Proceedings of the 25th Belgian-Dutch Machine Learning Conference (Benelearn) pages:1-2
Annual Machine Learning Conference of Belgium and The Netherlands edition:25 location:Kortrijk, Belgium date:12-13 September 2016
Recently, we developed the 3D Neighborhood Kernel (3DNK), which acts on 3D structures of small molecules and proteins. We showed its state-of-the-art performance on several biological datasets. However, 3D data are in many cases difficult to obtain. For this reason, we adopt a different strategy: instead of requiring actual 3D structures, we use as input protein sequences, of which we approximate the 3D structure through homology modelling. Then, we apply 3DNK on the approximated 3D protein structures and show that, on the task of predicting HIV resistance, we obtain better results than when using a kernel function based on the protein sequences alone.