However, the major challenge when comparing samples that were sequenced using barcoded amplicon sequencing is that it is difficult to control the number of sequences per sample, AKA the library depth. The problem is not unique to analysing SIP experiments and poses a major analytical challenge in the field of microbiome studies and comparative transcriptomics (RNA-Seq). In essence, most statistical methods used for comparison assume that across different samples templates with identical relative abundance should have equal chances of being sequenced and thus any observed in abundance are an indication that the true abundance of the given sequence differs between the samples. In macroecology the issue is known as "sampling effort". Traditionally, the most common way to alleviate the problem of unequal sequencing depths was to randomly subsample sequences from each sample down to the smallest sample size so that all samples become equal (a process sometimes called "rarefaction"). This practice, however, came under heavy scrutiny in recent years have sparked some heated polemic papers in the field of microbiome studies \cite{McMurdie_2014}. While the severity of the bias caused by random subsampling is debated, it is generally accepted that this is a suboptimal way to deal with the problem. Another common approach is to convert all abundances to relative abundances and compare the different sequences on a fraction (or percentage) basis. This, however, leads to other problems since it maintains the correlation between sequencing depth and the number of unique sequences (or OTUs) while at the same time drastically reducing the number of degrees of freedom by coercing the sum of abundance in each sample to 100% \cite{Faust_2012}. More recent methods try to "eat the cake and leave it whole" by attempting to equalise the variance between samples through a scaling factor while not discarding any data \cite{Weiss_2017}. Whichever method is chosen eventually it is important to remember that no statistical trick can solve the inherent problems that stem from large differences in library sizes and these should be handled at the level of sample preparation or sequencing and not data analysis.
The most common methods for comparing fractions in SIP experiments were developed for analysing RNA-Seq datasets. The parallels are obvious; typical RNA-Seq experiments are designed as a case-control study and the analytical challenge is to identify which sequences are differentially expressed (either up-regulated or down-regulated) compared to the control, while overcoming the natural variance and differences in library sizes. Similarly, in SIP experiments one would like to identify which sequences are "differentially abundant" in the fractions where labelled nucleic acids are expected to be present compared to those where unlabelled nucleic acids are present. An important difference to RNA-Seq experiments is however that only enriched sequences in the labelled fraction are of interest, while depleted sequences should only occur when labelling is strong enough to essentially displace unlabelled sequences.
Nearly all existing data analysis methods should apply to both DNA- and RNA-SIP, albeit with some differences. This book offers two recent and very promising ways to analyse SIP datasets: quantitative SIP (qSIP; Chapter XX) and High-Resolution SIP (HR-SIP; Chapter YY). Both yield similar results but they nevertheless differ in some important details. Since the two methods are carefully detailed in this book, repeating the steps here would be redundant. However, because the methods were published for DNA-SIP, some differences to RNA-SIP should be noted. In principle, both methods rely on a comparison of the gradient fractions from labelled samples to those from unlabelled control samples (between-gradient comparison). Moreover, both assume and make use of the fact that while DNA and RNA will concentrate around their theoretical BD they diffuse throughout the gradient in a Gaussian shape so that amplifiable amounts of nucleic acids are present in every fraction in the gradient \cite{Angel_2017,Youngblut_2018}. However, because the course of development of a microbial community is controlled by stochastic processes in addition to deterministic ones, parallel incubations from the same parent community often lead to different communities after a while, even if conditions are kept as similar as possible. Consequently, it was demonstrated that these stochastic variations reduce the detection accuracy and it was recommended that the Bray-Curtis dissimilarity between communities of labelled and unlabeled samples that are being compared should ideally be >0.2 \cite{Youngblut_2018}. Between-gradient comparisons are crucial for DNA-SIP because as mentioned above, the DNA of different taxa will migrate in the gradient also based on their G+C content. In RNA-SIP however, RNA is believed to migrate based on BD only and one can assume that in a gradient from an unlabelled sample the relative abundance of each taxon should remain relatively constant throughout the different fractions. In contrast, in a gradient from a labelled sample some taxa will be more abundant in the heavy fractions compared to the lighter ones, while the relative abundance of unlabelled taxa will remain constant throughout the gradient or decline in the heavy fractions, if the labelled taxa make up a significant proportion of the entire community. In any case, since in RNA-SIP differential migration of taxa is only expected as a response of labelling, detection of labelled taxa can also be done in a within-gradient fashion by comparing the relative abundances of taxa in the heavy fractions (i.e. ca. 1.72--1.76 g ml-1 for DNA-SIP or 1.80--1.84 g ml-1 for RNA-SIP) with those in the light fractions (i.e. ca. 1.68--1.72 g ml-1 for DNA-SIP or 1.77--1.80 g ml-1 for RNA-SIP). However, some label-free controls should nevertheless be set up (e.g. paralleling the beginning and end time points or the highest and lowest treatment extremes) and analysed because they can help to fine-tune the statistical cutoff parameters so that false positives can be avoided \cite{Angel_2017}.
differential abundance
qSIP
network analysis
metagenome/transcriptome