CNV regions (CNVRs) are defined as the union of all overlapping CNVs across subjects. As these regions are very long and therefore inadequate for the analysis, we divided them (based on the rules defined in 18) into several sub-CNVRs (Figure 2A). The frequency of a sub-CNVR is defined as the percentage of subjects which have a CNV inside the boundaries. Only those with a frequency higher than the user-defined threshold are selected and saved in the database (Figure 2B).

More precisely, our CNVR algorithm performs the following steps to detect sub-CNVRs with a frequency greater than the threshold:

1. A list is created that contains only SNPs from all study individuals on a specific chromosome that define the borders of individual CNVs; upstream SNPs are designated as "starting" SNPs "S", downstream SNPs are designated as "ending" SNPs "E"

2. The list is sorted by the physical position of those SNPs (note: if several individuals have a CNV with the same starting or ending SNP, this SNP is listed for each individual separately; thus, the same SNP could be listed several times, sometimes as starting SNP, sometimes as ending SNP)

3. A counter is initiated which increments on each CNV-starting SNP and decrements on each CNV-ending SNP.

4. When two consecutive SNPs within this sorted list have different (ascending) physical positions, a next sub-CNVR could begin or previous sub-CNVR would end. The frequency of this potential sub-CNVR is determined with the help of the counter, and only if the frequency is greater than the user-specified threshold, the specific sub-CNVR is actually designated as CNVR.

5. When two consecutive SNPs within this sorted list have exactly the same physical position, the counter actualizes to the frequency of the respective CNVR as defined under step 3.

Note: the boundaries of each sub-CNVR are only approximated by the physical positions of its bracketing SNPs.

If the number of subjects is very huge and their CNVs are highly interlaced with each other, the algorithm will detect many regions with almost all of the calculated CNVRs having a length of only two SNPs. Therefore we implemented a second algorithm which extends the former one by merging consecutive sub-CNVRs with a frequency greater than the threshold and building a single one on their basis. This leads to regions with greater length, but has the consequence that the state of a subject (e.g. deletion or amplification) in a region is no longer unique. Thus we have introduced a second threshold which is used to define the state of a subject: if the CNV is the longest in the given region and its physical length is greater than this threshold, then the state of this CNV is used for the association analysis (see Figure 3).

A multiple linear regression model is used to discover associations between extracted CNVRs and phenotypes. The regression analysis is performed for each CNVR separately; the dependent variable is the phenotype for which an association should be calculated. As independent variables we use the state of the subject in the region and a user defined list of covariates. Covariates are phenotypes that are used for adjustment. After the estimated coefficients and the standard variations are calculated using the Gaussian Algorithm, we determine the significance (p-value) of each region using Student's t-test. A sub-CNVR is genome-wide significant if the calculated p-value is below the Bonferroni-threshold.

CONAN has a very clear and simple interface (Figure 5): on the left side of the main window, all imported datasets, their calculated CNV regions and associated analyses are organized in a tree structure with different symbols. In the center, all CNV regions of the current selected dataset or association analysis are shown as a table (with a short summary of the parameters) and as a graphical representation. By selecting a certain CNVR a new dialog box appears which provides information about its position, its SNPs and its associations (with p-values). There is also the possibility to view the respective genomic region in the UCSC Human Genome Browser 19, HapMap Genome Browser 20 or ensemble Genome Browser 21 by just another mouse click. All algorithms and functions can be executed through well-structured menus and all required parameters can be set step by step. Moreover, the user always has the full control over the execution of each algorithm and can monitor its current progress and status.

CONAN supports copy number variations which are determined using Affymetrix 500K SNP Arrays. Additionally, our solution supports the import of the "Copy Number Segment Summary" and the "Copy Number Segment Data" file format which can be exported from the frequently used Affymetrix Genotype Console software. There also exists a generic importer for CNVs that were detected from any other platform with any other software tool; however, then the CNVs need to be stored in a comma separated values file format (an example can be downloaded at http://genepi-conan.i-med.ac.at).

After the data is uploaded, a dataset is created which covers all information about the loaded CNVs and subjects. For the association analysis, different phenotypes for the same subjects are required and can be easily and at any time imported into an existing dataset. At present, CONAN allows the import of phenotypes saved in a tabular data format (Microsoft Excel or CSV) in which each row represents a certain person and contains its related phenotypic information. In addition to spreadsheet and statistic software, the efficient phenotype management software eCOMPAGT 15 can also export phenotypic data for import into CONAN. CONAN automatically checks the input files to ensure that they are corresponding to the subjects and only numerical values are contained.

For further analysis with statistical software such as R 22 and SPSS, all results can be exported as CSV (comma separated values) or Microsoft Excel files. Visualizations can be saved as high quality PNG images.

Once all data are stored in the database, the analysis process starts with the determination of CNV regions. For this purpose we have implemented the procedure described in 1823 for the detection of regions where the number of subjects which have a CNV (with either a gain or a loss) therein is greater than a given threshold. In addition to this threshold, the user can also control the minimal number of consecutive SNPs which is used to define a CNV (CNVs, which involve less SNPs than the threshold, are discarded).

