Loading…
Multimodal contrastive model aligning protein structure and sequence with ligand conformation and graph to retrieve binders without docking.
No providers recorded yet. Browse all providers
Glutathione peroxidase 4 scores 0.196 for druggability on a scale of 0 to 1: its surface is essentially flat, with no concave site for a docking program to place a ligand into — and docking is the step most virtual screening pipelines are built around. That geometry, shared by disordered proteins such as c-Myc and by the featureless interfaces of the RAS family, is why roughly 85% of clinically significant targets are treated as undruggable in practice.
Uni-Clip removes the pocket from the loop. Developed by DP Technology and posted as a preprint in August 2024, it is a CLIP-style contrastive model that embeds an entire protein and an entire ligand into one shared space and ranks compounds by embedding similarity, never by a scored pose. Retrieval asks only whether a protein and a molecule look compatible in that space, so the same checkpoint applies to targets whose shape defeats structure-based methods.
Its second idea addresses how contrastive training goes wrong on binding data. Standard InfoNCE assumes each protein in a batch has exactly one true ligand and treats every off-diagonal pair as a negative. The training set holds far more ligands than proteins, so a batch routinely contains several genuine binders of the same target being pushed apart — on the authors' data, 15% of off-diagonal labels on average outweighed the diagonal one. Their collision-free InfoNCE (CF-InfoNCE) loss builds a label matrix from assay values, keeps only the strongest pair per row and column, and normalizes both ways so protein-to-ligand and ligand-to-protein alignment are guided separately.
Uni-Clip totals roughly 100 million parameters. Both protein encoders are detached from training and their representations precomputed; the two ligand encoders are pretrained and then fine-tuned during alignment. Atom- and residue-level features are mean-pooled, then passed through a sparse-activation projection head before cosine similarity. Training used MMBind, assembled from ChEMBL33 binding assays paired with AlphaFold DB structures and RDKit-generated MMFF94 conformers: 853,932 protein–ligand pairs over 533,656 distinct ligands but only 5,245 distinct proteins. Every DUD-E and LIT-PCBA target was excluded, making both benchmarks zero-shot. On LIT-PCBA, Uni-Clip exceeds its baselines' average by 240%, 147% and 70% in enrichment factor at 0.5%, 1% and 5%; on DUD-E, by 162%, 218% and 216%, with the highest AUROC on both. Screening throughput is at least 300 times that of GPU AutoDock.
Two prospective campaigns show the intended workflow. For GPX4, five retrieval branches scored the ChemDiv (1,566,730 compounds) and Enamine (2,056,498) catalogs plus a 1,900-compound in-house set; 5,000 top-ranked molecules were reduced by drug-likeness filters and ECFP clustering to 36 purchased candidates, of which DP021 inhibited the enzyme with an IC50 of 4.17 µM and an SPR KD of 1.83 µM — a non-covalent chemotype, where known GPX4 inhibitors are largely covalent. Capping the carboxyl group that alanine scanning implicated in a salt bridge with Arg152 raised the IC50 above 100 µM. Run in reverse, the same embeddings did target fishing: ranking 330 tumor-associated human proteins against benzbromarone, an approved uricosuric drug, nominated c-Myc, which benzbromarone inhibited in an E-box reporter assay at 38.96 µM.
Uni-Clip's clearest contribution is CF-InfoNCE, since adopted by unrelated groups building contrastive screening models — a sign that the collision problem it names is general to affinity data rather than specific to this architecture. Reaching wet-lab confirmation is unusual for a screening paper, but both hits are micromolar starting points rather than leads, and benzbromarone's measured affinity for a c-Myc peptide is millimolar. The training corpus is the sharper limit: 5,245 proteins is narrow coverage of the proteome, and the authors name expanding it as the first requirement for better generalization. The work remains a preprint awaiting peer review, and the source code has not been released — the paper states it will follow acceptance — leaving non-commercial hosted servers on DP Technology's Bohrium platform as the only way to run the model.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.