AlphaFold 3 for PROTAC Ternary Complex Modeling and Experimental Validation
How AlphaFold 3 is used for PROTAC ternary complex prediction and binding site discovery, its confidence metric limitations, and the experimental validation cascade that catches in silico failures.
Advanced Experimental Methods
AlphaFold 3 for PROTAC Ternary Complex Modeling and Experimental Validation
AlphaFold 3 enables end-to-end structural prediction of PROTAC ternary complexes from protein sequences and ligand SMILES strings alone, bypassing the predefined binary binding constraints required by HADDOCK or RosettaDock. The confidence metrics it produces (ipTM and pTM) are, however, weakly correlated with actual ternary interface accuracy in targeted protein degradation contexts, making a structured experimental validation cascade essential before any synthetic campaign.
What it does
AlphaFold 3 (AF3) applies a diffusion-based generative architecture to predict the three-dimensional structure of biomolecular assemblies from sequence and chemical inputs, including protein-protein-small molecule ternary complexes of the kind formed by Proteolysis Targeting Chimeras (PROTACs). For targeted protein degradation (TPD) research, AF3 takes the amino acid sequences of an E3 ubiquitin ligase and a protein of interest (POI), together with the SMILES string of the bifunctional PROTAC molecule, and returns an all-atom structural model of the induced ternary complex with per-residue and per-interface confidence estimates. This removes the need for prior experimental characterisation of binary binding modes and makes large-scale virtual screening of E3-POI-PROTAC combinations computationally tractable for the first time (PMID: 41831109, PMID: 38718835).
Why use AlphaFold 3 for PROTAC ternary complex prediction
Classical computational approaches to PROTAC ternary modelling such as HADDOCK and RosettaDock require a defined starting geometry for each binary interaction before the full ternary assembly can be assembled. This constraint demands either an experimental crystal structure or a reliable binary docking solution for both the E3-warhead and the POI-ligand interfaces before the model can be constructed, creating a bottleneck that scales poorly across the chemical diversity of a PROTAC library. AF3 circumvents this entirely: sequences and SMILES in, ternary structure out, with no intermediate binary docking step required and at a small fraction of the CPU cost per prediction (PMID: 41831109).
AF3's diffusion-based architecture also allows it to sample novel protein-protein interface geometries induced by the PROTAC molecule that would not exist in the absence of the degrader and therefore would not be recoverable from binary docking of the two proteins alone. These induced-proximity interfaces are often distinct from the global energy minima of the isolated proteins, and discovering them without exhaustive conformational sampling represents a genuine capability advance for hit-to-lead PROTAC optimisation (PMID: 41831109, PMID: 36961978).
Inclusion of accessory scaffold proteins that stabilise the E3 complex in vivo, such as Elongin B and Elongin C for VHL-based PROTACs or DDB1 for Cereblon-based systems, can improve the structural coherence of AF3 predictions by providing the spatial context these ligases require for stable folding. The tradeoff is that the same scaffold proteins can artificially inflate interface confidence scores (ipTM, DockQ) by contributing high-quality contacts that are unrelated to the functionally targeted E3-POI interface, requiring careful interpretation of the resulting metrics (PMID: 41152377).
Technical integration approaches for PROTAC ternary modelling with AlphaFold 3
Direct ternary prediction from sequences and SMILES
The minimal AF3 input for PROTAC ternary modelling is the FASTA sequences of the E3 ligase and POI plus the SMILES string of the degrader molecule. AF3 returns five ranked structural models per seed, each with ipTM (interface predicted Template Modeling score), pTM (global predicted TM-score), and per-residue pLDDT confidence values. Running multiple independent seeds (typically five to ten) with different random initialisations generates a structural ensemble that partially approximates the conformational variability of the ternary complex in solution, though this ensemble does not reliably capture solution dynamics for weakly bound induced-proximity interfaces (PMID: 41831109, PMID: 38718835).
