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

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:

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

Ternary Complex Prediction Tools

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

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)

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

Common Errors

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