In silico PROTAC modeling is the computational effort to design, assemble, model, score, and prioritize degrader candidates before synthesis or experimental testing. Unlike conventional small-molecule docking, PROTAC modeling must reason about a ternary system: protein of interest, PROTAC, and E3 ligase.
Binary affinity is not enough. The productive intermediate is a POI-PROTAC-E3 ternary complex whose behavior can depend on linker geometry, cooperativity, protein-protein contacts, lysine accessibility, residence time, permeability, and cellular context. Computational methods help triage and explain candidates, but they do not guarantee degradation.
In silico PROTAC modeling uses computational methods to generate and evaluate degrader candidates by combining component selection, linker feasibility, ternary complex construction, pose scoring, molecular dynamics, machine learning prediction, and experimental-prioritization logic.
The key design object is not just a ligand-protein complex but a ligand-induced protein-protein complex. A PROTAC can bind both proteins and still fail if the ternary orientation is unproductive, if the linker enforces strain, or if the resulting geometry does not support ubiquitination-competent presentation.
Best for: structure-first modeling, linker feasibility, mechanistic interpretation, ternary pose refinement, and explaining SAR.
Inputs: protein structures, ligand binding modes, anchor atoms, linker structures, and explicit geometric constraints.
Outputs: ternary poses, bridgeability checks, energy or stability layers, and ensemble behavior.
Risks: sampling cost, scoring sensitivity, brittle anchors, and incomplete handling of flexibility.
Best for: rapid triage, degradation prediction, pose re-ranking, candidate generation, and large-library prioritization.
Inputs: chemical structures, protein sequences or structures, assay labels, generated poses, and descriptors.
Outputs: degradation probability, pose viability, generated linkers, synthetic feasibility, and activity predictions.
Risks: dataset bias, weak out-of-domain generalization, limited interpretability, and dependence on training-data quality.
Physics-based workflows treat the degrader as a real geometric object instead of a loose protein-complex prompt. They are most useful when you know the warhead pose, recruiter pose, and plausible attachment atoms well enough to enforce bridgeability and interpret failures.
Early ternary docking and PRosettaC-style workflows use anchor atoms, distance restraints, and linker-compatible search to keep the pose space chemically meaningful.
MD can relax poses, reveal instability, and add MM/GBSA-style rescoring as one interpretive layer rather than a final truth metric.
HAPOD-style heating-accelerated pose departure stress-tests whether a pose persists under perturbation instead of only scoring a minimized snapshot.
DegraderTCM-style lower-cost pipelines support larger screens by trading structural nuance for speed and tractability.
Bridgeability, cyclic-coordinate linker placement, and solvation or energy-landscape mapping help identify protein orientations that a real linker can actually span.
These methods prioritize interpretability, trend-mapping, and linker-length or shape reasoning over atomistic precision.
BOTCP-style Bayesian optimization explores difficult pose spaces iteratively instead of relying on one docking run.
SILCS-xTAC-style strategies map favorable linker and interaction regions across ensembles to score geometric and energetic complementarity.
Mechanistic interpretability, explicit geometric feasibility, direct connection between linker and ternary pose, and usefulness for explaining SAR and failures.
Sampling cost, force-field sensitivity, incomplete flexibility, limited transferability across targets or E3 ligases, and no guarantee that static stability means cellular degradation.
Data-driven workflows shift the emphasis from explicit geometric construction to learned prediction, ranking, or generation. They can be extremely useful for throughput, but they inherit the quality, coverage, and biases of the data used to train them.
AlphaFold or AlphaFold3-style attempts can propose protein arrangements, while DeepTernary-style SE(3)-equivariant models aim to predict ternary geometry directly.
MAPD-style models ask whether a target is intrinsically suitable for degradation, which is a target-level question rather than a candidate-PROTAC question.
DeepPROTACs and DegradeMaster-style models predict degradation outcomes from molecular, protein, and sometimes 3D features, but remain sensitive to assay context and domain shift.
ET-PROTACs and PROTACable-style pipelines can re-rank ternary poses or connect docking outputs to learned structure-activity scoring.
PROTAC-Invent, ProLinker-Generator, DAD-PROTAC, DiffPROTAC-like, and graph-based models can propose new chemistry but still need feasibility and validation filters.
DeepPSA-style synthetic accessibility, property filters, and developability proxies help prevent large-scale generative outputs from becoming chemically unrealistic dead ends.
