PaperlayerView exact revision on arXiv

Source-linked AI summary

ProMeta: Few-shot PROTAC-targeted degradation prediction across E3 ligases

Yuansheng Liu, Yufei Ye, Tao Tang, Jiawei Luo, Wen Tao, Xiao Luo

arXiv:2609.09891v1cs.LGq-bio.QM

TL;DR

PROTAC degradation prediction remains constrained by scarce, imbalanced labels concentrated in a few E3 ligases, motivating methods that generalize across ligases from minimal supervision. ProMeta reformulates the task as episodic few-shot meta-learning with support-conditioned prototype inference, and its evaluated CRBN/VHL experiments show bidirectional transfer with direction- and data-regime-dependent performance. The evidence is limited to controlled support/query protocols, concentrated source-ligase data, and small rare-E3 query sets.

  • Problem

    Existing PROTAC predictors face scarce and highly imbalanced labeled data concentrated in CRBN and VHL, leaving cross-ligase generalization under minimal supervision insufficiently addressed.

  • Method

    ProMeta is a prototype-based graph neural network trained episodically on source-E3 tasks and used for held-out E3 prediction by constructing class prototypes from minimal target-ligase support samples without encoder updates.

  • Results

    ProMeta supports bidirectional CRBN/VHL transfer, achieving AUROC values of 0.796 under K = 2, Q = 3 and 0.883 under K = 2, Q = 5 for CRBN-to-VHL transfer, while performance varies by transfer direction and data regime.

  • Takeaways & Limitations

    ProMeta provides a practical strategy for cross-ligase few-shot PROTAC degradation prediction within the evaluated support/query protocols.

  • Takeaways & Limitations

    Cross-ligase evidence is bounded by source data concentrated in CRBN and VHL, small rare-E3 query sets, and controlled few-shot support/query evaluation rather than exhaustive validation across E3 ligase families.

Abstract

from arXiv · show

Proteolysis-targeting chimeras (PROTACs) have emerged as a transformative therapeutic strategy that selectively degrades historically ''undruggable'' targets via the ubiquitin-proteasome system. Despite growing efforts to develop computational predictors of PROTAC degradation activity, existing supervised approaches remain severely challenged by data scarcity and imbalance across E3 ligases, limiting their ability to generalize beyond well-studied ligase contexts. In practice, labeled data are heavily concentrated on a few ligases (e.g., CRBN and VHL), while the majority of E3 ligases remain underexplored yet are critical for expanding the design space of targeted degraders. Developing methods that enable robust cross-ligase generalization with minimal labeled data is therefore essential for improving the practical utility of computational PROTAC discovery. We reformulate PROTAC degradation activity prediction across E3 ligases as a few-shot meta-learning problem and present ProMeta, a prototype-based graph neural network trained through episodic meta-learning on source-E3 tasks and evaluated on held-out target-E3 tasks through support-conditioned inference. ProMeta performs inference without updating the encoder by dynamically estimating class prototypes from minimal target-ligase support samples. On the CRBN-to-VHL benchmark, ProMeta achieves AUROC values of 0.796 under K=2, Q=3 and 0.883 under K=2, Q=5, improving by 19.9% and 6.8%, respectively, over the corresponding supervised GNN baseline. Reverse VHL-to-CRBN transfer under the same protocol yielded AUROC values of 0.702 (K=2, Q=3) and 0.821 (K=2, Q=5), confirming bidirectional applicability while revealing direction and data-regime dependence. Together, these results support ProMeta as a practical framework for cross-ligase few-shot prediction under the evaluated support/query protocols.

Introduction

PROTAC degradation prediction is difficult because effective degradation depends on multi-component molecular and ternary-complex properties, while labeled data are scarce, imbalanced, and concentrated in a few E3 ligases. ProMeta addresses this setting by treating cross-ligase prediction as few-shot meta-learning with support-conditioned prototype inference.

  • Motivation: PROTACs recruit an E3 ubiquitin ligase to a protein of interest, inducing ubiquitination and proteasomal degradation through an event-driven mechanism.This catalytic mechanism can enable sustained target depletion at substoichiometric doses.
  • Motivation: Effective PROTAC degradation depends on warhead and E3-ligase ligands, linker properties, and ternary-complex cooperativity and structural compatibility.These dependencies make rational design and large-library screening challenging.
  • Research gap: Existing predictors show encouraging molecular and protein-level modeling performance, but their transfer to heterogeneous, label-scarce cross-ligase settings remains insufficiently established.The cited approaches include GNNs, DeepPROTAC, PROTAC-STAN, DegradeMaster, and SE(3)-PROTACs.
  • Research gap: Over 80% of raw PROTAC-DB entries lack DC50 and Dmax annotations, while labeled data are predominantly concentrated in CRBN and VHL.This distribution motivates evaluation of methods that can generalize across E3 ligases with minimal labeled data.
  • ProMeta: ProMeta is a prototype-based graph neural network trained with episodic meta-learning across source-E3 tasks and evaluated on held-out E3 tasks.It estimates class prototypes from minimal target-ligase support samples without updating the encoder, using molecular and protein-sequence representations with ECFP4 augmentation at inference.
  • Evaluation: The framework is evaluated on bidirectional CRBN/VHL transfer and exploratory extreme low-data transfer across rare E3 ligases.These evaluations support its use as a practical framework for data-scarce PROTAC degradation prediction within the reported settings.

