For complex multidimensional systems, Monte Carlo methods are useful for sampling probable regions of a configuration space and, in the context of annealing, for determining "low energy" or "high scoring" configurations. Such methods have been used in protein design as means to identify amino acid sequences that are energetically compatible with a particular backbone structure. As with many other applications of Monte Carlo methods, such searches can be inefficient if trial configurations (protein sequences) in the Markov chain are chosen randomly. Here a mean-field biased Monte Carlo method (MFBMC) is presented and applied to designing and sampling protein sequences. The MFBMC method uses predetermined sequence identity probabilities w(i)(alpha) to bias the sequence selection. The w(i)(alpha) are calculated using a self-consistent, mean-field theory that can estimate the number and composition of sequences having predetermined values of energetically related foldability criteria. The MFBMC method is applied to both a simple protein model, the 27-mer lattice model, and an all-atom protein model. Compared to conventional Monte Carlo (MC) and configurational bias Monte Carlo (BMC), the MFBMC method converges faster to low energy sequences and samples such sequences more efficiently. The MFBMC method also tolerates faster cooling rates than the MC and BMC methods. The MFBMC method can be applied not only to protein sequence search, but also to a wide variety of polymeric and condensed phase systems. (C) 2003 American Institute of Physics.