Neural network modeling of serum protein fractionation using gel filtration chromatography

A neural network model as a macroscopic model is proposed to estimate the impurity and yield of fractionated BSA monomer in gel filtration chromatography (GFC) when the viscosity of an injected bovine serum is high and its volume is large. In the neural network model, input variables, i.e. the partition coefficient of BSA, the injection interval, and the fractionation time coefficient, are selected referring to physical models published. The trained network model shows sufficient predictive performance and applicability when data points are estimated by interpolation. Furthermore, it is seen that the sensitivity of output variables to the injection interval and the fractionation time coefficient can be easily estimated using the model. Consequently the neural network model is found useful enough to easily predict the results of changing the operation, especially when a large injection volume of a viscous sample causes nonlinear and unstable behavior in the GFC fractionation process.