A confidence score is a model-produced numerical estimate of how reliable a prediction is. In structural drug discovery, it may rank protein–ligand poses, while an affinity probability estimates whether a ligand binds at all. These signals answer different questions: a confident pose supports geometric plausibility, not potency. section 13.2.2
Interpret confidence scores in context:
- Use pose confidence to select among sampled binding geometries, then validate against structural evidence such as RMSD and known binding contacts. section 13.2.2
- Do not treat pose confidence as a potency estimate: DiffDock confidence showed no meaningful potency trend across the TNIK series. section 13.2.2
- Gate affinity values by binder probability before converting predicted log10 IC50 values into pIC50 or micromolar IC50. section 13.2.2
- Calibrate and validate scores within each target: Boltz-2 ranked ligands within TNIK and TYK2, but target-specific offsets prevented direct pooled comparison. section 13.2.2
Listing 13.1 Boltz-2 NIM to Request a Complex Plus an Affinity Estimate
import os, requests
URL = "https://health.api.nvidia.com/v1/biology/mit/boltz2/predict"
headers = {"Authorization": f"Bearer {os.environ['NVIDIA_API_KEY']}",
"Content-Type": "application/json"} #A
payload = {
"polymers": [{"id": "A", "molecule_type": "protein",
"sequence": TNIK_KINASE_DOMAIN}], #B
"ligands": [{"id": "LIG", "smiles": RENTOSERTIB_SMILES,
"predict_affinity": True}], #C
"recycling_steps": 3, "sampling_steps": 50,
"diffusion_samples": 1,
"sampling_steps_affinity": 200,
"diffusion_samples_affinity": 5,
}
r = requests.post(URL, headers=headers, json=payload, timeout=600)
r.raise_for_status()
aff = r.json()["affinities"]["LIG"] #D
print(aff["affinity_pred_value"][0], aff["affinity_probability_binary"][0])Request a Boltz-2 complex and affinity estimate, then inspect the returned binder probability and predicted affinity fields. section 13.2.2
A structural confidence score that succeeds for geometry
For rentosertib bound to TNIK, DiffDock's rank-1 pose reached 0.92 Å symmetry-corrected heavy-atom RMSD against the crystallographic pose and reproduced key hinge contacts. This demonstrates pose recovery, not blind potency prediction. section 13.2.2















