Abstract
<jats:p>Predicting drug-target interactions (DTI) for entirely unseen drugs or proteins---the cold-start problem---remains a critical challenge in computational drug discovery. While sequence-based methods naturally support zero-shot generalization, they often ignore relational topology, and existing graph-based approaches either rely on global diffusion that blurs the boundary between inductive and transductive evaluation or require a few known interaction samples at test time (few-shot). We present EBD-DTI, a framework that enables zero-shot inference in graph-based DTI models without requiring any known interactions for unseen entities. The key innovation is episodic cold-start training: at each epoch, a random subset of training entities is masked and treated as pseudo-cold, forcing the model to learn cold-start inference with explicit gradient supervision. A bridge-conditioned local subgraph, together with multi-hop diffusion, provides cold entities with relational context from their nearest observed neighbors. Experiments on three benchmarks (BioSNAP, BindingDB, and DrugBank) demonstrate that EBD-DTI achieves competitive or superior performance compared to state-of-the-art methods under strict zero-shot evaluation, with episodic training improving AUC by up to 12%.</jats:p>