Differentiating ulcerative colitis from Crohn’s colitis among patients with indeterminate colitis (IC) is a major challenge. The definitive diseases share demographic and clinical features, yet differ in tissue inflammation and damage suggesting distinct mechanisms. Since treatments differ, a molecular diagnostic from accessible clinical samples would greatly benefit IC patients.

by Amanda Williams and Dr Amosy M’Koma

Background
Predominantly, colonic inflammatory bowel disease (IBD), or the colitides, encompasses ulcerative colitis (UC) and Crohn’s colitis (CC) [1, 2], and (when state-of-the-art diagnostic criteria for either are inconclusive) indeterminate colitis (IC) [3]. UC and CC share many demographic and clinical features yet present significant differences in tissue inflammation and damage, suggesting a distinct etiopathogenic trigger [4]. It is believed theoretically that IBD is caused by inappropriate activation of the mucosal immune system against commensal bacteria in the intestinal lumen [4]. Differentiating UC and CC among patients with IC has remained a major challenge in endoscopic precision medicine [5]. Disease unpredictability, treatment side-effects, potential surgery, interim morbidity and acute incapacitation are individual and system burdens [6]. Because treatments for the two diseases are different, identifying phenotype-specific molecular markers would be invaluable for developing diagnostic and prognostic tools, and for precise treatment [7–9].

The need for IC classification into UC and CC is urgent for patients suffering from IBD [10]. Patients diagnosed with IC are young [11], with onset of symptoms before or shortly after the age of 18 years [11, 12] and have an equal gender distribution [13]. This contrasts to UC where there is a male predominance and a mean age of onset at 36–39 years [14]. These figures have persisted despite the introduction of newer diagnostic modalities [15]. Even after long-term follow-up, a substantial number of patients with IC still retain the diagnosis [15]. The continued presence of an IC diagnosis over time supports part of our hypothesis that IBD may represent a spectrum of diseases rather than just two the entities of CC and UC. In order to understand and resolve this challenge, an exclusion tool for differential diagnosis is needed.

To date there is no diagnostic gold standard tool for IBD. Clinicians use an inexact classification system which combines clinical, endoscopic, radiological, and histopathological techniques in order to diagnose CC and UC [15]. Even with a combination of these methods, IBD patients are mistakenly diagnosed 30% of the time [15], resulting in inappropriate pharmacologic and surgical interventions, with correspondingly significant complications [16]. The most difficult and painstaking post-operative experience is when patients pouch-operated for definitive UC change in their diagnosis to de novo Crohn’s ileitis (CI) of the ileal pouch [15]. Currently, little is known about the molecular differences distinguishing UC and CC [7, 8]. Trends in the IBD field focus on genetic susceptibility, role of normal flora, inflammatory processes, and interactions between normal flora and the immune response [17]. Even though current research is promising [8, 15], there have been no definitive answers to help clinicians differentiate between the two diseases when current diagnostics prove inadequate and result in a diagnosis of IC [3]. Rising incidence and prevalence of IBD (Fig. 1) across the world [18] is accompanied by an increase in cases of IC [11, 19]. It is becoming even more important to find molecular markers of disease to distinguish between CC and UC in patients with IC [7, 8].

Transcriptome analysis
Recently, we have quantitated the global expression profiles of RNA levels using oligonucleotide microarray/genome-wide transcriptome analysis [20, 21] to investigate transcriptional signatures present in colonic tissues obtained from UC and CC mucosa and submucosa. We used genomic data mining from pragmatic studies to demonstrate how biomedical studies can use the technology. By extracting new and useful biomedical knowledge, we hope to develop significant momentum for applications that may have medical diagnostic potential in IBD laboratories. The genomic patterns we noted show greater intensity in CC versus UC, perhaps indicative of a greater degree or different type of inflammation in the tissues underlying the layers [8]. It is possible also that these differing genes may represent candidate biomarkers that could delineate the inflammatory colitides. Specifically, these genes were noted to show greater intensity in the CC submucosa, perhaps indicative of the greater degree or different type of inflammation in the underlying tissue [20, 21]. These studies identified genes involved in inflammatory responses generally overexpressed in IBD and demonstrate that the colonic tissue transcriptomes obtained from UC/CC patients were quite different. The gene sets identified appear to distinguish UC from CC, and may serve as an excellent resource for professionals involved with gene expression data mining in a variety of clinical settings (Table 1).

