Skip to content

split-evaluation.md

Latest commit

 

History

History
80 lines (54 loc) · 3.44 KB

split-evaluation.md

File metadata and controls

80 lines (54 loc) · 3.44 KB

CryoCore Split Evaluation

Last reviewed: 2026-05-15

CryoCore and Structure Factory boundary

Decision

Create cryocore as a sidecar repo for cryo-EM evidence production and structural review. Start by duplicating the contracts that need to exist in both repos, then promote concrete campaigns or scripts only after each lane has a clear owner.

Why This Split Is Useful

The cryo-EM and AI design stacks are complementary, but operationally different.

CryoCore needs heavy system tools:

  • RELION, Warp/M/WarpTools, MotionCor3, Topaz, CTFFIND
  • ModelAngelo, Coot, Phenix, gemmi, validation utilities
  • cryoDRGN/RECOVAR-style heterogeneity tooling
  • ChimeraX and other render/review tools

Structure Factory design lanes need ML/prediction tools:

  • Boltz, Chai, RFdiffusion
  • ProteinMPNN and LigandMPNN
  • Genie-style generation stacks and downstream jury scoring

Trying to put both classes into one image makes the system brittle. The dependency surface becomes a collision between C++/MPI/CUDA cryo tools and Python/PyTorch/JAX design tools. The licensing posture also gets harder because GPL cryo tools, no-cost noncommercial GUI tools, public model weights, and restricted design tools all end up in one packaging decision.

Complementarity

The handoff should be artifact-based:

CryoCore
  maps, models, state assignments, validation metrics, figure packs
    -> Structure Factory
       prediction/design candidates, Boltz/Chai re-predictions, RFdiffusion designs
    -> CryoCore or Structure Factory review
       structural plausibility, figure dossier, claim ledger

Experimental evidence outranks predicted structure output. Design lanes can propose hypotheses or candidates, but CryoCore artifacts define the evidence boundary.

Initial Repo Shape

The first version should contain:

  • a smaller software registry focused on cryo and model-build tools
  • explicit duplicate-tool policy for ChimeraX and visualization tools
  • provider-neutral lane modules
  • no-download smoke examples
  • local validators
  • templates for tracker-neutral issue contracts

It should not contain raw movies, maps, heavy public datasets, model weights, installers, or license records.

Migration Strategy

  1. Keep Structure Factory intact until CryoCore validators pass.
  2. Duplicate shared posture docs for tools that really span both repos.
  3. Move cryo-only campaign scaffolds after the new validators pass.
  4. Keep cross-repo handoffs through artifact contracts, not Python imports.
  5. Leave RFdiffusion/Boltz/Chai in Structure Factory unless a CryoCore campaign only consumes their finished outputs as comparison evidence.

Success Bar

CryoCore is useful when a worker can answer these questions from this repo alone:

  • What tools are allowed, gated, or blocked for cryo-EM processing?
  • What artifacts prove a cryo-EM lane succeeded?
  • What can be run locally without paid compute or raw-data download?
  • How does a map/model evidence pack hand off to Structure Factory?
  • Which tools are duplicated intentionally, and how are their license gates kept synchronized?

Related