Computational PROTAC modeling is advancing quickly, but the field cannot compare methods reliably without shared benchmarks, common molecular representations, consistent metrics, and transparent reporting.
Benchmarking is not about declaring one universal winner. It is about making methods comparable, reproducible, and useful across targets, E3 ligases, linker chemotypes, and assay contexts. For PROTACs, that means evaluating both geometry and downstream biological relevance without pretending one score can summarize everything.
A useful PROTAC benchmark should separate at least four tasks:
The key modeled object is a PROTAC-induced ternary complex, not a simple binary ligand-protein pose. Degradation depends on more than static geometry, and different papers define “success” differently depending on whether they are predicting structure, ranking poses, forecasting degradation, or generating chemistry.
Question: can the method recover experimentally observed ternary complex geometry?
Metrics: DockQ, interface RMSD, ligand RMSD where applicable, fraction of native contacts, interface precision or recall, and clash or linker-feasibility checks.
Cautions: global alignment can hide degrader-relevant interface errors, accessory proteins can inflate scores, and static crystal structures are snapshots rather than full ensembles.
Question: can the method rank productive or near-native poses above decoys?
Metrics: top-1, top-5, or top-k success, enrichment factors, score calibration, rank correlation against reference quality, and decoy rejection rate.
Cautions: a method may generate a good pose but fail to rank it well, and raw docking or Rosetta energy should not be treated as uncalibrated truth.
Question: can the method predict whether a candidate PROTAC degrades the target in a defined biological context?
Metrics: AUROC, AUPRC, accuracy, balanced accuracy, calibration curves, regression to DC50, Dmax, or percent degradation, and assay-stratified or target-E3 split performance.
Cautions: labels are noisy, assay context matters, and a plausible ternary structure does not guarantee cellular degradation.
Question: do generated linkers or degraders survive downstream feasibility and validation filters?
Metrics: validity, uniqueness, novelty, synthetic accessibility, property distribution matching, conformer pass rate, bridgeability pass rate, modeling pass rate, and experimental follow-up rate when available.
Cautions: a molecule can be valid in 2D and still impossible or poor in 3D, so novelty is not enough without feasibility.
A practical PROTAC benchmarking standard is still a proposed community direction rather than an established standard. It should help groups compare results without forcing every method into the same mold.
Common input schemas, common output formats, task-specific metrics, required metadata, recommended negative controls, domain-shift splits, and reproducibility assets.
A universal standard does not mean one universal score. It means a shared language for what was modeled, how it was prepared, how it was evaluated, and where it should or should not be trusted.
Include PDB or structure source, POI and E3 chains, accessory proteins, PROTAC structure, ligand poses, anchor atoms, linker identity, construct boundaries, missing residues, biological assembly choice, and the scored interface definition.
Include target, E3, warhead, linker, recruiter structures, cell line, assay type, treatment time, concentration series, DC50 and Dmax when available, label definition, selectivity context, and inactive analogues when available.
Include split definition, seed molecules or conditioning inputs, allowed chemistry constraints, validity and novelty metrics, downstream structural-feasibility metrics, and experimental confirmation status when available.
Benchmarks need binders that fail to degrade, linker-incompatible designs, nonproductive ternary poses, and decoy poses that are chemically plausible but geometrically wrong. Without negatives, learned models can latch onto shortcuts and pose-ranking methods cannot be evaluated fairly.
PROTACs are representation-sensitive because they combine long flexible linkers, explicit attachment points, stereochemistry, tautomer or protonation ambiguity, conformer-generation failures, and multiple possible head-group definitions. Without consistent molecular and workflow representations, it becomes difficult to compare docking, learned scoring, and generative outputs in a reproducible way.
To improve reproducibility, cross-method comparison, and prospective usability, computational PROTAC studies should report the following whenever applicable.
Protein structures, constructs, chain definitions, modeled domains, and E3 complex composition.
Warhead and E3-ligand binding modes, linker attachment atoms, anchor definitions, stereochemistry, protonation, and tautomer states.
Conformer-generation procedures, constraints, failure handling, protein-preparation steps, missing residues, and unresolved atoms.
