Our comparative analyses highlights PRosettaC's improved performance over AF3 in predicting ternary complex geometries that more closely resemble experimentally resolved structures. While the observed differences are modest in magnitude, they are consistent across multiple systems and suggest that PRosettaC may currently offer more reliable structural predictions for PROTAC-focused modeling. These findings are intended to inform the selection of in silico tools during the early stages of PROTAC development, particularly for applications where accurate ternary complex modeling can impact downstream design decisions. By contributing a systematic benchmarking of two widely used structural prediction platforms, we aim to support the broader community in refining computational strategies for targeted protein degradation.
We queried the RCSB Protein Data Bank (PDB) on September 03, 2023 using the Advanced Search/Search API to identify experimentally determined PROTAC-mediated ternary complexes. The Boolean criteria were:
Full Text CONTAINS PHRASE "ternary complex PROTAC" AND Chemical Component Molecular Weight > = 450 AND Number of Protein Instances (Chains) per Assembly > 1 AND Structure Determination Methodology = experimental AND Experimental Method = X-RAY DIFFRACTION AND Polymer Entity Type = Protein.
The search returned 2,500 records (PDB entries). A custom results table was generated via the PDB Custom Report service and downloaded in tabular format for downstream screening. The table included the following fields: Entry ID, PubMed ID, DOI, Deposition Date, Release Date, Ligand, Value, PDB ID, Resolution (Å), Journal Name (Abbrev), Title, Publication Year, Structure Title, Sequence, Polymer Entity Sequence Length, Entity Macromolecule Type, Source Organism, Gene Name, Molecular Weight (Entity), Macromolecule Name, Asym ID, Auth Asym ID, Entity ID, Accession Code(s), Database Name, Ligand MW, Ligand ID, Ligand Name, Ligand SMILES.
All retrieved records were examined. First, we programmatically screened the results to confirm (i) the presence of at least one non-polymer ligand meeting the ≥ 450 Da threshold (proxy for a heterobifunctional PROTAC-sized molecule) and (ii) assemblies containing > 1 protein chain. Next, we annotated protein chains to identify putative E3 ubiquitin ligase components (e.g., CRBN, VHL, MDM2; gene-based and macromolecule-name pattern matching) and candidate target proteins distinct from the E3 component. Entries lacking both components in the same assembly or lacking a qualifying ligand were flagged for exclusion.
All remaining candidates were manually curated. For each, we visually inspected the biological assembly to confirm that the putative PROTAC ligand simultaneously engages (or is positioned to engage) both the E3 ligase and target protein, that ligand chemistry was intact (no missing atoms required for bifunctionality), and that the complex corresponded to an experimentally determined X-ray model. After manual curation, 36 non-redundant, high-confidence PROTAC ternary complex crystal structures were retained along with the corresponding reference publications. These curated structures served as the empirical benchmarks for our computational predictions and were pivotal for subsequent comparative analyses.
Complex models were generated using the latest AlphaFold-3 (AF3) server (https://alphafoldserver.com/), which supports multimeric protein assembly predictions with high structural fidelity. For each of the 36 benchmarked systems, we generated two distinct model variations: (1) a Minimal Complex consisting solely of the target protein and E3 ligase, and (2) a Full Complex, which additionally included accessory proteins known to stabilize the E3 ligase complex, such as Elongin B/C in VHL systems or DDB1 in CRBN systems.
Inputs were prepared by concatenating relevant amino acid sequences without template guidance or manual restraints. Five models were generated per complex using default AF3 multimer settings, with efficient prediction runtimes typically completing within 10-30 min per submission.
Due to input size constraints imposed by the AF3 server, larger scaffold proteins such as cullin ring ligases (CUL2, CUL4A) and RING-box domains (RBX1) were excluded, focusing predictions explicitly on minimal functional components relevant to degrader binding and interface evaluation.
PRosettaC, a Rosetta-based protocol designed specifically for modeling PROTAC-induced ternary complexes, served as the second computational strategy in our benchmarking. We refer to PRosettaC predictions explicitly as Ternary Complexes. This method enforces geometric constraints derived from known warhead binding modes, facilitating the structure-guided assembly of ternary complexes involving an E3 ligase, a target protein, and a bifunctional degrader. We used a local implementation of PRosettaC for this work after necessary adjustments based on their GitHub repository (https://github.com/LondonLab/PRosettaC).
Pipeline inputs included experimentally resolved or modeled structures of target proteins and E3 ligases with their respective bound warhead and ligase recruiter. The PROTAC linker was input as a SMILES string, enabling PRosettaC to generate three-dimensional linker conformations compatible with the binding pocket geometries.
To enhance sampling depth beyond the original PRosettaC implementation, the protocol was modified to generate up to 1000 models per system, surpassing the default 200-model limit. The actual number of models generated varied by system based on convergence and constraint compatibility, ranging from 54 to 878 models. Specifically, the following numbers of models were generated per system: 5T35 (n = 199), 6BN7 (n = 216), 6BOY (n = 400), 6HAX (n = 200), 6HAY (n = 200), 6HR2 (n = 878), 6W7O (n = 200), 6ZHC (n = 204), 7JTO (n = 200), 7KHH (n = 400), 7PI4 (n = 400), 7Q2J (n = 400), 7S4E (n = 400), 7Z6L (n = 54), 7Z76 (n = 400), 7Z77 (n = 400), 7ZNT (n = 400), 8BDT (n = 400), 8BDX (n = 244), 8G1P (n = 400), 8G1Q (n = 800), 8PC2 (n = 214), 8QVU (n = 400), 8QW6 (n = 399), and 8QW7 (n = 400).
