C-G.C.18: AD-CLIP: Adenylation Domain-Substrate Contrastive Learning for Virtual Screening
Authors
Keywords
NRPS, Adenylation domain, specificity, retrieval
[sponser-meet-now-chat][/sponser-meet-now-chat]
Non-Ribosomal Peptide Synthetases (NRPSs) are modular enzymatic assembly lines that synthesise Non-Ribosomal Peptides (NRPs), secondary metabolites often used as therapeutics. Within NRPSs, Adenylation (A) domains dictate substrate specificity, governing the chemical composition of the NRP. Computational efforts have traditionally framed A-domain specificity prediction as a supervised classification problem, producing discrete predictions over a closed substrate vocabulary. This is at odds with NRPS research, where novel specificities are routinely characterised, requiring output-layer extension. We propose AD-CLIP, the first retrieval-based approach to A-domain specificity, embedding A-domains and substrates in a shared chemical-protein space where compatibility is a continuous similarity score and substrate vocabulary remains open. We curate 4,250 A-domain sequences spanning 43 substrates. We use 16 specificity-determining residues, partially overlapping with canonical Stachelhaus residues, identified in our prior classification work. The dual encoder pairs an MLP for substrates with an A-domain module of per-residue MLPs, an E(3)-equivariant graph neural network (EGNN) over Boltz2-predicted Cα coordinates, and mean pooling. We evaluate AD-CLIP using recall and Normalised Discounted Cumulative Gain (NDCG) under two scenarios: zero-shot retrieval, and n-shot continual pretraining (n ∈ {1, …, 5}), where pretraining continues on incrementally added examples per held-out substrate. Zero-shot retrieval achieves mean recall@200 = 0.61 and NDCG@200 = 0.45. AD-CLIP adapts rapidly under minimal supervision: a single example raises NDCG@200 / recall@200 to 0.74 / 0.85, averaging to 0.97 / 0.91 with five while avoiding catastrophic forgetting. AD-CLIP is thus effective and extensible, accommodating new substrates at minimal cost.
Co-authors: Kenan Bozhüyük, Olga Kalinina, Zhengjian Li, Dietrich Klakw
Contact Attendee
Warning: Attempt to read property "user_email" on string in /home/1276969.cloudwaysapps.com/ydbgzhdjeq/public_html/wp-content/plugins/my-conference-now/functions.php on line 2826