BACKGROUND: Machine learning and deep learning are powerful tools for analyzing electronic health records (EHRs) in healthcare research. Although family health history has been recognized as a major predictor for a wide spectrum of diseases, research has so far adopted a limited view of family relations, essentially treating patients as independent samples in the analysis. METHODS: To address this gap, we present ALIGATEHR, which models inferred family relations in a graph attention network augmented with an attention-based medical ontology representation, thus accounting for the complex influence of genetics, shared environmental exposures, and disease dependencies. RESULTS: Taking disease risk prediction as a use case, we demonstrate that explicitly modeling family relations significantly improves predictions across the disease spectrum. We then show how ALIGATEHR's attention mechanism, which links patients' disease risk to their relatives' clinical profiles, successfully captures genetic aspects of diseases using longitudinal EHR diagnosis data. Finally, we use ALIGATEHR to successfully distinguish the 2 main inflammatory bowel disease subtypes with highly shared risk factors and symptoms (Crohn's disease and ulcerative colitis). CONCLUSION: Overall, our results highlight that family relations should not be overlooked in EHR research and illustrate ALIGATEHR's great potential for enhancing patient representation learning for predictive and interpretable modeling of EHRs.

Abstract Summary: Mixed model-based approaches to genome-wide association studies (GWAS) of binary traits in related individuals can account for non-genetic risk factors in an integrated manner. However, they are technically challenging. GLOGS (Genome-wide LOGistic mixed model/Score test) addresses such challenges with efficient statistical procedures and a parallel implementation. GLOGS has high power relative to alternative approaches as risk covariate effects increase, and can complete a GWAS in minutes. Availability: Source code and documentation are provided at http://www.bioinformatics.org/~stanhope/GLOGS. Contact: [email protected] Supplementary information: Supplementary data are available at Bioinformatics online.

MicroRNAs (miRNAs) posttranscriptionally regulate targeted messenger RNAs (mRNAs) by inducing cleavage or otherwise repressing their translation. We address the problem of detecting m/miRNA targeting relationships in homo sapiens from microarray data by developing statistical models that are motivated by the biological mechanisms used by miRNAs. The focus of our modeling is the construction, activity, and mediation of RNA-induced silencing complexes (RISCs) competent for targeted mRNA cleavage. We demonstrate that regression models accommodating RISC abundance and controlling for other mediating factors fit the expression profiles of known target pairs substantially better than models based on m/miRNA expressions alone, and lead to verifications of computational target pair predictions that are more sensitive than those based on marginal expression levels. Because our models are fully independent of exogenous results from sequence-based computational methods, they are appropriate for use as either a primary or secondary source of information regarding m/miRNA target pair relationships, especially in conjunction with high-throughput expression studies.

Multivariate-to-anything methods for partially specified random vector generation work by transforming samples from a driving distribution into samples characterized by given marginals and correlations. The correlations of the transformed random vector are controlled by the driving distribution; sampling a partially specified random vector requires finding an appropriate driving distribution. This paper motivates the use of evolution strategies for solving such problems and compares evolution strategies to conjugate gradient methods in the context of solving a Dirichlet-to-anything transformation. It is shown that the evolution strategy is at least as effective as the conjugate gradient method for solution of the parameterization problem.

We analyze the fitness dynamics of a (1+1) mutation-only genetic algorithm (GA) operating on a family of simple time-dependent fitness functions. Resulting models of behavior are used in the prediction of GA performance on this fitness function. The accuracy of performance predictions are compared to actual GA runs, and results are discussed in relation to analyses of the stationary version of the dynamic fitness landscape and to prior work performed in the field of evolutionary optimization of dynamic fitness functions.