An end-to-end structural bioinformatics workflow evaluating local model confidence (pLDDT and PAE) in the oncogenic KRAS G12D mutant (residues 1–169 assay construct). This project assesses whether deep-learning structural models can reliably guide the identification of residue-level design anchors for growing small-molecule inhibitor series targeting the nucleotide-binding interface.
Developed as part of the Data-Driven Life Sciences (DDLS) programme (SciLifeLab & KTH Royal Institute of Technology).
KRAS is one of the most frequently mutated oncogenes in human cancers, with the G12D substitution accounting for a substantial portion of pancreatic ductal adenocarcinomas, colorectal tumours, and non-small cell lung cancers. Developing targeted small-molecule inhibitors requires identifying stable pocket neighbourhoods that can accommodate chemical substituents without causing steric clashes or relying on highly flexible loop regions.
- Model Identity & Boundary Checks: How does the canonical AlphaFold Database reference (wild-type G12, 189 residues) compare with the experimental assay construct (G12D, 169 residues)?
- Local Confidence vs Global Fold: Can global model confidence metrics (e.g., pTM) be trusted when making residue-level medicinal chemistry design choices?
- Design Anchor Shortlisting: Which residues in the nucleotide-pocket neighbourhood exhibit high, consistent structural confidence (pLDDT) to serve as design anchors for series expansion?
Local structural confidence was assessed across three candidate functional regions using the construct-matched KRAS G12D (1–169) fold:
| Pocket Region | Functional Context | Minimum pLDDT | Mean pLDDT | Structural Interpretation |
|---|---|---|---|---|
| Residues 10–17 | Phosphate-binding P-loop | 80.13 | 86.94 | Consistently high local confidence; rigid scaffold. |
| Residues 30–38 | Effector-binding Switch I | 62.00 | 68.27 | Moderate/low confidence; elevated conformational flexibility. |
| Residues 57–67 | Switch II catalytic region | 45.12 | 60.30 | Low confidence; highly dynamic and disordered state. |
Global Fold Metrics: Overall mean pLDDT =
85.45, pTM =0.9034. While global pTM indicates an overall well-predicted tertiary topology, local pLDDT reveals marked regional divergence in conformational reliability.
Five residues in the P-loop region (10–17) were identified as high-confidence design anchors:
| Residue | Position | Construct-Matched pLDDT | Local Structural Role |
|---|---|---|---|
| Gly | 10 | 87.70 | P-loop backbone entry; rigid geometric anchor |
| Val | 14 | 89.92 | Hydrophobic side-chain bordering nucleotide cleft |
| Gly | 15 | 90.79 | Highest confidence score; tight turn coordination |
| Lys | 16 | 87.06 | Crucial basic side-chain involved in pocket electrostatics |
| Ser | 17 | 87.03 | Conserved hydroxyl coordination motif |
P-Loop Sequence: 10 - G A G G V G K S - 17
Selected Anchors: * * * * *
pLDDT Scores: 87.7 89.9 90.8 87.1 87.0
Note: These residues serve as structural design anchors for hypothesis generation and analogue prioritisation; computational model confidence does not substitute for empirical mutational or crystallographic binding validation.
An interactive 3D structural viewer built with Reveal.js and 3Dmol.js allows visual inspection of:
- Residue-level pLDDT coloring mapped onto the construct structure.
- Side-by-side comparison between AlphaFold canonical reference and G12D assay construct.
- Full sequence pLDDT bar charts highlighting design anchor candidates.
👉 Launch Interactive 3D Viewer (GitHub Pages)
- Python 3.9+
piporuv
git clone https://github.com/RidaHayat/ddls-week4.git
cd ddls-week4
pip install -r requirements.txtTo parse the structure files, extract per-residue pLDDT metrics, and generate results/results.json:
python analysis.pyTo launch the interactive FastAPI/Tailwind interface:
uvicorn app:app --reload --port 8000Open your browser at: http://127.0.0.1:8000/
ddls-week4/
├── analysis.py # Computational pipeline parsing mmCIF, PDB, and PAE JSON
├── app.py # FastAPI application serving interactive dashboard
├── requirements.txt # Core dependencies (FastAPI, Uvicorn)
├── .gitattributes # Language statistics configuration
├── data/
│ ├── my_construct.fasta # Assay construct sequence (KRAS G12D 1–169)
│ ├── KRAS_G12D_construct_1-169.pdb # Sequence-matched predicted fold
│ ├── KRAS_alphafold_model.cif # Canonical AlphaFold DB reference (1–189 WT)
│ └── KRAS_alphafold_pae.json # Predicted Aligned Error (PAE) matrix
├── docs/
│ └── index.html # Standalone interactive 3D viewer for GitHub Pages
├── results/
│ └── results.json # Computed regional and residue-level pLDDT metrics
└── report.md # Detailed technical and methodological report
- Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2
- Canon, J., Rex, K., Saiki, A.Y. et al. The clinical KRAS(G12C) inhibitor AMG 510 drives anti-tumour immunity. Nature 575, 217–223 (2019).
- Data-Driven Life Sciences (DDLS) Programme, SciLifeLab & KTH Royal Institute of Technology.