Mps1/TTK In Silico Drug Discovery Pipeline

Ensemble docking, graph neural network QSAR, and first computational allosteric targeting of the Mps1 NTD TPR domain — MSc Bioinformatics internship, Oxford Brookes University, 2026.

Author: Enya Gavilan Esquitino
Supervisor: Prof. Victor M. Bolanos-Garcia
Institution: Oxford Brookes University and Université Clermont Auvergne

---

Biological Context

Mps1/TTK (Monopolar Spindle 1 / Threonine Tyrosine Kinase) is the master regulator of the Spindle Assembly Checkpoint (SAC), overexpressed in triple-negative breast cancer, glioblastoma, pancreatic ductal adenocarcinoma, and colorectal cancer. This pipeline targets two complementary sites:

• Kinase domain — ATP-binding hinge region (Gly605/Glu603 hydrogen bond pharmacophore)
• NTD TPR domain — Hec1/Ndc80 protein-protein interaction interface (novel allosteric strategy)

---

Pipeline Overview

Campaign 1 — Kinase Domain: Validation & Baseline (Phase 1)

1. 45 co-crystallised Mps1 inhibitors retrieved from RCSB/KLIFS
2. Ensemble receptor preparation (75 KLIFS structures → 7 diverse conformations)
3. AutoDock Vina docking (exhaustiveness=16)
4. PLIP interaction analysis (Gly605/Glu603 H-bond detection, geometric criteria)
5. ADME filtering (Lipinski Ro5, TPSA, PAINS, Brenk, SA score)
6. ML Model 1 — Ridge Regression (n=45, LOOCV, R²=0.706)
7. ML Model 2 — SVR (n=2,352 ChEMBL, 10-fold CV, R²=0.729)
8. ML Model 3 — D-MPNN GNN ensemble (n=3,607, R²=0.816, σ=0.186)

Campaign 2 — Kinase Domain: Novel Candidate Discovery (Phase 2B)

1. 5 diverse ChEMBL seeds (pIC50 ≥ 10.4, max pairwise Tanimoto ≤ 0.4)
2. PubChem similarity search (Tanimoto ≥ 0.7) → 395 candidates
3. Smina ensemble docking (3 receptors) + GNINA CNN rescoring (top 50)
4. PLIP hinge interaction analysis → 233/354 hinge binders
5. Full ADMET filtering (SwissADME) → 163/233 pass
6. D-MPNN GNN pIC50 prediction + uncertainty quantification
7. Multi-task GNN (Mps1 + Aurora B + hERG) → selectivity and safety
8. Aurora B counter-screen (docking into 4TND)
9. Multi-objective Pareto optimisation (5 objectives) → 25 front-1 candidates
10. Resistance mutation docking (C604Y mutant, 5NTT)

Campaign 3 — NTD Domain: First Allosteric Campaign

1. 4H7Y TPR domain (1.8 Å, apo) receptor preparation
2. fpocket binding site detection → Hec1 interface (Pocket 42)
3. AlphaFold2-Multimer complex prediction (TPR pLDDT=96.1)
4. 988 ZINC fragments screened (MW ≤ 320 Da, LogP ≤ 3.5)
5. GNINA docking (exhaustiveness=16) → 964 scored
6. PLIP interface analysis (residues 104,107,108,111,140,141,144,145,173,174,177,178)
7. Top 10 interface binders identified (Lys107/Arg108 contacts)

---

Key Results

Metric	Value
Self-docking RMSD (5LJJ)	0.666 Å
Cross-docking mean RMSD	1.903 ± 0.633 Å
Homologous docking ρ	0.833 (p=0.005)
PLIP Phase 1 hinge binders	39/45 (86.7%)
PLIP Phase 2B hinge binders	233/354 (65.8%)
Ridge Regression R²	0.706 (LOOCV)
SVR R²	0.729 (10-fold CV)
GNN R²	0.816 (held-out test)
Y-randomisation Z-score	15.88 (p < 0.01)
Phase 2B ADMET pass rate	163/233 (70.0%)
Pareto front 1 candidates	25
Top candidate pIC50	9.228 (IC50 = 0.59 nM)
Top candidate SI vs Aurora B	1353
NTD top fragment score	-8.943 kcal/mol (zinc_div_0277)
NTD interface binders	10/20 (Lys107 or Arg108)
C604Y resistance Δ	+2.56 to +2.70 kcal/mol (all susceptible)