Proteomics
More recently, we have developed a proteomic approach to delineating UC versus CC. Using histologic mucosal and submucosal tissue layers for analyses, we used MALDI MS for proteomic profiling along with bioinformatics technologies (Fig. 2) [7, 8]. We profiled surgical pathology resections of colonic mucosal and submucosal layers of patients with IBD undergoing colectomy in connection with pouch surgery [restorative proctocolectomy (RPC) and ileal pouch-anal anastomosis (IPAA)] [7, 8, 21]. We identified and compared protein profiles which had the necessary: (1) specificity; (2) sensitivity; (3) discrimination; and (4) predictive capacity to determine the heterogeneity of IBD7, and we were able to delineate UC and CC molecularly [7]. These molecular fingerprints are independent of tissue (mucosa, submucosa, or both) and appear to represent disease-specific markers (Table 1) [7]. Once these markers are further tested, we can potentially develop IBD screening tools which will rely on antibodies to the protein(s) of interest (Fig. 3). The distinction between UC and CC is of the utmost importance when determining candidacy for a pouch surgery [22–24]. Approximately 30% of IBD patients [7] face potential morbidity from an incorrect diagnosis with consequently inappropriate and unnecessary operative surgeries, underscoring the necessity of research efforts aimed at a more accurate diagnosis of the colitides [7, 20].

Peripheral blood biomarkers
In contrast to colon surgical pathology tissue resections, peripheral blood is a much more accessible source of cells that might be used to distinguish between CC and UC. Circulating peripheral blood cytokines are responsible for surveying the body for signs of disease. Cytokines may, therefore, serve as surrogates for disease-induced gene expression as biomarkers of disease status or severity. In pursuit of this, we studied differences in the serum cytokine behaviours between UC and CC patients [9]. We aimed so that, if successful, such analysis could lead to an assay that could be applied as an easy, accurate, affordable, non-invasive and fast screening test. However, although certain cytokines were found to differ between diseases and controls, no cytokine could clearly distinguish UC from CC [9]. An analysis of the literature has shown that although several attempts have been made to define the serum cytokines profile in IBD, the contradictory results of these studies do not indicate the possibility of finding the biomarker(s) among the serum cytokines at this time.

Differential diagnosis and treatment
These studies are highly relevant for creating a molecular differentiator for IC. Curative treatment for UC is often surgical, involving RPC and IPAA [6, 22]. Successful surgery removes the entire diseased colon while preserving bowel evacuation, continence and fertility [22]. This is largely a result of careful patient selection combined with meticulous surgical technique, but most importantly correct diagnosis [16, 22]. Clinical observations and experience suggest that it is difficult to identify patients with CC who are likely to have a successful outcome after RPC and IPAA surgery [6, 16, 23]. Thus, pouch surgery should be widely contraindicated by CC, but be an acceptable intervention for patients with UC and for those with IC who are likely to develop UC.

Despite the increased use of cutting-edge technologies, there is no single, straight- forward explanation for the heterogeneous results, and current approaches still require validation, and subsequently confirmation on patient outcomes in a large-scale clinical cohort.

Conclusion
Our multilevel transcript observations by proteomics and genomics in tissue and blood suggest that the development of a molecular biometric-based tool that can complement the inexact classification system for diagnosis of UC and CC with precision in IBD is still preliminary.

References
1. M’Koma AE, et al. Annual Congress – Digestive Disease Week, Chicago, IL, 2009; M1096 P600
2. Burakoff R. J Clin Gastroenterol. 2004; 38: S41–43.
3. Ballard BR, et al. World J Gastrointest Endos. 2015; 7: 670–674.
4. Podolsky DK. N Engl J Med. 2002; 347: 417–29.
5. North American Society for Pediatric Gastroenterology, Hepatology, and Nutrition, et al. J Pediat Gastroenterol Nutr. 2007; 44: 653–674.
6. Keighley MR. Acta Chir Iugosl. 2000; 47: 27–31.
7. M’Koma AE, et al. Inflamm Bowel Dis. 2011; 17: 875–883.
8. Seeley EH, et al. Proteomics Clin Appl. 2013; 7: 541–549.
9. Korolkova OY, et al. Clin Med Insingts Gastroenterol. 2015: 8: 29–44.
10. Telakis ET. Ann Gastroenterol. 2008; 3: 173–179.
11. Malaty HM, et al. J Pediat Gastroenterol Nutr. 2010; 50: 27–31.
12. Kugathasan S, et al. J Pediatrics 2003; 143: 525–531.
13. Lindberg E, et al. J Pediat Gastroenterol Nutr. 2000; 30: 259–264.
14. Lee KS, et al. Arch Pathol Lab Med. 1979; 103: 173–176.
15. M’Koma AE. World J Gastrointest Surg. 2014; 6: 208–219.
16. Shen B. Inflamm Bowel Dis. 2009; 15: 284–294.
17. Corfield AP, et al. Bioch Soc Trans. 2011; 39: 1057–1060.
18. M’Koma AE. Clin Med Insights Gastroenterol. 2013; 6: 33–47.
19. Malaty HM, et al. Clin Exp Gastroenterol. 2013; 6: 115–121.
20. M’Koma A, et al. Gastroenterology 2010; 138: S-525.
21. M’Koma AE, et al. Oral presentation at the annual congress of The American Society of Colon and Rectal Surgeons, Minneapolis, MN, USA 2010: 117.
22. M’Koma AE, et al. Int J Colorectal Dis. 2007; 22: 1143–1163.
23. Shen B, et al. Inflamm Bowel Dis 2008;14:942–948.
24. Shen B, et al. Clin gastroenterol Hepatol. 2008; 6: 145–158.

