bio-protac-degraders

Designs PROTACs, molecular glues, and bivalent degraders with explicit handling of E3 ligase choice (VHL, CRBN, IAP, MDM2, KEAP1), linker design (length, composition, rigidity), ternary complex predic

SKILL.md

GitHub

Version Compatibility

Reference examples tested with: PRosettaC (web service), DeepTernary 1.0+, AlphaFold3 (constraints-enabled), Boltz-1 / Boltz-2, RDKit 2024.09+, OpenMM 8.1+ (for ternary MD).

Before using code patterns, verify installed versions match. If versions differ:

  • Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

PROTAC and Bivalent Degrader Design

Design bifunctional molecules (PROTACs) that recruit an E3 ubiquitin ligase to a target protein, inducing target ubiquitination and proteasomal degradation. PROTACs differ from traditional drugs: a stable ternary complex (target + PROTAC + E3) is required, not just target binding. The PROTAC field exploded post-2020 with clinical successes (ARV-471 estrogen-receptor degrader, ARV-110 androgen-receptor degrader). Postdoc-grade PROTAC design balances target ligand binding, E3 ligand binding, linker geometry (length, rigidity, chemistry), cooperativity (positive = ternary stable; negative = hook effect), and cell permeability (PROTACs are 800-1500 Da, often Lipinski-violating).

For target ligand design, see chemoinformatics/virtual-screening and chemoinformatics/admet-prediction. For linker-only enumeration, see chemoinformatics/reaction-enumeration. For generative linker design, see chemoinformatics/generative-design.

E3 Ligase Choice

E3 ligase Ligand series Best at Limitations
VHL VL-269 (Salami 2018) Surface-exposed targets Tissue-restricted expression
CRBN (cereblon) thalidomide, pomalidomide Broad tissue expression Off-target neosubstrates (IKZF1, SALL4)
IAP (XIAP, cIAP1) SMAC mimetics (LCL161, etc.) Apoptotic / IAP targets Limited target scope
MDM2 nutlin / idasanutlin TP53 pathway Limited target diversity
KEAP1 DDB1-DCAF15-Keap1 NRF2 pathway Specialized use
RNF114 EN450 Newer; under exploration Limited tooling
RNF4 EN450-RNF4 Newer; growing Limited tooling
Cyclin E - SKP2 New ligands emerging Targeted cancers Discovery-phase

Decision: For first-generation PROTAC, CRBN (cereblon) is the most-developed (thalidomide-derived; broad tissue distribution). VHL is second-most-developed (more selective; tissue-restricted). For specialty targets, consider IAP / MDM2.

Linker Design Principles

Linkers tune ternary complex geometry and stability:

Property Range Effect
Linker length 8-30 atoms Critical; geometry-dependent
Linker rigidity Flexible (PEG) vs rigid (piperazine, pyridine) Higher rigidity often reduces entropy penalty
Linker chemistry PEG, alkyl, piperazine, triazole, ether, amide PEG common; rigid for tighter binding
Click chemistry compatibility Triazole compatible Easy synthesis
MW range PROTAC 800-1500 Da Lipinski-violating but accepted
Polar atoms 1-5 per linker Permeability vs solubility balance

Critical: The "Goldilocks linker length" is target-specific. Too short = ternary clash; too long = ternary entropy too high. Typically 12-20 atoms for surface-exposed targets.

Decision Tree by Scenario

Goal E3 / linker Tools
First-generation PROTAC, surface-exposed target CRBN + PEG linker (10-15 atoms) PRosettaC for ternary prediction
Selective degrader (avoid off-target) VHL + rigid linker PRosettaC + cellular validation
BTK / IAP family targets IAP-based PROTAC Standard pipelines
Targeted protein degradation in cancer CRBN or VHL Standard clinical track
Novel target, no cryptic Multiple E3 / linker variants Combinatorial design + PRosettaC
Molecular glue (non-PROTAC) CRBN-based Distinct mechanism; pomalidomide-class
Optimize cooperativity (alpha > 1) PRosettaC iterative; experimental Ternary MD + ITC
Cell-active candidate Standard development PK + degradation cellular assays

Ternary Complex Prediction Tools

Tool Approach Strength Fails when
PRosettaC Rosetta + REMD on PROTAC Mature; reliable License Rosetta
DeepTernary Equivariant deep learning Fast; SE(3) OOD chemistry
AlphaFold3 + constraints Foundation model + restraints High accuracy Public access limited
Boltz-1 / Boltz-2 + constraints Foundation model Fast; flexible Limited PROTAC training
HADDOCK Restraint-based MD Veteran Manual restraint specification
Schrodinger Phase + Glide Commercial Production-ready License cost

Decision: For first-pass ternary modeling, PRosettaC is the standard. For prospective generative + ranking, AlphaFold3 with constraints (if access) or Boltz-1/2 (open). DeepTernary is a fast alternative when validated against benchmark.

