We further calculated standard deviations (sd) obtained from the one way ANOVA, using the 5 level factor Type (normal, stage01, stage2, stage3, stage4).  The adjusted R-squared values representing the proportion of variation explained by Type are also reported.Type was statistically significant for every gene; all p-values were less than 0.000001 (no adjustments for multiple comparisons). These data are tabulated in Table 3, and​ shown graphically in Figure 1, below.

For each gene on the graph in Figure 1, we have shown the min and max in order to make the presentation clearer. At top left is high exxpression Value of 9985, which is the maximum value for that gene, at the bottom one finds the value for the minimum The colors range from dark blue (control) to orange (stage 4). The groups are also distinguished by line type: control (solid), stage 0-1 (long dash), stage 2 (dash), stage 3 (dot), stage 4 (dash nd dot). The figure is a parallel coordinate plot made in R 51, using the package MASS 52.

Stool was obtained from 15 participating subjects {three healthy controls and twelve colon cancer patients of all the colon cancer stages (three TNM stage 0-1 (e.g., polyps◻³ 1 cm, villous or tubvillous, or with high grade dysplasia), three stage 2, three stage 3, and three stage) 23. All stools were collected with sterile, disposable wood spatulas in clean containers, after stools were freshly passed, it was then placed for storage into Nalgene

screw top vials (Thermo Fisher Scientific, Inc., Palo Alto, CA, USA), each containing 2 ml of the preservative RNA later (Applied Biosystems/Ambion, Austin, TX, USA), which prevents the fragmentation of the fragile mRNA molecule 22, and vials were stored at - 80 °C until samples were ready for further analysis. Total small RNA, containing miRNAs, was extracted from all frozen samples at once, when ready, and there was no need to separate mRNA containing small miRNAs from total RNA, as small total RNA was suitable to make ss c-DNAs.

A procedure used for extracting small total RNA from stool was carried out using a guanidinium-based buffer, which comes with the RNeasy isolation Kit^®, Qiagen, Valencia, CA, USA, as we have previously detailed 22. DNase digestion was not carried out, as our earlier work had demonstrated no difference in RNA yield or effect on RT or PCR after DNase digestion 23414243. The time taken to purify aqueous RNA from all of the 15

frozen stool samples was ~ two hours. Small RNA concentrations were measured spectrophotometrically at λ 260 nm, 280 nm and 230nm, using a Nano-Drop spectrophotometer (Themo-Fischer Scientific). The integrity of total RNA was determined by an Agilent 2100 Bioanalyzer (Agilent Technologies, Inc., Palo Alto, CA, USA) utilizing

the RNA 6000 Nano LabChip^®. RNA integrity number (RIN) was computed for each sample using instrument's software 222342435354555657.

The RT² miRNA First Strand Kit^® from SABiosciences Corporation (Frederick, MD, USA) was employed for making a copy of ss-DNA in a 10.0 µl reverse transcription (RT) reaction, for each RNA samples in a sterile PCR tube, containing 100 ng total RNA, 1.0 µl miRNA RT primer & ERC mix, 2.0 /µl 5X miRNA RT buffer, 1.0 µl miRNA RT enzyme mix, 1.0 µl nucleotide mix and Rnase-free H₂O to a final volume of 10.0 µl. The same amount of total RNA was used for each sample. Contents were gently mixed with a pipettor, followed by brief centrifugation. All tubes were then incubated for 2 hours at 37^oC, followed by heating at 95^oC for 5 minutes to degrade the RNA and inactivate the RT. All tubes were chilled on ice for 5 minutes, and 90 µl of Rnase-free H₂O was added to each tube. Finished miRNA First Strand cDNA synthesis reactions were then stored overnight at -20^oC 2223425354555657.

