PRECISION MEDICINE: FROM DIPLOTYPES TO DISPARITIES TOWARDS IMPROVED HEALTH AND THERAPIES

Abstract

Precision medicine research efforts both in basic science discovery and clinical implementation are well underway and promise to provide individualized preventions and treatments, improving overall health care delivery. To achieve these goals, advances in data capture and analysis are needed spanning different types of ‘omic and clinical data. The efforts to enhance precise treatments for all may accentuate healthcare disparities unless specific challenges are identified and addressed. This session of the 2018 Pacific Symposium on Biocomputing presents the latest developments in this transdisciplinary research space of genomics, medicine, and population health.

1. Introduction

Precision medicine is often described as providing a patient the optimal tailored treatment the first time as opposed to standard treatment or trial and error. Precision medicine has arguably been practiced since the emergence of modern medicine¹. The definition of precision medicine has recently evolved from the 20^th century’s addition of genetic variants that impact drug response^2–4 to the 21^st century’s recognition that lifestyle, social, and environmental factors interact with the patient’s genome⁵, impacting a range of health consequences including conferring risk to disease and differential response to treatmen⁶.

Although the concept of precision medicine is not new, the implementation of precision medicine is relatively nascent. Recent key advances in genomics⁷ and the emergence of electronic health records (EHRs) in part through the HITECH Act^{8, 9} make it feasible to put precision medicine to practice¹⁰. Despite these advances, the adoption of genomic data in routine clinical practice has been slow likely due to a variety of reasons including costs or inconsistent reimbursement policies (such as changing and limited coverage by Medicare) and access to genomic testing¹¹.

Another major driver behind the lack of implementation is the lack of data. The lack of data is distributed across active research areas that fuel precision medicine including basic discovery and functional or biological characterization¹². Inherent in these broad areas of research are needed expansions of informatics approaches to extract health-informative data from various big data sources coupled with advances in novel statistical and computational methodology to integrate and interpret disparate data types for predictive modeling or further discovery and functional characterization^{13, 14}. Beyond the traditional boundaries of basic research are data needs related to precision medicine implementation ranging in topics from cost-effectiveness and clinical utility to clinical decision support^{15, 16}, among other topics¹⁷.

Precision medicine research and implementation programs will undoubtedly evolve rapidly with the recent infusion of support from the previous administration¹⁸. The initial structure and recruitment sites for the Precision Medicine Initiative Cohort Program, an ambitious effort to ascertain one million participants in the United States for precision medicine research, have been established as of 2016, and recruitment and collection of these data are anticipated this year. Of note is the Program’s emphasis on racial/ethnic, geographic, and economic diversity, variables that continue to be underrepresented in the basic discovery studies^19–21 despite their known influence on human health^{22, 23}. The absence of these data may lead to misdiagnoses and missed opportunities for many patients not represented in discovery studies²⁴. The new Cohort endeavor, now known as “All of Us,” is expected to both generate new data but also to inspire new regulations and guidelines based on safety and bioethics^25–27. Given innovative biocomputing is the engine of precision medicine’s implementation vehicle, PSB is the optimal forum for the presentation, discussion, and debate of these diverse topics that eventually fuel true individualized health for all.

The investigators and research featured in this session each represent a facet of precision medicine research highlighting needs and gaps that must be addressed to achieve the goals of translational research. Topics covered in this session include the problem of finding sufficient numbers of patients or participants with similar characteristics required to achieve adequate power to identify important biomarkers that distinguish subtypes including genetic, metabolomic or other phenotypic features that have a molecular or mechanistic relationship. Without careful attention, this can lead to health disparities, as less information may be available from vulnerable groups, and thus leading to less effective diagnosis and treatment. Such challenges need to be identified and actively addressed.

This session also addresses development of incorporating social network information to recruit patients in context where they are and better understand the variables that are operative for those individuals in disadvantaged coverage and difficult complex environments, all of which opens up for new possibilities in research and care.

2. Podium presentations

A group of four manuscripts describe novel and emerging approaches to tackle ‘omic data generated by metabolomics and transcriptomics. A Orlenko and colleagues²⁸ present a case study of Tree-based Pipeline Optimization Tool (TPOT)²⁹, an Automated Machine Learning (AutoML) tool, applied to the clinical metabolic profiling of patients exposed to metformin. In their assessment of TPOT in 546 samples and 42 metabolites, the investigators identified a putatively novel association between increased homocysteine and long-term exposure to metformin. The investigators also developed several considerations for future studies including suggestions for adjustments for confounding features and recommendations for the ideal study design or dataset characteristics that minimize bias and improve AutoML performance.

