Skip to content

README.md

Latest commit

 

History

History
91 lines (62 loc) · 3.35 KB

README.md

File metadata and controls

91 lines (62 loc) · 3.35 KB
Raw

Real Cryo-EM Map Tutorial: 9UGC_A

pdb_transition_real_data_compressed.mp4

EmbedOpt inference guided by a real cryo-EM density map (EMD-64136, 3.52 Å).

Inputs already in this folder:

File Description
emd_64136_cropped.mrc Cropped cryo-EM map (from EMDB EMD-64136)
9UGC.pdb Reference PDB for Kabsch alignment during reward computation
9UGC_A.fasta Protein sequence

Step 0: Crop your map (already done here)

If starting from a raw EMDB download, crop the map with the embedopt-crop-map CLI:

embedopt-crop-map emd_64136.map --thresh 0.35 --pad 10 -o emd_64136_cropped.mrc

--thresh is map-dependent — inspect the map in ChimeraX to find a density cutoff that tightly encloses the structure. --pad 10 adds a 10-voxel buffer on each side.

Step 1: Make minimal JSON

pixi run python examples/real_map_tutorial/01_make_minimal_json.py

Reads 9UGC_A.fasta and writes 9UGC_A_minimal.json (sequence only, no MSA), which is the input for Step 2.

Step 2: Generate MSA

bash examples/real_map_tutorial/02_make_msa.sh

Fetches MSA via the Protenix/ColabFold MMseqs2 server and writes results to 9UGC_A_msa/9UGC_A/msa/0/. Also produces 9UGC_A_minimal-update-msa.json with MSA paths embedded, which Step 3 uses as its basis.

Step 3: Generate inference configs

Edit the variables at the top of 03_make_cfg.py:

  • ALGORITHMS — any subset of "base" (unconditional), "dps" (Denoising Posterior Sampling), "embedopt" (our method, embedding optimization)
  • LEARNING_RATES — list of floats; one config per (algorithm, lr) combination
  • REF_PDB, REF_MAP, RESOLUTION — map reward inputs
  • OUT_DIR — where configs are written (default: inference_configs/)

Then run:

pixi run python examples/real_map_tutorial/03_make_cfg.py

Configs are written to inference_configs/, e.g. 9UGC_A_embedopt_lr{LEARNING_RATES}.json.

Step 4: Run inference

Edit the variables at the top of 04_run_inference.sh:

  • MODEL_NAME — model checkpoint (protenix_base_default_v0.5.0, protenix_base_default_v1.0.0, or protenix-v2); auto-downloaded on first run. Note: protenix-v2 weights currently cannot be auto-downloaded outside China (Bytedance CDN returns 403; see issue #296); use v0.5.0 or v1.0.0 instead, or set CHECKPOINT_DIR to a local copy
  • INPUT_JSON — path to a config from inference_configs/
  • DUMP_DIR — output directory (default: output/)
  • N_STEPS — number of diffusion denoising steps
  • SEEDS — comma-separated seeds; one output structure per seed

Then run:

bash examples/real_map_tutorial/04_run_inference.sh

Energy relaxation runs automatically on the final diffusion step.

Outputs under output/<config_name>/seed_<N>/:

  • trajectory/step_<last>_relaxed.pdb — energy-relaxed final structure (recommended for downstream use)
  • predictions/*.cif — raw model output (before Kabsch alignment and energy relaxation)
  • trajectory/step_<N>.pdb — intermediate snapshots every save_pdb_every_n_steps steps
  • reward_history.json — per-step map reward values
  • profile.json — timing and peak GPU memory