How are AlphaFold 3 predictions being used for PROTAC ternary modeling and binding site discovery, and what experimental validation catches the designs that look good computationally but fail to fold

AlphaFold 3 (AF3) facilitates PROTAC (Proteolysis Targeting Chimera) research by enabling end-to-end structural prediction of ternary complexes and potential binding sites without requiring predefined binary binding modes. However, computational models are treated as structural hypotheses that require rigorous experimental validation through kinetic, cellular, and high-resolution imaging techniques to identify designs that appear viable in silico but fail to function biologically.

AlphaFold 3 in PROTAC Modeling and Discovery

AF3 significantly advances targeted protein degradation (TPD) research by modeling the spatial arrangement of the E3 ligase, the protein of interest (POI), and the small-molecule degrader.

• Direct Ternary Prediction: Unlike previous docking methods (e.g., HADDOCK or RosettaDock) that require prior knowledge of binary binding modes and extensive CPU hours, AF3 directly predicts complete ternary complexes from protein sequences and ligand SMILES strings (Direct, High; PMID: 41831109).
• Confidence Scoring: AF3 reports intrinsic metrics such as ipTM (interface predicted Template Modeling score) and pTM to estimate the accuracy of inter-chain interfaces and global topology (Direct, High; PMID: 41831109, PMID: 38718835).
• Structural Context: Inclusion of accessory scaffold proteins (e.g., Elongin B/C for VHL-based systems or DDB1 for Cereblon-based systems) can stabilize AF3 predictions by providing essential spatial context, though this may sometimes result in inflated interface accuracy scores due to contacts with non-target subunits (Direct, High; PMID: 41152377).
• Binding Site Discovery: AF3's diffusion-based architecture allows for the discovery of novel protein-protein interfaces (PPIs) induced by the PROTAC, which are often distinct from the global energy minima of the two proteins interacting in isolation (Derived, Medium; PMID: 41831109, PMID: 36961978).

Catching Computational Failures

Computational designs may "look good" in static models but fail due to non-physical linker geometries, lack of cooperative stability, or poor cellular availability.

• Linker and Geometric Failures:

  • Linker Constraints: AF3 often lacks explicit structural constraints for linkers, sometimes resulting in protein orientations that preclude physical linker placement despite high confidence scores (Direct, High; PMID: 41152377).
  • Twisting and Flexibility: Failures often stem from "twisting" or misaligned orientations around the flexible PROTAC linker, as static models struggle to capture the conformational variability of these systems in solution (Direct, High; PMID: 41831109).
  • Stereochemistry: AF3 may disregard ligand stereochemistry, such as chirality, which is critical for bioactive PROTAC binding (Direct, Medium; PMID: 41831109, PMID: 38718835).

• Dynamic and Stability Validation:

  • Molecular Dynamics (MD): All-atom MD simulations are used to test the stability of predicted poses. Benchmarking shows that AF3 models often do not spontaneously converge toward experimentally resolved geometries during MD relaxation, suggesting the models may occupy inaccessible conformational regions (Direct, High; PMID: 41152377).
  • SPR (Surface Plasmon Resonance): This technique measures the dissociative half-life of the ternary complex. A long ternary half-life (residence time) often correlates with more effective polyubiquitination and degradation, catching designs that bind but dissociate too rapidly to function (Direct, High; PMID: 30721025).

• Cellular and In Vivo Functionality:

  • Environment-Sensitive Reporters (ESR): Novel reporters (e.g., JQ1-NR) use environment-sensitive fluorophores that "light up" only when bound in a nonpolar hydrophobic pocket. These tools allow non-invasive in vivo quantification of protein levels, identifying PROTACs that fail to induce degradation in living systems despite favorable modeling (Direct, High; PMID: 39987172).
  • NanoBRET and CETSA: NanoBRET evaluates target engagement in both live and permeabilized cells to identify models that fail due to poor cellular permeability or transporter-mediated efflux. CETSA (Cellular Thermal Shift Assay) confirms physical target engagement within complex cellular lysates (Direct, High; PMID: 39987172, PMID: 32297626).
  • High-Resolution Imaging: X-ray crystallography and Cryo-EM remain the ultimate arbiter, providing the ground truth for atomic arrangements that AF3 and other tools attempt to hypothesize (Direct, High; PMID: 38036854, PMID: 41831109).

