What is pLDDT?

The LDDT (Local Distance Difference Test) was originally developed as a model-quality assessment metric for comparing predicted protein structures to experimentally determined ones. It evaluates how well local interatomic distances in a model match those in the reference structure, making it less sensitive to global structural errors than RMSD. AlphaFold2 was trained to predict its own LDDT scores, creating the 'predicted' variant — pLDDT. This self-assessment proved remarkably well-calibrated: regions with pLDDT above 90 typically match experimental structures to within 1 angstrom, while regions below 50 are usually intrinsically disordered or poorly modeled.

Interpreting pLDDT requires understanding its four commonly used confidence bands. Scores above 90 indicate very high confidence — the backbone and side-chain positions are reliable. Scores between 70 and 90 represent confident predictions where the backbone trace is accurate but side-chain orientations may have minor errors. Scores between 50 and 70 suggest low confidence, often corresponding to flexible loops or surface-exposed regions. Scores below 50 typically indicate disordered regions that do not adopt a stable 3D structure in isolation. For drug design, binding sites usually need pLDDT above 70 to be structurally informative.

In PepFold's 10-dimensional scoring system, pLDDT contributes to the structural confidence dimension. When ESMFold predicts the 3D structure of a candidate peptide, the mean pLDDT and the minimum pLDDT across the binding interface residues are both factored into the overall score. Peptide candidates with uniformly high pLDDT across their sequence are ranked higher because their predicted structures are more likely to reflect the actual fold, making downstream analyses like binding affinity estimation and synthesis feasibility assessment more reliable.