The authors
Amanda Williams¹ MS; Amosy M’Koma*^2,3,4 MD, PhD
¹School of Medicine, Meharry Medical College, Nashville, TN, USA
²Department of Biochemistry and Cancer Biology, School of Medicine, Meharry Medical College, Nashville, TN, USA
³Department of Surgery, Vanderbilt University School of Medicine, Nashville, TN, USA
⁴Vanderbilt-Ingram Cancer Center, Vanderbilt University Medical Center, Nashville, TN, USA

*Corresponding author
E-mail: amkoma@mmc.edu

Advances in cellular pathology techniques will improve diagnostic medicine. However, such improvements have to overcome many challenges including variations in pre-analytical sample processing, bioinformatics data analysis and clinical interpretation of data. In order to resolve such challenges, bioinformatics needs to become more tightly coupled to the experimental methodology development.

by Dr Rifat Hamoudi, Dr Joshua Kapp, Sevgi Umur and Michael Gandy

Introduction
Molecular diagnostics within cellular pathology have been performed since the late 1990s and have developed to include a range of techniques including short tandem repeat (STR) identity analysis, classification of tumours and clonality determinations in hematopathology. More recently, with the introduction of qPCR and more recently of next generation sequencing (NGS) as shown in Figure 1, precision medicine testing for targeted therapies has rapidly gained access to daily practice and become a challenge for molecular biologists and pathologists to provide the most accurate and relevant information. As part of this testing process we discuss two major challenges which have developed, these are:

• Firstly, pre-analytical processing of formalin-fixed paraffin-embedded (FFPE) tissue, has shown to be a critical determinant in the accuracy of downstream molecular testing in specialities such as mutational screening for targeted therapies.
• Secondly, bioinformatics has become a bottleneck in data processing and interpretation, with the processing, analysis and reporting of the data shown variability between different laboratories.

This article looks to raise the awareness of these issues and presents possible areas for consideration to aid in their resolution.

Variation in pre-analytical sample processing of FFPE samples may lead to discrepancies in mutational testing of actionable genes
Within cellular pathology, the majority of molecular diagnostic clinical sample testing is now carried out on FFPE samples. Generally the tissue is screened using hematoxylin and eosin stained sections to estimate the tumour content before the preparation process of material for subsequent molecular testing, as shown in Figure 2.

Recent studies have shown that variations in pre-analytical processing of samples lead to discrepancies in downstream molecular diagnostic testing [1–3]. The variations using singleplex mutational screening were largely due to the DNA extraction system used [2, 3], quantitation using spectrophotometry and training of laboratory staff as one study showed that pre-analytical variation was significant even among experienced laboratories [3]. In addition both DNA quantitation and integrity measurements play important roles in the accuracy of downstream multiplex testing using NGS.

In order to resolve some of those issues it is important to include control series of diagnostic samples, prepared according to the diagnostic operating procedures of the laboratory with a variety of known mutations comprising missense mutations, simple and complex deletions and insertions. Assay control using known representative DNA samples from the FFPE tissue is also essential to ensure that the process of DNA extraction, quantitation and integrity measurements are performed correctly and consistently. This is important as DNA quality has a major effect on NGS performance, i.e. poor quality DNA causes a higher error rate [3].

In addition, differences in quantitation measurements need to be accounted for, since the different instruments used have different ways of measuring the concentration of DNA. For example, variations can be seen between systems such as Nanodrop spectrophotometry and Qubit fluorometry. Measurement of DNA integrity is also important and most labs use assays such as BIOMED [4, 5] or qPCR as the ‘gold standard’ measure.

Also European external quality assurance (EQA) programmes for mutation detection of solid tumours such as European Society for Pathology (ESP, www.esp-pathology.org), European Molecular Genetics Quality Network (EMQN, www.emqn.org), and United Kingdom National External Quality Assessment Scheme UK NEQAS for Molecular Pathology (www.ukneqas.org.uk and www.ukneqas-molgen.org.uk) may consider including pre-analytical (e.g. pre-PCR) component in their assessment for mutation detection from FFPE samples.