J Westra and colleagues³⁰ take a different approach in the analysis of metabolomic data with the observation that these data are better represented by Gaussian mixture distributions rather than the linear models that assume normality commonly used to test for SNP-trait associations. The investigators present an adaptation of Kim et al where tests of association between copy number variants and traits were performed with a likelihood ratio test³¹. The adapted test presented here is a likelihood ratio test that can be constrained based on a priori biological data describing the genotype-phenotype relationship, if available. This adapted tested was applied to simulated data to evaluate performance of the test versus liner models as well as to natural data from 20,315 SNPs on chromosome 11 for 5,936 Framingham Heart Study participants with gas-chromatography-measured red blood cell fatty acid levels³². A total of 28 SNPs from five different regions of chromosome 11 were associated with the metabolic trait, including 19 SNPs containing the FADS gene complex known to enzymatically desaturate arachidonic acid to dihoo-gamma-linoleneic acid. While further work is needed to extend the model, the present study suggests that powerful analysis of metabolomics data may require different models compared with the traditional linear models so popular in GWAS.

Both J Berghout and S Rachid Zaim offer methods applicable to single-subject transcriptomics. J Berghout and colleagues³³ describe a mixture mole for transcript fold-change clustering from isogenically paired samples known as the “N-of-1 pathways MixEnrich”³⁴ The Gene Ontology Biological Processes (GO-BP)^{35, 36} is applied to both reduce dimensionality as well as to identify functional attributes. The method was validated using a microarray dataset of inbred mouse strains exposed to different diets. The MixEnrich results for the paired mouse liver transcriptomes are compared with results from two other methods. Linear Models for Microarray (limma)³⁷ and Gene Set Enrichment Analysis³⁸, both of which require a minimum of three pairs as opposed to a single sample pair required by MixEnrich. Results suggest that MixEnrich reproduces GO-BP signals in similar priority order compared with the other approaches, thus offering an efficient alternative to cost prohibitive cohort-derived gold standards used for validation.

S Rachid Zaim and colleagues³⁹ use simulations to address limitations in using transcriptomics as a clinical biomarker. Here, the investigators assume that a transcriptional signal or association can be detected at the pathway level regardless of patient-level heterogeneity in gene expression. The simulations are performed to assess 54 different scenarios using single-subject and cohort-based techniques describing variability at various levels including pathway gene set size, fraction of expressed gene responsive within the gene set, fraction of up and down-regulated gene expression response, and sample size. Results of these simulations suggest that single-subject pathway detection methods that include patient-level variability can detect transcriptional dysregulation at the pathway level, scenarios likely relevant to heterogeneous clinical populations expected in precision medicine.

Of all the ‘omics, genomics, represented by genotyping and sequencing, is the most mature and readily available in the clinic. Whole genome sequencing has been particularly groundbreaking in diagnosing previously undiagnosed rare diseases^40–42, but these data remain an analysis challenge for more common, complex diseases, A Gupta and colleagues⁴³ note, as have others⁴⁴, that the genetic architecture of some complex diseases such as autism spectrum disorder (ASD) remains unknown despite evidence of a strong genetic component. While it is now recognized that complex diseases are often a result of many variants across multiple genes with independent and interacting effects, and sequencing methods are available to capture genome-wide variability, few statistical methods have emerged to detect these complex genetic architectures. The investigators posit that Coalition Game Theory (CGT) can be applied to genomic data to identify individual genes (“players”) who improve the performance of the coalition or, in this case, the relationship to ASD. The investigators apply CGT to 2,710 whole-genome sequences from ASD multiplex families and identify eight genes with significantly elevated “player scores,” All eight genes are in biological pathways known to be affected by ASD and directly interact with genes previously associated with ASD. Although further follow-up is needed, these results suggest CGT is a promising method for large-scale genome data generated for complex diseases.

