Loading…
Molecular glue degrader activity prediction from SMILES and protein sequence, with two-stage cross-attention that follows E3-then-substrate binding.
No providers recorded yet. Browse all providers
A molecular glue does not grip both of its proteins at once. A small molecule first drops into cereblon's tri-tryptophan pocket and remodels the surface above it; only that remodelled surface is competent to recruit a neo-substrate, and the reverse order does not occur. Models of degradation activity have largely ignored this asymmetry, fusing glue, E3 ligase and target simultaneously — a representation that carries statistical association but not the direction that defines the mechanism.
MG2Act writes that direction into the architecture. Developed at Ocean University of China with the Shanghai Institute of Materia Medica, it runs two cross-attention stages in the biological order: functional-group embeddings of the ligand attend first to the cereblon thalidomide-binding domain, and the resulting CRBN-conditioned representation then attends to the recruited target, before an MLP reads out a continuous degradation activity score. No ternary complex geometry is modelled — the inputs are a SMILES string and two protein sequences.
That structure-independence sets it apart from TriGlue and DeepTernary, which model ternary complex geometry, and from AiPROTAC and SynGlue, which score or design bifunctional PROTACs rather than monovalent glues. In the paper's own comparison, ipTM confidence from Boltz-2, Protenix and AlphaFold 3 tracked measured degradation inconsistently across targets.
Both protein inputs are embedded by a frozen ESM-C 300M protein language model (ESM Cambrian) and projected to 64 dimensions by a 1D convolution; ligands are parsed with RDKit into atom graphs and encoded by a three-layer graph attention network with hidden dimension 128. Fusion uses 4 attention heads over 2 transformer decoder layers, and the prediction head is an MLP of 128 and 64 units with dropout 0.1. Training data came from the group's MolGlueDB, standardised into 1,207 continuous-scored compound-target pairs across 47 neo-substrates; targets with fewer than 10 entries were dropped, leaving 1,159 pairs across 21 targets split 8:1:1. On the held-out test set MG2Act reaches R² = 0.5950 and RMSE = 0.1751, against R² = 0.5550 for the strongest baseline (XGBoost over MACCS fingerprints plus the same ESM-C embeddings), with an enrichment factor of 2.27 in the top 10%. Swapping the pretrained embeddings for amino-acid composition or one-hot features costs accuracy (R² 0.5673 and 0.5348); disabling functional-group boosting costs more (0.4840).
MG2Act is aimed at medicinal chemists triaging CRBN-directed libraries before synthesis. Applied prospectively to roughly 80 in-house CRBN-biased compounds, each retaining a glutarimide or dihydrouracil moiety, its scores correlated with measured degradation at Pearson r = 0.660 for IKZF1 and 0.595 for CK1α, and its top 40% score tier converted at a 50% hit rate for both targets, against stratified library base rates of 30% and 20%; the prioritised compounds included YFC-605, which degrades IKZF1 with a DC50 of 0.67 nM, and the sub-nanomolar CK1α degrader YFC-403. A harder test screened a 167-compound library against CDK4, a target absent from training: scaffold clustering over the top-scoring 20% surfaced SWC-202, which depletes endogenous CDK4 while largely sparing the paralog CDK6 through a proteasome-, neddylation- and CRBN-dependent mechanism.
MG2Act shows that a known biochemical ordering can be spent as an architectural inductive bias where data are too scarce for a model to find that structure unaided, and its falsification-oriented checks set a more searching standard than the single held-out split usual in this literature. The limits are stated by the authors: training covers cereblon alone, with VHL and DCAF15 named as future work; the corpus is about a thousand pairs; prospective correlations are moderate, so the practical claim is early enrichment rather than accurate ranking; and accuracy degrades on out-of-distribution targets. The work is a preprint that has not been peer reviewed, and characterisation of the reported compounds is deferred to supplementary information at publication. Source code, training scripts and the train, validation and test CSVs are on GitHub under an MIT license, with the trained checkpoint pointed at a HuggingFace repository that is not publicly accessible, despite the paper's code-availability statement placing the weights on GitHub.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.