If the number of subjects is huge (>1000) and their CNVs are highly interlaced with each other, the algorithm will detect many regions with almost all of the calculated CNVRs having a length of only two SNPs. Therefore, we developed a second algorithm which extends the former procedure by merging consecutive regions and building a single one on their basis.

Table 1 summarizes several algorithm runs with different parameters to demonstrate their impact on the resulting regions. The results suggest that the number of CNV regions and the execution time depend on the total number of subjects, total number of CNVs and the threshold parameters (see "Validation" for a description of the dataset).

After CNV regions were calculated, the user is provided with the ability to perform GWA analysis on their basis. For this task we provide a multiple linear regression model (assuming an additive genetic model) which enables to discover associations between the detected regions and the imported phenotypes. A second association analysis method combines the genotyping data from SNPs with the states of detected CNVRs in order to discover associations between cumulative effect of SNPs and CNVS and phenotypes. In both cases the user can select the dependent phenotype (e.g. blood sugar level) and a list of phenotypes which should be used for adjustment (e.g. sex, age, BMI). The software automatically calculates the corresponding p-values for all selected regions and checks for genome-wide significance after Bonferroni-correction for multiple testing (p < 0.05/total number of CNV regions). In some cases it is necessary to perform the analysis only on subjects with particular properties (for example only subjects whose blood was collected after an overnight fasting period or only male subjects). Therefore, it is possible to build user defined filters in order to perform the association analysis only on a subset of all available data.

In addition, to save CPU time, already detected CNV regions can be reused for several studies. These can be compared quickly to see their difference and to identify the impact of each changed parameter.

The interpretation of tables with thousands of regions is a complex and time-consuming task. Therefore, to assist the user, we have implemented several interactive visualizations to discover regions of interest in a fast manner and to show their attributes on demand. CONAN depicts the distribution of all detected CNV regions on each chromosome (Figure 6A). The results of an association study can be visualized with a Manhattan-Plot in which all p-values are plotted using log₁₀-transformation and each chromosome has a different color; genome-wide significant hits can be found above the Bonferroni-threshold line which is automatically drawn considering the number of tests performed (Figure 6B). Every plotted p-value can be addressed by a mouse click, and a short overview of its properties appears. As a special feature, CONAN compares the detected regions with already known and published associations from the GWAS database 24. Genomic regions that are known to be associated with the phenotype or disease in question are highlighted in yellow (Figure 5).

In order to verify the implemented algorithms we have tested CONAN with sample data consisting of 1,644 KORA subjects 25. The Affymetrix 500K SNP Chip data were analyzed by DARVIN, our in-house software solution for CNV detection using a Hidden Markov Model after identification of chromosomal gains and losses by comparing the intensity of the probe sets of all subjects with a reference set (manuscript under review). The software detected about 7,130 CNVs per sample on average. As phenotypes we used BMI, gender and age. CONAN has discovered the same associations between BMI and CNVs as previously suggested: (1) nearby gene KCTD15 6 we have discovered a CNVR on 19q13.11 with a p-value of 0.003; (2) on 5p15.33 18 we have discovered a CNVR with a p-value of 0.009.

The open source command line tool Birdsuite 10 enables the detection of CNVs and provides several scripts in order to perform GWAS on the results using PLINK 17. Visualizations are possible with gPLINK. SCIMMkit 26 is also open source and provides a command line tool which enables the targeted interrogation of CNVs using Illumina Infinium II and GoldenGate SNP assays; association analysis with phenotypes is not yet provided. Helixtree and Array Studio are both commercial solutions and support a variety of input formats (CNVs detected by analyzing Affymetrix SNP arrays and Illumina arrays). GWAS are performed through a user-friendly GUI and different graphical representations enable a rapid interpretation. However, most of those approaches are client-oriented and perform their calculations locally; this leads to poor scalability and all results are stored on different workstations and not on a central machine. This is an important aspect because the amount of the genotyping data for GWAS is increasing continuously and in a non-linear manner; thus high-performance data retrieval is an important issue. CONAN solved this problem by outsourcing all tasks to a central database server and by using the client workstation only for the presentation of the results.

Our software has several strengths: (1) Extensive tests with real data demonstrated that the analysis of a study with about 1,600 subjects and hundreds of thousands of CNVs can be performed with CONAN without any problems and in reasonable time frames. (2) Due to an intuitive user interface and a detailed user manual, no knowledge in computer science and statistics is required to perform the association analysis. (3) With the help of the Manhattan Plot it is possible to spot within seconds which regions are genome-wide significant. In addition, various export functions enable the further usage of the newly-detected information in other software packages such as R or SPSS (see Table 2 for a complete list of all key features). CONAN has limitations as well, as it supports only phenotypes with numerical values; phenotypes at nominal level must be pre-processed and encoded numerically before they can be imported. However, the next version of CONAN is conceived to provide functions for labelling nominal variables automatically with numbers. Moreover, CONAN is presently limited to the analysis of Affymetrix SNP Arrays, but an extension to the import of Illumina data is planned for the next release. CNVs generated by QuantiSNP 11 or PennCNV 12 must be converted into a CSV file before they can be used in the software. However, a direct support of those data formats is planned. Finally, an interface between CONAN and eCOMPAGT 1516 should eliminate the error prone export and import tasks of phenotype-data through files.