How do AlphaFold 3 ipTM and pTM scores correlate with the structural accuracy of PROTAC ternary complexes?

What role does Molecular Dynamics play in identifying non-physical linker geometries in AlphaFold 3 predictions?

Which experimental readouts are most effective for validating PROTAC-induced ternary complex stability?

Unverified Citations

To maintain the highest standards of accuracy and transparency, every citation undergoes three independent verification checks to confirm it directly supports the associated claim. The references below did not satisfy all verification stages. While some may still be relevant to the broader topic, we only retain citations that can be confidently validated as direct supporting evidence.

• PMID:40208910 — CETSA (Cellular Thermal Shift Assay) confirms physical target engagement within complex cellular lysates
  Failed: entities,conclusion — The paper makes no mention of 'CETSA' or 'Cellular Thermal Shift Assay'.

AlphaFold 3 (AF3) confidence metrics, specifically ipTM (interface predicted Template Modeling score) and pTM (predicted TM-score), show a weak to moderate correlation with the actual structural accuracy of PROTAC ternary complexes, often failing to reliably predict success (Direct, High; PMID: 41831109).

Specific Correlation and Reliability

In the context of Targeted Protein Degradation (TPD), the relationship between these metrics and structural accuracy is characterized by several limitations:

• Weak Predictive Power: Benchmarking of PROTAC-mediated ternary complexes shows a weak to moderate negative correlation between complex RMSD (root-mean-square deviation) and model output metrics (Direct, High; PMID: 41831109).
• Insensitivity to Misorientation: Both pTM and ipTM are largely insensitive to "rigid-body rotational errors," where the E3 ligase and target protein are individually modeled with high accuracy, but their relative orientation is incorrect (Direct, High; PMID: 41831109).
• Low Score Predictive Value: While high scores do not guarantee accuracy, very low ipTM or pTM scores are more reliable indicators of model failure or very low structural accuracy (Direct, High; PMID: 41831109).
• Internal Consistency vs. Correctness: These metrics are often a better reflection of the model's internal consistency than its actual correctness, particularly for systems with weak protein-protein interfaces common in PROTAC complexes (Direct, High; PMID: 41831109).

Factors Affecting Score Accuracy

The accuracy of these scores is frequently compromised by the composition of the modeled system:

• Accessory Protein Inflation: Including scaffold proteins like Elongin B/C (for VHL) or DDB1 (for Cereblon) can artificially inflate DockQ and confidence scores. This "high-quality" assessment often arises from contacts with these accessory subunits rather than the specific E3-target interface (Direct, High; PMID: 41152377).
• Physical Implausibility: AF3 may produce high-confidence global assemblies that are physically impossible because the predicted orientation of the proteins precludes the placement of the small-molecule linker (Direct, High; PMID: 41152377).
• Ranking Failures: In systematic assessments (e.g., CASP16), AF3's internal ranking (based on these confidence scores) selected the best-quality model as the top choice in only approximately 25% of cases for protein complexes (Direct, High; PMID: 41170922).

Comparison to General Performance

While AF3 confidences generally track accuracy well for standard biomolecular interactions, the unique flexibility and lack of natural evolutionary pressure in PROTAC-induced interfaces make these scores significantly less reliable for TPD than for traditional protein structures (Derived, Medium; PMID: 41831109).

How does the inclusion of scaffold proteins in AlphaFold 3 models affect the accuracy of predicted E3-target interfaces?

What specific structural failures in PROTAC modeling are AlphaFold 3 confidence metrics unable to detect?