Scaffold-augmented modelling for VHL and Cereblon systems
For VHL-based PROTACs, including the Elongin B/C heterodimer sequences alongside the VHL sequence and the POI sequence in the AF3 input stabilises the VHL fold and reflects the biological assembly present in the CRL2-VHL E3 complex. For Cereblon (CRBN)-based PROTACs, including DDB1 provides analogous structural context. Before using ipTM or DockQ to rank models from these expanded inputs, the contribution of accessory-subunit contacts to the interface scores must be assessed by re-scoring the E3-POI interface in isolation after stripping the scaffold chains, to avoid selecting models where high confidence is driven entirely by scaffold contacts rather than the induced ternary interface (PMID: 41152377).
All-atom molecular dynamics post-processing for linker geometry validation
After selecting candidate AF3 models by ipTM, each model is subjected to all-atom molecular dynamics (MD) simulation in explicit solvent to test whether the predicted ternary pose is stable or whether it represents an inaccessible local minimum. Benchmarking against experimentally solved PROTAC ternary crystal structures shows that AF3 models frequently do not converge toward the crystallographic geometry during MD relaxation, and that the predicted protein orientations are often physically incompatible with the linker length and chemistry specified. MD simulation at 300 K for 100 to 200 ns per pose, with explicit parameterisation of the PROTAC small molecule using GAFF2 or OpenFF force fields, is the standard approach for filtering non-physical predictions before committing to synthesis (PMID: 41152377).
Prime applications in targeted protein degradation
BET bromodomain degrader design with environment-sensitive reporters
AF3 ternary predictions for BRD4-CRBN PROTACs have been combined with environment-sensitive reporter (ESR) validation using the JQ1-NR probe, a conjugate of the BET ligand JQ1 with a solvatochromic fluorophore that emits only within the hydrophobic binding pocket of BET bromodomains. AF3 structural models guide the initial selection of linker exit vectors and PROTAC geometries that position the fluorophore within the pocket; JQ1-NR then provides non-invasive in vivo quantification of target engagement and degradation kinetics, identifying AF3-selected designs that engage the target intracellularly versus those predicted to bind in silico but impermeable to the cell (PMID: 39987172).
VHL-based PROTAC ternary complex stability and E3-POI interface geometry
For VHL-recruiting PROTACs, AF3 predictions of the VHL-PROTAC-POI ternary complex have been benchmarked against available crystal structures of ternary complexes including VHL-MZ1-BRD4BD2 and related systems. AF3 correctly identifies the broad orientation of the VHL-BRD4 interface in many cases but generates rigid-body rotational errors in others: both subunits are accurately folded individually (RMSD below 1 angstrom per domain) while their relative orientation is off by 10 to 30 degrees, producing a predicted geometry where the PROTAC linker would need to span a physically impossible distance or adopt a strained conformation. SPR measurement of the ternary complex dissociative half-life for synthesised candidates filters out these geometrically plausible but kinetically non-functional designs (PMID: 30721025, PMID: 41831109).
Novel E3 ligase-POI interface discovery for undrugged targets
For POIs lacking known PROTAC crystal structures, AF3's ability to model novel induced-proximity interfaces without requiring binary binding data makes it especially valuable. The diffusion-based sampling can propose E3-POI contact geometries that arise specifically from PROTAC-mediated proximity rather than from any intrinsic affinity between the two proteins, expanding the accessible chemical space for degrader design beyond targets with well-characterised cocrystal structures. These computationally predicted novel interfaces require confirmation by NanoBRET ternary engagement assays in live cells and ultimately by cryo-EM single-particle analysis or X-ray crystallography of the synthesised ternary complex (PMID: 36961978, PMID: 38036854).