Cooperativity (Alpha)

Cooperativity quantifies how the ternary complex stabilizes (or destabilizes) the binary binding:

alpha = (Kd_binary,target) / (Kd_ternary,target)
  • alpha > 1: positive cooperativity (ternary stronger than binary)
  • alpha = 1: no cooperativity (independent binding)
  • alpha < 1: negative cooperativity (mutual destabilization)

Positive cooperativity (alpha > 2-3) is the gold standard for PROTAC design. It means the ternary complex is stabilized by protein-protein interactions induced by the linker geometry.

Measure with ITC (isothermal titration calorimetry) or SPR/BLI titrations of binary vs ternary.

DC50 / Dmax Characterization

In cellular assays:

  • DC50: PROTAC concentration for 50% degradation (analogous to IC50)
  • Dmax: maximum fraction degraded at any concentration
Property Good value Notes
DC50 < 100 nM Clinically meaningful
Dmax > 80% Sufficient depletion for phenotype
Hook effect < 10x range Linear dose-response over 100x
Hook concentration typically > 10x DC50 Above which binary complexes dominate

Hook effect: at high PROTAC concentrations, binary complexes (PROTAC-target alone, PROTAC-E3 alone) dominate, and ternary complex formation drops. Dose-response curves are bell-shaped.

Ternary Complex Modeling Workflow

Goal: Predict 3D structure of target-PROTAC-E3 ternary complex.

Approach:

  1. Start with binary co-crystals: target + target-ligand pose; E3 + E3-ligand pose
  2. Connect via linker enumeration (combinatorial)
  3. Score by geometric feasibility (linker length, no clashes)
  4. Refine with energy minimization
# Pseudo-code workflow
def predict_ternary(target_pdb, target_ligand_sdf,
                    e3_pdb, e3_ligand_sdf, linker_smiles):
    # 1. Place binary complexes in same coordinate frame
    # 2. Enumerate linker connectivity from target-ligand exit vector to e3-ligand entry vector
    # 3. Score by total linker length, RMSD to expected geometry
    # 4. Minimize with restraint MD (5-20 ns)
    return ternary_poses

For production workflow, use PRosettaC web service or AlphaFold3 with chain-chain constraint inputs.

Linker Length Calculation

from rdkit import Chem
from rdkit.Chem import AllChem

def linker_distance(target_pose, e3_pose, target_ligand_smi, e3_ligand_smi, n_conf=20):
    """
    Compute distance between target-ligand exit vector and e3-ligand entry vector.
    Linker must span this distance with flexibility.
    """
    # Place target + e3 in same frame (from PRosettaC or HADDOCK)
    # Find target-ligand attachment atom and e3-ligand attachment atom
    # Compute distance
    pass

def required_linker_atoms(distance_A, rigidity='flexible'):
    """
    Estimate atoms needed in linker to span distance.
    Flexible (CC bond ~1.5 A) -> distance / 1.5
    Rigid (sp2 C-C ~1.4 A) -> distance / 1.4 + curvature factor
    """
    if rigidity == 'flexible':
        return int(distance_A / 1.4 * 1.2)  # 20% slack
    return int(distance_A / 1.4 * 1.5)

Generative Linker Design

Combine REINVENT 4 linker mode with ternary prediction:

[run]
type = "linker"
prior_file = "priors/linker.prior"

[parameters]
input_smiles = "[*]NC(=O)..."  # Target ligand with attachment
linker_target_smiles = "[*]NC(=O)..."  # E3 ligand with attachment
n_steps = 200

[scoring_function.components]
type = "linker_length"
weight = 0.3
target_atoms = 14

type = "ternary_score"
weight = 0.5
ternary_predictor = "deepternary"  # API call

type = "qed"
weight = 0.2

REINVENT 4 supports linker generation; scoring needs ternary prediction (currently expensive). Modern best practice: combinatorial enumeration with PRosettaC scoring.

Per-Tool Failure Modes

PRosettaC -- inaccessible E3 in selected ligase

Trigger: Target's known binding mode incompatible with E3 ligase orientation.

Mechanism: PRosettaC samples linker conformations; can't bridge if E3 surface is occluded.

Symptom: Low ternary complex scores; high RMSD across replicates.