How do AlphaFold 3 confidence metrics compare to Rosetta energy scores for ranking PROTAC ternary complexes?

Unverified Citations

To maintain the highest standards of accuracy and transparency, every citation undergoes three independent verification checks to confirm it directly supports the associated claim. The references below did not satisfy all verification stages. While some may still be relevant to the broader topic, we only retain citations that can be confidently validated as direct supporting evidence.

• PMID:38718835 — While AF3 confidences generally track accuracy well for standard biomolecular interactions, the unique flexibility and l...
  Failed: conclusion — The paper actually states that AF3 confidence measures are well calibrated with accuracy and provides data showing they track accuracy across various complex types, contradicting the claim that they are significantly less reliable.

AlphaFold 3 (AF3) confidence metrics such as ipTM and pTM primarily reflect internal model consistency and are often unable to detect fundamental structural failures in PROTAC ternary complexes, including incorrect rigid-body orientations, non-physical linker geometries, and mispredicted ligand stereochemistry.

Rigid-Body Rotational and Positional Failures

AF3 metrics are frequently insensitive to the relative positioning of subunits even when those subunits are individually modeled with high precision.
* Misorientation of Correct Subunits: AF3 can model the E3 ligase and the protein of interest (POI) with high structural accuracy (RMSD < 1 Å individually), yet place them in an incorrect relative orientation. In these "rigid-body rotational error" cases, pTM and ipTM scores often remain high, failing to signal the global assembly failure (Direct, High; PMID: 41831109).
* Non-Physical Linker Accomodation: The model may generate protein arrangements that are physically impossible for the specific PROTAC linker to bridge. These orientations are often proposed despite high confidence scores and global structural coherence (Direct, High; PMID: 41152377).

Ligand and Stereochemical Failures

AF3's architecture often overlooks fine-grained chemical details essential for drug-like molecule function.
* Stereochemistry and Chirality: Model outputs do not always respect ligand chirality or stereochemistry, which is critical for the binding and activity of many PROTACs (Direct, High; PMID: 41831109, PMID: 38718835).
* Twisting and Linker Misalignment: Metrics fail to reliably flag errors stemming from "twisting" or misaligned orientations around the flexible PROTAC linker (Direct, High; PMID: 41831109).

Artifacts and Environmental Limitations

Confidence scores can be misleading due to the context of the model or the static nature of the prediction.
* Accessory Protein Inflation: In complexes including scaffold proteins (e.g., DDB1 or Elongin B/C), high interface confidence scores may be driven by trivial contacts with these accessory subunits rather than the functionally relevant E3-target interface. Stripping these accessories often reveals a collapse in the actual accuracy of the ternary interface that the scores previously masked (Direct, High; PMID: 41152377).
* Conformational Bias: AF3 exclusively predicts specific states (e.g., the "closed" conformation of Cereblon) regardless of whether the system is in an apo or holo state, failing to capture the full range of biologically relevant conformational coverage (Direct, High; PMID: 38718835).
* Ensemble Insensitivity: PROTAC complexes are often dynamic ensembles in solution. Static AF3 confidence scores from multiple random seeds do not provide a reliable approximation of this solution ensemble or the stability of transient interfaces (Derived, Medium; PMID: 38718835).

In summary, while AF3 metrics are useful for standard protein folding, they often reflect internal model consistency rather than biological correctness in the context of induced proximity, especially when protein-protein interfaces are weak (Direct, High; PMID: 41831109).

Unverified Citations

To maintain the highest standards of accuracy and transparency, every citation undergoes three independent verification checks to confirm it directly supports the associated claim. The references below did not satisfy all verification stages. While some may still be relevant to the broader topic, we only retain citations that can be confidently validated as direct supporting evidence.

• PMID:41416887 — Static AF3 confidence scores from multiple random seeds do not provide a reliable approximation of this solution ensembl...
  Failed: conclusion — The paper does not test or present data on the performance of AlphaFold 3 (AF3) confidence scores; it focuses on validating the SILCS-xTAC method.