PROTAC degradation prediction from molecular graphs of the target, linker, and E3 ligase, combining cross-attention with contrastive learning.
No providers recorded yet. Browse all providers
Proteolysis-targeting chimeras degrade a protein rather than inhibit it, which lets them reach targets with no druggable pocket. Whether a given chimera actually works, though, depends on a ternary complex forming between the target protein, the bifunctional ligand, and an E3 ligase — a property that is hard to reason about from structure and expensive to establish experimentally. More than two decades after the first proof-of-concept PROTAC, no degrader has reached market approval, and degradation capacity remains the property that most often decides whether a series is worth pursuing.
AiPROTAC predicts that degradation capacity directly. Led by Li Zhang and Xing Chen with collaborators at Tsinghua University, A*STAR, Gluetacs Therapeutics, and several Chinese universities, and posted as a preprint in March 2025, it represents all three components as molecular graphs and passes them through dedicated encoders before a two-stage cross-attention decoder produces a binary degradation call. The design is deliberately shaped like the wet-lab logic it replaces: the first attention stage models the proximity the chimera induces between target and ligase, the second reads the resulting ternary representation from both partners' perspectives.
Labeled degradation data is scarce, and the model's second idea addresses that shortage. Two auxiliary contrastive branches, active only during training, learn from graph augmentations of unlabeled PROTAC-DB molecules and proteins, so the supervised classifier benefits from data that carries no degradation label at all. Where SynPROTAC generates candidate chimeras, AiPROTAC scores them.
Evaluation used PROTAC-DB 2.0 plus PROTAC-ZL, a manually curated set assembled by the authors from molecules not present in PROTAC-DB 2.0. Labeled PROTAC-DB samples were split 8:2 into training and test sets across five independent runs with different random seeds. Averaged over those runs, AiPROTAC reached accuracy 0.8502, precision 0.8298, recall 0.8467, F1 0.7935, AUROC 0.9192, and AUPR 0.8278, with standard deviations at or below 0.0505 on every metric. On the fully external PROTAC-ZL set it held accuracy 0.7928, AUROC 0.8156, and AUPR 0.7504; the mean of those three metrics, 0.7862, compares with 0.5450 for DeepPROTACs. Ten baselines were implemented for comparison, including random forest and SVM models over MACCS keys and Morgan fingerprints, DeepPROTACs, and four ablated AiPROTAC variants that swap the encoder for a GCN or GAT or strip out the attention and contrastive machinery.
In the prospective campaign, 18 cereblon-recruiting androgen receptor degraders were designed, synthesized, and assayed in LNCaP cells; AiPROTAC's predictions matched 8 of 12 in the first series and 5 of 6 in the second. A nineteenth compound, GT19, was designed by moving the linker attachment point on the cereblon ligand's benzene ring, and was predicted and confirmed to degrade the androgen receptor — at 100 nM within 24 hours, outperforming bavdegalutamide (ARV-110), with a cereblon binding IC50 of 0.15 μM against greater than 10 μM for a methylated negative control. Across all 19, 14 labels were called correctly, a 74% hit rate.
The intended use is triage. A medicinal chemistry group designing a degrader series can score every candidate before committing to synthesis, discarding the ones predicted not to degrade and carrying forward only the plausible ones — the AR campaign is a worked example of exactly that loop. Because the input is three molecular graphs, the model applies to any target and any E3 ligase for which structures or sequences are available, not only the cereblon and VHL systems that dominate published data.
AiPROTAC is a preprint and has not been peer reviewed. The code, including the training scripts, model definitions, preprocessed PROTAC-DB and PROTAC-ZL data, and predictor entry points, is public under Apache-2.0; the predictor loads a saved state dictionary, but no trained checkpoint ships with the repository, so users must retrain before running inference. The prospective AR result is the entry's strongest evidence — a degrader designed with model guidance, synthesized, and shown by Western blot to outperform a clinical-stage comparator — while the benchmark margins rest on datasets of a few hundred labeled chimeras, which is the standing constraint on every model in this area.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.