Affinity prediction
OpenDDE uses Boltz-2 to predict the binding affinity of small molecules against protein targets. Unlike a docking score, Boltz-2 produces a physically interpretable affinity head trained on experimental binding data.
What it predicts
For each ligand the model returns several numbers:
- pIC50 / pKi — the negative log of the predicted IC50 in molar units. A value of 7 means a predicted IC50 of 100 nM. Higher is better.
- IC50 (nM) — the same number expressed back in concentration units, derived from pIC50.
- Binder probability (
affinity_probability_binary) — the model’s confidence that the ligand is a true binder versus a decoy. This is the number to trust most for triage. - Confidence — iPTM (interface predicted-TM, 0–1) and complex pLDDT. These describe how confident the model is in the predicted binding pose, not in the affinity itself.
How to read the numbers
OpenDDE colours the predicted pKi cell green when iPTM ≥ 0.7, slate when iPTM is between 0.5 and 0.7, and amber with an asterisk (*) when iPTM < 0.5. A high pKi with low iPTM means the model is unsure about the pose, which usually means you should not trust the magnitude.
Hover any predicted cell to see the full breakdown:pKi 7.2 · IC50 63 nM · iPTM 0.82 · pLDDT 87 · binder 91%.
Binary classifier vs quantitative ranking
Boltz-2 is a strong binary classifier — it tells you whether a molecule is likely to bind your target at all. It is much weaker at fine-grained ranking of close structural analogs whose true pKi values differ by less than one log unit.
The practical rule of thumb:
- For hit discovery across diverse compounds — use
affinity_probability_binaryas your primary signal. - For lead optimization within a series — trust the ordering only weakly; rely on experimental SAR for pKi differences smaller than 1 unit.
Honest limitations
- Metals in SMILES: ligands containing metal atoms (e.g.
[Fe],[Pt]) fail standardization and are rejected. The UI surfaces this with a friendly error. - Pockets with cofactors: targets that need a metal cofactor, essential structural water, or undergo large allosteric motions are systematically harder. Treat predictions on those targets as exploratory.
- Top-of-list collapse: a March 2026 evaluation (arXiv:2603.05532) found that within the top-performing compounds, Boltz-2’s ranking signal is much weaker than the overall AUROC suggests. Always validate top hits experimentally.
Tips for best accuracy
- Keep
use_msa_serverenabled in the Boltz service — the accuracy gain over no-MSA inference is large. - Pre-validate SMILES with
POST /api/v1/validatebefore submitting big batches; the screening UI does this automatically. - For production use, point
BOLTZ_SERVICE_URLat a GPU instance. See Microservices for swap-in options.
Next
- Virtual screening — how to run hundreds of predictions as a single campaign.
- Complex prediction — promote a top hit to an AlphaFold 3 / Vina pose.