Descriptor and property filtering
Fast first-pass triage for MW, cLogP or logD, HBD/HBA, TPSA, rotatable bonds, charge, and basic developability risk. Useful early, but it does not prove ternary formation.
PROTAC Downstream Modeling
Downstream PROTAC Modeling Tools
PROTAC Builder outputs are starting points for deeper computational and experimental decision-making. Once a candidate degrader is assembled, downstream modeling asks whether the molecule is chemically valid, geometrically plausible, bridgeable in 3D, compatible with a target-E3 ternary complex, and worth prioritizing for synthesis or biological testing.
The strongest workflows move in stages: preserve the right metadata, screen for obvious feasibility issues, build ternary-complex hypotheses, refine only the most credible poses, and interpret every score as one layer of evidence rather than a guarantee of degradation.
Builder output to 3D handoff
Geometry before heavy scoring
Physics plus ML
Experiments still decide
Quick answer: what happens after PROTAC Builder?
Downstream modeling starts once a candidate has been assembled and exported with enough chemical and structural context to support reproducible 3D evaluation.
- Export a chemically valid candidate structure.
- Preserve warhead, linker, recruiter, and attachment-atom metadata.
- Generate or validate 3D conformers.
- Check basic descriptors and property burden.
- Test linker bridgeability and exit-vector geometry.
- Build candidate ternary-complex poses.
- Refine and score prioritized poses.
- Use ML predictors or re-rankers where appropriate.
- Benchmark and document the workflow.
- Prioritize candidates for experimental degradation assays.
Takeaway: downstream modeling is a filtering and hypothesis-generation layer, not a replacement
for degradation experiments.
What should leave PROTAC Builder?
Downstream workflows are only as reliable as the handoff. Preserve enough chemistry and metadata so the molecule can be reconstructed, interpreted, and benchmarked without guesswork.
Full assembled PROTAC structure
SMILES, MOL, SDF, or other supported export format
Warhead identity and boundary
E3 recruiter identity and boundary
Linker identity and chemistry
Warhead-side attachment atom
Recruiter-side attachment atom
Atom mapping or equivalent connection notation when available
Stereochemistry
Protonation and tautomer assumptions when known
Intended POI structure or model
Intended E3 ligase or recruiter-bound structure
Notes on solvent-exposed exit vectors
Builder version or export date where possible
Reproducibility warning: if attachment atoms are lost during export, downstream geometry-aware
modeling becomes much less reproducible.
{
"protac": "assembled structure",
"warhead_anchor_atom": "documented",
"recruiter_anchor_atom": "documented",
"component_boundaries": ["warhead", "linker", "recruiter"],
"stereochemistry": "documented",
"protonation_notes": "documented when known",
"target_structure": "PDB or model reference",
"e3_structure": "PDB or model reference"
}Downstream method families at a glance
Linker bridgeability and geometry checks
Tests whether anchors and exit vectors can plausibly be connected in 3D before expensive docking or MD. Strongly dependent on correct anchor definitions.
Restrained or tethered ternary docking
Builds POI-PROTAC-E3 poses under distance or linker constraints. Useful for structure-first design, but sensitive to sampling and scoring.
PRosettaC-style modeling
Uses anchor atoms and linker constraints to guide Rosetta-based ternary modeling. Strong when binding modes are known; brittle when anchors or linker hypotheses are wrong.
Efficient ternary construction workflows
DegraderTCM-style lightweight assembly and minimization support rapid screens, but may miss subtler protein-interface rearrangements.
Molecular dynamics refinement
Relaxes poses and tests local stability for prioritized candidates. Useful, but expensive and force-field sensitive.
Dynamic stability scoring
HAPOD-style or short-MD stress tests help rank pose persistence rather than just minimized snapshots. Requires careful setup and interpretation.
Grid or field-based ensemble methods
SILCS-xTAC-style approaches evaluate favorable linker or interaction regions across ensembles and can add field-based reasoning to pose review.
ML re-ranking and prediction
DeepPROTACs, DegradeMaster, ET-PROTACs, PROTACable, and related models can prioritize larger sets quickly, but only within their training domain.
Generative design feedback loops
PROTAC-Invent, ProLinker, DAD-PROTAC, diffusion, and graph generators can propose new candidates after failure analysis, but all outputs need chemistry and geometry filters.