Because the use of 96- or 384-well plates for a single sample is nether practical or affordable, nor very accurate, widespread implementation of dPCR technology has necessitated the introduction of nanofluidic techniques and/or emulsion chemistries. Three enhancements associated with dedicated instruments have helped promote the use of dPCR: (a) Partition volumes have been lowered to as little as 5 picoliter (pl); (b) The partitioning process has been automated, and (c) The number of partitions has been increased to over 100,000 for a single experiment. These innovative elements have simplified dPCR, and increased its precision, while holding down the total reaction value of a single experiment, compared to that of a conventional qPCR 48495051.

Digital PCR is a new approach to miRNAs quantification that offers alternate method to qPCR for absolute quantification, by partitioning a sample of DNA or cDNA into many individual, parallel PCR reactions; some of these reactions contain the target molecule (positive), while others do not (negative). A single molecule can be amplified a million-fold or more. During amplification, TaqMan chemistry with dye-labeled probes is used to detect sequence-specific targets. When no target sequence is present, no signal accumulates. Following PCR analysis, the fraction of negative reactions is used to generate an absolute count of the number of target molecules in the sample, without the need for standards or endogenous controls. In conventional qPCR, the signal from wild-type sequences dominates and obscures the signal from rare sequences 555657. By minimizing the effect of competition between targets, dPCR overcomes the difficulties inherent to amplifying rare sequences and allows for sensitive & precise absolute quantification of the selected miRNAs.

Applied Biosystem QuantStudio™ 3D instrument used in this research study only performs the imaging and primary analysis of the digital chips. The chips themselves must be cycled offline on a Dual Flat Block GeneAmp® 9700 PCR System. or the ProFlex™ 2x Flat PCR System. The QuantStudio™ 3D Digital PCR System can read the digital chip in less than 1 minute, following thermal cycling 48. It allows for one sample per chip; although, duplexing allows for analsis of two targets per chip. Sample prep for digital PCR is no different than for real-time PCR, when using the QuantStudio™ 3D Digital PCR System. To figure out the concentration of cDNA stock from results, if one includes all of the necessary dilution factors into the AnalysisSuite™ software, the software will give the copies/µL in the stock.

There are 2 dilutions that one needs to take into account: (a)The first is the dilution of the sample in the reaction,. and (b) The second is the dilution of the stock that one makes before adding it to the digital PCR reaction. For example, if one wants to add 1 µL of a sample that has been diluted 1:10 from the stock. Thus, if one adds 1 µL of his/her sample to a 16 µL (final volume) reaction, the dilution factor of the sample is 1:16 or 1/16 = 0.0625. Since the stock has also been diluted 1:10 (0.1), one also need to factor this in. The final dilution factor to enter into the software is 0.0625 x 0.1 = 0.00625 (1:160). One can use either annotation to indicate the dilution factor in the AnalysisSuite™ software. If one enters that value into the Dilution column, the software will give the copies/µL in the starting material (stock). The Poisson Plus algorithm for projects that contain QuantStudio™ 3D Chips with target, quantities >2000 copies/μL. The Poisson Plus algorithm corrects for well-to-well load volume variation, on a per Chip basis. This becomes important at higher target concentrations. There is also an option to export the Chip data as XML on the Export tab-thousands of discrete subunits prior to amplification by PCR, each ideally containing either zero or one (or at most, a few) template molecules 50.

Each partition behaves as an individual PCR reactions -as with real-time PCR—fluorescent FAM probes (or others, as VIC fluorescence). Samples containing amplified products are considered positive (1, fluorescent), and those without product -with little or no fluorescence (i.e., are negative, 0). The ratio of positives to negatives in each sample is the basis of amplification. Unlike real-time qPCR, dPCR does not rely on the number of amplification cycles to determine the initial amount of template nucleic acid in each sample, but it relies on Poisson Statistics to determine the absolute template quantity. The unique sample partitioning step of dPCR, coupled with Poisson Statistics allows for higher precision than both traditional and qPCR methods; thereby allowing for analysis of rare miRNA targets quantitativley and accuratley 5051.

