An atom-level restraint can be interpreted as noise by the structure-prediction model during reverse diffusion. Without conformer restraints, local geometry is not guaranteed to remain valid.
If you do not want to use conformer restraints, define restraints between appropriate domains or other sufficiently large atom groups. Domain-level guidance allows the model to treat the restraint as a coherent structural change rather than local noise.
Possible remedies are:
- Move domains or other suitable atom groups instead of individual atoms.
- If necessary, use
moveto choose whichatom_selectiongroups are allowed to move; the other selected groups are pinned for that restraint. - Combine the distance restraint with conformer restraints.
- Set
stop_sigmaso that the restraint is released during the final denoising steps.
Note that releasing a restraint with stop_sigma may cause the final structure to no longer
satisfy the restraint target.
Try increasing restraints_config.max_iter to allow more optimizer iterations per denoising
step. Also check whether any structure can satisfy all active restraints with zero loss.
Incompatible targets or an infeasible combination of restraints can make zero loss impossible,
regardless of the iteration limit.
In some cases, the model produces an intermediate, distorted conformation of a bridged compound.
This can leave the minimizer trapped in a local minimum with nonzero loss, even when a zero-loss
conformation is possible. Increasing max_iter alone may not resolve such a local minimum.
Consider combining the RMSD restraint with conformer restraints or setting stop_sigma to release
the RMSD restraint during the final denoising steps.
Why does applying a distance restraint only near the end (start_sigma: 1) break the protein structure?
During reverse diffusion, larger sigma values allow guidance toward larger-scale structural
changes. As sigma decreases, the scale of changes that can be induced also decreases; around
sigma = 1, guidance is approximately limited to atom-scale changes.
If the restraint requires a domain motion or another large rearrangement, applying it only at low sigma can force the model to satisfy it through local distortion instead. Choose the restraint's activation timing according to the scale of the desired structural change and any conformer correction that is required. See Sigma gating for configuration details.
RGI minimizes the restraint energy at every diffusion step where restraints are active, so it is generally more computationally expensive than standard guidance. The cost increases as the number of restraints grows, both from evaluating the additional restraints and from resolving conflicts among them. Conformer and RMSD restraints are especially prone to longer runtimes because they often involve many restrained terms or atoms.
Consider reducing the number of restraints, using start_sigma and stop_sigma to limit the
diffusion steps on which they are active, and ensuring that GPU execution and JIT compilation are
enabled appropriately for your integration.
Please use GitHub Issues.