Reference-Guided 3D Classification (classify3D_refs)¶
Problem¶
Given particles and a complete external set of state volumes, classify the particles without importing incompatible amplitude scaling, noise statistics, or pose history from those references. The common multi-state estimator, bandwidth schedule, and final reconstruction are described in the heterogeneity-analysis overview.
This workflow is for references that may come from another dataset or docking.
When the states and particles already share a refinement lineage, use
refine3D_states.
Algorithm¶
Direct Euclidean comparison is unsafe before the external maps and particle data share a noise model and pose scaffold. The workflow therefore creates data-derived checkpoint maps first:
external maps, low-pass 15 A, nspace = 2500, at most 100,000 particles
-> one greedy correlation-objective pass assigning (state, pose)
-> noise power estimated from that pass's residuals
-> checkpoint maps reconstructed from the assigned particles
-> Euclidean probabilistic multi-state refinement from the checkpoints.
The correlation objective is scale-free, so the initialization tolerates references whose amplitude spectrum differs from the data. The external maps remain unchanged; after the checkpoint, only data-derived maps act as references. Coverage is accepted only when at least 99 percent of sampled particles were updated and every state is non-empty.
The subsequent state decisions use the shared whitened Euclidean loss at a common bandwidth. Balanced fractional updates and the final all-particle reconstruction produce the delivered state and half maps.
Rationale¶
The correlation pre-pass changes the initialization objective, not the final estimator. It supplies the data-derived noise model and pose scaffold that the Euclidean likelihood requires while retaining the external maps only as an initial source of structural hypotheses.
Implementation¶
- Workflow:
src/main/commanders/simple/simple_commanders_refine3D.f90. - Correlation initialization:
src/main/simple_external_reference_pose_initialization.f90. - Frequency schedule:
src/main/simple_refine3D_stage_plan.f90. - State assignment:
src/main/strategies/search/probabilistic/simple_eul_prob_tab.f90andsrc/main/strategies/search/probabilistic/simple_strategy3D_prob.f90. - Policy:
doc/policies/3D/classify3D_refs_policy.md.