ADCLIP: Adenylation Domain-Substrate Contrastive Learning for Virtual Screening

Within non-ribosomal peptide synthetases, adenylation domains govern substrate selection and, subsequently, the chemical composition of the resulting peptide. Adenylation domain specificity has been predominantly approached as a closed-set classification problem, limiting predictions to substrates seen during training. We propose ADCLIP, the first bidirectional retrieval-based approach to adenylation domain specificity. Adenylation domains and substrates are embedded in a shared space where compatibility is a continuous similarity score, enabling retrieval in both directions: given a substrate, ADCLIP ranks compatible adenylation domains, and given an adenylation domain, it can rank substrates beyond those seen during training. ADCLIP is computationally lightweight, relying solely on one-hot encodings and a contrastive learning objective. It adapts rapidly under minimal supervision, raising mean NDCG@K, Recall@K, and MRR from 0.16/0.28/0.33 at zero-shot to 0.85/0.81/0.87 with only three labeled examples per test substrate.