Dataset Construction
Training a highly expressive framework like ProMaya requires a massive, perfectly balanced, and extremely difficult dataset. If the model is only shown "easy" negative examples (like proteins that don't even belong in the same cell), it will fail to learn true binding physics.
The 88,740 Protein Pairs
ProMaya is trained on exactly 88,740 protein pairs, curated primarily from PDBbind and DIPS-PLUS. - Positive Pairs: 44,370 experimentally validated interacting chains. - Negative Pairs: 44,370 non-interacting chains, carefully constructed (see below).
CD-HIT Splitting Strategy
To prevent the model from simply memorizing sequence homology, the data is split 70% Train / 30% Test at the CD-HIT Sequence Cluster level. This guarantees that no protein cluster spans multiple splits; if Protein A is in the training set, no protein with \(>30\%\) sequence identity to A will appear in the test set.
The 5-Tier Negative Sampling
The 44,370 negative pairs are not just random pairings. They are heavily engineered across 5 categories of increasing difficulty to force ProMaya to learn true geometric and biochemical binding rules.
| Negative Category | Count | Description | Difficulty |
|---|---|---|---|
| Subcellular Localization | ~14,000 | Proteins from completely different cellular compartments (e.g., Nucleus vs Extracellular). | Low |
| Random Pairing | ~21,700 | Randomly paired chains known not to interact. | Low |
| Interface Ablated | ~7,500 | True positive pairs where the specific binding interface residues have been heavily mutated or deleted in silico. | High |
| Context-Aware Hard | ~1,170 | Proteins that belong to the same complex or pathway but do not physically touch. | Very High |
| Docking-Derived | ~4,000 | False-positive pairings generated by rigid-body docking software with high geometric fit but incorrect electrostatics. | Extreme |
By forcing the model to distinguish between a True Positive and a Docking-Derived Negative, ProMaya learns that geometric shape fit alone is insufficient—the atoms must actually possess complementary electron mass densities (LSMD).
Next: View the loss functions in the Training Pipeline →