The use of a nanofluidic chip, shown below, provides a convenient and straight forward mechanism to run thousands of PCR reactions in parallel. Each well is loaded with a mixture of sample, master mix, and Applied Biosystems TaqMan Assay reagents, and individually analyzed to detect the presence (positive) or absence (negative) of an endpoint signal. To account for wells that may have received more than one molecule of the target sequence, a correction factor is applied using the Poisson model. It features a filter set that is optimized for the FAM™, VIC®, and ROX™ dyes, available from Life Technologies 49, (Figure 2).

A workflow of the dPCR procedure by the QuantStudio^TM 3D Digital PCR System Chip is presented in Figure 3, below. Digital PCR, however, has several tips to follow: 1) A rough estimate of the concentration of miRNAs of interest has to be first carried out, in order to make appropriate dilutions, so that not too many partitions will get multiple copies that prevent accurate calculation of the copy number of miRNAs of interest; 2) Non-template controls and a RT negative control must be set up for each miRNA, when using a primer pool method for retro-transcription; 3) A chip-based dPCR method requires less pipetting steps, which reduces potential PCR contamination compared to another type of dPCR marketed by Bio-Rad Laboratories, thus called Bio-Rad's droplet digital PCR , which requires multiple pipette transfers that potentially increase the risk of contamination 50, and 4) Quant Studio^TM 3D chip has 20,000 fixed reaction wells, whereas Bio-Rad's droplet PCR relies upon the generation of droplets; a step that could be extremely variable, as reported by Miotto et al1148.

Our collaborating clinicians are aware of the constraints imposed by working with RNA and the need to preserve it so it does not ever fragment thereafter. 2223414243. Participating clinics will consent prospective individuals when they report to the clinic for consult. Those individuals not showing any polyps, or inflammatory bowel diseases, such as colitis or diverticulitis, will be asked by their physicians if they wish to participate in the study. If they agree, they/their guardian will be consented, each given a stool collection kit and detailed collection instructions. Each study subject will collected one 10 g stool sample, in a standardized fashion, in a large 40 cc plastic jacket given to each participating, consented individual, prior to any bowel preparation. The study nurse will show and ask participants to brush both the mucinous layer, which is rich in colonocytes, and the non-mucinous parts of stool in order to have a representation of the entire colon (both right and left side colon) 12754555657, to be preserved overnight at room temperature in the fixative RNALater® (Invitrogen) added at 2.5 ml per 1 g of stool, followed by calling the laboratory personnel to pick up the sample by next morning. Samples will be stored at -80^oC in small aliquots until needed. International Collaborators will also give study participants these written instructions in their native languages to ensure standardization, and will explain to them what's needed to collect samples representing both right and left side colon. When ready for analysis, samples are defrosted at room temp, filtered through a nylon mesh by laboratory personnel, in order to remove the preservative, and any debris prior to extraction of total small RNA. All laboratory work will be carried out and standardized under blind conditions and, in accordance with organization s Standard Operating Procedures (SOPs) for handling of biohazardous waste material 1212223.

To avoid bias, and ensure that biomarker selection and outcome assessment will not influence each other, a prospective specimen collection retrospective blinded evaluation (PRoBE) design randomized selection 31 of control subjects and case patients from our consented cohort population, will be employed, without nopriori knowledge of who has what diagnosis, and stool specimens collected prior to removal of the lesion on patients undergoing colonoscopy, which will form the cohort.

By the 8^th months of each year, we would have a cohort of 135 subjects, who are representative of the entire cohort, to select 6 control subjects and 30 CC patients. This will undoubtedly be the least common of the three groups (normal, adenoma & cancer) by far. We will then match 1 to 1 adenoma cases to the cancer cases for age (+/- 5 years), gender, clinic and month of diagnosis. Similarly, we will then match the normal controls from among the collected specimens to the cancer and adenoma cases. If there is no match, we will liberalize the data restriction to allow +/- 2 months. Thus, we will collect a case-case-control group nested in the overall colonoscopy cohort that is collected. The absolute quantitative dPCR miRNA expression analysis will be carried out on all coded samples at once during the last three months of each study year as shown in the time line Table 4, with the investigators blind to knowledge of the patients diagnosis, so that no analytical bias is introduced to the study.