---

Repository Structure

mps1-drug-discovery/
│
├── kinase_domain/
│   ├── data/
│   │   ├── raw/                        # PDB structures (5LJJ, 5NTT, etc.)
│   │   ├── receptor/
│   │   │   └── ensemble/               # 7 KLIFS-validated receptor PDBQTs
│   │   ├── ligands/
│   │   │   ├── raw/                    # 45 Phase 1 SDF files
│   │   │   └── pdbqt/                  # 45 Phase 1 PDBQT files
│   │   ├── phase2/                     # Phase 2A candidates (PDB-seeded)
│   │   ├── phase2b/                    # Phase 2B candidates (ChEMBL-seeded)
│   │   │   ├── raw/                    # 395 SDF files
│   │   │   └── pdbqt/                  # 385 PDBQT files
│   │   ├── libraries/                  # Chemical libraries
│   │   └── bindingdb_ttk.tsv           # BindingDB TTK activity data
│   │
│   ├── docking/
│   │   ├── results/                    # Phase 1 docked poses + scores
│   │   ├── phase2_results/             # Phase 2A docked poses + scores
│   │   ├── phase2b_results/            # Phase 2B docked poses + scores
│   │   ├── selectivity_results/        # Aurora B counter-screen (4TND)
│   │   └── resistance_results/         # C604Y mutant docking (5NTT)
│   │
│   ├── analysis/
│   │   ├── validation/                 # RMSD, cross-docking, homologous
│   │   ├── interactions/               # Legacy distance-based (deprecated)
│   │   ├── interactions_plip/
│   │   │   ├── phase1/                 # PLIP results (39/45 hinge binders)
│   │   │   └── phase2/                 # PLIP results (Phase 2A)
│   │   ├── adme/                       # Phase 1 ADME results
│   │   ├── adme_plip/                  # ADME + PLIP combined (Phase 1+2A)
│   │   ├── admet/                      # Full ADMET results
│   │   ├── ic50/                       # ChEMBL + BindingDB training data
│   │   ├── ml_model/
│   │   │   ├── best_model.pkl          # Ridge Regression (Model 1)
│   │   │   ├── ridge_coefficients.csv
│   │   │   └── chembl/                 # SVR + validation (Model 2)
│   │   ├── gnn_model/
│   │   │   ├── optimised/              # D-MPNN GNN ensemble (5 models)
│   │   │   └── multitask/              # Multi-task GNN (Mps1+AuroraB+hERG)
│   │   ├── phase2/                     # Phase 2A results
│   │   ├── phase2b/
│   │   │   ├── plip/                   # PLIP hinge binders (233/354)
│   │   │   ├── admet_candidates_clean.csv  # 163 ADMET-passing candidates
│   │   │   ├── phase2b_multitask_predictions.csv
│   │   │   └── swissadme_top5.csv
│   │   ├── pareto/                     # Pareto optimisation results
│   │   ├── selectivity/                # Selectivity analysis
│   │   ├── resistance/                 # Resistance mutation analysis
│   │   ├── consensus_docking/          # Consensus docking results
│   │   └── md_simulations/             # MD analysis results
│   │
│   ├── notebooks/
│   │   ├── Mps1_ChemProp_Training.ipynb   # D-MPNN hyperparameter optimisation
│   │   ├── GNINA_top50.ipynb              # Phase 2B GNINA CNN rescoring
│   │   └── Multi_task_GNN.ipynb           # Multi-task GNN training
│   │
│   └── scripts/
│       ├── download_ligands.py
│       ├── cleanup_ligands.py
│       ├── prepare_ligands.py
│       ├── ensemble_receptor_prep.py
│       ├── run_docking.py
│       ├── run_ensemble_docking.py
│       ├── run_smina_ensemble.py
│       ├── plip_analysis.py
│       ├── filter_hinge_binders.py
│       ├── admet_filter.py
│       ├── ml_model.py
│       ├── ml_chembl.py
│       ├── expand_training_data.py