Discrepancies in variant-calling pipelines and high-throughput sequencing clinical interpretation
Most diseases such as cancer and inherited diseases are driven by genomic alterations. Recent advances in high-throughput sequencing technologies have enabled the identification of somatic mutations at very high resolution. However, accurate somatic mutation-calling using high-throughput sequence data remains one of the major challenges in genomics. For somatic mutation-calling, one looks for a site in which a variant allele exists in the tumour sample but not in the normal sample. Even with the sequence data from a normal sample, variant-calling in high-throughput sequencing data is challenging due to the multiple potential sources of errors. For example, artefacts occurring during PCR amplification or targeted capture (e.g. exome-capture), machine sequencing errors, and incorrect local alignments of reads are all well documented sources of error [6–8]. Tumour heterogeneity and normal contamination contribute additional challenges for the tumour samples [9].

Various studies have shown low concordance between different variant callers and bioinformatics analysis pipelines. Wang et al. [10] compared six variant callers on whole exome sequencing melanoma sample and matched blood of 18 lung tumour–normal pairs and seven lung cancer cell lines carried out on the Illumina HiSeq 2000. The results showed discordance between the six variant callers, and the top two performing callers could only detect 86% and 71% of validated mutations respectively. O’Rawe et al. [11] compared the analysis of five different Illumina alignment and variant-calling pipelines on 15 exome sequencing data carried out using Illumina HiSeq 2000 and Agilent SureSelect version 2 capture kit at 120X mean coverage. Results showed variant-calling concordance of 57.4% between the five different Illumina pipelines across all 15 exomes with the authors urging more caution when analysing individual genomes in genomic medicine. In addition, comparison of the two most prominent cancer genome sequencing databases; catalogue of somatic mutations in cancer (COSMIC) [12] and Cancer Cell Line Encyclopaedia (CCLE) [13] revealed marked discrepancies in the detection of missense mutations in identical cell lines (57.4% conformity), where the main reason for such discrepancy is inadequate sequencing of GC-rich areas of the exome [14].

In addition to the above, various studies have shown discrepancies in the interpretation of genomic data between the clinician and diagnostic laboratory. Shashi et al. [15] tried to follow up the results of 93 patients who underwent exome sequencing. They investigated how the clinical interpretation of the lab results changed the diagnosis and its conformity with it. Overall, the results showed that in 25% of patients (24/93), exome sequencing showed a positive result and in 80% (19/24) of cases, the clinicians agreed with the molecular diagnosis of the lab. However, in 20% of patients reported to be positive by the diagnostic lab, the clinicians thought that the suggested molecular diagnosis was not correct. In addition, 5% of patients that were considered negative by the exome lab or had a lower confidence diagnosis, were eventually found to be positive when the exome data was reviewed by clinicians. In summary the results showed 20% false positives and 5% false negatives when comparing the interpretation of genomic data between different healthcare staff.

However, it is worth noting that all the above studies used samples with high molecular weight DNA from cell lines, fresh frozen tissue or blood and carrying out the same studies above using FFPE samples has the potential to lead to further discrepancies due to the degraded DNA inherent to those samples increases the variation at the pre-analytical steps resulting in downstream discrepancies in mutational profiling. This crates it a big challenge in the development of bioinformatics pipelines required to produce consistent clinically reliable data.

One way to resolve some of the bioinformatics related issues is to exchange the raw datasets between laboratories that preferentially use different software as part of the software validation process to establish the ability of the various laboratories to detect identical gene mutations. In addition, new software updates need to be validated by analysis of prior NGS datasets covering simple and complex mutations. Finally, raw NGS datasets need to be included in EQA programmes as in silico assessment.

Conclusion
Although the above discussion very briefly surveys the current landscape in cellular pathology, the future of molecular diagnostics will undoubtedly develop to include integrated RNA expression analysis, DNA amplification and epigenetics. Each methodology will have its own idiosyncrasies and will require the development of new clinically validated bioinformatics pipeline. Additionally, the need for a novel bioinformatics system to support integrative analysis will become essential. Although previously attempted [16], new systems need to be developed to support integrative high-throughput sequencing analysis.

However, before novel bioinformatics software solutions can be devised for big data, concerns about bioinformatics software development need to be addressed. A potential starting point to address this is via supporting new bioinformatics courses that use software engineering, computer programming and mathematical modelling of biological complexity at their core, supporting the education of future bioinformaticians in the art of bioinformatics software development. This will help support a change in the current paradigm where much of the current bespoke bioinformatics software today has been developed by local institutions in relative isolation, often in conjunction within the framework of a specialist area experimental research program [17].

