Clonealign 2.0 contains several updated changes, both to the core model and inference, as well as to package functionality.
- While the modelled expectation remains identical, the likelihood employed is now multinomial rather than negative binomial. This speeds up inference and removes the need for size factor and variance calculation.
- The preferred package entrypoint is the run_clonealign function, that will run clonealign across a range of initial parameter values and return the fit that maximizes the ELBO.
- Post-hoc calculated correlations between the copy number and gene expression are calculated and assigned to the $correlations slot
- Multinomial likelihood rather than negative binomial
- mu[1] no longer constrained to be 1 to improve optimization in some cases
- It is no longer necessary to specify size factors as the multinomial distribution implicitly conditions on the total counts per cell
- Convergence is monitored by looking at the average change in the previous 10 iterations rather than the single previous iteration, which can be sensitive to random fluctuations