pRosettaC — PROTAC Ternary Complex Modelling Pipeline

A computational pipeline for modelling PROTAC-induced ternary complexes between a kinase (POI), the VHL E3 ligase, and the bifunctional degrader molecule SJF8240. The goal is to predict which kinases are geometrically compatible with productive ubiquitination and therefore likely to be degraded.

---

Background

PROTACs (PROteolysis TArgeting Chimeras) are bifunctional small molecules that recruit an E3 ubiquitin ligase to a protein of interest (POI), inducing its ubiquitination and subsequent proteasomal degradation. Whether degradation actually occurs depends critically on the geometry of the resulting ternary complex — the PROTAC must bridge the two proteins in a way that places a lysine residue on the POI within reach of the E2 enzyme (UBE2D).

This pipeline uses pRosettaC (Zaidman et al. 2020) to sample ternary complex poses via global docking and constraint-guided local optimisation with Rosetta, then analyses the resulting poses for biological plausibility.

---

Repository Structure

pRosettaC/
├── Structure_Prep/          # Step 1 — prepare input structures
│   ├── 01_download_structures.py   # Download and chain-extract PDB files
│   ├── 02_prepare_full_kinase.py   # Fix residue-number overlaps for RIPK2/MAPK14
│   └── 03_place_warhead.py         # Place SJF8240 warhead into each kinase ATP site
│
├── Rosetta_utils/           # Step 2 — run pRosettaC
│   ├── 01_install_rosetta          # Rosetta + pRosettaC installation guide
│   ├── 02_sds_to_params.py         # Generate Rosetta .params from SDF (RDKit-based)
│   ├── 03_run_local_prosettac.py   # Run pRosettaC locally (no cluster required)
│   ├── 04_run_all_pois.py          # Batch-run all POIs sequentially
│   ├── 05_gen_local_fasc.py        # Score docking PDBs → local.fasc with PyRosetta
│   └── 06_resume_pipeline.py       # Resume from constraint_generation + clustering
│
└── Analysis /               # Step 3 — analyse poses
    ├── 01_find_feasible_poses.py   # Scan Patchdock_Results for bridging poses
    ├── 02_recluster_feasible.py    # Re-cluster bridging poses by Cα RMSD
    ├── 03_batch_check_protac.py    # Quick PASS/MARGINAL/FAIL screen across all POIs
    └── 04_run_poi_full_analysis.py # Full feature extraction + pre-MD report

---

Prerequisites

Dependency	Purpose
Rosetta	Global docking and local optimisation (free academic licence)
pRosettaC	PROTAC-specific ternary complex sampling
PyRosetta	Python interface to Rosetta (scoring, params)
RDKit	SDF→params conversion, warhead coordinate handling
BioPython	Sequence alignment, Cα superposition
NumPy / SciPy	Distance matrices, hierarchical clustering

All Python dependencies can be installed into a conda environment:

conda create -n protac python=3.11
conda activate protac
conda install -c conda-forge rdkit biopython scipy numpy
pip install pyrosetta-installer   # or install PyRosetta wheel directly

Set environment variables before running:

export ROSETTA_HOME=/path/to/rosetta_src_<version>
export ROSETTA_BIN=$ROSETTA_HOME/main/source/bin
export PROSETTAC_HOME=~/pRosettaC       # upstream pRosettaC repo

---

Pipeline Walkthrough

Step 1 — Structure Preparation

# 1a. Download PDB files and extract the relevant chain for each POI and VHL
python Structure_Prep/01_download_structures.py

# 1b. Fix residue-number overlaps for RIPK2 and MAPK14
#     (their N-lobe numbering clashes with VHL residues 1-103)
python Structure_Prep/02_prepare_full_kinase.py

# 1c. Place the SJF8240 kinase warhead into each POI ATP site via MCS + superposition
python Structure_Prep/03_place_warhead.py
# Add --force <TARGET> to re-generate specific targets

After this step each POI directory contains:

• <POI>_protein_clean.pdb — cleaned kinase structure
• <POI>_warhead_placed.sdf — warhead in the ATP-site frame (no H)
• <POI>_warhead_placed_H.sdf — same with explicit hydrogens

Step 2 — Rosetta / pRosettaC

# 2a. Generate Rosetta .params for the PROTAC (if not already done)
python Rosetta_utils/02_sds_to_params.py data/structures/PROTAC/SJF8240/SJF8240.sdf SJF data/ligand_params/SJF8240.params