Two manuscripts centered on the use of automated clinical systems in the delivery or practice of precision medicine. In the first, C-L Chi and colleagues⁴⁵ apply a treatment simulation and optimization approach to develop decision support for warfarin dosing. Warfarin is a commonly prescribed anti-coagulant, and variability in initial dosing has a strong and known genetic component^{46, 47} Genetically-guiding warfarin dosing was an early poster-child for pharmacogenomics in precision medicine, but lackluster clinical trials^48–50 save for one⁵¹ among other issues have dampened enthusiasm. Logistically, the development of algorithm-based dosing delivered via EHR-automated decision support has been a challenge for this and other drugs impacted by genetics, particularly for diverse populations⁵². The investigators present results of simulations that employed the property of minimal entropy to minimize overall risks for the largest patient groups. The investigators further discuss these results through the lens of ease of implementation, a factor highly relevant for this⁵³ and other potential precision medicine clinical applications.

S Poole and colleagues⁵⁴ address alarm fatigue, an unfortunate consequence of easy and constant automated vital sign monitoring. The investigators aim to improve default vital sign alarm thresholds to decrease the number of unnecessary alarms. The investigators develop personalized vital sign thresholds based on a large heart rate database used to identify the 1^st and 99^th percentiles of a patient’s heart rate on the patient’s first day of monitoring. Results suggest that these new thresholds would decrease low and heart rate alarms while preserving sensitivity and boosting specificity. Overall, the investigators suggest these thresholds will reduce alarm fatigue thereby improving both patient care and hospital costs, major goals underlying the original HITECH Act⁵⁵.

3. Posters with published papers

This year’s poster session with papers published in the proceedings will feature two research groups. In the first, A Fish and colleagues⁵⁶ present evidence that genetic associations are modified by local ancestry transitions inferred from genome-wide association study (GWAS) fine-mapping Metabochip data available on ~10,000 African Americans with de-identified EHRs⁵⁷. African Americans are considered admixed with an average 75-93% their genomes originating from West African ancestral populations and the remaining from European ancestral populations^58–61. Local ancestry, as opposed to inferred genome-wide global ancestry, is the inference of ancestral state at the locus-level. Local ancestry estimates have been instrumental in admixture mapping efforts⁶² as well as the estimation of recombination rates and the identification of recombination hotspots⁶³. The investigators leveraged this admixed population with clinical data to identify SNP × transition interactions where the transition represent a switch in ancestral backgrounds between nearby variants. Five Bonferroni-corrected significant interactions were identified, and subsequent statistical follow-up suggested that a European to African transition modifies the association rs16890640 between mean corpuscular hemoglobin and mean corpuscular volume. Bioinformatic data coupled with model organism data suggest that alterations in the region chromatin conformation are the biological basis for the modifying effect of the ancestral transition. More broadly, this study offers an example of epistasis where the interaction, and the not variant itself, is associated with the phenotype⁶⁴. Furthermore, this study highlights yet another example of the importance of diverse populations in the search for all genetic variants and their modifiers important in human health and health disparities.

In the second, B Li and colleagues⁶⁵ test two functions of PrediXcan: 1) its ability to predict gene expression and 2) its ability to prioritize GWAS results. PrediXcan is a gene-based association method rooted in the observation that phenotypic variability can be explained by regulatory variants that modulate the expression levels of genes⁶⁶. PrediXcan uses reference transcriptome data to infer gene expression in GWAS data by estimating the genetically determined component of gene expression. This gene-based approach reduces multiple testing and proffers biological insights or mechanisms compared with standard single SNP tests of association common in GWAS. This study evaluates PrediXcan using genotypic and transcriptomic datasets available from the 1000 Genomes Project (Yoruba; YRI) and GWAS data from the AIDS Clinical Trials Group (ACTG) protocol A5202^{67, 68}. To characterize the accuracy in predicting gene expression levels, the investigators compared the PrediXcan-inferred YRI data based on whole blood models and transcriptomic data from the Genotype-Tissue Expression (GTEx) Project⁶⁹ and Depression. Genes and Networks (DGN)⁷⁰ to the actual YRI expression data and found that the slopes of correlation between predicted and actual were negative for almost one-half of the genes tested. Despite these differences, PrediXcan identified 19 genes in the A5202 cohort dataset associated with triglyceride change from baseline to 24 or 48 weeks that were not previously identified using phenome-wide association testing, attesting to PrediXcan’s potential ability to prioritize GWAS findings. Of note is the poor transcriptome prediction in YRI despite the fact that the GTEx cohort includes African Americans. These data suggest that testing the limits of PrediXcan in gene-trait associations will require more, larger, and diverse populations with both GWAS and transcriptome-level data.