Ki Predictions from Molecular Structure

A reproducible Applied Pharmaceutical Bioinformatics project for modelling molecular binding affinity from SMILES strings. The target is Ki (nM), transformed to log10(Ki) for regression and thresholded for activity classification.

The repository is structured as a code-first PhD-application portfolio project: clean source modules, reproducible environment files, a runnable smoke check, and a notebook workflow. It avoids inflated claims: random cross-validation is useful internal validation, not proof of prospective drug-discovery performance.

Portfolio summary

Item	Detail
Research task	QSAR-style prediction of Ki from molecular structure
Dataset	7,562 SMILES/Ki records supplied for an exam project
Curation	RDKit parsing, largest-fragment selection, canonical SMILES, median aggregation of repeats
Final modelling set	4,820 unique canonical molecules
Descriptors	ECFP4, MACCS keys, RDKit descriptors, custom SMARTS-derived structural features
Best regression model	ExtraTrees regressor, R² = 0.722 ± 0.055 on log10(Ki)
Main limitation	Random CV does not prove scaffold-level generalisation

Project structure

.
├── data/                         # example data; full exam dataset can be added locally
├── notebooks/                    # notebook workflow location
├── src/apb_ki_modelling/         # reusable cleaning and descriptor code
├── scripts/                      # smoke check and notebook execution helpers
├── environment.yml
├── requirements.txt
├── pyproject.toml
└── Makefile

Methods

1. Load raw SMILES and Ki values.
2. Remove missing and non-positive Ki values.
3. Parse molecules with RDKit.
4. For salts and multi-fragment records, keep the largest covalent fragment.
5. Convert molecules to canonical isomeric SMILES.
6. Aggregate repeated canonical molecules using median Ki.
7. Transform Ki to log10(Ki) for regression.
8. Generate ECFP4, MACCS, RDKit, and custom descriptors.
9. Compare regression and classification models using 10-fold cross-validation.
10. Run descriptor ablation to test where predictive signal comes from.

Key results

Cleaning

Step	Result
Raw records	7,562
Valid Ki records after cleaning	7,556
Invalid RDKit SMILES	0
Multi-fragment SMILES	552
Unique canonical molecules after aggregation	4,820

Regression on log10(Ki)

Model	R²	MSE
ExtraTrees regressor	0.722 ± 0.055	0.476 ± 0.102
HistGradientBoosting regressor	0.694 ± 0.065	0.524 ± 0.124
Random Forest regressor	0.669 ± 0.055	0.569 ± 0.106
Ridge regression	0.463 ± 0.106	0.909 ± 0.116

Classification

Threshold	Best model by AUC	AUC	F1
Ki < 1000 nM	HistGradientBoosting / ExtraTrees	0.942 ± 0.013 / 0.942 ± 0.018	0.966 ± 0.005 / 0.965 ± 0.006
Ki < median Ki	ExtraTrees	0.891 ± 0.017	0.802 ± 0.025

The Ki < 1000 nM task is chemically interpretable but imbalanced: about 90% of canonical molecules are active. The median split is less chemically meaningful but gives a more honest classifier stress test.

Descriptor ablation

Descriptor set	Features	R² mean	MSE mean
ECFP4 + MACCS + RDKit + custom	1236	0.721753	0.475985
ECFP4 + RDKit core	1037	0.707132	0.502200
ECFP4 only	1024	0.696819	0.520654
RDKit core only	13	0.504880	0.851078

ECFP4 carries most of the signal. The extended descriptor set helps, but the gain is incremental, not transformative.

Reproducibility

Use Conda or Mamba. RDKit is the dependency most likely to break in a casual pip-only setup.

mamba env create -f environment.yml
mamba activate apb-ki-modelling
make smoke

Run the notebook after adding it to notebooks/Ki_modelling.ipynb:

make notebook

Critical limitation

Random 10-fold cross-validation after duplicate aggregation is acceptable for an exam project, but it can still leak analogue-series information across folds. A research-grade next step would use Bemis-Murcko scaffold splitting, external validation, uncertainty estimates, and error analysis by chemical series.

Data note

The dataset was supplied for an exam project. The code is reusable; dataset reuse outside the course context may require checking the original data rights.