Background: Newly-evolved multiplex sequencing technology has been bringing transcriptome sequencing into an unprecedented depth. Millions of transcript tags now can be acquired in a single experiment through parallelization. The significant increase in throughput and reduction in cost required us to address some fundamental questions, such as how many transcript tags do we have to sequence for a given transcriptome? How could we estimate the total number of unique transcripts for different cell types (transcriptome diversity) and the distribution of their copy numbers (transcriptome dynamics)? What is the probability that a transcript with a given expression level to be detected at a certain sampling depth?
Methodology/principal findings: We developed a statistical model to evaluate these parameters based on transcriptome-sampling data. Three mixture models were exploited for their potentials to model the sampling frequencies. We demonstrated that relative abundances of all transcripts in a transcriptome follow the generalized inverse Gaussian distribution. The widely known beta and gamma distributions failed to fulfill the singular characteristics of relative abundance distribution, i.e., highly skewed toward zero and with a long tail. An estimator of transcriptome diversity and an analytical form of sampling growth curve were proposed in a coherent framework. Experimental data fitted this model very well and Monte Carlo simulations based on this model replicated sampling experiments in a remarkable precision.
Conclusions: Taking human embryonic stem cell as a prototype, we demonstrated that sequencing tens of thousands of transcript tags in an ordinary EST/SAGE experiment was far from sufficient. In order to fully characterize a human transcriptome, millions of transcript tags had to be sequenced. This model lays a statistical basis for transcriptome-sampling experiments and in essence can be used in all sampling-based data.