NCGAMI: Domain-Adaptive Drug-Target Interaction Prediction with Graph and Sequence Models

Ellen Yi-Ge, Taric Chen

Keywords: Graph Representation, Domain Adaptation, Distrbution Alignment

Abstract: Drug-target interaction (DTI) prediction is of central importance in computational pharmacology, but how robust these predictions are in the face of distribution shift (e.g. between chemical scaffolds, or protein families) remains challenging. We introduce a well-behaved, label-free objective and encoder fusion recipe for DTI, which is referred to as NCGAMI, under an unsupervised domain adaptation (UDA) training paradigm. NCGAMI only uses a graph encoder for the drug molecule graph and a sequence encoder for the proteins, and uses a three-term objective, consisting of (i) the supervised source-domain risk, (ii) an explicit cross-domain representation alignment, and (iii) regularization of the target domain through conditional-entropy minimization and prediction-consistency. The implementation is possible without using target labels. Throughout, we take great care to avoid using target labels during training and to clarify situations where there may be overlap between source and target entities. Using two widely used DTI datasets (Human and DrugBank), and using a protocol (random split) that we applied strictly for feasibility check, NCGAMI achieves good AUC/AUPR and is competitive against representative baselines. On Human our model achieves AUC 0.895 and AUPR 0.852 on Drugbank our model achieves AUC 0.733 and AUPR 0.675. Ablations can be seen to contribute the graph encoder, sequence encoder, and UDA regularization. We also incorporate non-operational geometric background so that theoretical assertion is motivating in your choice of design without overwhelming you with strong theoretical claims.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 9865