Through sequence-based classification, this paper tries to accurately predictthe DNA binding sites of transcription factors (TFs) in an unannotated cellularcontext. Related methods in the literature fail to perform such predictionsaccurately, since they do not consider sample distribution shift of sequencesegments from an annotated (source) context to an unannotated (target) context.We, therefore, propose a method called "Transfer String Kernel" (TSK) thatachieves improved prediction of transcription factor binding site (TFBS) usingknowledge transfer via cross-context sample adaptation. TSK maps sequencesegments to a high-dimensional feature space using a discriminative mismatchstring kernel framework. In this high-dimensional space, labeled examples ofthe source context are re-weighted so that the revised sample distributionmatches the target context more closely. We have experimentally verified TSKfor TFBS identifications on fourteen different TFs under a cross-organismsetting. We find that TSK consistently outperforms the state-of the-art TFBStools, especially when working with TFs whose binding sequences are notconserved across contexts. We also demonstrate the generalizability of TSK byshowing its cutting-edge performance on a different set of cross-context tasksfor the MHC peptide binding predictions.
展开▼