The future landscape highly likely see the validation of wet chemistries (laboratory and clinical based) and dry (computational based) experiments carried out in more tightly coupled format than is currently performed, supporting clinical product development in the commercial market. Also, the future will see more focus on the development of more efficient adaptive algorithms that address the clinical questions, leading to faster analysis and improving the clarity in the interpretation of the data.

In conclusion, within cellular pathology the incremental development of pre-analytical processing from FFPE samples coupled with more efficient adaptive bioinformatics algorithms implementation are key areas of focus and crucial to the further advancement of next generation molecular pathology.

References
1. Carrick DM, Mehaffey MG, Sachs MC, Altekruse S, et al. Robustness of Next Generation Sequencing on older formalin-fixed paraffin-embedded tissue. PLoS One 2015; 10: e0127353.
2. Heydt C, Fassunke J, Kunstlinger H, Ihle MA, et al. Comparison of pre-analytical FFPE sample preparation methods and their impact on massively parallel sequencing in routine diagnostics. PLoS One 2014; 9: e104566.
3. Kapp JR, Diss T, Spicer J, Gandy M, et al. Variation in pre-PCR processing of FFPE samples leads to discrepancies in BRAF and EGFR mutation detection: a diagnostic RING trial. J Clin Pathol. 2015; 68: 111–118.
4. Johnson NA, Hamoudi RA, Ichimura K, Liu L, et al. Application of array CGH on archival formalin-fixed paraffin-embedded tissues including small numbers of microdissected cells. Lab Invest. 2006; 86: 968–978.
5. van Dongen JJ, Langerak AW, Bruggemann M, Evans PA, et al. Design and standardization of PCR primers and protocols for detection of clonal immunoglobulin and T-cell receptor gene recombinations in suspect lymphoproliferations: report of the BIOMED-2 Concerted Action BMH4-CT98–3936. Leukemia 2003; 17: 2257–2317.
6. Meacham F, Boffelli D, Dhahbi J, Martin DI, et al. Identification and correction of systematic error in high-throughput sequence data. BMC Bioinformatics 2011; 12: 451.
7. Nakamura K, Oshima T, Morimoto T, Ikeda S, et al. Sequence-specific error profile of Illumina sequencers. Nucleic Acids Res. 2011; 39: e90.
8. Nielsen R, Paul JS, Albrechtsen A, Song YS. Genotype and SNP calling from next-generation sequencing data. Nat Rev Genet. 2011; 12: 443–451.
9. Gerlinger M, Rowan AJ, Horswell S, Larkin J, et al. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. N Engl J Med. 2012; 366: 883–892.
10. Wang Q, Jia P, Li F, Chen H, et al. Detecting somatic point mutations in cancer genome sequencing data: a comparison of mutation callers. Genome Med. 2013; 5: 91.
11. O’Rawe J, Jiang T, Sun G, Wu Y, et al. Low concordance of multiple variant-calling pipelines: practical implications for exome and genome sequencing. Genome Med. 2013; 5: 28.
12. Forbes SA, Beare D, Gunasekaran P, Leung K, et al. COSMIC: exploring the world’s knowledge of somatic mutations in human cancer. Nucleic Acids Res. 2015; 43: D805-D811.
13. Barretina J, Caponigro G, Stransky N, Venkatesan K, et al. The Cancer Cell Line Encyclopedia enables predictive modelling of anticancer drug sensitivity. Nature 2012; 483: 603–607.
14. Hudson AM, Yates T, Li Y, Trotter EW, et al. Discrepancies in cancer genomic sequencing highlight opportunities for driver mutation discovery. Cancer Res. 2014; 74: 6390–6396.
15. Shashi V, McConkie-Rosell A, Schoch K, Kasturi V, et al. Practical considerations in the clinical application of whole-exome sequencing. Clin Genet. 2015; doi: 10.1111/cge.12569.
16. Watkins AJ, Hamoudi RA, Zeng N, Yan Q, et al. An integrated genomic and expression analysis of 7q deletion in splenic marginal zone lymphoma. PLoS One 2012; 7: e44997.
17. Prins P, de Ligt J, Tarasov A, Jansen RC, et al. Toward effective software solutions for big biology. Nat Biotechnol. 2015; 33: 686–687.

