Using non-linear regression to predict bioresponse in a combinatorial library of biodegradable polymers

We have developed an empirical method to model bioresponse to the surfaces of biodegradable polymers in a combinatorial library using Artificial Neural Networks (ANN) in conjunction with molecular modeling and machine learning methodology. We validated the procedure by modeling human fibrinogen adsorption to 22 structurally distinct polymers. Subsequently, the method was used to model the more complicated phenomena of rat lung fibroblast and normal human fetal foreskin fibroblast proliferation in the presence of 24 and 44 different polymers, respectively. In each case, the root mean square (rms) percent error of the prediction was substantially less than the experimental variation, showing that the models can distinguish high and low performing polymers based on structure/property information. Using this method to screen candidate materials in terms of specific bioresponse prior to extensive experimental testing will greatly facilitate materials development for biomedical applications.