First Ascent by
Ethan Truong
⚠️ Pending written proposal
The LVMOF Crystal Face
Materials Science / Metal-Organic Frameworks
The Proposed Route
A deep feature-engineering and ML pipeline predicting crystallinity of metal-organic frameworks (MOFs) from a massive feature matrix: Morgan fingerprints, RAC descriptors, ChemBERTa embeddings, TEP calculations, and more. An ordinal XGBoost classifier with SHAP interpretability identifies the molecular drivers of MOF crystalline quality.
🧗 The Crux
64,739 features for ~756 samples is an extreme high-dimensional setting — feature selection (VarianceThreshold + SelectKBest) is doing a lot of work and its validity should be interrogated. The feature name mismatch bug already encountered in the notebook (608-feature offset) suggests the pipeline is complex enough to have subtle errors.
⚠️ Pre-Climb Checklist
⚠️ Written proposal not yet submitted — required. ✅ Notebook is clearly advanced and well-structured. ⚠️ Validate that the feature selection pipeline is not inadvertently leaking test set information.
Guidance
- Submit written proposal for formal approval
- With 64K features / 756 samples: overfitting and data leakage are real risks
- SelectKBest must be fit only on training data — report test set performance separately
- SHAP analysis is the key deliverable — top features must make chemical sense for MOF crystallinity
- Measure AU-ROC and AU-PRC curves to evaluate model accuracy
- Consider contrastive learning to separate positives/negatives before the classification head — see R030 for an example
Source proposal: Ethan_Truong_LVMOF_Surrogate.ipynb
CHEM 169/269 · Applied AI & Machine Learning for Biochemistry