Seed and result replication

[0Proma0]

Hi, I tried to regenerate my trajectory due to crash on MPNN step (trajectory had gread quality). I extracted seed and length which are randomny generated in bindcraft.py and passed them for another run.
Results are inconsistent within runs.

How do you reproduce specific generation? What is the purpose of seeds if they alone are insufficient for reproduction?

Extraction from pdb name (my code):

length = int(name_parts[-2].replace('l', ''))
seed = int(name_parts[-1].replace('s', ''))

Original randomisation:

seed = int(np.random.randint(0, high=999999, size=1, dtype=int)[0])
length = np.random.choice(samples)

Probable cause could be Collab ulits where random mutations are used without any seeds. I used 4stage generation.

uses gradient descend to get a PSSM profile and then uses PSSM to bias the sampling of random mutations to decrease loss
        af_model.design_pssm_semigreedy(soft_iters=advanced_settings["soft_iterations"], hard_iters=advanced_settings["greedy_iterations"], tries=greedy_tries, models=design_models, 
                                        num_models=1, sample_models=advanced_settings["sample_models"], ramp_models=False, save_best=True)

advanced_settings["design_algorithm"] == 'greedy':
        # design by using random mutations that decrease loss
        af_model.design_semigreedy(advanced_settings["greedy_iterations"], tries=greedy_tries, num_models=1, models=design_models,
                                sample_models=advanced_settings["sample_models"], save_best=True)

advanced_settings["design_algorithm"] == 'mcmc':
        # design by using random mutations that decrease loss
        half_life = round(advanced_settings["greedy_iterations"] / 5, 0)
        t_mcmc = 0.01
        af_model._design_mcmc(advanced_settings["greedy_iterations"], half_life=half_life, T_init=t_mcmc, mutation_rate=greedy_tries, num_models=1, models=design_models,
                                sample_models=advanced_settings["sample_models"], save_best=True)