│       ├── gnn_predict.py
│       ├── multitask_predict.py
│       ├── get_multitask_data.py
│       ├── prepare_multitask_data.py
│       ├── y_randomisation.py
│       ├── leverage_ad.py
│       ├── applicability_domain.py
│       ├── synthetic_accessibility.py
│       ├── phase2_similarity_search.py
│       ├── phase2b_seed_selection.py
│       ├── phase2_predict.py
│       ├── merge_docking_scores.py
│       ├── pareto_optimisation.py
│       ├── selectivity_screen.py
│       ├── rmsd_validation.py
│       ├── crossdock_5N7V.py
│       ├── crossdock_validation.py
│       ├── homologous_docking.py
│       ├── prepare_crossdock_receptor.py
│       ├── get_ic50.py
│       └── run_phase2b_pipeline.sh
│
├── ntd_domain/
│   ├── data/
│   │   ├── raw/                        # 4H7Y PDB + fixed structure
│   │   ├── receptor/                   # 4H7Y PDBQT receptor
│   │   └── libraries/
│   │       ├── zinc_1000_diverse.smi   # 988 ZINC fragment library
│   │       └── pdbqt/                  # Prepared fragment PDBQTs
│   ├── docking/
│   │   ├── results/                    # All GNINA docking results
│   │   └── top20_results/              # Top 20 docked poses
│   ├── analysis/
│   │   ├── ntd_docking_scores.csv      # All 964 fragment scores
│   │   ├── ntd_top_candidates.csv      # Top 10 fragments
│   │   ├── alphafold_model.png         # AlphaFold2-Multimer figure
│   │   ├── admet/                      # NTD fragment ADMET results
│   │   ├── binding_site/               # fpocket binding site analysis
│   │   ├── fragments/                  # Fragment analysis
│   │   ├── leads/                      # Lead fragment results
│   │   └── plip_top20/
│   │       └── ntd_plip_interactions.csv
│   └── scripts/
│       ├── ntd_receptor_prep.py
│       ├── prepare_zinc_library.py
│       ├── ntd_docking.py
│       └── ntd_plip_analysis.py
│
├── md_simulations/
│   ├── kinase_domain/
│   │   ├── systems/                    # 5 holo systems (EM-minimised)
│   │   ├── topol/                      # GROMACS topologies (GAFF2/AMBER99sb-ildn)
│   │   ├── mdp/                        # MDP files (EM, NVT, NPT, MD)
│   │   ├── analysis/                   # MD trajectory analysis
│   │   └── scripts/                    # MD preparation scripts
│   └── ntd_domain/
│       ├── systems/                    # NTD apo system
│       ├── topol/                      # NTD topology
│       ├── mdp/                        # NTD MDP files
│       ├── results/                    # NTD MD results (future work)
│       └── analysis/                   # NTD trajectory analysis
│
├── shared/
│   ├── figures/                        # Shared publication figures
│   └── models/                         # Shared ML model files
│
├── report_figures/
│   ├── poses/                          # PDB files for PyMOL visualisation
│   ├── kinase_top3_structures.png      # Chemical structures (top 3 kinase)
│   ├── ntd_top3_structures.png         # Chemical structures (top 3 NTD)
│   ├── kinase_docking_pose.png         # PyMOL — kinase binding pose
│   ├── ntd_docking_pose.png            # PyMOL — NTD binding pose
│   ├── resistance_WT_pose.png          # PyMOL — WT binding pose
│   ├── resistance_C604Y_pose.png       # PyMOL — C604Y mutant pose
│   ├── resistance_superimposed.png     # PyMOL — WT vs C604Y overlay
│   └── pkcsm_input.smi                 # SMILES for ADMET submission
│
├── README.md
└── environment.yml                     # Conda reproducible environment

