Compound preference, the LSL layout: ligand, sequence, ligand. The steps below follow that order: the first compound, then the target, then the second compound. Add more compounds to rank a series. How it was built and what it scores.
One target, and every compound is compared at it.
Targets are named by UniProt accession. Gene symbols are many-to-many against accessions and the failure is silent, so the accession is what the model is given. This model can score 2,079 targets across 32 families; the targets page lists them.
Up to 16 compounds per run. Every pair is scored, so n compounds is n(n−1)/2 comparisons.
Two compounds drawn in the editors never needs an email. A longer series does, and the report is emailed back.
Unlocks the extra compound box below ↓
Results go to that address, and we use it to say when the model changes.
Every answer carries one number, its strength, from 0.5 to 1.0, and the strength is its own confidence. Each is labeled with its band and that band's measured accuracy.
| Strength at or above | Share kept | Accuracy |
|---|---|---|
| 0.50, answer everything | 100.0% | 0.71 |
| 0.60 | 42.8% | 0.84 |
| 0.70 | 17.4% | 0.92 |
| 0.80 | 6.1% | 0.96 |
| 0.90 | 1.2% | 0.98 |
A near-tie is a near-tie: where the two measured values sit within half a log the model is right 0.57 of the time.
The model answers which of two compounds a target prefers, that is which is active at the lower concentration there. Both readings behind every training comparison share one endpoint, so the ordering is between two readings of the same kind. It returns an ordering, and a compound against itself returns exactly 0.50.