Earlier independent capstone · Flatiron School

The question

Can molecular fingerprints and protein sequence representations support prediction of drug–kinase bioactivity measurements?

The workflow

The project brought together drug–target records and compound/protein information from sources including Drug Target Commons, ChEMBL, and UniProt.

  • Filtered and prepared interaction records for modelling.
  • Retrieved compound SMILES and protein sequences.
  • Explored ECFP4 molecular fingerprints and PSSM / Sequence Graph Transform protein representations.
  • Compared Random Forest, XGBoost, and MLP regression models for predicting pChemBL values.
  • Explored protein clustering and its effect on predictive performance.

What this demonstrates

The project connects biological data retrieval, molecular feature preparation, and model comparison in an end-to-end exploratory workflow. It also provides a foundation for my more recent work on biomedical graphs and evaluation.

Scope and limitations

This is an educational capstone, not a validated drug-discovery platform. Protein features here are sequence-derived representations; the project does not predict three-dimensional ternary complexes or molecular-glue activity.

Interpreting the reported results requires checking the underlying assay endpoints and the train/test split. Predicting activity for unseen compounds or proteins requires evaluation designed specifically for those settings.

View the original code and project README · Back to projects