Stage 1: descriptor and property triage
Before heavy modeling, check whether the assembled candidate is chemically valid and carries a plausible property burden for the intended context. PROTACs often live beyond classic Rule-of-Five space, so these filters are best used as risk indicators rather than hard pass or fail rules.
MW
cLogP or logD
HBD and HBA
TPSA
Rotatable bonds
Formal charge
Chiral centers
Linker length and polarity
Structural alerts
Synthetic accessibility or developability score where supported
This is where large enumerations and generative outputs should be pruned before expensive structure-based work. DeepPSA-style synthetic accessibility or related feasibility modules can support this stage, but they should be interpreted as triage signals rather than final answers.
Stage 2: geometry and bridgeability checks
A 2D PROTAC can look perfectly reasonable while being impossible or highly strained in 3D. Bridgeability checks ask whether the linker can plausibly connect the warhead and recruiter anchors across candidate protein orientations without buried paths, impossible distances, or severe clashes.
Are both anchor atoms solvent-exposed?
Does the linker exit toward the partner protein?
Is the linker too short, too long, or overconstrained?
Are key binary binding interactions preserved?
Are alternate attachment atoms worth testing?
Interpretation: bridgeability is a feasibility test, not a proof of productive degradation.
Constraint-driven design
Use anchor-aware and exit-vector-aware reasoning before expensive ternary-complex modeling.
Review constraintsLinker design
Compare flexible, rigid, and polarity-balanced linker hypotheses before final handoff.
Review linkersStage 3: ternary complex construction
Restrained or tethered docking
Uses geometric restraints to keep the PROTAC linker compatible with both bound ligands. Useful for POI and recruiter pairs with known structures, but it needs decoys and ranking checks.
PRosettaC-style workflows
Use anchor atoms and distance constraints to guide Rosetta sampling. Especially useful when the input binding modes are reliable and you want interpretable linker-geometry hypotheses.
DegraderTCM-style efficient construction
Faster and lighter ternary modeling for larger candidate sets. Strong as a triage layer, but not definitive for subtle protein-interface rearrangements.
Energy landscape and linker-feasibility approaches
Restrict evaluation to linkable protein-protein orientations and help explain why some linker lengths, vectors, or chemotypes fail before deeper refinement.
Stage 4: refinement and dynamic evaluation
Once you have a smaller set of plausible poses, use refinement and ensemble methods to test whether those poses hold together under more realistic conditions.
Energy minimization, explicit-solvent MD, MM/GBSA-style rescoring, HAPOD-style dynamic stress testing, short-MD persistence checks, ensemble evaluation, and SILCS-xTAC-style field-based scoring.
A stable pose does not guarantee degradation. Cell context, E3 expression, lysine accessibility, permeability, ubiquitination geometry, and assay conditions still matter.
Caution: a beautiful MD trajectory can still be biologically irrelevant if the cell context, E3
expression, target lysines, permeability, or assay design are unfavorable.
Stage 5: scoring, ranking, and interpretation
Downstream modeling should not rely on one score. The goal is to build a layered argument for why a candidate is worth keeping, not to hide score disagreement.
Chemical validity
Linker bridgeability
Severe clash check
Interface plausibility
Buried surface and contact quality
Ligand and linker strain
Solvent exposure
Lysine proximity or ubiquitination-plausibility proxies
Ensemble stability
Learned pose viability or degradation probability
Experimental feasibility
- DockQ and interface RMSD are most useful when reference ternary structures exist.
- Docking scores and Rosetta energies can support ranking, but they need calibration.
- ML predictions should be interpreted with training-domain and uncertainty limits in mind.
- Score disagreement is informative and should be documented, not hidden.
Stage 6: ML re-ranking and predictive layers
Learned predictors can help triage large candidate sets after geometry filters. They are most useful when you need throughput, uncertainty-aware prioritization, or learned re-ranking on top of structure-first generation.
DeepPROTACs and related degradation predictors
Useful for candidate-level degradation prioritization from molecular and protein features.
DegradeMaster and geometry-aware predictors
Useful when spatial arrangement matters and the training data covers similar systems.
ET-PROTACs and PROTACable-style re-ranking
Useful for pose-ensemble prioritization or integrative 3D modeling pipelines.
DeepTernary-style structure prediction
Useful when direct ternary-geometry prediction is the main question rather than classic restrained docking.
