MIL-based QSAR modeling
Conformational-ensemble QSAR using multi-instance learning and chirality-aware 3D pharmacophore descriptors.
Scientific problem
Single-conformer 3D QSAR depends on choosing one molecular geometry even though the biologically relevant conformation is usually unknown.
Motivation
The study tested whether representing each molecule by an ensemble of conformers could retain useful 3D and stereochemical information without selecting a single presumed bioactive conformation.
My role
- Implemented the conformational-ensemble modeling approach and prepared and curated the chiral/achiral datasets.
- Built and evaluated QSAR models, interpreted the results, and contributed to manuscript preparation.
- Automated collection of stereochemically sensitive activity datasets.
- First author of the Molecular Informatics publication.
Methodology
- Generation of conformational ensembles for each molecule.
- Chirality-aware, alignment-independent 3D pharmacophore descriptors.
- MIL-k-means, MIL-max, and related ensemble representations.
- Comparison with single-conformer 3D models and strong 2D QSAR baselines.
Results
- Published in Molecular Informatics (2021).
Software
- Python
- RDKit
- pmapper
- scikit-learn
- NumPy
- pandas
- Matplotlib