---

Reproducing the Pipeline

Setup

git clone https://github.com/engavilane/mps1-drug-discovery
cd mps1-drug-discovery
conda env create -f environment.yml
conda activate docking

Campaign 1 — Kinase Domain Phase 1

python kinase_domain/scripts/download_ligands.py
python kinase_domain/scripts/cleanup_ligands.py
python kinase_domain/scripts/prepare_ligands.py
python kinase_domain/scripts/ensemble_receptor_prep.py
python kinase_domain/scripts/run_docking.py
python kinase_domain/scripts/rmsd_validation.py
python kinase_domain/scripts/crossdock_validation.py
python kinase_domain/scripts/homologous_docking.py
python kinase_domain/scripts/plip_analysis.py
python kinase_domain/scripts/admet_filter.py
python kinase_domain/scripts/ml_model.py
python kinase_domain/scripts/ml_chembl.py
python kinase_domain/scripts/y_randomisation.py
python kinase_domain/scripts/leverage_ad.py
# GNN training: run Mps1_ChemProp_Training.ipynb on Google Colab

Campaign 2 — Kinase Domain Phase 2B

python kinase_domain/scripts/phase2b_seed_selection.py
python kinase_domain/scripts/phase2_similarity_search.py
python kinase_domain/scripts/prepare_ligands.py \
    --raw kinase_domain/data/phase2b/raw \
    --pdbqt kinase_domain/data/phase2b/pdbqt
# GNINA docking: run GNINA_top50.ipynb on Colab GPU
bash kinase_domain/scripts/run_phase2b_pipeline.sh
# Multi-task GNN: run Multi_task_GNN.ipynb on Colab
python kinase_domain/scripts/selectivity_screen.py

Campaign 3 — NTD Domain

python ntd_domain/scripts/ntd_receptor_prep.py
python ntd_domain/scripts/prepare_zinc_library.py
# NTD docking: run on Colab GPU (see ntd_domain/scripts/ntd_docking.py)
python ntd_domain/scripts/ntd_plip_analysis.py

---

Top Candidates

Kinase Domain (Phase 2B)

Candidate	pIC50	IC50 (nM)	SI vs Aurora B	hERG risk
phase2b_seed_4_44199470 ★	9.228	0.59	1353	moderate
phase2b_seed_4_66555847	8.870	1.35	270	high
phase2b_seed_4_90072314	8.808	1.56	1117	moderate
phase2b_seed_4_86713879	8.587	2.59	290	moderate
phase2b_seed_4_44217530	8.431	3.71	150	moderate

★ Best overall — sub-nM potency, SI=1353, moderate safety profile.

NTD Domain

Fragment	Score (kcal/mol)	Contact	MW (Da)	LogP
zinc_div_0277 ★	-8.943	Arg108	318.3	-1.24
zinc_div_0276	-8.862	Lys107	318.3	-1.39
zinc_div_0300	-8.422	Lys107	318.4	-1.06

★ Unique Arg108 contact — distinct binding mode within TPR groove.

---

Software

Tool	Version	Purpose
AutoDock Vina	1.2.5	Phase 1 docking
Smina	latest	Phase 2B ensemble docking
GNINA	1.0.3	CNN rescoring + NTD docking
Meeko	latest	Receptor/ligand preparation
PLIP	2.x	Protein-ligand interaction analysis
Chemprop	2.x	D-MPNN GNN (Models 3 + multi-task)
RDKit	2024.x	Cheminformatics + ADMET filtering
scikit-learn	1.x	Ridge + SVR models
GROMACS	2026.0	MD simulations (future work)
PyTorch	2.x	GNN training
AlphaFold2-Multimer	v3	NTD-Hec1 complex prediction
SwissADME	web	Extended ADMET panel
PyMOL	2.x	Docking pose visualisation

---

References

Biology & Target

1. Musacchio A, Salmon ED. The spindle-assembly checkpoint in space and time. Nat Rev Mol Cell Biol. 2007. DOI: 10.1038/nrm2163
2. Pugh KM et al. In silico identification of potential Mps1 inhibitors from natural isoflavones. Int J Mol Sci. 2022. DOI: 10.3390/ijms232214228
3. Nijenhuis W et al. A TPR domain–containing N-terminal module of MPS1 is required for its kinetochore localization by Aurora B. J Cell Biol. 2013. DOI: 10.1083/jcb.201210117
4. Bolanos-Garcia VM. Trends Biochem Sci. 2025. (in press — ask supervisor for DOI)