While we believe that the 135 stool samples collected every year are representative of the overall cohort, there may be some volunteer bias, which we will not know how it would affect the studied miRNA markers. Therefore, we will collect demographic & clinical data on both groups (those who participated & those who did not) and compare for the following factors: age, gender, race/ethnicity, reason for colonoscopy, diagnoses, so that an assessment can be made at study conclusion as to what degree selection may have affected the study results.

Approximately 1 g of thawed stool is homogenized in a Stomacher^® 400 EVO Laboratory Blender (Seward, UK) at 200 rpm for 3 minute, with 40 ml of a buffer of Hank s solution, containing 10% fetal bovine serum (FBS) and 25 mmol/L Hepes buffer (pH 7.35). The homogenates is filtered through a nylon filter (pore size 512 µm), followed by addition of 80 µl of Dynal superparamagnetic polystyrene beads (4.5 µm diameter) (Invitrogen, Carlsbad, CA, USA) coated with a mouse IgG1 monoclonal antibody (Ab) Ber-Ep4 (Dako, Glostrup, Denmark) specific for an epitope on the protein moiety of the glycopolypeptide membrane antigen Ep-CAM, which is expressed on the surface of human epithelial cells, including colonocytes and colon carcinoma cells 5859, at a final concentration of 12 ng of Ab/mg magnetic beads (1 µg Ab/106 target cells). The mixtures is incubated for 30 min on a shaking platform at room temperature. To visualize colonocytes. a drop of the solution is spread on a glass slide, dried and stained with Diff-Quick stain (Fisher Scientific, Pittsburgh, PA), and another drop is placed on a hemocytometer, and counted under the microscope to estimate the number of colonocytes form which total small RNA will be extracted. The supernatant is removed and the pellet containing colonocytes will be stored at -80^oC until small RNA extraction 225960.

By the 9 ^th month of each study year, isolation of colonocytes from stool, and comparing the Agilent electrophoretic (18S and 28S) patterns to those obtained from total RNA extracted from whole stool, and differential lysis of colonocytes by RT lysis buffer (Quagen), could be construed as a validation that the electrophoretic pattern observed in stool (18S and 28S) is truly due to the presence of human colonocytes, and not due to stool contamination with Escherichia coli (16S and 23S) 24. One must also take into account that some exsosomal RNA 30 will be released from purified colonocytes into stool, and correction is made for that effect.

The expression of individual genes may be altered by mutations in the DNA, or by a change in their regulation at the RNA or protein levels. Epigenetic silencing is an important mechanism that contributes to gene inactivation in CRC 21. Analysis of promoter methylation of hypermethylated in cancer 1 (HIC1) gene in stool showed it to be highly specific (98%) for colon adenoma and carcinoma, but sensitivity was quite low (31% for adenoma & 42% for all cancer) 61, which suggested that an epigenetic marker only is not good enough for screening, but a combination of genetic and epigenetic markers would be required to reliably identify CRC at an early stage.

Working with the stable DNA has been relatively easy. A study by scientists affiliated with Exact Sciences Corp., Marlborough, MA, which markets a mutation-based DNA test, assessed a newer version of a fecal DNA test for CRC screening using a vimentin methylation marker and another mutation DY marker plus non degraded DNA in a limited sample of 44 CRC patients and 122 normal controls. It cited a sensitivity of 88% and a specificity of 82% only for advanced cancer, but not adenoma 6263. Besides, DNA mutation tests are not cost-effective, as screening for multiple mutations is expensive because these demanding mutation tests are not automated and are labor intensive. In addition, mutation detection in oncogenes and suppressor genes suffers from: a) the detection of mutations in these genes in fewer than half of large adenomas and carcinomas, b) the detection of gene mutations in non-neoplastic tissues, c) mutations found only in a portion of the tumor, and d) mutations often produce changes in the expression of many other genes 6364.

