SynGlue

Proteolysis-targeting chimeras (PROTACs) degrade a protein of interest by recruiting an E3 ubiquitin ligase to it, and their potency depends on three interdependent parts: the warhead, the E3 ligase ligand, and the linker joining them. Small structural changes in any one component shift degradation efficiency in ways that are hard to anticipate, yet most machine-learning approaches represent a PROTAC as a single whole molecule, which obscures which substructure drives the observed activity.

SynGlue, developed by the Department of Computational Biology at IIIT-Delhi together with Aurigene Oncology and posted to bioRxiv in August 2025, takes a fragment-centric view instead. It couples two interoperable modules: a data-driven module that maps query molecules onto a large curated corpus of protein-ligand interactions fragment by fragment, and a structure-guided module that designs cavity-specific warheads de novo from a protein structure. A generative module then assembles warhead, linker, and E3 ligand into complete degraders, with trained predictors ranking candidates by degradation potency (DC50) and maximal degradation (Dmax).

That pairing lets one system serve two very different regimes: ligand-rich targets, where interaction data can be mined directly, and ligand-poor targets with no useful chemical precedent. It sits alongside degrader-design models such as SynPROTAC, which enforces synthetic accessibility during generation, TriGlue, which models the glue-bound ternary complex, and LinkLlama, which designs the linker alone — but SynGlue is distinguished by carrying its designs through to in vitro and in vivo experiments.

#Key Features

• Fragment-indexed interaction corpus: MagnetDB holds 6.37 million high-confidence experimental protein-ligand interactions distilled from over 37.5 million raw entries, covering 1.94 million compounds and 20,129 protein targets.
• TRIE-based fragment retrieval: RECAP decomposition yields roughly 2.4 million unique terminal fragments indexed in a prefix tree, giving near-linear substructure search rather than quadratic brute-force matching.
• Component-aware potency prediction: An attention model learns separate weights for warhead, linker, and E3 ligand, attributing predicted DC50 and Dmax to specific substructures rather than the molecule as a whole.
• Two design modes: Data-driven fragment mining handles targets with abundant ligand data; structure-guided cavity detection generates warheads for targets with little chemical precedent.
• Polypharmacology classification: Query molecules are sorted into monovalent or multivalent and monotarget or multitarget classes, extending the framework beyond degraders to non-degrading multi-target molecules.

#Technical Details

The Generator uses the Link-INVENT architecture — an encoder-decoder recurrent network with LSTM units (embedding size 256, three hidden layers of 512 units) trained by reinforcement learning — and scores linkers with a multi-objective function balancing hydrogen bond donors, ring content, and graph length for ternary-complex compatibility. The DC50/Dmax regressor computes 4,800-dimensional GROVER graph embeddings for each component plus the full molecule, applies component-wise attention, and passes the result through a three-layer feedforward network (512, 256, 128 units) with ReLU and dropout of 0.2, trained with MSE loss and Adam over 200 epochs. A complementary multiclass model stratifies 2,320 PROTACs with measured DC50 into quartile-defined potency classes using five orthogonal descriptor families (Signaturizer, GROVER, Mordred, ImageMol, and ChemBERTa-77M-MLM), Boruta feature selection, and twelve classifiers, reaching a median test AUC of about 0.80 under 10-fold cross-validation. Against a benchmark of 6,935 compounds spanning DrugCentral drugs, known PROTACs, and multi-target entities, SynGlue classified 4,621 (66.7%) into defined interaction types and reported higher target recovery than LigAdvisor, SwissTargetPrediction, and SuperPred.

#Applications

SynGlue is aimed at medicinal chemistry and targeted protein degradation teams choosing warheads, tuning linkers, and prioritising degrader candidates before synthesis. In the data-driven case study, mining MagnetDB for BRD4 fragments and filtering for paralog selectivity and drug-likeness yielded 22 structurally novel warheads coupled to a VHL ligand; two prioritised BRD4 PROTACs degraded with DC50 of 0.19 nM in vitro and were efficacious in mouse models. The structure-guided workflow, starting from the CRBN-GSPT1 complex, produced GSPT1 degraders with DC50 near 0.0011 uM that were also active in vitro and in vivo, and the framework additionally designed non-degrading multi-target molecules for EGFR/CDK1 and PRKX/DRD1 pairs. It is distributed as a Python client on PyPI that submits design and screening jobs to a REST API, with Colab notebooks for both workflows.

#Impact

SynGlue is one of the few generative degrader-design systems to report a complete path from computational design to animal efficacy for two mechanistically unrelated targets, using the same trained models rather than a per-target refit, and its fragment-level treatment of interaction data makes predictions interpretable in terms medicinal chemists act on. Several caveats apply: the work is a preprint that has not been peer reviewed, the wet-lab results are self-reported and not independently replicated, and the authors note that DC50 and Dmax are assay-dependent, bounding generalisation by inter-study variability in degradation assays. The source code is on GitHub under an MIT license, but MagnetDB is not distributed and the pip package is a thin client to a server the authors host, so full reproduction depends on infrastructure that is free for academic use and requires a commercial license key otherwise.