The authors
Rifat Hamoudi*¹ PhD, Joshua Kapp1 MBBS, Sevgi Umur² BSc and Michael Gandy³ MSc
¹Division of Surgery and Interventional Science, University College London, London, UK
²Genonymous Sciences, Küçükbakkalköy, Defne Sokak, Flora Residence Istanbul,Turkey
³Health Services Laboratories, 60 Whitfield Street, London, UK

*Corresponding author
E-mail: r.hamoudi@ucl.ac.uk

The two aims of this article are to review some current methods of early diagnosis of Alzheimer’s disease (AD) and to discuss a new integration method proposed in Kong et al. [1]. We focus on how to combine the whole genome single nucleotide polymorphism (SNP) data and high-dimensional whole-brain imaging data to offer predictive values to identify subjects at risk for progressing to AD.

by Dr Dehan Kong, Prof. Kelly S. Giovanello, Eunjee Lee, Prof. P. Murali Doraiswamy and Prof. Hongtu Zhu

Introduction
Mild cognitive impairment (MCI), which commonly arises as a result of underlying neurodegenerative pathology, is a clinical syndrome characterized by insidious onset and gradual progression. Recently, much research has focused on delineating a set of biomarkers that provide evidence of such neurodegenerative pathology in living individuals and increasing attention has been devoted to combine the information from imaging, genetic, clinical, behavioural and fluid data to predict the conversion from MCI to AD. In this article, we first review some of the literature on MCI-to-AD conversion and their limitations. Subsequently, we review the results in Kong et al. [1]. According to the best of our knowledge, it is the first paper that addresses the question of how to combine the whole genome single nucleotide polymorphism (SNP) data and high-dimensional whole-brain imaging data to offer predictive values to identify subjects at risk for progressing to AD.

Methods for predicting progression from MCI to AD
There are several studies assessing the relative importance of different modalities in predicting the diagnostic change from MCI to AD by using a small subset of biosignatures [2–6]. For example, Cui et al. simultaneously examined multiple features from different modalities of data [2]. In particular, they combined structural magnetic resonance imaging (MRI) morphometry, cerebrospinal fluid biomarkers and neuropsychological measures to access an optimal set of predictors of conversion from MCI to AD. Their findings suggested that structural changes within the medial temporal lobe (MTL), particularly the hippocampus, and the performance on cognitive tests that rely on MTL integrity provided strong prediction for MCI-to-AD conversion.

Recent studies focused on the analysis of longitudinal data to assess the dynamic changes of various biomarkers associated with the MCI-to-AD conversion. A prominent neural correlate of MCI–AD is volume loss within the MTL, especially within the hippocampus and entorhinal cortex [7, 8], with increasing atrophy in these structures from normal aging to MCI to AD [9, 10]. The importance of assessing MTL changes in tracking the progression of MCI to AD has been highlighted in various longitudinal studies of individuals with MCI-AD conversion. For instance, an increased likelihood of progressing to clinical dementia has been linked with documented diminished baseline hippocampal and entorhinal volumes in several studies [11, 12].

The aforementioned studies focused on the question of whether the MCI subjects progress to AD or not, i.e. treating the conversion as a binary response. However, an important question remains, namely, how can we predict the time to conversion in MCI individuals, as well as determine the early markers of conversion? In Tarbert et al., the authors used 148 MCI subjects to identify the most predictive neuropsychological measures [13]. In Li et al., the authors used 139 MCI subjects from the Alzheimer’s disease neuroimaging initiative phase 1 (ADNI-1) to evaluate the prediction power of brain volume, ventricular volume, hippocampus volume, apolipoprotein E (APOE) gene status, cerebrospinal fluid (CSF) biomarkers, and behavioural scores [14]. They found that baseline volumetric MRI and behavioural scores were selectively predictive, and their model can achieve a moderately accurate prediction with the value of an area under the curve of 0.757 at 36 months. In Da et al., the authors used 381 MCI subjects from ADNI-1 to evaluate how several biomarkers for predicting MCI-to-AD conversion including spatial patterns of brain atrophy, Alzheimer’s disease assessment scale-cognitive subscale (ADAS-Cog) score, APOE genotype, and cerebrospinal fluid (CSF) biomarkers [15]. They have found that a combination of spatial patterns of brain atrophy and ADAS-Cog score offers a good predictive power of conversion from MCI to AD.

To the best of our knowledge, none of the previous studies have leveraged both genome-wide association study (GWAS) SNP data as well as high-dimensional whole-brain imaging data to examine their combined value in identifying subjects at greatest risk for progressing to AD.

Predicting AD using combined imaging–whole genome SNP data
In Kong et al., the authors focused on the MCI patients and combined information from whole brain MR imaging and whole genome data to predict the time to onset of AD in a 48-month national study of subjects at risk [1]. This study considered 343 subjects with MCI enrolled in ADNI-1. The patients were followed over 48 months, with 150 participants progressing to AD. The data can be treated as time-to-event data before those MCI subjects without conversion are censored data. One of the most popular models for the time-to-event data is the Cox proportional hazards model. The authors used this model to account for the covariates that are associated with the time of the events, i.e. conversion from MCI to AD.