Docking & Scoring

5. Trott O, Olson AJ. AutoDock Vina: improving the speed and accuracy of docking. J Comput Chem. 2010. DOI: 10.1002/jcc.21334
6. McNutt AT et al. GNINA 1.0: molecular docking with deep learning. J Cheminform. 2021. DOI: 10.1186/s13321-021-00522-2
7. Quiroga R, Villarreal MA. Vinardo: A Scoring Function Based on Autodock Vina Improves Scoring, Docking, and Virtual Screening. PLoS ONE. 2016. DOI: 10.1371/journal.pone.0155183

Structure Databases

8. Berman HM et al. The Protein Data Bank. Nucleic Acids Res. 2000. DOI: 10.1093/nar/28.1.235
9. Kanev GK et al. KLIFS: an overhaul of the kinase-ligand interaction fingerprints and structures database. Nucleic Acids Res. 2021. DOI: 10.1093/nar/gkaa895

Interaction Analysis

10. Salentin S et al. PLIP: fully automated protein-ligand interaction profiler. Nucleic Acids Res. 2015. DOI: 10.1093/nar/gkv315

Machine Learning & QSAR

11. Heid E et al. Chemprop: A machine learning package for chemical property prediction. J Chem Inf Model. 2024. DOI: 10.1021/acs.jcim.3c01250
12. Yang K et al. Analyzing learned molecular representations for property prediction. J Chem Inf Model. 2019. DOI: 10.1021/acs.jcim.9b00237
13. Liaw A, Wiener M. Classification and Regression by randomForest. R News. 2002. (scikit-learn implementation)
14. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995. DOI: 10.1007/BF00994018

Cheminformatics & ADMET

15. RDKit: Open-source cheminformatics. rdkit.org 16.Ertl P, Schuffenhauer A. Estimation of synthetic accessibility score. J Cheminform. 2009. DOI: 10.1186/1758-2946-1-8
16. Daina A et al. SwissADME: a free web tool to evaluate pharmacokinetics and drug-likeness. Sci Rep. 2017. DOI: 10.1038/srep42717
17. Baell JB, Holloway GA. New substructure filters for removal of pan assay interference compounds (PAINS). J Med Chem. 2010. DOI: 10.1021/jm901137j

Activity Data

19. Gaulton A et al. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 2012. DOI: 10.1093/nar/gkr777 20.Gilson MK et al. BindingDB in 2015: A public database for medicinal chemistry. Nucleic Acids Res. 2016. DOI: 10.1093/nar/gkv1072
20. Irwin JJ et al. ZINC20 — A free ultralarge-scale chemical database for ligand discovery. J Chem Inf Model. 2020. DOI: 10.1021/acs.jcim.0c00675

Binding Site & Structure Prediction

22. Le Guilloux V et al. Fpocket: An open source platform for ligand pocket detection. BMC Bioinformatics. 2009. DOI: 10.1186/1471-2105-10-168
23. Mirdita M et al. ColabFold: making protein folding accessible to all. Nat Methods. 2022. DOI: 10.1038/s41592-022-01488-1
24. Evans R et al. Protein complex prediction with AlphaFold-Multimer. bioRxiv. 2022. DOI: 10.1101/2021.10.04.463034

Ligand Preparation

25. Eberhardt J et al. AutoDock Vina 1.2.0: New docking methods, expanded force field, and Python bindings. J Chem Inf Model. 2021. DOI: 10.1021/acs.jcim.1c00203

MD Simulations

26. Abraham MJ et al. GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX. 2015. DOI: 10.1016/j.softx.2015.06.001
27. Wang J et al. Development and testing of a general amber force field. J Comput Chem. 2004. DOI: 10.1002/jcc.20035
28. Lindorff-Larsen K et al. Improved side-chain torsion potentials for the Amber ff99SB protein force field. Proteins. 2010. DOI: 10.1002/prot.22711

Visualisation

29. The PyMOL Molecular Graphics System, Version 2.0, Schrödinger LLC.

Multi-objective Optimisation

30. Deb K et al. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Trans Evol Comput. 2002. DOI: 10.1109/4235.996017

---

Licence

This project is licensed under the MIT Licence — see LICENSE for details.

AI Disclosure

This project used Claude (Anthropic) as an assistance tool for code development and results analysis. The author has reviewed, edited, and assumes full responsibility for all content.