# 2b. Run pRosettaC for a single POI (from that POI's run directory)
cd pRosettaC/runs/high_affinity_degraders/DDR2
python /path/to/Rosetta_utils/03_run_local_prosettac.py Protac_params.txt \
    --n_dock 200 --n_workers 8

# 2c. Or batch-run all configured POIs
python Rosetta_utils/04_run_all_pois.py

# 2d. If the pipeline was interrupted after docking but before clustering
python Rosetta_utils/06_resume_pipeline.py   # run from the POI run directory

Output per POI lands in pRosettaC/runs/<category>/<POI>/:

• Patchdock_Results/ — raw docking PDBs (pd.*_docking_????.pdb) and constraint-generated ternary complexes (combined_*_0001.pdb)
• Results/cluster*/ — pRosettaC clustered representatives

Step 3 — Analysis

# 3a. Quick bridging screen: PASS / MARGINAL / FAIL for each POI
python "Analysis /03_batch_check_protac.py" DDR2 MET EPHB2
python "Analysis /03_batch_check_protac.py" --all

# 3b. Find all bridging poses (writes <poi>_feasible.txt to $HOME)
python "Analysis /01_find_feasible_poses.py" DDR2 MET

# 3c. Re-cluster the bridging poses by Cα RMSD and pick top-3 candidates for MD
python "Analysis /02_recluster_feasible.py" DDR2

# 3d. Full feature extraction + pre-MD structural report
python "Analysis /04_run_poi_full_analysis.py" DDR2
python "Analysis /04_run_poi_full_analysis.py" --all

Outputs of 04_run_poi_full_analysis.py:

File	Contents
cluster_features.csv	Per-cluster ML feature table (contacts, clashes, lysine geometry)
cluster_features.json	Same data in nested JSON
premd_descriptors.csv	Features for the top-3 MD candidates
premd_report.txt	Human-readable pre-MD structural check (chain completeness, clash counts, PROTAC geometry, verdict)

---

Configured POIs

Target	Category	Notes
MET	high_affinity_degraders
DDR2	high_affinity_degraders
RIPK2	high_affinity_degraders	N-lobe absent; renumbered to res 200+ (_full run)
EPHB2	high_affinity_degraders
MAPK14	high_affinity_degraders	N-lobe absent; renumbered to res 200+ (_full run)
AXL	high_affinity_no_degradation
ABL1	high_affinity_no_degradation
EPHA2	high_affinity_no_degradation
MAP4K5	high_affinity_no_degradation
SLK	high_affinity_no_degradation
TNIK	low_affinity_degraders
PIP4K2C	low_affinity_degraders	Non-kinase phosphatase

---

Bridging Criteria

A ternary complex pose is scored:

Verdict	Criterion
PASS	Closest PROTAC atom < 5 Å from both VHL (chain A res 1–103) and POI
MARGINAL	Closest PROTAC atom < 10 Å from both
FAIL	Outside both thresholds

For N-lobe–truncated kinases (RIPK2, MAPK14, SLK, TNIK, MAP4K5) a stricter poi_check_min residue lower bound is used to avoid a systematic artefact where the VHL–kinase junction creates trivial false-positive bridging poses.

---

Key Design Notes

• Rosetta chain layout: VHL = chain A res 1–103, POI = chain A (original or renumbered), ElonginB = chain B, ElonginC = chain C, PROTAC = chain X.
• Score shifting: pRosettaC's clustering.py filters out scores ≥ 0. If PyRosetta's PackRotamersMover produces positive scores, 06_resume_pipeline.py shifts the entire score.sc by max_score + 100 before clustering.
• N-lobe renumbering: RIPK2 and MAPK14 have kinase domain numbering that overlaps with VHL (res 1–103). 02_prepare_full_kinase.py adds an offset (+192 / +196) so the full domain is retained without silently losing the N-lobe.

---

Citation

If you use pRosettaC in your work, please cite:

Zaidman, D. et al. PROTACable integrates modeling and deep learning to enable data-driven PROTAC design. Nature Chemical Biology (2023). Zaidman, D. et al. PRosettaC: Rosetta based modeling of PROTAC mediated ternary complexes. J. Chem. Inf. Model. 60, 4894–4903 (2020).