Complexa Design

Complexa Design Skill

Drive the full four-stage complexa design pipeline: generate (flow matching + search) -> filter (top-N by reward) -> evaluate (refold with AF2/RF3) -> analyze (success rate, FoldSeek/MMseqs diversity). Pick the right pipeline config for the design intent, validate the run upfront so the user does not discover a missing ckpt mid-folding, run it, and emit a replayable manifest + per-design success CSV.

What this skill enables

• Protein binder design for protein targets (AF2 reward + ColabDesign refold).
• Ligand binder design for small-molecule targets (RF3 reward + RF3 refold).
• AME motif scaffolding with ligand context (motif + ligand features, RF3).
• Search-based optimization: single-pass, best-of-n, beam-search, fk-steering, mcts.
• Refold backends: ColabDesign (AF2), RF3, Boltz2, ESMFold (fast iteration).
• Pass-rate + diversity analysis with per-result_type thresholds.

Step 1: Pre-flight

Always run the shared preflight before launching a design — generation needs the GPU and the right checkpoint, evaluation needs AF2/RF3 weights and tool binaries. Bail early if the host cannot run the chosen pipeline.

bashCopy code
bash .claude/skills/_shared/scripts/preflight.sh

Read ./complexa_setup/preflight.json and bail if any of these are missing for the chosen pipeline:

• gpu.available: false -> all pipelines fail.
• gpu.vram_gb < 40 -> generation OOMs at default batch_size: 16; lower to 8.
• ckpts.complexa[.ckpt] -> required for protein binder.
• ckpts.complexa_ligand[.ckpt] -> required for ligand binder.
• ckpts.complexa_ame[.ckpt] -> required for AME.
• env.AF2_DIR missing -> protein binder default eval (colabdesign) fails.
• env.RF3_CKPT_PATH or env.RF3_EXEC_PATH missing -> ligand binder / AME default eval (rf3_latest) fails.

If a ckpt is missing, point at complexa-setup and have the user run complexa download --complexa-<variant> first.

Step 2: Pick the pipeline

Complexa has one default pipeline (protein binder) and two extensions (ligand binder, AME / enzyme). The pipeline is selected entirely by the configs/search_*_pipeline.yaml you pass to complexa design — each YAML pins its own model checkpoint, autoencoder, targets dict, default reward, and default refold backend. Switching pipelines is just "swap the config path and the target name comes from a different dict".

Default — protein binder

bashCopy code
complexa design configs/search_binder_local_pipeline.yaml \
    ++run_name=pdl1_v1 ++generation.task_name=02_PDL1

Use this when the user says "design a binder for X", "PDL1 binder", "de novo binder", "design proteins for a target", etc. — i.e. the target is a protein surface. The config pins complexa.ckpt + complexa_ae.ckpt, reads targets from configs/targets/targets_dict.yaml, rewards with AF2 (af2folding), inverse-folds with SolubleMPNN, and evaluates with ColabDesign / AF2. If the user did not specify, this is what they want.

Extension A — ligand binder (small-molecule pocket)

bashCopy code
complexa design configs/search_ligand_binder_local_pipeline.yaml \
    ++run_name=v11_v1 ++generation.task_name=39_7V11_LIGAND \
    ++metric.binder_folding_method=rf3_latest

Switch to this when the user says "ligand binder", "small-molecule pocket", "SMILES target", "ATP-binding protein", or names a target ending in _LIGAND / from the FAD / SAM / OQO / 7V11 / etc. families. The config also activates LoRA (r=32, lora_alpha=64) which is required for the released ligand checkpoint — leave the lora: block alone.

Extension B — AME / motif + ligand (enzyme scaffolding)

bashCopy code
complexa design configs/search_ame_local_pipeline.yaml \
    ++run_name=ame_chm ++generation.task_name=M0096_1chm

Switch to this when the user says "scaffold a motif near a ligand", "active-site design", "enzyme scaffolding", "AME", or names a target like M0024_1nzy, M0096_1chm (the M####_<pdb> AME task naming). The config sets env_vars.USE_V2_COMPLEXA_ARCH=True (the CLI runner injects it into the subprocess) and uses both MotifFeatures and LigandFeatures. Default search is single-pass; switch to best-of-n only if you also enable the CompositeRewardModel (commented out in ame_generate.yaml).

Pipeline cheat sheet — what changes when you switch

For the full per-pipeline breakdown (reward weights, success thresholds, analysis modes), see reference/pipelines.md.

Step 3: Gather parameters

Use AskUserQuestion to fill in the four parameters that vary every run. Default to sensible production settings if the user has no preference.