Dataset collection and preprocessing

The study uses PROTAC-DB records containing molecular structures, POI and E3 annotations, and degradation evidence, then applies activity labeling, filtering, sequence processing, and class-balancing procedures to construct modeling data. The balanced CRBN and VHL dataset contains 860 and 560 samples, respectively.

  • Data source: The raw PROTAC-DB export contains 9,384 rows with molecular structures, POI and E3 annotations, and available DC50 and Dmax evidence.The raw export is distinguished from the smaller filtered labeled tables used for modeling.
  • Activity labeling: Binary activity labels combine quantitative DC50/Dmax thresholds with explicit qualitative activity calls from experimental descriptions.Complete quantitative records are labeled high activity when DC50 < 100 nM and Dmax ≥ 80%; qualitative-only rows require an explicit activity call.
  • Class balancing: CRBN and VHL dominate the raw export with 6,041 and 2,858 rows, respectively, before strict activity-evidence filtering.Random majority-class down-sampling was applied independently to these ligases.
  • Class balancing: The balanced labeled dataset contains 860 CRBN samples and 560 VHL samples, with equal high- and low-activity class counts within each ligase.CRBN contains 430 high-activity and 430 low-activity samples; VHL contains 280 of each.
  • Sequence preprocessing: POI and E3 amino acid sequences were retrieved from UniProt, with 89 unmapped POI entries removed and E3 identities resolved through a manually curated mapping.E3 sequences were truncated to a maximum length of 2,000 residues.
  • Dataset construction: The 1,420-record CRBN/VHL subset is reported as a candidate pool before duplicate and conflict cleanup for the main strict transfer experiments.Supplementary Table S1 summarizes the preprocessing flow from the raw export to the strict context-level dataset.

Episodic meta-learning dataset construction

The benchmark uses ligase-disjoint episodic partitions to test transfer from source E3 ligases to held-out targets without compound overlap. CRBN→VHL and rare-E3 experiments use CRBN as the source, while the reverse benchmark uses VHL as the source.

  • Source-target partitioning: CRBN-associated molecules form the source pool for CRBN→VHL and rare-E3 transfer, whereas VHL-associated molecules form the source pool for reverse VHL→CRBN transfer.The target ligase is excluded from source training and validation partitions.
  • Data partitioning: Each source pool is split into 80% training and 20% validation partitions, with no compound overlap between source and target partitions.Ligases lacking both positive and negative samples are excluded.
  • Rare-E3 benchmark: The rare-E3 targets are cIAP1, IAP, MDM2, XIAP, and FEM1B, and all experiments use random seeds 42, 2025, and 3407.The benchmark therefore evaluates transfer across explicitly held-out ligase contexts.
  • Benchmark rationale: CRBN and VHL were selected because their recruiting PROTACs occupy chemically distinct, largely non-overlapping chemical and target spaces.CRBN was chosen as the primary source because it is the most data-rich ligase in PROTAC-DB.

Framework overview

ProMeta combines graph, protein-sequence, and fingerprint representations with prototype-based episodic meta-learning for cross-ligase few-shot prediction. It meta-trains on source-E3 episodes and classifies held-out target-E3 queries from support-derived prototypes without encoder updates.

  • Cross-ligase inference: At cross-ligase meta-test, the encoder is frozen and target-ligase prototypes are recomputed from the available support set without gradient updates.This support-conditioned inference is applied to held-out E3 tasks whose query compounds are reserved for evaluation.
  • Representation: Molecules are represented as attributed graphs whose node embeddings are updated by neighborhood aggregation and summarized through global mean pooling.Stacked message-passing layers integrate structural information from progressively larger atom neighborhoods.
  • Representation: ProMeta integrates learned graph features, POI and E3 protein-sequence embeddings, and a fixed ECFP4 fingerprint into the representation used for prototype construction.The fingerprint is concatenated during target-ligase inference with weight λfp = 0.5.
  • Prototype-based classification: During episodic meta-training, class prototypes are computed from support embeddings, query molecules are classified by prototype distance, and encoder parameters are optimized with query loss.Episodes contain active and inactive classes with K support and Q query samples per class.
  • Cross-ligase inference: The episodic design trains embeddings for separated within-episode class prototypes, supporting prototype-based prediction when the E3-defined task changes at test time.Source-E3 episodes train the encoder, while held-out target-E3 support/query episodes define meta-test tasks.

Evaluation metrics

Performance was evaluated with five complementary metrics, with AUROC and AUPRC serving as primary metrics because they are threshold-independent and robust to class imbalance.

  • AUROC and AUPRC were designated as the primary evaluation metrics because they are threshold-independent and robust to class imbalance.
  • Accuracy, balanced accuracy, and macro F1-score were reported as secondary metrics for reference.
  • Macro F1-score may favor majority-class prediction in the class-imbalanced episodic setting and therefore requires cautious interpretation.

