AstraAFFINITY
Rank Ligands by Predicted Affinity
Given a target sequence and a small-molecule structure, AstraAFFINITY predicts how tightly they bind — enough to rank a compound series for a known target and focus screening on the front of the list. It is coarse triage, not a precise dissociation constant.
Holds Up on a Time-Split
The honest test for an affinity model is a temporal split — train on what was known, predict what came later. Many models that look strong in-domain collapse to near-zero there. AstraAFFINITY holds its correlation on the time-split, well clear of a strong sequence-and-chemistry baseline.
Real Structure · Kinase–Inhibitor Complex (PDB 1IEP)
A small molecule in the pocket
What It Does
Focus a Screen on the Front of the List
For a target you already know, AstraAFFINITY scores a compound series so the assay budget goes to the candidates most likely to bind.
Rank a Series
Every compound gets an affinity score against the target, so a large series sorts down to a shortlist to test.
Known Targets
Strongest where the target is well-characterised and you are exploring new chemistry around it, not a brand-new protein.
Sequence + Ligand
The read comes from the target sequence and the small-molecule structure — no co-crystal required.
Performance
Measured the Honest Way
Evaluated on a temporal held-out set — the split that exposes memorization — and against a strong sequence-and-chemistry baseline on the same data.
AstraAFFINITY is strong for new chemistry against a known target, where it ranks a series reliably. It is not a novel-target predictor, and it gives a coarse affinity band rather than an exact binding constant. We report it as an early triage capability, sized to what it has been shown to do.
Put Astra on Your Targets
Every Astra model runs inside the Orbion platform. Bring a sequence and get the full read-out — structure, function, modifications, binding, and stability.