infercnv i3 HMM type
The inferCNV i3 HMM is highly similar to that used by HoneyBADGER, containing three states representing deletion, neutral, and amplification. The state model is illustrated below.
The 'preliminary inferCNV object' (reference cells subtracted and smoothed) is used as the target for HMM prediction of CNV. The expression intensity distributions corresponding to each of the CNV states are set based on the residual tumor cell expression intensities.
The assumption is that most of the tumor cell residual expression intensities are not found in CNV regions. We model the residual expression intensities according to a Gaussian distribution parameterized by Normal(mu, sigma) as estimated based on all residual tumor expression intensities.
Those residual expression intensities that would be considered significantly different from this distribution would be considered evidence for potential amplification or deletion. As in HoneyBADGER, we determine the mean value for amplification or deletion state expression intensities based on a KS-test, requiring that the alternative mean expression intensity fit a distribution with the same variance and be significantly different from the neutral distribution at a p-value of 0.05 (default, 'infercnv::run(HMM_i3_pval=0.05)'). This parameterization depends on the number of (k) cells being analyzed within a CNV block, where we model the distribution of the mean for (k) cells. The greater the number of (k) cells, the tighter the variance for our estimate of the mean (as per the Central Limit Theorem).
The HMM state emissions do not directly reflect emission probabilities from these distributions, but rather the reflect the relative likelihood of being emitted from the corresponding distribution.
note, insert math equation