Learning patterns in combinatorial protein libraries by Support Vector Machines

Recent advances such as directed evolution and high throughput experiments can generate recombinant protein libraries and screen them for properties of interest. However it is impractical to span the theoretical range of combinatorial library and hence predictive models using the limited experimental data are of invaluable use. In this work, we have developed a novel machine learning strategy using Support Vector Machine (SVM) to predict the folding nature of recombinant proteins from Cytochrome P450 family using available experimental data. The folding-status is determined by an empirical energy model based on pair-wise interactions. It is shown that applying similarity-kernel function to the SVM formulation enables inclusion of many body interaction terms without additional computational effort. This approach can be generalized to other recombinant families and different properties of interest. The inferences derived by analyzing the data using the new method are in agreement with published results.