Structure-Aware Prediction of PROTAC-Mediated Protein Degradability via Graph Neural Networks Conference: 17th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics; June 30 – July 3, 2026; Calabria, Italy ACM-BCB ’26: 17th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics, June 30 – July 3, 2026, Calabria, Italy CCS: Computing methodologies Neural networks CCS: Applied computing Bioinformatics CCS: Applied computing Computational biology

Bryan Cheng, Austin Jin · 2026

Read Structure-Aware Prediction of PROTAC-Mediated Protein Degradability via Graph Neural Networks Conference: 17th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics; June 30 – July 3, 2026; Calabria, Italy ACM-BCB ’26: 17th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics, June 30 – July 3, 2026, Calabria, Italy CCS: Computing methodologies Neural networks CCS: Applied computing Bioinformatics CCS: Applied computing Computational biology by Bryan Cheng, Austin Jin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-04021.txt.

Introduction

Cheng and Jin present DegradoMap, a graph neural network for PROTAC planning. PROTACs can degrade disease-causing proteins in a selective way. Guessing which targets will degrade is a known bottleneck. Older tools need the full PROTAC structure before scoring. That structure does not exist before synthesis, at target choice time.

DegradoMap scores a target from protein structure and E3 ligase identity alone. Lysine-weighted pooling and cross-attention carry biophysical priors into the model. Cancer Dependency Map data adds cellular context. On the PROTAC-8K benchmark, target-unseen AUROC reached 0.646 on average. E3 transfer scores were higher, at 0.811. It also picks good E3 ligases, with 74 percent Hit-at-3 accuracy. High seed variance means ensembles are needed for firm use, they warn.

Catalog record

CREATORBryan Cheng, Austin Jin
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE76 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.04021