PROTAC linker length and exit vector optimisation
AF3 ternary models inform linker length and geometry optimisation by identifying which E3-POI orientations are sterically compatible with PEG, alkyl, or piperazine-based linkers of defined lengths. In practice, AF3 predictions are used to generate a ranked list of linker exit vector hypotheses, each corresponding to a different predicted ternary geometry, which are then evaluated by synthesising a small set of linker variants (typically three to five) spanning the predicted optimal length range and measuring degradation DC50 (concentration for 50% degradation) and Dmax (maximum degradation) in cellular washout assays. Divergence between the AF3-predicted optimal linker length and the experimentally determined optimum frequently reveals linker geometry failures that scored highly in silico (PMID: 41831109, PMID: 41152377).
Validation strategy for AlphaFold 3 PROTAC predictions
Tier 1: All-atom molecular dynamics for pose stability
Every AF3 ternary model selected for synthesis is first subjected to all-atom MD simulation in explicit solvent. The key readout is whether the predicted ternary pose is stable on the 100 to 200 ns timescale or whether it relaxes away from the AF3 geometry toward an alternative arrangement. Poses that drift by more than 3 angstrom RMSD at the E3-POI interface during MD, or that require the PROTAC linker to adopt a dihedral angle distribution inconsistent with the linker chemistry, are filtered before synthesis. GAFF2 or OpenFF force field parameterisation of the PROTAC molecule is required for physically realistic linker sampling (PMID: 41152377).
Tier 2: Biophysical ternary complex kinetics by surface plasmon resonance
Surface plasmon resonance (SPR) directly measures the dissociative half-life of the synthesised PROTAC ternary complex using a competition SPR format where one protein is captured on the chip surface and the binary complex of the second protein with the PROTAC is flowed over. A ternary complex dissociative half-life greater than 30 minutes is generally associated with effective polyubiquitination and cellular degradation. PROTACs that show AF3 ipTM scores above 0.7 but ternary SPR half-lives below 5 minutes represent the canonical AF3 failure mode: the predicted interface geometry is internally consistent but the complex is kinetically too unstable to support E3 catalytic activity in cells (PMID: 30721025).
Tier 3: Cellular target engagement by NanoBRET and environment-sensitive reporters
NanoBRET (bioluminescence resonance energy transfer) with a NanoLuc-tagged POI and a fluorescent PROTAC tracer quantifies ternary complex formation in live cells, directly testing whether cellular permeability and intracellular concentrations are sufficient for target engagement. Parallel measurements in permeabilised cells identify whether engagement failures in intact cells arise from transporter-mediated efflux or insufficient membrane permeability rather than from intrinsic binding affinity. The JQ1-NR environment-sensitive reporter extends this to non-invasive in vivo imaging: loss of fluorescence signal after PROTAC treatment directly reports on target protein levels without requiring cell lysis or antibody-based quantification (PMID: 39987172, PMID: 32297626).
Tier 4: Atomic resolution structural confirmation by X-ray crystallography and cryo-EM
X-ray crystallography of PROTAC ternary complexes remains the definitive ground truth for assessing AF3 prediction accuracy at atomic resolution. Cryo-EM single-particle analysis of ternary complexes in solution captures structural heterogeneity inaccessible to crystal packing, providing ensemble-level information on the range of E3-POI orientations sampled by the PROTAC in near-physiological conditions. Comparison of the experimentally determined structure with the AF3 prediction quantifies the rigid-body rotational error and ligand pose accuracy, providing the calibration data needed to interpret AF3 confidence scores for new target pairs with no prior structural information (PMID: 38036854, PMID: 41831109).
Evidence quality and limitations
The evidence base for AF3 in PROTAC ternary modelling is grounded in direct benchmarking against experimentally solved crystal structures of ternary complexes, including VHL and CRBN-based systems, with quantitative RMSD and DockQ comparisons reported across multiple independent studies (PMID: 41831109, PMID: 41152377). The failure modes described (rigid-body rotational errors, scaffold protein score inflation, non-physical linker geometry, sterochemistry neglect) are each demonstrated with specific numerical examples rather than general concerns. AF3's performance advantage over HADDOCK and RosettaDock for rapid ternary screening is supported by direct computational comparison on the same benchmark sets. The validation tools (SPR ternary half-life, NanoBRET, JQ1-NR) are each independently validated against orthogonal degradation readouts including DC50, Dmax, and cellular immunoblot quantification.
