Computational prioritization pipeline for BBB-penetrant EGFRvIII-targeting peptides (glioblastoma). Produces ordinal InterfacePriorityScore (IPS) and robustness calibration — without claiming binding affinity, dissociation constants, free energies, or experimental BBB permeability.
- What this pipeline does: Heuristic ranking of candidate peptides by interface contact count, contact density, IPS composite score, and BBB heuristic criteria.
- What this pipeline does NOT do: Predict binding affinity, dissociation constants (Kd), binding free energies (ΔG), experimental BBB permeability, or biological efficacy.
- Scoring models are heuristic (0.20 per IPS component) and exploratory. No experimental calibration was performed.
Two distinct control groups — never merged statistically:
| Control type | Source | N | Contacts | Label in figures |
|---|---|---|---|---|
| Primary control | P5_Scrambled (ESMFold structure) | 1 | 30 | Scrambled (primary control) |
| Ensemble control | 5 calibration scrambles (extended backbone) | 5 | [0,0,0,0,0] | Scrambled ensemble (n=5) |
- Primary control = baseline screening comparator (ranking context).
- Ensemble control = robustness null model (variance/calibration context).
- The ensemble distribution is degenerate under current CA-contact threshold (all 0 contacts, zero variance). Cohen's d and raw Z-scores are undefined for this baseline.
- CA-contact model: Interface contacts computed from Cα distance thresholding. Does not capture side-chain orientation, electrostatics, solvation, or entropy.
- No ΔG or docking affinity interpretation: Contact counts are NOT proxies for binding affinity. The docking scoring function is heuristic (clash penalty + distance reward), not physics-based.
- No experimental validation: BBB heuristic criteria are four simple physicochemical filters (molecular weight, HBD, HBA, LogP). Not validated against experimental permeability data.
- Scrambled control n=1 in primary pipeline: Primary control is a single sequence. No standard-deviation-based statistics computable from primary alone.
Only /figures and /results_final constitute manuscript inputs. All other directories are archived analytical outputs retained for reproducibility.
.
├── pipeline.py # Primary end-to-end pipeline (ESMFold structures, docking, scoring)
├── calibration.py # Robustness analysis (extended backbone, 5 scrambles, 3 scoring models, sensitivity)
├── final_analysis.py # Statistical lock, methods definitions, claim boundaries, figure datasets
├── generate_figures.py # Final figure generation (5 manuscript figures from figure_datasets/)
├── figure_provenance.csv # Figure ID, source dataset, control type, pipeline stage, inclusion status
├── results_final/ # MANUSCRIPT DATASETS (locked statistical outputs, claim boundaries, verdict)
│ └── figure_datasets/ # CSV datasets for figure generation
├── figures/ # MANUSCRIPT FIGURES (5 files — only ones used in manuscript)
├── archive_full_analysis/ # Archived non-manuscript outputs (see below)
└── bbb_penetrant.py # Original Colab notebook (reference only)
Only final calibrated figures are used in manuscript:
All other figures (7 from pipeline.py, 4 from calibration.py) are archived in archive_full_analysis/ and not referenced in manuscript.
- All figures: matplotlib only, 300 DPI minimum
- No external web servers or manual steps required to reproduce
- All code runs end-to-end locally