Implementation details

The implementation used graph and fingerprint representations with standard deep-learning libraries, while evaluation trained on source-ligase episodes and tested on a completely held-out target ligase.

  • Molecular graphs and 512-bit ECFP4 fingerprints were generated with RDKit, while models were implemented using PyTorch and PyTorch Geometric.
  • The GNN used three message-passing layers with hidden dimension 128 and global mean pooling.
  • Training used Adam with learning rate 1 × 10^-3 and weight decay 1 × 10^-5, with hyperparameters selected on a 20% source-ligase validation partition.
  • Meta-training sampled K = 2, Q = 3 episodes from the source E3, then evaluated support/query episodes on a target E3 excluded from meta-training.

Baseline evaluation protocol

Baselines were compared under shared support/query splits and bidirectional CRBN/VHL episodic settings, with matched-head analyses separating representation effects from decision-rule effects.

  • All baselines were re-evaluated or adapted using shared support/query splits to enable cross-ligase comparison.
  • Fig. 2 compares mean AUROC, AUPRC, and balanced accuracy across three seeds for both transfer directions and two K/Q settings.
  • The same support/query rows were provided to every method within each transfer direction and K/Q setting.
  • Matched-head analyses used a shared support-set prototype classifier for GNN, PROTAC-STAN, and DegradeMaster representations, while ECFP4 used nearest-centroid inference and ProMeta used a support-trained linear head.

Results

ProMeta supported cross-ligase few-shot prediction across CRBN, VHL, and rare-E3 settings, with performance varying by transfer direction and data regime. Ablations and retrospective case studies indicate that episodic representation learning combined with support-set prototype inference contributed to the evaluated results.

  • CRBN-to-VHL transfer: 0.796 AUROC under K = 2, Q = 3 and 0.883 under K = 2, Q = 5 for CRBN-to-VHL transfer, exceeding the supervised GNN baseline.RF + ECFP achieved 0.785 and 0.842 under the corresponding settings.
  • Bidirectional transfer: 0.702 AUROC under K = 2, Q = 3 and 0.821 under K = 2, Q = 5 for reverse VHL-to-CRBN transfer, with lower performance than CRBN-to-VHL.The authors describe the difference as direction- and data-regime-dependent under distribution shift.
  • Rare-E3 evaluation: 0.700 aggregate AUROC across 90 strict common-query rows made ProMeta the best-performing method in the rare-E3 evaluation.PROTAC-STAN and DegradeMaster achieved 0.550 and 0.564, respectively, under the same coverage-matched evaluation.
  • Rare-E3 evaluation: ProMeta achieved the highest AUROC across all five rare ligases, but the benchmark was treated as a low-data stress test because several groups had few query rows.One-shot support selection, repeated query compounds, and small query pools made some per-ligase estimates unstable.
  • Ablation study: Removing episodic training changed mean AUROC from 0.840 to 0.818 for CRBN-to-VHL and from 0.761 to 0.554 for VHL-to-CRBN.The larger reverse-transfer gap indicates that the contribution of episodic representation learning was not uniform across data regimes.
  • Retrospective case studies: The VZ185 case study reached 91.7% accuracy, correctly classifying 11 of 12 compounds and outperforming the evaluated baseline models.Additional series achieved 78.6% for CRBN-recruiting CDK6 degraders and 100.0% for VHL-recruiting BCLxL degraders, with small-sample scope limitations.

Conclusion and discussion

ProMeta targets cross-ligase degradation prediction with limited labeled data, but its evidence remains bounded by the evaluated protocols and representation choices. The results support few-shot transfer without establishing general superiority in native prediction settings or exhaustive validation across E3 families.

  • Method and scope: ProMeta constructs active/inactive prototypes from small target-ligase support sets without updating model parameters.This support-conditioned inference addresses cross-ligase prediction under limited labeled data.
  • Interpretation: The evaluated results do not claim general superiority over DeepPROTAC, PROTAC-STAN, or DegradeMaster in their native prediction settings.The comparison instead concerns whether architectures optimized for data-rich prediction solve cross-ligase few-shot generalization.
  • Limitations: ProMeta combines 2D molecular graph and ECFP4 representations with protein-sequence context, without explicit 3D conformational or ternary-complex information.The paper identifies 3D and structure-based ternary-complex features as future extensions.
  • Case study: In the VZ185 case study, C1–C12 were evaluated while C13–C16 were excluded because they lacked a POI-binding warhead.The heatmap records whether each model’s prediction matches the experimental label across four model columns.
  • Limitations: The cross-ligase generalization evidence is limited by CRBN/VHL-concentrated source data and small rare-E3 query sets.Accordingly, the authors restrict the conclusion to the controlled few-shot support/query evaluation rather than exhaustive validation across E3 ligase families.
  • Evidence across settings: ProMeta’s cross-ligase evidence is strongest for CRBN-to-VHL transfer, with more heterogeneous behavior in reverse VHL-to-CRBN and rare-E3 settings.The authors characterize these results as informative but direction- and data-regime-dependent.
Loading 2609.09891v1…