The candidate covariates include demographic covariates (age, gender, handedness, mean education length, retirement percentage, and three dummy variables for the marital status), the APOE4 genotype, the AD assessment scale-cognitive subscale (ADAS-Cog) score, the hippocampus surface data, the region of interest (ROI) volume data, the chromosome-wise information and the significant SNP information. For each subject, the radial distance was obtained from the baseline hippocampal surfaces data for each subject, which yielded two sets of 15 000 dimensional vectors denoting the surface data from both parts of the hippocampus. For better illustration, we have plotted the hippocampus surface in Figure 1.

In Kong et al., the authors treated each part of the hippocampi as a functional predictor, and applied functional principal component analysis, and selected seven functional principal component scores for each functional predictor, which explain approximately 70% of the variance [1]. These functional principal component scores were taken as the summary measures for the hippocampus surface and put into the Cox regression model as predictors. For the chromosome-wise information, they extracted the top two principal components of the SNP data along each chromosome as predictors. For the significant SNP information, the top 101 significant SNPs were picked up using a kernel machine method, and then their top five principal components (PCs) were calculated and been used as predictors.

Specifically, they considered three candidate models. The first model is to fit a Cox regression model with demographic, clinical and ADAS-Cog score as predictors as well as APOE. This model did not include any other imaging and genetic data. The second model is to fit a Cox regression model with demographic, imaging and chromosome-wise predictors, but without the ADAS-Cog score and significant SNPs information. As a comparison, they also considered the genetic information from genome-wide association analysis (GWAS). The third model is to fit a Cox regression model with demographic, imaging and significant SNP information, but without the ADAS-Cog score and chromosome-wise information.

They compared the predictive value of the first and second model, and receiver operating characteristic (ROC) analysis indicated that the first model had a lower predictive value at 48 months than the second model. For the first model, they only identified APOE4 and ADAS-Cog score as the significant predictors. For the second model, they found that combining full genetic SNP and high-dimensional imaging data had a much higher predictive value. In particular, SNPs on chromosomes 2, 10, 11, 15, 17 and 18, APOE4 genotype, surface morphology data of both hippocampi and volumes of hippocampus, amygdala and thalamus contributed significantly. The findings support those from previous MRI studies of volumetric hippocampal changes in prodromal AD and extend them by finding that the possible prognostic value of combining information from high-dimensional imaging and genetics may be superior to that provided by routine clinical cognitive testing data. The findings also confirm the association between APOE4 status and AD, and identify additional new markers on chromosomes 2, 10, 11, 15, 17 and 18 as significant predictors for conversion. For the third model, they found that the predictive value is lower than that of the second model. This finding indicates that using the chromosome-wise information instead of the traditional significant SNPs information would be favoured. They suggest it may be due to the pitfalls of prediction using significant SNPs [16].

There are some limitations to their analysis. First, there are no test data for this study, and their findings are based on the interval cross validation. Second, they did not include measures of pathology in the models, as cerebrospinal fluid and amyloid-PET were available only in a small subset of individuals in ADNI-1. It would be beneficial to combine the data from ADNI-GO and ADNI-2 for future research.

Summary
In this article, we first reviewed some current methods of diagnosis of AD, and discussed their limitations. Then we reviewed the method and findings in Kong et al. and discussed the novelty and advantages of their study and how their proposal can be used for early diagnosis of AD by combined imaging–whole genome SNP data [1].

Acknowledgements
This material was based upon work partially supported by the NSF grant DMS-1127914 to the Statistical and Applied Mathematical Science Institute. The research of Dr Zhu was supported by NSF grants SES-1357666 and DMS-1407655 and NIH grants MH086633, T32MH106440, and 1UL1TR001111. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH and NSF.