ML caution: random-split performance can look strong while failing under new targets, new E3
ligases, new linker classes, or different assay contexts.
Stage 7: generative feedback loops
Downstream failure analysis should feed back into design. If bridgeability fails, descriptors look poor, or ternary geometry is repeatedly unconvincing, the next step may be a new linker, a different attachment vector, or even a different recruiter or warhead.
If bridgeability fails, generate or enumerate new linkers.
If descriptors are poor, try property-balanced linker alternatives.
If ternary geometry is weak, revisit attachment vectors, linker rigidity, or E3 choice.
If learned activity is weak, test alternate warhead or recruiter combinations.
PROTAC-Invent-style 3D linker generation, ProLinker-style language-model workflows, DAD-PROTAC, diffusion, and graph-based generators can propose new molecules, but every output still needs chemical sanity, synthetic feasibility, bridgeability, and biological validation filters.
Recommended downstream workflow
- Builder export: export candidate structure and metadata.
- Representation audit: confirm atom mapping, anchors, stereochemistry, and protonation assumptions.
- Descriptor triage: remove obvious chemistry or property failures.
- Geometry triage: check exit vectors, linker bridgeability, and severe clashes.
- Rapid ternary construction: generate candidate poses with docking, restrained modeling, or low-resource pipelines.
- Pose filtering: rank by interface plausibility, linker strain, and geometry.
- Selective refinement: use MD, HAPOD-style scoring, or ensemble checks for top candidates.
- Learned prioritization: apply ML re-ranking or degradation predictors with uncertainty.
- Benchmark-ready reporting: document inputs, parameters, scores, and failure modes.
- Experimental follow-up: test degradation, DC50, Dmax, selectivity, target engagement, and hook effect.
Benchmark-ready handoff checklist
If you want downstream results to be reproducible, report the handoff as carefully as the score.
System definition, protein structures and chain IDs, E3 complex composition, PROTAC representation, warhead-recruiter-linker boundaries, attachment atoms, anchor definitions, and stereochemistry or protonation state.
Conformer-generation protocol, software and versions, random seeds, restraints, scoring settings, decoys or negative controls, and the exact output poses and scores kept for review.
Domain of applicability, failure modes, validation status, and whether the workflow is being used for structure recovery, re-ranking, degradation prediction, or generative redesign.
PROTAC workflows are highly sensitive to attachment atoms, protonation, conformers, construct definitions, and restraint choices, so incomplete reporting undermines comparison across methods.
Common downstream modeling mistakes
Exporting a molecule without anchor metadata
Treating 2D validity as 3D feasibility
Running docking before checking linker bridgeability
Ignoring solvent-exposed exit vectors
Using one score as the final answer
Overtrusting ML predictions outside the training domain
Ignoring failed conformer generation
Running expensive MD on poorly prepared systems
Forgetting E3 expression or cellular context
Claiming degradation from a model without experimental validation
How this page connects to the ecosystem
How to Build a PROTAC
Use the staged assembly workflow before you ever hand a candidate into downstream modeling.
Read guideLinker Design
Review linker length, rigidity, polarity, and exit-vector tradeoffs before bridgeability analysis.
Review linkersE3 Recruiter Discovery
Inspect recruiter binding poses, solvent exposure, and attachment sites before ternary modeling.
Explore recruitersConstraint-Driven PROTAC Design
Ground handoff decisions in anchor-aware geometry and bridgeability reasoning.
Review constraintsIn Silico PROTAC Modeling
See the broader computational landscape beyond this practical handoff guide.
Read modeling guideBenchmarking
Use benchmark-ready reporting and task-specific metrics when comparing methods or publishing results.
Open benchmarking guideAPI Builder
Support scripted and batch-oriented handoffs for larger enumerations and reproducible pipelines.
Open API BuilderExamples
Use case-study pages to show what an end-to-end builder-to-modeling workflow can look like.
View examplesReference note
This page is based in part on the Schürer Lab perspective manuscript From Ternary Modeling to Predictive PROTAC Design: A Computational Perspective. It is intended as an educational workflow guide for downstream computational handoffs.
The focus here is practical workflow design: what to export, what to check first, when to use heavier physics, when to use learned prioritization, and how to document the whole process so results remain interpretable and reproducible.