This paper presents a flexible approach to the predictive modeling and optimal design of extractive bioseparations in aqueous two-phase systems (ATPS) using polyethylene glycol (PEG) and dextran. We combine the qualitative reasoning skills of an extractive-separation expert system with the quantitative modeling capabilities of a protein-partitioning neural network. The resulting hybrid system, called an expert network, provides an accurate and efficient tool for quantitative predictions of phase diagrams and partition coefficients (i.e., separation factors). We demonstrate the use of the neural-network and response-surface modeling as a bioseparation optimizer to facilitate experimental design and process development of extractive separation flowsheets for multicomponent protein mixtures in ATPS. We describe the advantages and limitations of our protein-partitioning network when compared to available theoretical models for protein partitioning. In particular, our protein-partitioning network is compartmentalized into physically identifiable subsystems. Therefore, research advances in specific areas can be easily incorporated without major development or reconstruction.