All generated models were scored using the standard Rosetta energy function and evaluated using DockQ without prefiltering. This comprehensive evaluation ensures that our benchmarking reflects the full conformational diversity produced by PRosettaC.
To evaluate the structural accuracy of predicted ternary complexes, we used DockQ v2, the Python-based reimplementation of DockQ, which supports multimeric systems and automatic chain mapping. Its improved efficiency and portability made it well suited for high-throughput evaluation of both AlphaFold-3 (AF3) and PRosettaC predictions.
DockQ scores were computed by comparing each predicted model to the corresponding experimentally resolved crystal structure for all 36 benchmarked systems. For each AF3-generated model, we computed DockQ scores using three evaluation strategies: Full Complex: Predicted models including accessory proteins were evaluated against the full experimental structure; Core Complex: The same Full Complex models were stripped of accessory proteins before scoring, and compared to similarly stripped experimental structures; Minimal Complex: Models predicted using only the E3 ligase and target protein were compared to experimental structures containing only those components.
For PRosettaC-based Ternary Complex predictions DockQ evaluation was performed on the two protein chains -- E3 ligase and target. These predictions were scored against experimental structures similarly stripped to contain only the E3 and target chains.
Each DockQ score reflects a composite of interface RMSD (iRMSD), ligand RMSD (LRMSD), and the fraction of native contacts (Fnat), producing a normalized value between 0 and 1, where higher values indicate closer structural agreement with the reference interface.
In addition to static comparisons, DockQ v2 was also used for time-resolved analysis by comparing predicted models to molecular dynamics (MD) trajectories of the corresponding crystal structures. This enabled us to assess transient structural alignment across thousands of simulation frames. For PRosettaC-generated models, which can include up to 1000 structures per system, this provided fine-grained insight into conformational diversity and model accuracy in dynamic solution contexts.
To assess the dynamic behavior and conformational stability of modeled ternary complexes, we performed all-atom MD simulations using GROMACS 2023.1. The CHARMM36-jul2022 force field was employed for all protein and ligand components to ensure compatibility with high-fidelity protein-ligand interactions. Ligand parameters were generated using the CGenFF server, with .str files converted to GROMACS-compatible .itp formats.
Each complex -- whether derived from experimental crystal structures or computational models -- was first converted into GROMACS format and assigned appropriate protonation states. Systems were solvated in a TIP3P water box with a 1.0 nm buffer and neutralized with Na⁺ and Cl⁻ ions to achieve an ionic strength of 0.15 M. Topologies were generated using standard CHARMM36 force field protocols.
After solvation, systems underwent steepest descent energy minimization until a maximum force threshold of 1000 kJ/mol/nm was reached. Equilibration was conducted in two phases: a 100 ps NVT (constant Number, Volume, Temperature) ensemble using the modified Berendsen thermostat at 300 K, followed by a 100 ps NPT (constant Number, Pressure, Temperature) ensemble using the Parrinello-Rahman barostat at 1 bar. Position restraints were applied to heavy atoms of the protein and ligand during equilibration to allow solvent relaxation.
Production runs were performed in the NPT ensemble for 50 ns per system using a 2 fs integration time step and LINCS constraints on all bonds. Long-range electrostatics were treated using the Particle Mesh Ewald (PME) method. Periodic boundary conditions were applied in all directions, and coordinates were saved every 10 ps, yielding 5000 frames per trajectory for downstream analysis.
All simulations were conducted on GPU-accelerated high-performance computing clusters to ensure efficient sampling. Trajectories were post-processed to correct for periodicity and protein displacement, allowing accurate computation of root-mean-square deviation (RMSD), root-mean-square fluctuation (RMSF), and frame-by-frame DockQ comparisons to predicted and experimental structures.
To comprehensively assess the structural accuracy of predicted ternary complexes, we conducted both descriptive and inferential statistical analyses across multiple modeling and evaluation strategies: AF3 Minimal, AF3 Full, AF3 Core, and PRosettaC Ternary. DockQ v2 scores were analyzed at the per-model and per-structure level to evaluate interface quality across 36 crystal benchmark systems.
Descriptive statistics -- including median, mean, standard deviation, and interquartile range -- were calculated for each method to summarize performance distributions. To assess significance in model performance, paired Wilcoxon signed-rank tests were applied to compare median DockQ values across methods (e.g., PRosettaC Ternary vs AF3 Minimal, AF3 Core vs AF3 Minimal, etc.), with Cohen's d used to estimate effect sizes.
Visualization of score distributions was achieved through box plots, swarm plots, scatter plots, and heatmaps of per-complex deltas, allowing intuitive comparison across prediction strategies. Additionally, ranking frequency plots and median-centered bar charts were used to compare relative method performance across systems.
For MD simulations, time-series DockQ trajectories were generated across 5000 MD frames for five benchmark systems. We analyzed per-frame fluctuations in DockQ scores for each AF3 Minimal model (models 0-4), annotating peak scores and comparing temporal trends to assess whether any spontaneous realignment toward the crystal interface occurred. Maximum observed DockQ scores were noted for each trajectory, and graphical overlays were used to highlight these transient structural alignments.
In a follow-up dynamic evaluation, we also compared static PRosettaC Ternary models against MD-resolved frames of the crystal structures to detect transient conformational matches, enabling a frame-by-frame similarity analysis that captured model -- receptor alignment beyond the static crystal pose.
All statistical analyses were performed using Python libraries including scipy.stats, pandas, and matplotlib, and visualizations were rendered using seaborn and plotly.