GeneTrail2 1.6

Statistical analysis of molecular signatures

Input data

GeneTrail2 is able to read various input file formats through which the user can provide measurement data or categories that should be analyzed. In general, GeneTrail2 will try to automatically detect the meta-data of the uploaded data. This means it attempts to detect the used data format, identifier type, and organism the data was derived from. If errors arise during this step, it is important to understand which input types are supported by GeneTrail2.

Thus, in the following we discuss the expected input formats and the assumptions bc GeneTrail2 makes about their contents.

As GeneTrail2 is able to process data not only from microarray experiments, but also from e.g. mass-spectrometry experiments, we use the term entity for talking about genes, protein, miRNA, etc. Similarly, we uses the term identifier whenever we mean the name of such an entity as it is used in some database such as Ensembl, UniProt, or NCBI Gene.

Identifier lists

The simplest way to provide input data to GeneTrail2 is to upload a list of identifiers. Identifier lists can contain both: a, typically short, list of relevant entities or a, typically long, list of entities sorted by relevance.
Different methods assume different properties for the input lists. For example the ORA method [1] requires a list of relevant entities. Methods such as the Kolmogorov-Smirnov statistics assume that the identifier list is sorted by relevance. Do not use a list prepared for one method to compute enrichments of the second kind.
The input format for identifier lists recognized by GeneTrail2 is a simple text file containing exactly one identifier per line: 
GDA
SCN3A
SCN3B
RPLP2
GFER
SNORA68
SNORA65
PIP5KL1
BTBD1
RPLP0
BTBD2
BTBD3
...

Identifier level scores

Similarly to identifier lists, score lists can be provided in a text based format containing one identifier per line. The difference to identifier lists is that a score, a numerical value measuring the relevance of the entity, is provided in an additional column. Both columns are separated by a whitespace, preferably by a tab character. 

GDA 0.05501
SCN3A   -0.017374
SCN3B   0.33427200000000046
RPLP2   -0.10048799999999997
GFER    0.08075766666666603
SNORA68 0.2532145
SNORA65 -0.289492
PIP5KL1 0.267125
BTBD1   -0.824291000000001
RPLP0   0.050174750000000046
BTBD2   -0.424771999999999
BTBD3   0.267594
RPLP1   -0.1359804999999995
ATP6    -0.2206155
...
If possible prefer score lists to identifier lists. A score list can be used in any scenario an identifier list can be used in and is much less likely to run into the difficulties frequently encountered with the former.
Note, that GeneTrail2 does not check whether the scores follow a certain distribution or not. While most of the implemented methods work surprisingly well if their assumptions are violated, we recommend that the user chooses an appropriate analysis technique. To this end, the (unweighted) Kolmogorov-Smirnov test and the Wilcoxon test are non-parametric enrichment methods that do not require a specific score distribution.

Measurements

GeneTrail2 provides support for directly analyzing matrices containing high-throughput measurements. These can be normalized expression values obtained from microarray or RNA-seq experiments or protein abundances from mass-spectrometry runs. Additionally we offer rudimentary support for analyzing count data obtained via RNA-seq.

Analyzing data from high-throughput experiments is not just applying a statistical test to each row of the dataset. In practice, quality control, batch effect removal, and normalization must be performed carefully. The features offered by GeneTrail2 are provided for convenience and assume, that the data has been properly prepared!

Measurements can be uploaded as a plain text, tab-separated matrix. Optionally, the first column of the file contains names for each of the contained samples. Each subsequent row contains the measurement data for one identifier in all samples. Thus each row except the first starts with an identifier followed by N numerical values, where N is the number of samples. 

Sample1	Sample2	Sample3
	GeneA	0.1	4.3	2.3
	GeneB	3.2	-1.2	1.1
	GeneC	2.7	9.1	0.3
	...
The advantage of uploading matrices of measurements is, that sample-based (sometimes called phenotype-based) permutation schemes can be used to determine p-values.

