Varying free energy estimates depending on sampler used

[thomas-elix]

Hello Yank team,
I am trying to run a benchmark using the data available here https://github.com/MCompChem/fep-benchmark .
I have been running simulations with different samplers and end up with widely varying free energy estimation depending on the sampler, and even on the run.
Running with REMD for about 5ns, I get binding energy values that are way too high (around -19kcal/mol) for all the ligand-protein complex I simulated so far (when experimental values are around -9kcal/mol). I am using around 30 to 40 lambda windows, and analysis of the simulations seems to indicate good mixing between replicas.
Running with SAMS sampler for a total of 36, 60 or 96ns (the number of iteration spent in each replica is roughly the same, so anywhere between 1 and 3ns per replicas), I respectively got - 6.804 ± 0.868 kcal/mol, - 6.512 ± 0.785 kcal/mol and - 2.350 ± 0.853 kcal/mol when the experimental value is of -8.73 kcal/mol. The simulations parameters and starting files (prepared by hand) are the same, only the number of iterations was modified.
I have troubling identifying the reason REMD is constantly overshooting the free energy estimates, and the inconsistencies with the SAMS sampler.
I'd be happy to share more informations or files if necessary! Thanks