By check the "Upload Biological Data" radio button, users can upload their data matrix to SSizer in various formats (csv, tab delimited, xls and xlsx). In particular, the first row of the data matrix should be the feature name, and the first column provides the sample name. The second column indicates the class label (case or control) of each sample. The standard format accepted by SSizer can be downloaded HERE ( Right Click to Save). By clicking the "Upload Data" button, the biological dataset provided by the users can be uploaded for further assessment.

Three sets of sample data are also provided in this step facilitating a direct access and evaluation of SSizer. These sample data are benchmark datasets collected from a variety of public databases. These sample data are all benchmark datasets collected from a variety of public databases. The first set of sample data MTBLS354 is metabolomics benchmark dataset collected from the MetaboLights database. This dataset contains 142 samples of community-acquired pneumonia (CAP) patients and 97 samples of people without CAP (non-CAP controls) (To K K W, et al. Diagn Microbiol Infect Dis. 85(2):249-54, 2016). The second set of sample data PXD005144 is proteomics benchmark dataset from the PRoteomics IDEntifications (PRIDE) database constructed by the European Bioinformatics Institute. This dataset contains 66 samples of patients with pancreatic cancer and 36 samples of people with chronic pancreatitis (Saraswat M, et al. Cancer Med. 6(7): 1738-1751, 2017). The last set of sample data GSE28702 is transcriptomics benchmark dataset from the Gene Expression Omnibus database dedicated by the National Center for Biotechnology Information. This dataset contains 42 samples of responders to FOLFOX therapy and 41 samples of non-responders (Tsuji S, et al. Br J Cancer. 106(1): 126-132, 2012). By clicking the "Load Data" button, the sample dataset selected by the users can be uploaded for further assessment.

Several factors, like unwanted experimental & biological variations and technical errors, can hamper the analysis of OMICs and other biological data, which requires normalization before further study (Li B, et al. Nucleic Acids Res. 45(W1): 162-170, 2017). Because of the huge amount of hypothesis tests during the analysis, it is also necessary to conduct a multiple testing adjustment. However, in case of large size of tests and low differential features, the above adjustment will lead to a substantially low power for identifying truly differential features. Data filtering is thus introduced to reduce the size of tests and increase the power (Hackstadt AJ, et al. BMC Bioinformatics. 45(W1): 10:11, 2009). In the current version of SSizer, the option to conduct data normalization and filtering before the assessment and determination of sample size is provided to the users. In particular, 15 normalization methods (Auto Scaling, Contrast, Cubic Splines, Cyclic Loess, EigenMS, Linear Baseline, Log-transform, Level Scaling, MSTUS, Pareto Scaling, Mean Normalization, Median Normalization, MSTUS, Pareto Scaling, Power Scaling, PQN, Range Scaling, Vast Scaling and VSN) frequently used to pre-process OMICs and other biological data are provided, and 2 popular filtering methods (Standard Deviation and Mean) in current biological research are further included. Users can select a specific normalization and filtering method by checking corresponding radio checkbox in the left side panel, and can also check the "NONE" option to skip one or both of these pre-processing. Moreover, the resulting data can be previewed and is fully downloadable from the website. The distributions of feature intensity before and after normalization are visualized by boxplots, and the distributions of feature intensity mean before and after data filtering are also viewable by histograms.

Three types of well-established statistical indexes for comprehensive assessment and determination of sample size are used. The colored bar shown below illustrated the assessment results of sample size (SS) by multiple types of statistical indexes. SS within the RED (not enough) region refers to the non-satisfaction of any type of indexes. SS within the ORANGE (passable) region indicates the satisfaction of only one type of indexes, and the specific type ID is also provided under this colored region. SS within the BLUE (good) region implies the satisfaction of two types of indexes, and the additional type ID is provided under the BLUE region. SS within the GREEN (very good) region denotes the satisfaction of all three types of indexes.

To assess the level of difference between comparative groups, Statistical Power Analysis is performed to assess whether the "statistical power" is at a desired level (≥ 0.8), which in turn helps to estimate the sample size required (Blaise B J, et al. Anal Chem. 88(10): 5179-5188, 2016). Under this index, users need not only to specify a false discovery rate and an expected power value, but also to select the minimum sample size per group and the number of increment and repeats. The larger the power value of a biological dataset, the higher the probability that an observed effect can pass the required threshold of claiming its discovery (Button K S, et al. Nat Rev Neurosci. 14(5): 365-376, 2013). Moreover, the sample outputs of "Line-plots of POWER Changing along with the Sample Size" that performs interactively in the same way as real output are provided.

Type II: Classification Accuracy Based on the Identified Markers

"AUC" Denoting the Overall Diagnostic Accuracy

The overall diagnostic accuracy is evaluated by the receiver operating characteristic curve together with the area under that curve (AUC) based on 3 popular machine learning algorithms. An adequate sample size is reported to be reflected by desired AUC value (≥ 0.9), and the users can also define their preferred cutoff of desired AUC value (Xia J, et al. Nucleic Acids Res. 43(W1): 251-257, 2015). Firstly, markers are identified by choosing feature selection method from 3 popular ones (Students‘ t-test, PLS-DA and OPLS-DA) based on the users‘ preference. Secondly, users are asked to select their preferred machine learning algorithms (support vector machine, random forest or diagonal linear discriminant analysis) for constructing the classification models based on those identified markers. After the k-folds cross validation on these models, results with higher AUC value are recognized as well-performed (Sing T, et al. Bioinformatics. 21(20): 3940-3941, 2005). The sample outputs of "Boxplots and Medium Values of AUC Changing along with Sample Size" performing interactively in the same way as real output are also provided.

