Thanks- personally, based on my experience in functional genomics, they have far too many assumptions to be trusted as "clocks", even relative ones.
Here's the second paper's methods:
a molecular clock signal was estimated on the ML tree using TempEst v1.532. BEAST v1.10.433 was used for Bayesian MCMC analysis to estimate temporal nodes. As recommended in Patrono et al., we selected the simplest model: strict clock and constant population size with a HKY substitution model and four Gamma categories7. An MCMC model was run with 50,000,000 generations and a burn-in of 25,000 sampled trees. All other parameters were by default. BEAST executions were completed checking chain convergence and sufficient sampling of the posterior space (ESS > 200) with Tracer v1.7.1. The final chronogram was generated using TreeAnnotator v1.10.433.
I'd call these "time-scaled trees", not molecular clocks.
Their accuracy is completely subject to stable mutation rate and frequency of sample -- with that second paper, they were working with ~10 samples from 15 years ago vs. thousands of samples with hourly granularity and entirely traceable transmission chains in the early days of Covid. Completely different capabilities arise with Covid-level specificity.
It's more a matter of decreasing the error bars on the rate of substitution, to have extremely specific dates for the oldest infections in each clade and the ability to run real-time backtesting to make sure your assumptions are solid. The rate of spread of Covid was a curse but was important to this type of work -- when we were looking at contact tracing and trying to assess the rate of substitution on mutations, it was useful to know that e.g. on Jan 1st there was no Covid in NYC and on Feb 1st there was.
My point is that if you find multiple samples with the same protein sequence, and those samples have independently diverged, you can easily assume those viruses did not originate by recombination from the Wuhan strain.
Here's the second paper's methods: a molecular clock signal was estimated on the ML tree using TempEst v1.532. BEAST v1.10.433 was used for Bayesian MCMC analysis to estimate temporal nodes. As recommended in Patrono et al., we selected the simplest model: strict clock and constant population size with a HKY substitution model and four Gamma categories7. An MCMC model was run with 50,000,000 generations and a burn-in of 25,000 sampled trees. All other parameters were by default. BEAST executions were completed checking chain convergence and sufficient sampling of the posterior space (ESS > 200) with Tracer v1.7.1. The final chronogram was generated using TreeAnnotator v1.10.433.
I'd call these "time-scaled trees", not molecular clocks.