Microarray data

A major use case of GeneTrail2 is the analysis of microarray data. For this experimental platform, well established normalization pipelines exist that usually generate normal or log-normal distributed expression values. GeneTrail2 can directly work with this kind of data and offers a range of statistics that can be used to derive scores from expression matrices.

RNA-seq data

RNA-seq data usually comes in the form of count data. This means, that for each transcript and sample the number of reads that were mapped to the transcript is reported. The distribution of this data is fundamentally different to the distribution of microarray data, and hence new methods for the analysis of count data have been developed. GeneTrail2 offers some basic support for directly analyzing count data. For this purpose it uses the DESeq2 [2], edgeR [3], and RUVSeq [4] R packages that can be used to compute scores from count data.

Note that currently for count data, no sample-based permutations can be performed due to the prohibitive runtime of the score computation process.

The used packages perform some level of normalization. However, GeneTrail2 performs no quality control or proper batch effect removal. Just as with microarray data, the web service relies on normalized or at least well-behaved input data.

Others

Data from other experimental platforms can also be used in GeneTrail2. Here, however, it is up to the user to select an appropriate scoring scheme.

Categories

While GeneTrail2 offers a large collection of categories that have been derived from a number of third-party databases (see List of categories), it can be desirable to create custom categories that should be checked for enrichment. Examples would be potential targets of a transcription factor that have been identified by a Chip-seq experiment. For specifying categories GeneTrail2 uses the Gene Matrix transposed (GMT) format [5]. In the GMT format every line represents a category. The first column corresponds to the name of the category, the second column to an optional description or source url. The following columns contain the members of the category. Each member occupies exactly one column. GeneTrail2 assumes, that the columns are tab-separated. 
CategoryA	http://test.url/A	GeneA	GeneB	GeneC	GeneD
CategoryB	http://test.url/B	GeneA	GeneD
CategoryC	http://test.url/C	GeneD	GeneE	GeneH
...
An additional description of the format can be found here.

Reference sets

Besides the list of relevant entities, the ORA method requires a second list of identifiers which represents the universe of identifiers that can be detected by an experiment. The input format is the same as for identifier lists.

Use a reference set that best fits your experimental platform. For microarrays, this would be the list of genes or transcripts for which probes are present on the array. For RNA-seq experiments this should be the list of all genes or transcripts that are present in the used genome annotation.

Troubleshooting

GeneTrail2 does not recognize my score list exported from Excel

MS Excel is a popular tool for managing biological datasets. However, there are some pitfalls especially when it comes to interoperability with other tools. It can happen that Excel reformats gene identifiers as dates. For example the gene Apr1 is routinely recognized as April the first. Please make sure, that no such conversions have taken place before exporting your data from Excel.

For more information see also Zeeberg et al. [6].

Bibliography

  1. Draghici, Sorin and Khatri, Purvesh and Martins, Rui P. and Ostermeier, G. Charles and Krawetz, Stephen A. Global functional profiling of gene expression Genomics Elsevier (View online)
  2. Love, Michael I and Huber, Wolfgang and Anders, Simon Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2 Genome Biol (View online)
  3. Robinson, Mark D and McCarthy, Davis J and Smyth, Gordon K edgeR: a Bioconductor package for differential expression analysis of digital gene expression data Bioinformatics Oxford Univ Press (View online)
  4. Risso, Davide and Ngai, John and Speed, Terence P and Dudoit, Sandrine Normalization of RNA-seq data using factor analysis of control genes or samples Nature biotechnology Nature Publishing Group (View online)
  5. Subramanian, Aravind and Kuehn, Heidi and Gould, Joshua and Tamayo, Pablo and Mesirov, Jill P GSEA-P: a desktop application for Gene Set Enrichment Analysis Bioinformatics Oxford Univ Press (View online)
  6. Zeeberg, Barry R and Riss, Joseph and Kane, David W and Bussey, Kimberly J and Uchio, Edward and Linehan, W Marston and Barrett, J Carl and Weinstein, John N Mistaken identifiers: gene name errors can be introduced inadvertently when using Excel in bioinformatics BMC bioinformatics BioMed Central Ltd (View online)