ProMeta: Few-shot PROTAC-targeted degradation prediction across E3 ligases
Yuansheng Liu, Yufei Ye, Tao Tang, Jiawei Luo, Wen Tao, Xiao Luo
cs.LG · q-bio.QM
Quick verdict
Cross-E3 few-shot PROTAC prediction via prototypical GNN; credible, well-hedged, reproducible, but the edge over RF+ECFP is thin and seed-sensitive.
Summary
The paper reformulates PROTAC degradation-activity prediction as cross-E3-ligase few-shot meta-learning. ProMeta combines a 3-layer GCN molecular encoder, lightweight POI/E3 sequence embeddings fused via learnable weights, and prototypical classification: during episodic meta-training on a source ligase (CRBN or VHL), class prototypes are built from small support sets; at test time the encoder is frozen and prototypes for a held-out ligase are computed from K=1–2 support compounds per class, with a fixed L2-normalized ECFP4 fingerprint concatenated at inference. Using PROTAC-DB 3.0 (raw export 9,384 rows; 1,386 clean CRBN/VHL contexts) under ligase-disjoint splits, the model reaches AUROC 0.796/0.883 (K=2, Q=3/5) for CRBN→VHL and 0.702/0.821 for VHL→CRBN, versus 0.785/0.842 and 0.613/0.702 for a tuned RF+ECFP baseline; on five rare E3 ligases under K=1 it scores 0.700 on 90 common-query rows versus 0.550/0.564 for adapted PROTAC-STAN/DegradeMaster. Ablations (Table 2) show episodic training matters mainly in the reverse direction (0.761 vs 0.554). Main limitations: three seeds only with no significance testing, a 10-seed control yielding 0.731±0.079 that questions the 0.796 headline, very small query sets, and baseline adaptation that may under-tune stronger models. Code and Zenodo-archived splits and checkpoints support high reproducibility.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Cross-E3-ligase PROTAC degradation prediction can be reformulated as episodic few-shot meta-learning, with prototypes estimated at a held-out ligase from K=1–2 support compounds per class and no encoder updates at test time.
- On the CRBN→VHL benchmark, ProMeta achieves AUROC 0.796 (K=2, Q=3) and 0.883 (K=2, Q=5), improvements of 19.9% and 6.8% over the supervised GNN baseline.
- Cross-ligase few-shot transfer is bidirectional but direction- and data-regime-dependent, with VHL→CRBN yielding 0.702/0.821 under the same protocol.
- Under K=1 on five rare E3 ligases, ProMeta attains 0.700 AUROC on 90 strict common-query rows, above adapted PROTAC-STAN (0.550) and DegradeMaster (0.564).
- Episodic representation learning and prototype inference contribute complementarily, with the episodic-training benefit concentrated in the VHL→CRBN direction (mean AUROC 0.761 vs 0.554 without).
Strengths
Unusually disciplined OOD benchmark design
Ligase-disjoint episodic splits with scaffold verification: cross-set Tanimoto similarity 0.145/0.163 vs within-CRBN 0.289 (Supplementary Fig. S2); identical support/query rows given to every method under three fixed seeds (42, 2025, 3407), with frozen split files archived on Zenodo.
Reciprocal ablation separates decision rule from representation
Table 2 shows ProMeta (0.840 CRBN→VHL / 0.761 VHL→CRBN mean AUROC) beats GNN encoder + prototype head (0.806/0.520), ECFP4 nearest-centroid (0.799/0.689), and ProMeta w/o episodic training (0.818/0.554), so neither component alone explains the result.
Confounds proactively addressed
High reproducibility package
Honest scope statements
Weaknesses
Thin inferential statistics in the main text
The headline number is seed-fragile
Baseline adaptation protocol may under-tune competitors
Extremely small evaluation sets limit evidential weight
Representation attribution incomplete
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ProMeta is a method paper reformulating PROTAC degradation-activity prediction as **cross-E3-ligase few-shot meta-learning**. The core claims: (1) a prototypical-network formulation where each E3 ligase defines a task, meta-trained on a source ligase and applied to held-out ligases via support-set prototypes without encoder updates; (2) on CRBN→VHL, AUROC 0.796 (K=2, Q=3) and 0.883 (K=2, Q=5), framed as 19.9%/6.8% improvements over the supervised GNN baseline; (3) bidirectional transfer with direction dependence (VHL→CRBN: 0.702/0.821); (4) K=1 rare-E3 AUROC of 0.700 over 90 common-query rows; (5) retrospective case studies (VZ185: 11/12 correct; CDK6: 11/14; BCL-xL: 6/6).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What performance does ProMeta achieve on PROTAC degradation prediction?
On the CRBN→VHL cross-ligase benchmark it reports AUROC 0.796 with K=2, Q=3 and 0.883 with K=2, Q=5; in the reverse VHL→CRBN direction it reports 0.702 and 0.821. Under one-shot (K=1) evaluation on five rare E3 ligases it reports 0.700 AUROC over 90 common-query rows, versus 0.550 (PROTAC-STAN) and 0.564 (DegradeMaster) for adapted baselines.
Is ProMeta's code and data publicly available?
Yes. The paper states that source code is on GitHub (github.com/yeyufeiyyf/prometa) and that datasets, experimental splits, model checkpoints, and released predictions are archived on Zenodo (doi 10.5281/zenodo.21371599).
What dataset does ProMeta use?
PROTAC-DB 3.0. The raw local export contains 9,384 rows, of which over 80% lack DC50/Dmax annotations. After filtering, balancing, and duplicate/conflict cleanup, the main CRBN/VHL transfer experiments use 1,386 clean compound–E3–target contexts (842 CRBN, 544 VHL; 690 positive, 696 negative).
Does ProMeta work for E3 ligases it was never trained on?
That is its design goal: the encoder is trained only on a source ligase (CRBN or VHL) and applied to held-out ligases by computing class prototypes from 1–2 labeled support compounds, with no parameter updates. Results are direction-dependent — stronger CRBN→VHL than VHL→CRBN — and the rare-E3 evidence is based on very small query sets, so the authors themselves frame it as a stress test rather than broad validation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- In the extracted Figure 2 AUROC heatmap, only one of DegradeMaster's four per-setting cell values is legible (0.668); DegradeMaster per-setting AUROCs were therefore not cited and narrative-prose numbers were used instead. AUPRC and BALACC heatmaps extracted completely.
- Supplementary materials (Tables S1–S11, Figs S1–S4) are referenced but not included in the provided text, so claims resting solely on supplement contents could not be verified. The main text through the references (item 32) is complete, with no end-of-text truncation.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.