DeepTernary

DeepTernary predicts the 3D structure of the ternary complexes that drive targeted protein degradation: an E3 ubiquitin ligase, a small molecule, and a target protein in one assembly. It came from the ReLER Lab at the University of Technology Sydney with co-authors at Nankai University, Renmin University of China, the University of Chinese Academy of Sciences, Zhejiang University, and UCLA, posted in February 2025 and published in Nature Communications that July.

Proteolysis-targeting chimeras (PROTACs) and molecular glue degraders (MGDs) work by inducing an interaction that does not otherwise exist, and whether a candidate degrader forms a productive ternary complex is the question that decides a program. Structure prediction has transformed the binary case — protein-protein and protein-ligand — but the ternary case stayed hard for a specific reason: only a few dozen PROTAC and MGD ternary structures have been solved experimentally, far too few to train on. Existing approaches lean on rule-based sampling with extensive random search, which is slow and inaccurate.

The paper's central bet is that ternary complexes obey the same atomic interaction principles whether or not the bridging molecule is a designed degrader. Acting on that, the authors assembled TernaryDB from ordinary PDB entries containing two proteins and a shared small molecule, deliberately excluding all known PROTACs and MGDs, then trained an SE(3)-equivariant graph neural network on it. The resulting model has never seen a PROTAC during training and still predicts them, which is strong evidence that it learned interaction physics rather than a target class.

#Key Features

• End-to-end ternary prediction: An encoder-decoder network outputs the assembled complex directly, replacing rule-based pose enumeration and random sampling.
• Ternary inter-graph attention: Each component is a graph, and attention operates within graphs and across all three, so ligase, ligand, and target representations are shaped by mutual context.
• Pocket points decoder: A query-based decoder reads the binding interface out of the learned ternary embedding, yielding both the small molecule conformation and the docking pose.
• Held-out degrader evaluation: MMseqs2 clustering removed every training cluster containing a known PROTAC or MGD, so benchmarks measure generalization rather than retrieval.
• Confidence ranking: A predicted aligned error score orders outputs from multiple RDKit-generated ligand conformers; an acceptable prediction is typically within the top four of 40 samples.

#Technical Details

TernaryDB comprises 22,303 filtered ternary complexes from the PDB, spanning 363 species and a chemically diverse ligand set whose members mostly carry fewer than 60 heavy atoms. Sequence clustering yields 1,398 PROTAC-safe clusters over 16,203 complexes and 1,982 MGD-safe clusters over 22,046 complexes, and cluster-wise sampling prevents large families from dominating batches. The network is an SE(3)-equivariant GNN over three graphs — protein 1, ligand, protein 2 — with RMSD to protein 2 as a training signal.

On the PROTAC benchmark DeepTernary reaches an average DockQ of 0.65 against 0.44 for BOTCP, and outperforms FRODock- and RosettaDock-based methods, PRosettaC, and other published pipelines on CAPRI quality rates and the fraction of predictions under 10 A RMSD, with success rates near 50%. Mean top-ranked ligand RMSD is 3.43 A, and 43% of PROTAC predictions fall under 2 A; the harder blind-docking MGD benchmark gives a DockQ of 0.21. Inference takes roughly 7 seconds per PROTAC complex and about 1 second per MGD complex. Comparison against AlphaFold 3 and Chai-1 required filtering to complexes outside their training cutoffs; on the three PROTACs AlphaFold 3 had not trained on, its best DockQ scores were 0.12, 0.25, and 0.48 against 0.56, 0.67, and 0.53 for DeepTernary. Buried surface areas of predicted PROTAC complexes fall in the 1,100-1,500 range reported for potent degraders and correlate with experimental potency metrics.

#Applications

The model serves targeted protein degradation programs, where the questions are which ligase to pair with a target, which linker geometry gives a cooperative complex, and which candidates to make first. Predicted structures give medicinal chemists a starting hypothesis for linker design, and the buried surface area correlation offers a cheap proxy for potency during triage. Because training used no degrader-specific data, it applies to novel ligase-target pairings with no analogous crystal structure — the situation motivating most degrader campaigns.

#Impact

DeepTernary is the first deep learning method applied to PROTAC ternary structure prediction and the first in-silico approach reported for molecular glue degrader complexes, and its data strategy is the transferable idea: when a target class is too rare to train on, train on the general phenomenon and let the specific case fall out. Code and the pretrained checkpoint are released under Apache-2.0 with a standalone prediction script, alongside the TernaryDB complex list, cluster assignments, and training logs. Two caveats matter. A DockQ of 0.21 under blind docking is weak in absolute terms, reflecting how much harder glue-induced interfaces are than PROTAC ones, and the potency correlation is a retrospective association across predicted structures, not prospective validation of a designed degrader.