Define the prediction before reading scores
Boltz-2 structure prediction and affinity prediction answer different questions. The documented YAML accepts protein, DNA, RNA and ligand entities; ligands are described by SMILES or CCD, and affinity must be requested separately. A structurally plausible complex does not supply an experimental binding measurement. Use this guide when interpreting results; the related Recipe covers operational setup.
Confidence is not affinity
| Output | Documented interpretation | Do not infer |
|---|---|---|
| pLDDT / complex_plddt | Local/aggregated structural confidence | Experimental function |
| pTM / ipTM and chain/interface variants | Global or interface structural confidence | Measured binding strength |
| PDE | Predicted distance error, in Å | A binding free energy |
| confidence_score | Score used to sort structure samples | Validated binder probability |
The documentation defines confidence_score as 0.8 × complex_plddt + 0.2 × iptm, using ptm for single chains. Aggregated JSON pLDDT-style scores are documented on a 0–1 scale. Preserve the field and file context rather than assuming every displayed pLDDT uses the same scale. High confidence cannot resolve assay context or experimental binding.
Two affinity heads, two uses
affinity_probability_binary is a model-predicted binder probability from 0 to 1, intended for separating binders from decoys. It is neither structural confidence nor proof of calibrated probability on a new screening set. affinity_pred_value is intended to compare active molecules and small modifications in ligand optimization. The two outputs have different supervision and should not be merged into one undifferentiated score.
Write its reported value as y = log10(IC50 / µM): y = −3 corresponds to 1 nM, y = 0 to 1 µM, and y = 2 to 100 µM. These are illustrative unit conversions, not measured results. Lower y indicates stronger predicted binding under this convention. It is not Kd. Dimensionless pIC50 is 6 − y; this does not establish a binding free energy. The documentation contains wording about pIC50 in kcal/mol that this guide deliberately does not adopt.
Stay inside the documented ligand scope
Only one small-molecule ligand chain can be selected for affinity calculation against a protein target. It cannot be a protein, DNA or RNA binder. The documentation sets a maximum of 128 atoms, counting heavy atoms and hydrogens retained by RDKit RemoveHs, and discourages ligands significantly larger than the 56-atom training limit under that counting rule. Passing an input check does not establish scientific applicability. RNA/DNA/cofactor targets may not crash the code but are documented as unreliable for this module.
Retain enough context to compare results
Record entity identities, chemical representation, target sequence, MSA source, software revision, both checkpoints, sampling settings, and affinity_mw_correction. With --use_msa_server, protein sequence data is sent to the configured service; local inference alone does not prove local-only handling. Preserve structure and affinity outputs together and keep failed inputs. Compare the same endpoint and settings, then assess predictions with independent assay evidence. No model execution or experimental validation is claimed here.
Related resources and reading
Boltz-2 Complex Prediction and Small-Molecule Affinity Analysis
Protein & Structural Biology AI Workflow · Evaluating Molecular Property Prediction Models
Sources and evidence boundary
Sources were inspected on 2026-10-08. Revision-pinned project documentation supports capability statements. Evaluation choices are editorial proposals. This article reports no executed workflow, measured performance, or experimental validation.