Fix: Try different E3 (CRBN vs VHL); use ARV-110 / ARV-471 family templates.

DeepTernary -- novel chemotype

Trigger: Target ligand or E3 ligand outside training distribution.

Mechanism: DeepTernary learns from TernaryDB; novel chemotypes extrapolate.

Symptom: Predicted ternary complex unrealistic.

Fix: Validate against PRosettaC; use as fast screening, validate top-N with PRosettaC.

Hook effect at low PROTAC concentration

Trigger: PROTAC linker too long; binary preferred energetically.

Mechanism: Negative cooperativity at low concentration; ternary unstable.

Symptom: Dose-response curve narrow; DC50 close to hook.

Fix: Shorter linker; increase cooperativity by adding protein-protein contact-promoting motifs.

Insufficient cell permeability

Trigger: PROTAC > 1200 Da, polar, high TPSA.

Mechanism: Lipinski violations; permeability bottleneck.

Symptom: Cellular DC50 1000x worse than biochemical DC50.

Fix: Optimize linker for amphipathic profile; PROTACs often violate Ro5 but successful ones are < 1200 Da with < 130 TPSA.

E3-target distance miscalculation

Trigger: Computing linker length from binary models without ternary refinement.

Mechanism: Binary structures don't capture ternary geometry.

Symptom: PROTACs synthesized at wrong linker length; no degradation.

Fix: Use PRosettaC-refined ternary distance; enumerate ±3 atoms around predicted optimal.

Molecular glue vs PROTAC confusion

Trigger: Designing as PROTAC when target lacks defined ligand.

Mechanism: Molecular glues (like thalidomide for IKZF1) don't have target ligand component.

Symptom: Design too rigid; no degradation despite ternary prediction.

Fix: For targets without known ligand, consider molecular glue discovery instead.

Reconciliation: PRosettaC vs AlphaFold3 (Constraint-Based)

Aspect PRosettaC AlphaFold3
Approach REMD on linker + binary Foundation model + restraints
Accuracy High (validated benchmark) Limited PROTAC data in training
Speed Hours per ternary Minutes
Open access Yes (web service) Public API, limited usage
Customization Limited Constraint flexibility
Decision Standard for PROTAC Promising; validate

For PROTAC design, PRosettaC + ITC/SPR experimental validation is the current best practice.

Common Errors

Symptom Cause Fix
PRosettaC fails to converge Linker too long or E3 surface buried Try different E3; check binary structure quality
DeepTernary returns clashing pose OOD chemotype Use PRosettaC for refinement
AlphaFold3 ternary unrealistic Constraint not specified properly Use chain-chain distance constraints
PROTAC cell-active but biochemical degradation poor Off-target neosubstrate (CRBN-IKZF) Confirm target specificity via mass spec
Hook effect at low PROTAC concentration Negative cooperativity Optimize linker geometry
Synthesis too complex Linker via 5+ step Use click chemistry (triazole linker)
MW > 1500 Da issue Excessive linker Reduce linker length; rigid linker

References

  • Békés et al., Nat. Rev. Drug Discov. 21:181 (2022) -- PROTAC clinical review.
  • Drummond ML, Williams CI 2019 J Chem Inf Model 59(4):1634-1644 -- PROTAC ternary-complex modeling principles.
  • Schapira et al., Nat. Rev. Drug Discov. 18:949 (2019) -- targeted protein degradation overview.
  • Zaidman D, Prilusky J, Kessler N, Hadar D, Zarivach R, London N 2020 J Chem Inf Model 60(10):4894-4903 -- PRosettaC for ternary modeling (DOI 10.1021/acs.jcim.0c00589).
  • Liu et al., Nat. Commun. (2025) -- DeepTernary deep learning.
  • Saunders et al., Sci. Rep. (2025) -- PRosettaC vs AlphaFold3 comparison.
  • Bondeson et al., Nat. Chem. Biol. 11:611 (2015) -- early PROTAC clinical readout.

Related Skills

  • chemoinformatics/molecular-io - Parse linker and ligand SMILES
  • chemoinformatics/reaction-enumeration - Linker enumeration combinatorial
  • chemoinformatics/generative-design - REINVENT linker mode
  • chemoinformatics/conformer-generation - Ternary conformer sampling
  • chemoinformatics/virtual-screening - Validate target ligand binding
  • chemoinformatics/free-energy-calculations - Ternary ABFE / cooperativity
  • chemoinformatics/admet-prediction - PROTAC ADMET specific challenges
  • structural-biology/structure-io - PDB / mmCIF for ternary complex