The number of samples is not equivalent across all states

When I run the tutorial example using my protein and two small molecules, I got the following error:

openfe/lib/python3.12/site-packages/openfe/protocols/openmm_utils/multistate_analysis.py", line 322, in get_forward_and_reverse_analysis
raise ValueError(errmsg)
ValueError: The number of samples is not equivalent across all states [ 0 878 878 878 878 878 878 878 878 878 878 878 0]

It seems that lambda=0 and 1 are not sampled properly. But, I don't know why this happens.

How can I fix this?

The below script shows my setting:

rbfe_settings = RelativeHybridTopologyProtocolSettings(
protocol_repeats=3, # Number of independent repeats of the Protocol transformation
forcefield_settings=equil_rfe_settings.OpenMMSystemGeneratorFFSettings(
constraints='hbonds', # 'hbonds': Use constraints for bonds involving hydrogen
rigid_water=True, # True: Use constraints for bonds in water
hydrogen_mass=3.024, # Perform hydrogen mass repartitioning
forcefields=[ # OpenMM force fields to use for solvents and proteins
'amber/ff14SB.xml',
'amber/tip3p_standard.xml',
'amber/tip3p_HFE_multivalent.xml',
'amber/phosaa10.xml'
],
# Small molecule force field to use with OpenMM template generator:
#small_molecule_forcefield='openff-2.1.1',
small_molecule_forcefield='gaff-2.11',

    # Nonbonded settings
    nonbonded_method='PME',            # Particle Mesh Ewald for long range electrostatics
    nonbonded_cutoff=1.0 * unit.nm,    # Cut off Lennard-Jones interactions beyond 1 nm
),
thermo_settings=equil_rfe_settings.ThermoSettings(
    temperature=298.15 * unit.kelvin,  # Set thermostat temperature
    pressure=1 * unit.bar,             # Set barostat pressure
    ph=7.4,                            # None: Do not keep pH constant
    redox_potential=None               # None: Do not keep redox potential constant
),
solvation_settings=equil_rfe_settings.OpenMMSolvationSettings(
    solvent_model='tip3p',             # Solvent model to generate starting coords
    solvent_padding=1.2 * unit.nm,     # Total distance between periodic image starting coords
),
partial_charge_settings=equil_rfe_settings.OpenFFPartialChargeSettings(
    partial_charge_method='am1bcc',    # Partial charge method applied - am1bcc
    off_toolkit_backend='ambertools',  # Toolkit to use for partial charge assignment - ambertools
    number_of_conformers=None,         # None: use input conformer for partial charge assignment
    nagl_model=None,                   # None: not using NAGL so no model needs to be chosen
),
lambda_settings=equil_rfe_settings.LambdaSettings(
    lambda_functions='default',        # Interpolation functions for force field parameters
    lambda_windows=11,                 # Split the transformation over n lambda windows
),
alchemical_settings=equil_rfe_settings.AlchemicalSettings(
    # False: Don't use unsampled (non-hybrid) endstates for long range correction
    endstate_dispersion_correction=True,
    use_dispersion_correction=True,   # False: Don't use dispersion correction
    softcore_LJ='gapsys',              # Use LJ potential from Gapsys et al. (JCTC 2012)
    softcore_alpha=0.85,               # Set soft-core Lennard-Jones potential parameter α
    # False: Keep all exceptions (1,4 or otherwise) at all λ
    tun_off_core_unique_exceptions=False,

    # Explicit charge correction settings
    # False: don't apply explicit charge correction using an alchemical water
    explicit_charge_correction=True,
    # Cutoff distance for choosing alchemical waters
    explicit_charge_correction_cutoff=0.8 * unit.nm,
),
simulation_settings=equil_rfe_settings.MultiStateSimulationSettings(
    # Simulation lengths
    minimization_steps=5000,                    # Minimize potential energy for n steps
    equilibration_length=1.0 * unit.nanosecond, # Simulation time to equilibrate for
    production_length=5.0 * unit.nanosecond,    # Simulation time to collect data for

    # Alchemical Space Sampling settings
    n_replicas=11,                           # Number of replicas sampling alchemical space
    sampler_method='repex',                  # Sample lambda with Hamiltonian Replica Exchange
    time_per_iteration=1*unit.ps,            # Time interval between state sampling (MCMC) attempts

    # SAMS sampling settings (used if sampler_method='sams')
    sams_flatness_criteria='logZ-flatness',  # Criteria for switch to asymptomatically optimal scheme
    sams_gamma0=1.0,                          # Initial SAMS weight adoption rate.

    # Settings to control free energy analysis
    # Time interval at which to perform an analysis of the free energies
    real_time_analysis_interval=250*unit.picosecond,
    # Minimum simulation time before energy analysis is carried out
    real_time_analysis_minimum_time=500*unit.picosecond,
    # Stop simulation if this target error is reached:
    early_termination_target_error=0.0*unit.kilocalorie_per_mole,
),
engine_settings=equil_rfe_settings.OpenMMEngineSettings(
    compute_platform='CUDA',              # Let OpenMM choose the best platform for your hardware
),
integrator_settings=equil_rfe_settings.IntegratorSettings(
    timestep=4 * unit.femtosecond,         # Integration timestep
    langevin_collision_rate=1.0 / unit.picosecond,  # Langevin integrator collision rate γ
    reassign_velocities=False,             # False: Velocities are not lost through MCMC moves
    n_restart_attempts=20,                 # Restart simulations the first n times they blow up
    constraint_tolerance=1e-06,            # Tolerance for holonomic constraints
    barostat_frequency=25 * unit.timestep, # Attempt MC volume scaling every n integration steps
    remove_com=False,                      # False: don't remove the center of mass motion
),
output_settings=equil_rfe_settings.MultiStateOutputSettings(
    output_filename='simulation.nc',            # Filename to save trajectory
    output_structure='hybrid_system.pdb',       # Filename to save starting coordinates
    checkpoint_storage_filename='checkpoint.chk',  # Filename for simulation checkpoints
    forcefield_cache='db.json',                 # Cache for small molecule residue templates
    output_indices='not water',                 # Do not save water positions
    checkpoint_interval=250 * unit.ps,          # Save a checkpoint every 250 picoseconds
),

)

Thanks in advance for your help.

Start with openfe/protocols/openmm_utils/multistate_analysis.py at get_forward_and_reverse_analysis, using the reported sample-count error as the reproduction target. Inspect simulation.nc and the listed multistate simulation settings to determine why the endpoint states have zero samples; done means identifying a reproducible cause and confirming that all states contain equivalent samples.