Model Architecture

The system employs ChemBERTa, a transformer-based language model pre-trained on SMILES representations of millions of molecules. A custom hierarchical multi-head classification layer is appended for three-level natural product taxonomy prediction.

Hierarchical Taxonomy

Predictions span three levels of biological–chemical ontology:

• Pathway — Broad biosynthetic origin (e.g., Terpenoids, Polyketides)
• Superclass — Structural family (e.g., Monoterpenoids, Flavonoids)
• Class — Specific structural subgroup (e.g., Iridoids, Isoflavones)

Training Data

The model was trained on a curated dataset of natural products with verified taxonomic annotations, ensuring broad coverage across known natural product chemical space.

Performance

Evaluated using stratified cross-validation with metrics including accuracy, macro-F1, and weighted-F1 across all three classification levels. Confidence scores accompany every prediction.