Optimization of Genomic Classifiers for Clinical Deployment: Evaluation of Bayesian Optimization to Select Predictive Models of Acute Infection and In-Hospital Mortality.

Acute infection, if not rapidly and accurately detected, can lead to sepsis, organ failure and even death. Current detection of acute infection as well as assessment of a patient's severity of illness are imperfect. Characterization of a patient's immune response by quantifying expression levels of specific genes from blood represents a potentially more timely and precise means of accomplishing both tasks. Machine learning methods provide a platform to leverage this host response for development of deployment-ready classification models. Prioritization of promising classifiers is dependent, in part, on hyperparameter optimization for which a number of approaches including grid search, random sampling and Bayesian optimization have been shown to be effective. We compare HO approaches for the development of diagnostic classifiers of acute infection and in-hospital mortality from gene expression of 29 diagnostic markers. We take a deployment-centered approach to our comprehensive analysis, accounting for heterogeneity in our multi-study patient cohort with our choices of dataset partitioning and hyperparameter optimization objective as well as assessing selected classifiers in external (as well as internal) validation. We find that classifiers selected by Bayesian optimization for in-hospital mortality can outperform those selected by grid search or random sampling. However, in contrast to previous research: 1) Bayesian optimization is not more efficient in selecting classifiers in all instances compared to grid search or random sampling-based methods and 2) we note marginal gains in classifier performance in only specific circumstances when using a common variant of Bayesian optimization (i.e. automatic relevance determination). Our analysis highlights the need for further practical, deployment-centered benchmarking of HO approaches in the healthcare context.

Comprehensive and quantitative mapping of RNA-protein interactions across a transcribed eukaryotic genome.

RNA-binding proteins (RBPs) control the fate of nearly every transcript in a cell. However, no existing approach for studying these posttranscriptional gene regulators combines transcriptome-wide throughput and biophysical precision. Here, we describe an assay that accomplishes this. Using commonly available hardware, we built a customizable, open-source platform that leverages the inherent throughput of Illumina technology for direct biophysical measurements. We used the platform to quantitatively measure the binding affinity of the prototypical RBP Vts1 for every transcript in the Saccharomyces cerevisiae genome. The scale and precision of these measurements revealed many previously unknown features of this well-studied RBP. Our transcribed genome array (TGA) assayed both rare and abundant transcripts with equivalent proficiency, revealing hundreds of low-abundance targets missed by previous approaches. These targets regulated diverse biological processes including nutrient sensing and the DNA damage response, and implicated Vts1 in de novo gene "birth." TGA provided single-nucleotide resolution for each binding site and delineated a highly specific sequence and structure motif for Vts1 binding. Changes in transcript levels in vts1Δ cells established the regulatory function of these binding sites. The impact of Vts1 on transcript abundance was largely independent of where it bound within an mRNA, challenging prevailing assumptions about how this RBP drives RNA degradation. TGA thus enables a quantitative description of the relationship between variant RNA structures, affinity, and in vivo phenotype on a transcriptome-wide scale. We anticipate that TGA will provide similarly comprehensive and quantitative insights into the function of virtually any RBP.

The organisational structure of protein networks: revisiting the centrality-lethality hypothesis.

Protein networks, describing physical interactions as well as functional associations between proteins, have been unravelled for many organisms in the recent past. Databases such as the STRING provide excellent resources for the analysis of such networks. In this contribution, we revisit the organisation of protein networks, particularly the centrality-lethality hypothesis, which hypothesises that nodes with higher centrality in a network are more likely to produce lethal phenotypes on removal, compared to nodes with lower centrality. We consider the protein networks of a diverse set of 20 organisms, with essentiality information available in the Database of Essential Genes and assess the relationship between centrality measures and lethality. For each of these organisms, we obtained networks of high-confidence interactions from the STRING database, and computed network parameters such as degree, betweenness centrality, closeness centrality and pairwise disconnectivity indices. We observe that the networks considered here are predominantly disassortative. Further, we observe that essential nodes in a network have a significantly higher average degree and betweenness centrality, compared to the network average. Most previous studies have evaluated the centrality-lethality hypothesis for Saccharomyces cerevisiae and Escherichia coli; we here observe that the centrality-lethality hypothesis hold goods for a large number of organisms, with certain limitations. Betweenness centrality may also be a useful measure to identify essential nodes, but measures like closeness centrality and pairwise disconnectivity are not significantly higher for essential nodes.