MG2Act: A Mechanism-Inspired Sequential Attention Framework for Molecular Glue Degradation Prediction
zhuang, z.; teng, d.; Xu, X.; Fang, S.; Wang, Y.; Hou, M.; Ge, L.; Yuan, S.; Yang, M.; Cheng, L.; Zhang, Z.; He, Q.; Li, Z.; Xu, X.; -Ma, S.; Zhang, S.; Wang, X.; Zheng, M.; Qin, C.
Show abstract
Molecular glue degraders act by inducing productive proximity between an E3 ligase and a substrate protein. For most characterized degradative glues, a small molecule first engages the E3, conditions its substrate-recognition surface, and only then enables recruitment of a compatible neo-substrate. This directionality is rarely encoded explicitly in computational models, which typically fuse molecule, E3 and target representations simultaneously. We present MG2Act, a structure-independent framework that translates this two-step logic into sequential cross-attention, using CRBN-mediated degradation as the most data-rich representative system. Starting from a curated continuous-valued benchmark of 1,207 pairs across 47 targets, a refined subset of 1,159 pairs was selected to train MG2Act after excluding rare targets. On identical processed data, MG2Act consistently outperforms machine learning baselines, robustly generalizes under strict redundancy-filtering, and responds coherently to mechanism-based perturbations. Prospective screening and zero-shot target-conditioned prioritization identified nanomolar degraders of IKZF1, CK1 and CDK4, including the non-classical IMiD-core CDK4 degrader SWC-202.
Matching journals
The top 7 journals account for 50% of the predicted probability mass.
Similar papers in this journal
- Context-aware geometric deep learning for protein sequence design 94%
- PreMode predicts mode-of-action of missense variants by deep graph representation learning of protein sequence and structural context 94%
- Higher-order epistasis creates idiosyncrasy, confounding predictions in protein evolution 94%
Similar papers in this journal
- PSICHIC: physicochemical graph neural network for learning protein-ligand interaction fingerprints from sequence data 95%
- Efficient protein structure generation with sparse denoising models 95%
- TrustAffinity: accurate, reliable and scalable out-of-distribution protein-ligand binding affinity prediction using trustworthy deep learning 93%
Similar papers in this journal
- Engineering of highly active and diverse nuclease enzymes by combining machine learning and ultra-high-throughput screening 94%
- Undersampling and the inference of coevolution in proteins 93%
- Learning multi-cellular representations of single-cell transcriptomics data enables characterization of patient-level disease states 93%
Similar papers in this journal
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.