Figure 8

a. Synthetic data with Poisson distributed length statistics is shown in the upper trace. Emission broadening is introduced with an emission variance amplification factor of 20. This effectively broadens the noise band (thickness) seen in the upper trace by a factor of 20, which leads to a blurring between the upper and lower levels of blockade since the noise bands now overlap (i.e., a toggling cross-over instability is introduced to challenge the projection method). The middle trace shows the clean, highly accurate Viterbi parsing into the appropriate levels that is still obtained with the HMM-with-Duration implementation. The lower trace shows the Viterbi parse with a simple HMM, that is uninformed about the underlying length distributions, thus giving rise to a Viterbi traceback parse that fails to penalize unlikely, very short duration, blockade events (seen as the unstable, rapid level-projection toggles). b. Synthetic data with Gaussian distributed length statistics is shown in the upper trace, with the suceessful HMM-with-Duration parsing shown in the middle trace. Emission broadening is introduced with an emission variance amplification factor of 20 as in Fig. 8a, with a similar failure in the HMM-without-Duration's ability to parse critical kinetic feature information.