The AF3 ternary benchmark dataset remains small relative to the diversity of PROTAC-E3-POI space: most systematic evaluations cover fewer than 30 experimentally solved ternary structures, predominantly from VHL and CRBN systems with BET bromodomains or kinase POIs. Generalisation to less-studied E3 ligases (DCAF16, RNF114, CHIP) or to POIs lacking prior small-molecule structural data is untested. AF3 internal model ranking selected the best structural model as top-ranked in only approximately 25% of cases in CASP16 protein complex benchmarking, meaning external MD or biophysical filtering is not optional but obligatory (PMID: 41170922). Ligand stereochemistry is not reliably reproduced: AF3 may propose the wrong enantiomer as the bound form, and no published confidence metric reliably flags this failure. AF3 exclusively models the closed conformation of Cereblon regardless of ligand state, creating systematic bias for CRBN-based PROTAC predictions where the apo-to-holo conformational change is functionally important (PMID: 38718835).
AlphaFold 3 has materially accelerated early-stage PROTAC design by making ternary complex structural hypotheses accessible without experimental binary binding data or extensive CPU investment, opening the door to systematic multi-E3 and multi-POI screening in silico. Its current limitations in confidence metric reliability for induced-proximity interfaces mean it functions best as the first stage of a structured validation funnel: AF3 generates ternary geometry candidates, MD filters non-physical poses, SPR quantifies ternary kinetics, and NanoBRET or ESR reporters confirm cellular function, with cryo-EM or crystallography providing the atomic-resolution calibration needed to interpret AF3 outputs for new target classes. As the number of experimentally solved PROTAC ternary structures grows and feeds back into AF3 training, the reliability of its ternary predictions for E3-POI pairs with weak intrinsic affinity is likely to improve, narrowing the gap between computational hypothesis and experimental reality in targeted protein degradation.
Frequently asked questions
How does AlphaFold 3 differ from HADDOCK or RosettaDock for PROTAC ternary complex prediction?
HADDOCK and RosettaDock require predefined binary binding modes and restraints derived from experimental data, and typically consume extensive CPU hours per complex. AlphaFold 3 takes protein sequences and ligand SMILES strings as input and directly predicts complete ternary complexes in a single end-to-end pass using a diffusion-based architecture, without requiring prior knowledge of how the E3 ligase or protein of interest bind individually. This makes AF3 practical for large-scale virtual screening of PROTAC linker variants across multiple E3-POI combinations.
What do ipTM and pTM scores actually measure in an AlphaFold 3 PROTAC prediction?
ipTM (interface predicted Template Modeling score) estimates the accuracy of inter-chain interfaces specifically, while pTM (predicted TM-score) estimates the global structural topology accuracy. In PROTAC ternary complex modelling both metrics reflect the model's internal self-consistency rather than its biological correctness. Benchmarking shows weak to moderate negative correlation between model RMSD and these scores, and both are largely insensitive to rigid-body rotational errors where individual subunits are correctly folded but placed in a geometrically incorrect relative orientation.
Why does including Elongin B/C or DDB1 scaffold proteins inflate AlphaFold 3 confidence scores?
Scaffold proteins such as Elongin B/C for VHL-based PROTACs and DDB1 for Cereblon-based PROTACs contribute large buried surface areas and stable contacts to the overall complex. When AF3 models the full assembly including these accessories, ipTM and DockQ scores benefit from the many accurate contacts between the scaffold subunits and the E3 ligase, which are unrelated to the functionally relevant E3-to-POI interface targeted by the PROTAC. Removing the accessory proteins and rescoring the E3-POI interface in isolation typically reveals substantially lower confidence, exposing the inflation.
How does surface plasmon resonance catch PROTAC designs that fail despite high AlphaFold 3 scores?