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Structure-Aware Prediction of PROTAC-Mediated Protein Degradability via Graph Neural Networks Conference: 17th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics; June 30 – July 3, 2026; Calabria, Italy ACM-BCB ’26: 17th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics, June 30 – July 3, 2026, Calabria, Italy CCS: Computing methodologies Neural networks CCS: Applied computing Bioinformatics CCS: Applied computing Computational biology Bryan Cheng, Austin Jin Abstract . Proteolysis-targeting chimeras (PROTACs) can selectively degrade disease-causing proteins, yet predicting which targets are amenable to degradation remains a critical bottleneck: existing computational methods require the complete PROTAC molecular structure, information unavailable before synthesis. We present DegradoMap , a graph neural network that predicts PROTAC-mediated degradability from protein structure and E3 ligase identity alone—the minimal information available at the target selection stage. The model encodes biophysical priors through lysine-weighted graph pooling with per-protein normalization, models protein–E3 compatibility via cross-attention, and integrates cellular context from the Cancer Dependency Map. On the PROTAC-8K benchmark (3,101 samples, 155 targets, 10 E3 ligases), DegradoMap achieves 0.646 ± \pm 0.124 AUROC on target-unseen evaluation (best seed: 0.7449) and 0.811 AUROC on CRBN → \to VHL E3-unseen transfer, outperforming GNN and machine learning baselines. The model additionally recommends optimal E3 ligases with 74% Hit@3 accuracy. Two findings carry broader implications: E(3)-equivariant architectures underperform the simpler invariant design for this scalar prediction task, and ESM-2 embeddings improve peak performance only with careful regularization—naive integration fails. DegradoMap provides pre-synthesis computational guidance for degradability assessment; its well-calibrated confidence scores (ECE = 0.029, target-unseen) enable practitioners to prioritize high-confidence predictions for experimental follow-up. However, the high seed variance (std = 0.124) and limited E3 coverage require ensembling for reliable deployment. 1. Introduction Proteolysis-targeting chimeras (PROTACs) invert the logic of drug design: rather than occupying a protein’s active site, they recruit the cell’s ubiquitin–proteasome machinery to destroy the protein entirely ( bekes2022protac ; sakamoto2001protacs ; pettersson2019pbd ) . A PROTAC bridges two binding events—its warhead engages the target protein while its E3 ligand recruits an E3 ubiquitin ligase—and the induced proximity triggers polyubiquitination of surface lysine residues, marking the target for proteasomal degradation ( lai2017induced ) . Because this mechanism is catalytic, a single PROTAC molecule degrades multiple target copies, enabling efficacy at sub-stoichiometric concentrations and access to proteins that lack traditional drug-binding pockets. With over 20 PROTACs now in clinical trials ( bekes2022protac ) , the modality’s therapeutic promise is clear. Yet development faces a critical bottleneck at its earliest stage: target selection . Is a given disease-relevant protein amenable to PROTAC-mediated degradation? Answering this currently requires synthesizing candidate PROTACs and conducting cell-based degradation assays—an expensive process with high failure rates. Computational prediction of degradability before PROTAC synthesis would dramatically accelerate the pipeline. Limitations of existing methods. Existing computational approaches all require the PROTAC molecular structure as input. DeepPROTACs ( li2022deepprotacs ) operates on voxelized binding pockets requiring complete PROTAC, target, and E3 coordinates. Ribes et al. ( ribes2024protac ) combine protein embeddings with PROTAC SMILES fingerprints. DegradeMaster ( liu2025degrademaster ) achieves 0.856 AUROC using full PROTAC molecular descriptors alongside protein features. While valuable for optimizing existing PROTACs, these methods cannot serve target selection because the PROTAC is unavailable at that stage. MAPD ( zhang2022mapd ) addresses target assessment from protein-intrinsic features but does not leverage AlphaFold structural information ( jumper2021alphafold ) . Our approach. We introduce DegradoMap , a graph neural network that predicts protein degradability from minimal inputs : a protein structure (from AlphaFold, requiring only a UniProt ID) and E3 ligase identity (e.g., “CRBN” or “VHL”). The design reflects the biophysics of ubiquitin transfer, which requires (1) surface-accessible lysine residues, (2) geometric compatibility between target and E3 ligase for ternary complex formation, and (3) sufficient cellular E3 expression and proteasome capacity. DegradoMap addresses these requirements through three modules and a gated fusion mechanism: (1) Structure–Ubiquitination Graph (SUG) encoder : An invariant message-passing GNN with lysine-weighted pooling that focuses representation on ubiquitination-relevant residues, incorporating ESM-2 embeddings ( lin2023esm2 ) , structural properties, and known ubiquitination site annotations from PhosphoSitePlus. (2) E3 compatibility module : Bidirectional cross-attention modeling compatibility between protein structure and E3 ligase identity, enabling generalization across E3 ligases. (3) Cellular context encoder : A residual MLP processing 59 tissue-specific features from the Cancer Dependency Map ( tsherniak2017depmap ) . (4) Gated fusion : Learned gating that dynamically weights module outputs for multi-task prediction of binary degradability and continuous efficiency. Evaluation and contributions. On the PROTAC-8K benchmark ( liu2025degrademaster ) (3,101 samples, 155 targets, 10 E3 ligases), DegradoMap achieves a 3-seed mean of 0.646 \pm 0.124 and 6-seed mean of 0.603 \pm 0.097 AUROC on target-unseen evaluation (best seed: 0.7449), 0.811 AUROC [0.785, 0.836] on CRBN \to VHL E3-unseen transfer, and 74% Hit@3 for E3 recommendation. We note the 6-seed improvement over gradient boosting baseline (0.607) is not statistically significant (p = 0.556), and VHL-targeted proteins fail below random (0.396 AUROC)—both limitations prominently acknowledged. Two architectural findings carry implications beyond PROTAC prediction: E(3)-equivariant architectures underperform the invariant design for scalar property prediction, and ESM-2 embeddings require task-specific regularization to improve performance. Our principal contributions include a structure-based pre-synthesis degradability predictor requiring only a UniProt ID and E3 ligase name, a lysine-weighted pooling mechanism encoding ubiquitin transfer biology, a cross-attention E3 compatibility module, well-calibrated confidence scores (ECE = 0.029) enabling threshold-based screening, and rigorous multi-seed evaluation revealing variance characteristics critical for deployment planning. 2. Related Work PROTAC prediction methods. Computational approaches to PROTAC modeling vary significantly in their input requirements and intended use cases, as summarized in Table 1 . DeepPROTACs ( li2022deepprotacs ) uses 3D CNNs on binding pocket voxelizations, requiring the complete PROTAC molecule plus target and E3 binding pockets in mol2 format—information available only after PROTAC synthesis and structural characterization. Ribes et al. ( ribes2024protac ) combine pretrained protein embeddings with molecular fingerprints, requiring target sequence and PROTAC SMILES, which presupposes PROTAC design. MAPD ( zhang2022mapd ) predicts degradation tractability from protein-intrinsic features using machine learning; while it shares our target-assessment framing, it does not leverage structural information. DegradeMaster ( liu2025degrademaster ) achieves 0.856 AUROC using full PROTAC molecular descriptors alongside target features, representing the state-of-the-art for PROTAC optimization but addressing a fundamentally different problem than target selection. DegradoMap uniquely combines structural awareness with minimal input requirements, filling the gap between sequence-only assessment (MAPD) and full-PROTAC optimization (DeepPROTACs, DegradeMaster). Table 1 . Input modality comparison of PROTAC prediction methods. DegradoMap requires only protein structure and E3 identity, enabling use before PROTAC synthesis. Method Protein PROTAC E3 Use Case DeepPROTACs Pocket (3D) Full mol. Pocket (3D) Optimization Ribes et al. Sequence SMILES – General MAPD Features – – Assessment DegradeMaster Features Descriptors Features Optimization DegradoMap Structure – Identity Assessment Geometric deep learning for proteins. GNNs operating on protein structures have achieved strong performance across diverse tasks including function prediction, binding site identification, and property estimation. GVP-GNN ( jing2021gvp ) maintains SO(3)-equivariance using geometric vector perceptrons that jointly process scalar and vector features, enabling the model to reason about directional relationships. GearNet ( zhang2023gearnet ) employs relational graph convolutions with multiple edge types (sequential bonds, spatial contacts) for protein representation pre-training. SchNet ( schutt2017schnet ) introduces continuous-filter convolutions with radial basis function expansions, originally designed for quantum chemistry but

Read the complete text