Protein-based methods are currently not suited for screening and early diagnosis either because proteins are not specific to one tumor or tissue type (e.g., CEA), their susceptibility to proteases, current lack of means to amplify proteins, no function is known for more than 75% of predicted proteins of multicellular organisms, there is not always a direct correlation between protein abundance and activity, and most importantly because detection of these markers exfoliately often signifies the presence of an advanced tumor stage. The dynamic range of protein expression in minimally-invasive body fluids (e.g., blood) is as large as 101065. Moreover, mRNA levels do not necessarily correlate with protein expressions 66. Protein microarray studies revealed that protein expression vastly exceeds RNA levels, and only posttranslationally modified proteins are involved in signal transduction pathways leading to tumorigenesis 67. There is no well-documented protein test that has been shown in clinical trials to be a sensitive and a specific indicator of colon neoplasia, especially in early stages. More recently, a serum proteomic study employing liquid chromatography (LC)-mass spectrometry (MS) carried out in a nonbiased fashion failed to differentiate between individuals with large adenoma (1 cm) and normal individuals 68. Proteomic research is a relatively new discipline, so it will take considerable time to identify and validate proteins suitable for use as clinical markers, and resolve issues of bias and validations 6569.

On the other hand, a transcriptomic mRNA approach has shown promise to detect adenomas and colon carcinomas with high sensitivity and specificity in preliminary studies 22, but no randomized, standardized, blinded prospective clinical study has been carried out to validate the superiority of the mRNA approach. A study indicated that a combination of a transcriptomic mRNA and miRNA expression signatures improves biomolecular classification of CRC 69. Furthermore, not only does miRNAs regulate mRNA, but they also regulate protein expression. Two studies have shown that a single miRNA act as a rheostat to fine tune the expression of hundreds of proteins 7071. Hence, for CRC screening, miRNA markers are much more comprehensive and preferable to a DNA-, epigenetic-, mRNA- or a protein-based markers23. An added advantage of the use of the stable, nondegradable miRNAs by PCR expression, or chip-based methods is being automatable, which makes them much more economical and more easily acceptable by laboratory personnel performing these assays 48495051.

Links between miRNAs and CRC have been reported in several studies in colon cancer cell lines, cells in culture, blood, colon tissue of CRC patients, and human stool 234354555657.

Stool testing has several advantages over other colon cancer screening methods as it is truly noninvasive and requires no unpleasant cathartic preparation, formal health care visits, or time away from work or routine activities. Unlike sigmoidoscopy, it reflects the full length of the colorectum and samples can be taken in a way that represents the right and left side of the colon. It is also believed that colonocytes are released continuously and abundantly into the fecal stream 7980, contrary to blood that is released intermittently as in guaiac FOBT 25; therefore, this natural enrichment phenomenon partially obviates the need to use a laboratory-enrichment technique to enrich for tumorigenic colonocytes, as for example when blood is used for testing. Furthermore, because testing can be performed on mail-in-specimens, geographic access to stool screening is essentially unimpeded. The American Cancer Society (ACS) has recognized stool-based molecular testing as a promising screening technology for CRC (www.cancer.org).

Our results and others have show that even the presence of bacterial E. coli DNA, non-transformed RNA and other interfering substances in stool does not interfere with measuring miRNA expression 1222351525354555657, when an enrichment method such as the immunological paramagnetic capture method is used 2526, when good ss-cDNA is produced 87, and when appropriate PCR primers are employed 23555657, as in this study. Besides, stool colonocytes 88 contain much more miRNA ( than that available in free circulation such as in plasma 535784. Considerable effort has gone into selecting a reasonable number of miRNA genes (fourteen) from among the many mature human miRNA sequences identified in a previous preliminary microarray data generation study, as a number that can be screened reliably by PCR in a subsequent quantitative dPCR study to ultimately validate smaller panel of miRNA diagnostic screening gene markers, preferably 10 or less, for routine use.

