 2017 combined data. It is found that K-Nearest Neighbors performs similarly to Random Forest with ± 12- fold dilution MIC predictions. Random Forests had higher F1-micro scores when predicting the exact MIC value. Training data separated by year outperforms training data separated by continent or species.

Antibiotic resistant bacteria are an ever-present threat in today's society and matter of great concern for the medical and scientific communities. With the high frequency of mutations, and the ability to transfer genes across species, bacteria can acquire resistance to currently used antibiotics. Identifying resistance to antibiotics requires growing the bacteria in various antibiotic concentrations to determine the Minimum Inhibitory Concentration (MIC) value for that bacteria-antibiotic combination which is a longer than two day process. Additionally, the testing to evaluate genetic resistance determinants is cumbersome and demands considerable expertise. This paper looks at Neural Networks (NN) and decision trees as viable options to predict the resistance category for an antibiotic, bacteria combination based in the beta-lactamases present in that organism. We show that decision trees have just as high accuracy as NN, and provide more detail and incite in how a decision was made.

Particle Swarm Optimization is an example of a trivially parallelizable algorithm where good performance gains can be achieved through the use of a few OpenMP pragmas. Writing an efficient parallel PSO algorithm, however, is much more challenging because although particle updates can occur independently, they rely on a shared global state (the globally best particle). The difficulty of maintaining this global state can be seen in the large body of work studying the parallelization of PSO - almost uniformly these algorithms rely on a global synchronization step, which can result in idle cores and reduced parallel efficiency. In this work, we explore two techniques for implementing a fast, cache-aware parallel PSO algorithm: batching the creation of the random weights and reducing critical section contention via a relaxed consistency guarantee. Our technique shows impressive performance improvements over prior work, seeing more than 60% speed-up over naive parallelization and more than 10% speed-up over the cache-aware algorithm. This speed comes at a cost; while our method quickly reaches an approximate solution, it struggles in environments requiring a high level of resolution. Despite these trade-offs, our method is both easy to understand and implement and is widely transferable to other swarm intelligence algorithms.

One of the most used methodologies in the field of education assessment is Item Response Theory (IRT). In this work, we propose the use of a novel Variational Autoencoder (VAE) architecture for a multidimensional IRT model. Our approach combines the advantages of the IRT model while allowing us to model high latent trait dimensions, previously unattainable in prior work. Additionally, it has the advantage of interpretability in the domain of educational assessment.Our experiments show that, given enough data, the new model is competitive with the state-of-the-art methods with respect to predictive power and is much faster in runtime performance. In our experiments, we achieve competitive results on a sample size 20× larger in a runtime that is 40× faster than the state-of-the- art model.

Basis Pursuit DeNoising (BPDN) is a version of sparse least squares, where there is an ℓ1penalty for the coefficients in addition to the sum of squares of the errors. To apply this technique to large data sets, parallel algorithms are needed. In this paper, we create a parallel version of a BPDN algorithm [11] that can compute exact solutions for an interval of values for the ℓ1penalty parameter that exploits the sparsity of the solutions, unlike stochastic gradient descent-type algorithms.

For a collection of sets in R d we consider the task of finding all sets in the collection that contain a given point. The parallel algorithm introduced in this paper is based on quadtrees and their generalizations to R d . Our algorithm solves a difficult problem faced by meshfree discretizations, and is based on the sequential algorithm of Han, Oliveira, and Stewart [3].

For a collection of sets in R d we consider the task of finding all sets in the collection that cover or contain a given point. The algorithms introduced in this paper are based on quadtrees and their generalizations to R d . The advantages of our new splitting algorithm to find the covering sets of a point over the basic algorithm are detailed by means of hit rates and the expected depth traversed in the quadtree search, numerically and theoretically. This solves a difficult problem faced by mesh-free discretizations of partial differential equations.