Ligand affinity prediction with multi-pattern kernels

We consider the problem of affinity prediction for protein ligands. For this purpose, small molecule candidates can easily become regression algorithm inputs if they are represented as vectors indexed by a set of physico-chemical properties or structural features of their molecular graphs. There are plenty of so-called molecular fingerprints, each with a characteristic composition or generation of features. This raises the question which fingerprint to choose for a given learning task? In addition, none of the standard fingerprints, however, systematically gathers all circular and tree patterns independent of size and the adjacency information of atoms. Since structural and neighborhood information are crucial for the binding capacity of small molecules, we combine the features of existing graph kernels in a novel way such that finally both aspects are covered and the fingerprint choice is included in the learning process. More precisely, we apply the Weisfeiler-Lehman labeling algorithm to encode neighborhood information in the vertex labels. Based on the relabeled graphs we calculate four types of structural features: Cyclic and tree patterns, shortest paths and the Weisfeiler-Lehman labels. We combine these different views using different multi-view regression algorithms. Our experiments demonstrate that affinity prediction profits from the application of multiple views, outperforming state-of-the-art single fingerprint approaches.