We have routinely carried out RNA isolation procedures (both manual and automatic) from colon tissue, blood and stool samples in our labs, manually, as well as by employing the Roche MagNA pure LC™ automated system, using Qiagen s RNeasy Isolation Kit® from Qiagen, Valencia, CA, containing RLT buffer (a guanidinium-based solution) and other commercial RNA extraction preparations, which provide the advantage of manufacturer s established validation and quality control standards, increasing the probability of good results 2122435354555657, to extract high quality total RNA from an environment as hostile as stool; thus, shattering the myth that it is difficult to employ RNA as a screening substrate. The trick has been to stabilize total RNA shortly after obtaining fresh stool by fixing samples in a chaotropic agent (RNALater® (Invitrogen)) and observing that RNA does not ever fragment thereafter. Fragmented RNA results in poor cDNA synthesis and ultimately in less than optimal PCR amplification.

We found total small RNA isolated from stool to be suited for dPCR analysis, without further mRNA purification 87 because: a) purified mRNA involves additional steps, and the increased sensitivity could be balanced by possible loss of material, b) not all mRNA molecules have polyA tail, and c) the concentration of mRNA may be insufficient to allow quality assessment using the Agilent 2100 Bioanalyzer 87. Good human stool preparations showed two sharp ribosomal 18S and 28S rRNA (28S/18S=0.33), with a small fraction of micro RNA and 5S rRNA or tRNA molecules in the Agilent capillary electrophoresis equipment fitted with a RNA 6000 Nano LabChip 22. However, E. coli shows 16S and 23S (23S/16S=1.8) 24. RNA will be quantified spectrophotometrically. Acquiring sufficient small mRNA to analyze from stool or isolated colonocytes is feasible, as each cell contains ~ 20 pg total RNA or 0.4 pg mRNA (equivalent to 0.36 pg ss-cDNA). Only few nanograms of that DNA are needed per PCR reaction] [92].

An Applied Biosystem kit (the TaqMan® MicroRNA Reverse Transcription Kit) that makes high quality ss-cDNA from total small RNA, and has been employed in earlier studies, will also be used in this study. It uses 50 nM RT primers that bind to the 3 portion of miRNA molecules, 1x RT buffer, 0.25 mM each of dNTPs, 3.33 U/µl RT in a 7.5 .µl reaction for 30 min at 96^oC, 2 min at 56^oC, 30 sec at 98^oC, 2 min at 60^oC and held at 10^oC, the chip is then processed, and results expresed in copies µl 4850, as shown in the workflow in Figure 3.

Rigid QC considerations are necessary to ensure the uniformity, reproducibility and reliability of dPCR amplification technology. Compared to real-time quantitative PCR (qPCR), dPCR clearly offers more sensitive and considerably more reproducible clinical methods that could lend themselves to diagnostic, prognostic, and predictive tests. But for this to be realized, the technology will need to be further developed to reduce cost and simplify application. Concomitantly the preclinical research will need be reported with a comprehensive understanding of the associated errors 4850.

