Antibody structure prediction

Predicting an Fv is a different problem from predicting a protein. The framework is conserved enough that any method gets it close. Five of the six loops fall into a small number of canonical conformations that are largely determined by length and by a few key residues. Then there is CDR-H3.

Why CDR-H3 resists

H3 sits at the center of the paratope, spans the junction where V, D and J segments recombine, varies from a few residues to more than twenty, and adopts conformations that do not cluster into a small canonical set. It also packs against the light chain, so its conformation depends on the VH and VL arrangement, which the model has to get right at the same time.

Dedicated antibody models, including IgFold and ImmuneBuilder, predict frameworks and canonical loops quickly and accurately and are honest about H3 being the weak point. General predictors are competitive and slower. For a large panel, the speed of a dedicated model is usually worth more than a small accuracy difference on loops that are already easy.

The orientation problem

VH and VL pack at a defined angle, and small changes in that angle move the loop tips by more than the loop prediction error. A model can produce correct loops on a slightly wrong interdomain arrangement and still misplace the paratope. When you are using a model to reason about which residues contact an antigen, the orientation matters as much as the loops, and it is rarely reported.

For VHHs the problem changes shape rather than disappearing: no light chain to orient against, but a long CDR3 that often folds back over the former interface, frequently pinned by a disulfide that the model has to predict correctly to place the loop.

What the models are good for

Panel-scale work: numbering-aware structural features, surface exposure for liability weighting, charge and hydrophobic patch analysis, rough paratope definition, and inputs to downstream models. These uses tolerate a couple of angstroms of error in a loop.

What they are not good for is deciding an epitope, reasoning about individual contacts, or serving as the structural claim in a report. If a decision rests on where a residue sits, that belongs to an experiment.

Reading a predicted Fv critically

• Check the H3 length against what the model does well; long H3s deserve more skepticism.
• Look at per-residue confidence across the loops, not the global score.
• Predict with more than one method and see whether H3 agrees. Disagreement is informative and common.
• If the model came from a pipeline that also designed the sequence, remember that the predictor and the designer may share failure modes.

Used this way, predicted Fvs earn their place: they make a panel analyzable, they focus the experiments, and they do not pretend to be structures.