Docking, restraints, MD, scoring, filtering, pose-selection parameters, software versions, and hyperparameters.
Definitions of pose success, ranking success, degradation-prediction success, generative-model success, top-k metrics, and enrichment metrics.
Binders that fail to degrade, linker-incompatible designs, nonproductive ternary poses, and other control cases.
Random seeds, model checkpoints, scripts, containers, benchmark inputs, and output files needed to reproduce the workflow.
Limitations for new targets, E3 ligases, linker chemotypes, molecular-glue-like systems, or biological contexts outside the validation set.
Random splits can overestimate performance. The field needs harder generalization tests that expose where methods collapse rather than only where they succeed.
Tests whether the model can move beyond familiar protein families.
Tests whether performance generalizes beyond dominant CRBN or VHL-heavy datasets.
Tests learned dependence on familiar POI-ligand chemistry.
Tests generalization beyond common PEG or alkyl linker motifs.
Reduce leakage from near-duplicate analogues or chronologically later structures.
Expose dependence on one cell line, one readout format, or one treatment window.
Geometry and stability do not uniquely determine degradation. DockQ or RMSD may help for structure recovery but do not measure cellular activity. Docking or Rosetta energy may not correlate with interface correctness. ML confidence may not be calibrated under domain shift. Generative validity does not mean biological usefulness.
This is a conceptual example, not a backend implementation requirement. The goal is to show the kind of structured metadata that makes benchmark outputs comparable across groups.
{
"benchmark_task": "ternary_structure_prediction",
"system_id": "example_target_e3_protac",
"target": {
"name": "POI",
"structure_id": "PDB_OR_MODEL_ID",
"chains": ["A"],
"construct_notes": "domain boundaries and missing residues"
},
"e3_ligase": {
"name": "VHL_or_CRBN_or_other",
"structure_id": "PDB_OR_MODEL_ID",
"chains": ["B"],
"accessory_proteins": ["optional"]
},
"protac": {
"smiles": "atom-mapped SMILES",
"warhead_anchor_atom": "atom_map_id",
"recruiter_anchor_atom": "atom_map_id",
"stereochemistry_documented": true,
"protonation_state_notes": "documented"
},
"modeling_protocol": {
"method": "PRosettaC_or_other",
"software_version": "version",
"random_seed": "seed",
"parameters": "summary_or_file_link"
},
"evaluation": {
"metrics": ["DockQ", "interface_RMSD", "top_k_success"],
"negative_controls": ["decoy_pose_set"],
"domain_split": "leave_one_E3_out"
},
"reproducibility": {
"input_files": "available",
"output_files": "available",
"container_or_environment": "available"
}
}PROTAC Builder is not a benchmark standard and not a validated benchmarking engine. It can, however, serve as a preparation and standardization layer for benchmark-ready computational workflows by helping users normalize component boundaries, attachment atoms, and assembly outputs before downstream modeling.
Component selection, warhead and recruiter boundaries, attachment atoms, linker choices, assembled candidate structures, export formats, and batch or API-ready handoffs.
It is not presented as the official global benchmark standard or as a guarantee of predictive modeling performance or biological degradation.
Read the broader computational guide that frames physics-based, ML, and hybrid workflows.
Open in silico guideReview geometry-aware assembly, anchor atoms, and bridgeability logic.
Read constraint-driven designSee where builder outputs feed into docking, MD, scoring, and hybrid modeling workflows.
Open downstream modelingZoom out to the full staged design and validation workflow.
Read build guideReview linker feasibility, bridgeability, and property tradeoffs.
Open linker guidePrepare benchmark-ready or batch-ready computational handoffs once candidates are standardized.
Open API BuilderInspect examples and cross-site handoffs across the broader PROTAC Builder ecosystem.
Open examplesContextual external commentary link supplied by the project owner; treat it as related reading rather than a validated benchmark source.
Open related commentary ↗Related published benchmarking study already referenced elsewhere in the site ecosystem.
Open study ↗This page is based in part on the Schurer Lab perspective draft From Ternary Modeling to Predictive PROTAC Design: A Computational Perspective. It is intended as an educational benchmarking guide & will be maintained by the lab and team.