The term absolute quantification used in dPCR refers to an estimate derived from the count of the proportion of positive partitions relative to the total number of partitions and their known volume. When the sample is sufficiently dilute, most partitions will not contain template and those that do are most likely to contain single molecules. As the sample becomes more concentrated, the chance of more than 1 molecule being present within a positive partition increases. This does not pose too great of a challenge, because the distribution of molecules throughout the partitions approximates a Poisson distribution, and a Poisson correction is applied. The dynamic range of a dPCR assay can extend beyond the number of partitions analyzed but the assay precision deteriorates at each end. In contrast, qPCR precision deteriorates only at low copy numbers 50.

dPCR benefits from a far more predictable variance than qPCR, but dPCR is susceptible to upstream errors associated with factors like sampling and extraction. dPCR can also suffer systematic bias, particularly leading to underestimation, and internal positive controls are likely to be as important for dPCR as they are for qPCR, especially when reporting the absence of a sequence. Calibration curves are frequently employed to reduce the error associated with qPCR, but they in turn are challenging to select, value assign, and apply in a manner that will be reproducible; their application also contains inherent error that is almost never considered. Arguably, a key problem with applying qPCR to areas such as the discovery of biomarkers that will eventually be translated to clinical care, is understanding whether poor reproducibility is biological, or if it is due to issues related the fact the qPCR technique is difficult to perform reproducibly. Taking all these arguments in consideration, we are therefore in the opinion that chip-based dPCR is more suited than qPCR in our proposed validation, 5-years study 50. dMIQUE Guidelines have been implement on dPCR data 93. Adopting these guidelines helps to standardize our experimental protocol, maximize efficient utilization of resources, and enhance the impact of this technology. Measuring miRNA by dPCR takes the last 3 months of every study year, after all stool samples have been collected.

If the difference in gene expression dPCR value in copies/µl between healthy and cancer patients and among the stages of cancer at the end of the proposed validation study is as large and informative for multiple miRNA genes as in the limited preliminary results, suggesting that classification procedures could be based on values exceeding a threshold, then sophisticated classification procedures would not be needed to distinguish between these two groups; otherwise, we will use predictive classification, as detailed below. The goal will be to assign cases to predefined classes based on information collected from the cases. In the simplest setting, the classes (i.e., tumors) are labeled. cancerous and .non-cancerous. Statistical analyses for predictive classification of the information collected (i.e., quantitative PCR results on miRNA genes) attempt to approximate an optimal classifier. Classification can be linear, nonlinear, or nonparametric 9495.

The miRNA expression data will be analyzed first with parametric statistics such as Student t-test or analysis of variance (ANOVA) test if the data distribution is random, or with nonparametric Kruskall-Wallis, Mann-Whitney and Fisher exact tests if the distribution is not random 96. If necessary, more complicated models such as multivariate analysis 97 and logistic discrimination will be employed. For the corrected index, cross-validation will be used to protect against overfitting. Efron and Tibshirani 98 suggested dividing the data into 10 equal parts and using one part to assess the model produced by the oter nine; this is repeated for each of the 10 parts. Cross-validation provides a more realistic estimate of the misclassification rate.

The area under the ROC curves, (in which sensitivity is plotted as a function of (1 - specificity)), will be used to describe the trade-off between sensitivity and specificity 99. We will also employ principal component analysis (PCA) method 100, which is a multivariate dimension reduction technique to simplify grouping of genes that show aberrant expression from those not showing expression, or a much reduced expression.

In cases where several genes by themselves appear to offer distinct and clear separation between control and cancer cases in either stool or tissue samples, a PMI 101102 may not be needed. If the miRNA gene panel (or a PMI) derived by the end of the study is better than existing screening methods, all of the data generated will be used to assess the model so over-fitting is not a concern.

Cross-validation will be used to protect against over-fitting. The level of gene expression will be displayed using parallel coordinate plots 103104105 produced by the lattice package in R (version 2.9.0, http://cran.r-project.org) 51. Other packages such as GESS (Gene Expression Statistical System) published by NCSS (www.ncss.com) will also be employed in the study.

Each subject will have his or her medical record number as the key ID for merging various tables in the database. A database will be established using widely available software like MS-Access, which output spreadsheets that will be analyzed with R (version 2.9.0, The R Foundation for Statistical Computing, http://www.r-project.org/) and S-plus software (Insightful Corporation, Seattle, WA).