References
1. Kong D, Giovanello KS, Wang Y, Lin W, Lee E, Fan YD Doraiswamy PM, Zhu H, Alzheimer’s Neuroimaging Initiative. Predicting Alzheimer’s disease using combined imaging-whole genome SNP data. J Alzheimer’s Dis. 2015; 46(3): 695–702.
2. Cui Y, Liu B, Luo S, Zhen X, Fan M, Liu T, Zhu W, Park. M, Jiang T, Jin SE. Identification of conversion from Mild Cognitive Impairment to Alzheimer’s disease using multivariate predictors. PLoS One 2011; 6: e21896.
3. Davatzikos C, Bhatt P, Shaw LM, Batmanghelich KN, Trojanowski JQ. Prediction of MCI to AD conversion, via MRI, CSF biomarkers, and pattern classification. Neurobiol Aging 2011; 32: 2322.e19–2322.e27.
4. Dickerson BC, Wolk DA, Alzheimer’s Disease Neuroimaging Initiative. Biomarker-based prediction of progression in MCI: comparison of AD signature and hippocampal volume with spinal fluid amyloid-β and tau. Front Aging Neurosci. 2013; 5: 1–9.
5. Young J, Modat M, Cardoso MJ, Mendelson A, Cash D, Ourselin S. Accurate multimodal probabilistic prediction of conversion to Alzheimer’s disease in patients with mild cognitive impairment. Neuroimage Clin. 2013; 2: 735–745.
6. Zhang D, Shen D. Predicting future clinical changes of MCI patients using longitudinal and multimodal biomarkers. PLoS One 2012; 7(3): e33182.
7. Dickerson BC, Goncharova I, Sullivan MP, Forchetti C, Wilson RS, Bennett DA, Beckett LA, deToledo-Morrell L. MRI-derived entorhinal and hippocampal atrophy in incipient and very mild Alzheimer’s disease. Neurobiol Aging 2001; 22: 747–754.
8. Xu Y, Jack CR, O’Brien PC, Kokmen E, Smith GE, Ivnik RJ, Boeve BF, Tangalos RG, Petersen RC. Usefulness of MRI measures of entorhinal cortex versus hippocampus in AD. Neurology 2000; 54: 1760–1767.
9. Du AT, Schuff N, Amend D, Laakso MP, Hsu YY, Jagust WJ, et al. Magnetic resonance imaging of the entorhinal cortex and hippocampus in mild cognitive impairment and Alzheimer’s disease. J Neurol Neurosurg Psychiatry. 2001; 71: 441–447.
10. Pennanen C, Kivipelto M, Tuomainen S, Hartikainen P, Hanninen T, Laakso MP, et al. Hippocampus and entorhinal cortex in mild cognitive impairment and early AD. Neurobiol Aging. 2004; 25: 303–310.
11. Jack CR, Petersen RC, Xu YC, O’Brien PC, Smith GE, Ivnik RJ, Boeve BF, Waring SC, Tangalos EG, Kokmen E. Prediction of AD with MRI-based hippocampal volume in mild cognitive impairment. Neurology 1999; 52: 1397–1403.
12. Killiany RJ, Gomez-Isla T, Moss M, Kikinis R, Sandor T, Jolesz F, et al. Use of structural magnetic resonance imaging to predict who will get Alzheimer’s disease. Ann Neurol. 2000; 47:430–439.
13. Tabert MH, Manly JJ, Liu X, Pelton GH, Rosenblum S, Jacobs M, Zamora D, Goodkind M, Bell K, Stern Y, Devanand DP. Neuropsychological prediction of conversion to Alzheimer disease in patients with mild cognitive impairment. Arch Gen Psychiatry. 2006; 63(8): 916–24.
14. Li S, Okonkwo O, Albert M, Wang M-C. Variation in variables that predict progression from MCI to AD dementia over duration of follow-up. American J Alzheimer’s Dis. 2013; 1: 12–28.
15. Da X, Toledo JB, Zee J, Wolk DA, Xie SX, Ou Y, Shacklett A, Parmpi P, Shaw L, Trojanowski JQ, Davatzikos C, Alzheimer’s Neuroimaging Initiative. Integration and relative value of biomarkers for prediction of MCI to AD progression: Spatial patterns of brain atrophy, cognitive scores, APOE genotype and CSF biomarkers. Neuroimage Clin. 2014; 4: 164–173.
16. Wray NR, Yang J, Hayes BJ, Price AL, Goddard ME and Visscher PM. Pitfalls of predicting complex traits from SNPs. Nat Rev Genet 2013; 14(7): 507–15.

The authors
Dehan Kong¹ PhD, Kelly S. Giovanello^2,3 PhD, Eunjee Lee⁴ MS, P. Murali Doraiswamy⁵ MD, Hongtu Zhu*^1,3,6 PhD
¹ Department of Biostatistics, University of North Carolina (UNC), NC, USA
² Department of Psychology, UNC, NC, USA
³ Biomedical Research Imaging Center, UNC, NC, USA
⁴ Department of Statistics, UNC, NC, USA
⁵ Departments of Psychiatry and Duke Institute for Brain Sciences, Duke University, Durham, NC, USA
⁶ Department of Radiology, UNC, NC, USA

*Corresponding author
E-mail: htzhu@email.unc.edu