AlphaFold2-RAVE combines reduced-MSA AlphaFold2 (ColabFold) structure prediction with metadynamics-biased molecular dynamics. ColabFold is modified by reducing MSA cluster sizes and using multiple random seeds to introduce stochasticity, generating a diverse pool of candidate structures. SPIB (State Predictive Information Bottleneck) then learns 2D collective variables from short unbiased MD trajectories seeded at clustered AF2 outputs. Those CVs drive a well-tempered metadynamics simulation with PLUMED+OpenMM to recover Boltzmann-weighted conformational ensembles.