High throughput, ability to learn patterns from assay and structure data, usefulness for prioritizing large enumerations, and capacity to propose new chemistry.
Dataset bias, limited negatives, out-of-domain generalization risk, lower interpretability, and strong dependence on molecular-representation consistency.
This table is a field map, not a ranking. The best method depends on the design question, available structures, throughput needs, and validation plan.
| Year | Method | Where It Helps | Major Limitation |
|---|---|---|---|
| 2019 | In silico ternary docking | Early Monte Carlo plus protein-protein docking workflows showed that ternary pose generation was feasible and useful for initial geometry exploration. | Recovered some near-native poses, but scoring produced many false positives and was not robust enough for confident ranking. |
| 2020 | PRosettaC | Anchor-constrained Rosetta docking improves geometry-aware ternary construction and supports rational linker SAR interpretation. | Needs reliable anchor definitions and prior structural knowledge; performance falls when those assumptions are wrong. |
| 2021 | MD-refined ternary modeling | Rosetta or docking poses can be relaxed in explicit solvent, then inspected with MM/GBSA-like rescoring and cooperativity analysis. | Computationally expensive and still sensitive to imperfect scoring functions. |
| 2022 | HAPOD scoring | Stress-tests pose persistence with heating-accelerated MD instead of trusting one minimized snapshot. | Requires repeated MD runs and can be force-field sensitive. |
| 2022 | MAPD | Target-level degradability or tractability prediction helps decide whether a protein is a plausible degradation candidate at all. | Not candidate-PROTAC-specific; it scores target biology more than degrader chemistry. |
| 2022 | DeepPROTACs | Supervised degradation prediction can triage candidate molecules from molecular and protein features. | Outcome quality depends heavily on training labels, assay context, and domain coverage. |
| 2022 | Graph-based generative models | Can propose new degraders or optimize candidate chemistry across large design spaces. | Generated outputs are only as good as the predictive model and constraints guiding them. |
| 2023 | Coarse-grained cooperativity | Useful for understanding linker-length and protein-shape trends in ternary cooperativity. | Interpretive rather than atomistically precise for a specific medicinal-chemistry decision. |
| 2023 | Energy landscape mapping | CCD-style linker placement and energy or solvation landscapes reveal which protein orientations are physically bridgeable. | Provides thermodynamic understanding but is not a standalone docking engine. |
| 2023 | PROTAC-Invent | 3D generative linker design expands beyond empirical linker sets and proposes new bridge chemotypes. | Generated linkers still need structure, property, and activity validation. |
| 2023 | BOTCP | Bayesian optimization over pose parameters helps when initial ternary docking is ambiguous. | Search and scoring can still miss the correct pose even after iterative refinement. |
| 2024 | DegraderTCM | Lower-resource ternary construction supports broader screens when exhaustive sampling is too expensive. | Speed comes at the cost of structural nuance and subtle rearrangement capture. |
| 2024 | PROTACable | Integrates structure generation and learned activity prediction into a more end-to-end pipeline. | Still depends on structural templates and model retraining for new target classes. |
| 2024 | AlphaFold-style ternary adaptations | Can produce fully automated protein-complex hypotheses with minimal manual setup. | Often cannot enforce actual PROTAC geometry and may misorient complexes or overtrust accessory interfaces. |
| 2025 | DeepTernary | SE(3)-equivariant ternary structure prediction aims to generate 3D ternary complexes directly from proteins and degrader inputs. | Needs curated structural training data and may weaken on novel targets, E3 ligases, or scaffolds. |
| 2025 | DegradeMaster | Semi-supervised E(3)-equivariant degradation prediction adds geometry-aware features to outcome models. | Complex training setup and ongoing dependence on curated degradation assay datasets. |
| 2025 | ProLinker-Generator | Transformer-style linker generation expands linker chemical space with high novelty and validity. | Generated linkers still require downstream feasibility, docking, and property filtering. |
| 2025 | DAD-PROTAC | Diffusion-style generation adapts general molecule models toward large PROTAC-like linker chemistry. | Heavier computation and sensitivity to how domain adaptation is calibrated. |
| 2025 | ET-PROTACs | Cross-modal learned scoring helps re-rank docking ensembles by pose viability or complex stability. | It scores existing poses rather than generating them and depends strongly on input-pose quality. |
| 2026 | SILCS-xTAC | Grid-based ensemble scoring captures geometric and energetic complementarity across ternary ensembles. | Score transferability across target, E3, and linker classes still needs broad validation. |
| 2026 | SE(3)-PROTACs | Geometry-aware transformer models extend degradation prediction with 3D molecular graphs and protein context. | Performance still depends on curated labels and may fall under harder out-of-domain evaluation splits. |
Standardize warheads, recruiters, attachment atoms, and linker candidates before heavier modeling.