"ACC" Indicating the Classification Accuracy on Independent Data

The classification accuracy on independent test data is assessed by accuracies (ACCs) of the constructed classification models based on 3 popular machine learning algorithms (support vector machine, random forest and diagonal linear discriminant analysisdiagonal linear discriminant analysis). According to the previous studies, an adequate sample size can be reflected by a desired ACC value (≥ 0.7), and this desired value can also be defined by users in SSizer (Billoir E, et al. Brief Bioinform. 16(5): 813-819, 2015). In particular, 3 popular feature selection methods (Students‘ t-test, PLS-DA and OPLS-DA) are applied in the first place to identify the markers based on feature intensities. Secondly, users are asked to select their preferred machine learning algorithm for sample classification. The higher value of ACC denotes better classification accuracy and indicates better performance of prediction capacity (Nyamundanda G, et al. BMC Bioinformatics. 14(1): 338, 2013).The sample outputs of "Boxplots and Medium Values of ACC Changing along with Sample Size" performing interactively in the same way as real output are also provided.

Type III: Robustness of the Identified Markers

"Overlap" between Lists of Markers Identified from Two Sub-datasets

In simple terms, overlap is the fraction of shared features that appear on both two lists of markers which determined the robustness of the identified markers by measure the similarity of two lists of identified markers. According to the previous studies, an adequate sample size can be reflected by a desired overlap value (≥ 0.5), and this desired value can also be defined by users in SSizer (Wang C, et al. Nat Biotechnol. 32(9): 926-932, 2014). The higher value of overlap denotes better classification accuracy and indicates robustness of the identified markers (Ein-Dor L, et al. PNAS. 103(15): 5923-5928, 2006). The sample output of "Boxplots and Mediums of OVERLAP Changing along with Sample Size" performing interactively in the same way as real output is also provided.

"Concordance" between Lists of Markers Identified from Two Sub-datasets

Concordance is reported to be more relevant than the mere overlap between markers in evaluating the similarity between different signatures (Fan C, et al. N Engl J Med. 355(6): 560-569, 2006). The classification of individual samples is the relevant measure of concordance which is assessed by Cramer‘s V statistic: different lists of markers could track a common set of biologic characteristics, and result in similar predictions of outcome. According to the previous studies, an adequate sample size and robust identified markers can be reflected by a desired concordance value (≥ 0.36), and this desire value could be defined by users in SSizer either (Fan C, et al. N Engl J Med. 355(6): 560-569, 2006). Similar to overlap, the higher value of concordance denotes better robustness of the identified markers. Sample output of "Boxplots and Mediums of CONCORDANCE Changing along with Sample Size" performing interactively in the same way as real output is also provided.

"Weighted Consistency" among Lists of Markers Identified from Multiple Sub-datasets

Weighted Consistency (CW) is a different kind of index which is distinguished from overlap and concordance for its overall markers statistics calculation principle. It counts the number of times of every single feature appeared in every single list of markers to represent the robustness of the identified markers from an overall perspective. According to previous studies, the higher weighted consistency indicates the better robustness of the identified markers (Somol P, et al. IEEE Trans Pattern Anal Mach Intell. 32(11): 1921-1939, 2010). But there is no well-established desired value of weighted consistency, we provided a recommended desired value (≥ 0.5) and this desire value could also be defined by users in SSizer. The sample outputs of "Linear Graph of CW Changing along with the Sample Size" that performs interactively in the same way as real output are also provided.

Hypothetical data is simulated in SSizer to enlarge the data size. The newly generated large dataset is then used to assess and further determine the adequate sample size for a specific biological problem. Data simulation starts from a relatively small cohort, variables are identified by the statistical recoupling of variables (SRV) procedure (Blaise B J, et al. Anal Chem. 88(10): 5179-5188, 2016;Navratil V, et al. Bioinformatics. 29(10): 1348-1349, 2013). A larger dataset is then generated based on the kernel density estimation of SRV (Rosenblatt M, et al. Annals of Mathematical Statistics. 27(3): 832-837, 1956; Parzen E, et al. Annals of Mathematical Statistics. 33(3): 1065-1076, 1962). Statistical significance of variables identified by SRV are assessed by Benjamini-Yekutieli correction for simulated datasets of variable sizes (Benjamini Y, et al. Annals of statistics. 1165-1188, 2001). Robustness of simulated model is evaluated by receiver operating characteristic analysis on an independent cohort and cross-validation.

According to the pre-case study of SSizer, overlap has been proved to be a stable and effective index for sample size evaluation. So, the sample size evaluation in SSizer is achieved by data simulation and subsequent statistic overlap analysis. The data are simulated using a multivariate log-normal distribution fit to the pilot data, which allows SSizer to maintain the long-tails and strong correlations that are typically seen. This data-driven data simulation approach that SSizer used is based on the assumption that a small random cohort is a good estimator of the general population. Users are required to impute the size of simulation data for data simulation and choose expected overlap, minimum sample size for per group, number of increment and number of repeats for overlap calculation.

The colored bar shown above illustrated the assessment results of sample size (SS) by multiple types of statistical indexes. SS within the RED (not enough) region refers to the non-satisfaction of any type of indexes. SS within the ORANGE (passable) region indicates the satisfaction of only one type of indexes, and the specific type ID is also provided under this colored region. SS within the BLUE (good) region implies the satisfaction of two types of indexes, and the additional type ID is provided under the BLUE region. SS within the GREEN (very good) region denotes the satisfaction of all three types of indexes.