Structure-Aware Prediction of PROTAC-Mediated Protein Degradability via Graph Neural Networks

Technical Summary: Structure-Aware Prediction of PROTAC-Mediated Protein Degradability via Graph Neural Networks

Problem Statement
Proteolysis-targeting chimeras (PROTACs) offer a catalytic mechanism for degrading disease-causing proteins by recruiting E3 ubiquitin ligases. However, a critical bottleneck exists in the early drug discovery pipeline: target selection. Determining whether a specific protein is amenable to PROTAC-mediated degradation currently requires synthesizing candidate molecules and conducting expensive cell-based assays. Existing computational methods (e.g., DeepPROTACs, DegradeMaster) rely on the complete PROTAC molecular structure as input, rendering them unusable for initial target assessment where the PROTAC has not yet been designed. The authors address the need for a pre-synthesis prediction tool that can assess degradability using only the minimal information available at the target selection stage: the target protein's structure and the identity of the E3 ligase.

Methodology: DegradoMap
The authors introduce DegradoMap, a Graph Neural Network (GNN) designed to predict binary degradability and continuous degradation efficiency without PROTAC structural data. The architecture consists of four primary modules:

1. Structure–Ubiquitination Graph (SUG) Encoder:

   • Input: Protein structure graphs constructed from AlphaFold-predicted structures (radius $r=8$ Å). Node features include optional ESM-2 embeddings (1,280-dim), physicochemical properties, pLDDT, SASA, disorder scores, and known ubiquitination site masks from PhosphoSitePlus.
   • Mechanism: An invariant message-passing GNN (SE(3)-invariant) updates node representations using scalar Euclidean distances between $C_\alpha$ atoms.
   • Key Innovation: A lysine-weighted global pooling mechanism. Since ubiquitin attaches to lysine residues, the model applies a per-protein softmax to upweight lysine residues based on a learned parameter. Crucially, this uses per-protein normalization to prevent data leakage across the batch and $1/\sqrt{N}$ size normalization to prevent protein size from acting as a trivial shortcut.
   • Design Choice: The authors deliberately chose an invariant architecture over E(3)-equivariant designs, finding that scalar distance features suffice for this scalar prediction task and that equivariant models underperform.

2. E3 Compatibility Module:

   • Mechanism: A bidirectional cross-attention layer that models the compatibility between the protein representation (query) and the E3 ligase identity (key/value).
   • Implementation: E3 ligases are represented as learned embeddings (rather than structural features) due to the limited number of distinct E3s (10) in the dataset. This allows the model to learn E3-specific "filters" over the protein structure.

3. Cellular Context Encoder:

   • Input: 59 tissue-specific features from the Cancer Dependency Map (DepMap), covering gene expression, protein levels, and pathway membership.
   • Mechanism: A residual MLP processes these features to account for cellular variables like E3 expression levels and proteasome capacity.

4. Gated Fusion and Prediction:

   • A learned gating mechanism dynamically weights the outputs of the SUG, E3, and Context modules.
   • The fused representation feeds into two heads: a binary classifier for degradability (primary task) and a regressor for continuous efficiency (auxiliary task).
   • Training: The model uses a multi-task loss (Binary Cross-Entropy, Huber, and Focal Loss) and a three-phase training strategy (SUG pre-training, joint SUG+E3 training, and full fine-tuning) to ensure stability.

Key Results
The model was evaluated on the PROTAC-8K benchmark (3,101 samples, 155 targets, 10 E3 ligases) using three distinct splits: Random, Target-Unseen (most rigorous), and E3-Unseen.

• Target-Unseen Performance:
  • The improved model (with ESM-2) achieved a 3-seed mean AUROC of 0.646 ± 0.124 (best seed: 0.7449).
  • A more extensive 6-seed evaluation yielded a mean AUROC of 0.603 ± 0.097, which was not statistically superior to a Gradient Boosting baseline (0.607, $p=0.556$).
  • The authors note high variance across seeds, suggesting that ensembling is required for reliable deployment.
• E3-Unseen Transfer:
  • The model achieved 0.811 AUROC when transferring from CRBN-dominated training data to a VHL-only test set, demonstrating successful cross-E3 generalization at the ligase level.
  • However, VHL-targeted proteins specifically failed on the target-unseen split (0.396 AUROC, below random chance), indicating that structural patterns learned for CRBN do not generalize well to VHL targets.
• E3 Recommendation:
  • The model successfully recommends optimal E3 ligases, achieving 74% Hit@3 accuracy (the correct E3 is in the top 3 recommendations) and a Mean Reciprocal Rank (MRR) of 0.641.
• Calibration:
  • The model exhibits excellent calibration on the target-unseen split (ECE = 0.029), allowing practitioners to trust confidence scores for threshold-based screening.
• Ablation Findings:
  • Removing the explicit "lysine indicator" feature actually improved performance, suggesting that the architectural lysine-weighted pooling is a more effective inductive bias than explicit feature annotation.
  • ESM-2 embeddings improved peak performance only with careful regularization (lower learning rate, reduced dropout); naive integration degraded performance.

Significance and Claims
The paper positions DegradoMap not as a definitive solution that outperforms all baselines in every metric, but as a pragmatic pre-synthesis tool that fills a specific gap in the drug discovery workflow.

• Minimal Input Requirement: Unlike existing methods, DegradoMap requires only a UniProt ID (for AlphaFold structure) and an E3 ligase name, enabling assessment before any PROTAC synthesis.
• Complementary Paradigm: The authors frame DegradoMap as complementary to full-information methods (like DegradeMaster). DegradoMap is for target screening (pre-synthesis), while full-information methods are for molecule optimization (post-synthesis).
• Honest Limitations: The authors explicitly acknowledge that the model's 6-seed average performance is not statistically significant compared to simpler baselines and that performance is highly variable and E3-dependent (specifically failing on VHL targets). They argue the model's value lies in its structural interpretability (via lysine attention weights), E3 recommendation capability, and well-calibrated confidence scores, rather than raw AUROC dominance.
• Architectural Insights: The paper contributes broader findings to geometric deep learning, specifically that invariant architectures can outperform equivariant ones for scalar property prediction in this domain, and that protein language model embeddings require task-specific regularization to be beneficial.

In summary, DegradoMap provides a computationally efficient, structure-aware framework for early-stage PROTAC target assessment, offering actionable guidance on E3 selection and degradability likelihood while transparently reporting its limitations regarding data scarcity and seed variance.