Launch builderStart from explicit geometry and bridgeability when structures and anchors are available.
Review constraint-driven designUse reporting and evaluation standards before trusting method comparisons.
Open benchmarkingMove prioritized candidates into restrained docking, MD, or learned re-ranking workflows.
Open downstream modelingStatic pose quality does not uniquely determine degradation. Energy scores may not correlate with biology. DockQ or RMSD require reference structures and may miss functional relevance. ML scores can be biased by training data. In cells, degradation depends on ternary formation, ubiquitination geometry, lysine accessibility, residence time, permeability, expression, and assay context.
The draft’s reporting checklist is a practical reminder that computational PROTAC studies are sensitive to chain definitions, attachment atoms, stereochemistry, protonation, conformer generation, missing residues, and software settings. Reproducibility requires more than final poses or a single summary score.
Report protein structures, constructs, chain definitions, modeled domains, and E3 complex composition.
Report warhead and recruiter binding modes, linker attachment atoms, anchor definitions, stereochemistry, protonation, and tautomer states.
Document conformer generation, constraints, failure handling, missing residues, and unresolved atoms.
Disclose docking, restraints, MD, scoring, filtering, pose selection, software versions, and relevant hyperparameters.
Define pose success, ranking success, degradation-prediction success, and any top-k or enrichment metrics.
Include binders that fail to degrade, linker-incompatible designs, nonproductive ternary poses, and other negative controls when possible.
Share random seeds, model checkpoints, scripts, environments, benchmark inputs, and output files where feasible.
State limitations around new targets, E3 ligases, linker chemotypes, molecular-glue-like systems, and biological contexts outside the validation domain.
Good individual binding does not ensure a productive degrader-induced target-E3 arrangement.
Exit vectors and anchor labels are not bookkeeping details; they define the geometry the linker must satisfy.
A linker that cannot physically connect the bound ligands can make every downstream score meaningless.
Overtrusting one docking, energy, or ML score hides uncertainty and often masks failure modes.
Applying predictors to new targets, E3 ligases, or chemotypes without checking domain fit is a common source of misleading confidence.
Missing protonation, tautomer, stereochemistry, and conformer protocols make studies hard to reproduce and compare.
Large generative outputs still need chemical sanity, synthetic feasibility, and geometry filters.
A promising ternary model still does not prove degradation, selectivity, exposure, or developability.
PROTAC Builder is not a full predictive modeling engine. It is a preparation and assembly layer that helps users select or import warheads, select or import E3 recruiters, define attachment atoms, choose or enumerate linkers, generate candidate PROTAC structures, standardize representation, and export candidates into downstream modeling or batch workflows.
Zoom out to the end-to-end design and validation workflow.
Read build guideUse anchor-aware geometric logic before heavier ternary modeling.
Review constraint-driven designRefine bridgeability hypotheses and linker-property tradeoffs.
Open linker guideInspect recruiter structures, solvent exposure, and attachment atoms.
Explore E3 recruitersSee how assembled candidates move into docking, MD, and scoring workflows.
Open downstream modelingReview what should be reported when comparing computational pipelines.
Open benchmarkingPrepare batch-ready payloads once candidate structures and attachments are standardized.
Open API BuilderSee how cross-site handoffs and assembly workflows are presented across the ecosystem.
Open examplesThis page draws in part from the Schurer Lab perspective From Ternary Modeling to Predictive PROTAC Design: A Computational Perspective, which frames modern computational PROTAC discovery as a staged workflow spanning candidate generation, linker feasibility, ternary complex modeling, learned scoring, ensemble validation, and experimental prioritization.
Figures and workflow concepts are presented here as project-owned educational content derived from that work and adapted for the PROTAC Builder platform. The goal of this page is to translate the publication’s computational perspective into a practical guide for researchers designing, modeling, and prioritizing PROTAC candidates in silico.