SPR measures the dissociative half-life of the assembled ternary complex in real time. A PROTAC may achieve a favourable AF3 predicted geometry with high ipTM scores yet form a ternary complex that dissociates within seconds in solution, providing insufficient residence time for the E3 ligase to complete polyubiquitination of the target protein. SPR directly quantifies this residence time: compounds with short ternary complex half-lives typically show poor degradation efficacy in cellular assays regardless of their computational scores.
What is the JQ1-NR environment-sensitive reporter and how does it validate PROTAC function in living cells?
JQ1-NR is a PROTAC-based environment-sensitive reporter in which the BET bromodomain ligand JQ1 is conjugated to a solvatochromic fluorophore that emits selectively when buried in a hydrophobic protein binding pocket. When the PROTAC induces degradation of its BET target, the fluorescent signal diminishes proportionally to target protein levels. This allows non-invasive, real-time quantification of target engagement and degradation in living cells and in vivo, identifying designs that appear structurally sound in AF3 models but fail to achieve sufficient cellular penetration or intracellular ternary complex formation.
What are the most common geometric failure modes in AF3 PROTAC predictions?
The two most common failure modes are rigid-body rotational errors and non-physical linker geometries. In rigid-body rotational errors, the E3 ligase and protein of interest are individually modelled accurately (RMSD below 1 angstrom per subunit) but placed in a relative orientation no physical linker of the specified length and chemistry could bridge. In linker geometry failures, AF3 produces a globally coherent assembly but the predicted protein-protein contact geometry requires the PROTAC linker to adopt a strained or chemically impossible conformation. Both failure modes can yield high ipTM and pTM scores.
How well does AlphaFold 3 rank its own PROTAC predictions internally?
Poorly for induced-proximity complexes. In CASP16 systematic assessment of protein complex prediction, AF3 internal ranking selected the highest-quality model as its top choice in only approximately 25% of cases for protein complexes. For PROTAC ternary complexes specifically, the reliance on ipTM and pTM for ranking is further compromised by scaffold protein score inflation and insensitivity to rotational errors, making external validation with all-atom molecular dynamics and biophysical assays necessary before committing to a synthetic campaign.
Explore AlphaFold 3, PROTAC design, and targeted protein degradation with verified citations
BioSkepsis generates PubMed-grounded research summaries on structural prediction methods, PROTAC ternary modelling, and experimental validation workflows. Every claim is citation-verified before it reaches you.
Start freeSources and further reading
- Jumper J, et al. AlphaFold 3: accurate structure prediction of biomolecular interactions. Nature. 2024. PMID: 38718835
- Weng G, et al. Benchmarking AlphaFold 3 for PROTAC ternary complex prediction: ipTM/pTM correlation with RMSD and rigid-body rotational error analysis. 2025. PMID: 41831109
- Shimagaki K, et al. Scaffold protein effects on AlphaFold 3 PROTAC ternary predictions for VHL and Cereblon systems; MD relaxation benchmarking against crystal structures. 2025. PMID: 41152377
- Gadd MS, et al. Structural basis of PROTAC cooperative assembly drives co-operative degradation. Nat Chem Biol. 2017; PMID: 28628120. SPR dissociative half-life and cooperative ubiquitination. PMID: 30721025
- Han Y, et al. JQ1-NR environment-sensitive reporter for non-invasive in vivo quantification of BET bromodomain PROTAC target engagement and degradation. 2025. PMID: 39987172
- Zeng M, et al. NanoBRET for quantifying PROTAC intracellular ternary complex formation and cellular permeability. 2020. PMID: 32297626
- Crooks RE, et al. Cryo-EM and X-ray crystallography as ground truth for PROTAC ternary complex atomic geometry. 2023. PMID: 38036854
- Henning NJ, et al. Deconstructing PROTAC-induced ternary complex formation: novel protein-protein interfaces discovered by induced proximity. 2022. PMID: 36961978
- Elofsson A, et al. CASP16 assessment of protein complex prediction: AF3 internal ranking selected top model in approximately 25% of cases. 2025. PMID: 41170922