• Target name — must be a key in the relevant dict (targets_dict.yaml for protein binder, ligand_targets_dict.yaml for ligand, ame_dict_v2.yaml for AME). If the user names a target that is not in the dict, hand off to complexa-target to add it first.
• Run name — a short identifier appended to the output dir (e.g. pdl1_v1).
• Search algorithm — default to beam-search with beam_width=8 and n_branch=4 for production. Use single-pass for a quick smoke test.
• Evaluation refold backend — protein binder defaults to colabdesign (AF2); ligand/AME default to rf3_latest. Use esmfold for fast iteration (worse but seconds per sample).

Step 4: Validate

Validate before running. This is cheap (seconds) and catches missing ckpts, missing env vars, unknown override keys, and missing target entries — all of which would otherwise abort the pipeline mid-evaluation after hours of generation.

bashCopy code
complexa validate design configs/search_binder_local_pipeline.yaml \
    ++generation.task_name=02_PDL1 \
    ++metric.binder_folding_method=colabdesign

The validator returns non-zero on failure and prints a pass/fail report. Re-run the command with the suggested overrides until it returns clean.

Step 5: Run the pipeline

complexa design is the right tool for the full 4-stage run — it orchestrates generate → filter → evaluate → analyze as sequential subprocesses with a shared run name, log dir, and multi-GPU split (see run_design_pipeline in src/proteinfoundation/cli/cli_runner.py). Re-implementing that manually loses the per-stage log routing and progress prints.

Use ++ (forced) Hydra overrides; they apply to all stages. The minimal production protein-binder invocation:

bashCopy code
complexa design configs/search_binder_local_pipeline.yaml \
    ++run_name=pdl1_v1 \
    ++generation.task_name=02_PDL1 \
    ++generation.search.algorithm=beam-search \
    ++generation.search.beam_search.beam_width=8 \
    ++metric.binder_folding_method=colabdesign

For ligand binder / AME, swap the pipeline YAML and target name per Step 2's cheat sheet — every other override above is pipeline-agnostic and can be reused as-is.

Add --verbose to stream logs to the terminal instead of ./logs/. The skill does not poll progress — the user re-invokes if they want a status; point them at complexa status and ./logs/design_pipeline_*/.

Direct module invocation (debug fallback)

To debug a single stage without pipeline orchestration, invoke the underlying Hydra module directly. complexa generate CONFIG is just a logged subprocess wrapper around this:

bashCopy code
python -m proteinfoundation.generate \
    --config-path "$(realpath configs)" \
    --config-name search_binder_local_pipeline \
    ++run_name=debug_pdl1 \
    ++generation.task_name=02_PDL1

Same pattern for proteinfoundation.{filter,evaluate,analyze}. Use this when you want to attach ipdb, run under nsys, or skip the pipeline log dir. Prefer complexa generate/filter/evaluate/analyze for normal one-shot runs (you get logging and parallel job splitting for free).

AME-specific gotcha (when running AME with RF3 refold): RF3 will try to complete missing atoms on the ligand based on its CCD code, which produces shape errors in RMSD calculations and the wrong structure. Before RF3 sees the PDB, rename the ligand residue to L:0 so RF3 treats it as a generic ligand and skips atom completion. If your AME generation outputs already encode the ligand this way (the canonical Complexa pipeline does), no extra step is needed; otherwise patch each PDB with atomworks.io:

pythonCopy code
from atomworks.io import load_any, to_pdb_file
atom_array = load_any("my_design.pdb")[0]
ligand_mask = atom_array.chain_id == "A"
atom_array.res_name[ligand_mask] = "L:0"
to_pdb_file(atom_array, "my_design_rf3_ready.pdb")

Same applies if you're piping AME outputs into the complexa-evaluate-pdbs skill with an RF3 backend. Skip the rename for AF2/colabdesign refold or for non-AME pipelines.

Wall-clock at default (nsteps=400, beam_width=8, batch_size=16, 100 designs, colabdesign eval) is ~30–120 minutes on a single A100/H100.

Step 6: Collect results

Outputs land in two directories. Surface both:

bashCopy code
ls ./inference/${CONFIG_STEM}_${TASK}_*${RUN_NAME}/   # generated PDBs + filter
ls ./evaluation_results/${RUN_NAME}/                  # per-design CSV + analysis

Read the combined results CSV and summarize:

bashCopy code
ls ./evaluation_results/*/binder_results_*_combined.csv
ls ./evaluation_results/*/motif_binder_results_*_combined.csv  # AME
ls ./evaluation_results/*/res_filter_*_pass_*.csv              # success rate
ls ./evaluation_results/*/res_div_foldseek_*.csv               # FoldSeek diversity

Pull the success rate from res_filter_binder_pass_*.csv, the per-design metrics (interface pAE, pLDDT, scRMSD) from the combined CSV, and FoldSeek TM-score diversity from res_div_foldseek_*.csv. Report top-N designs by i_pAE (protein binder) or min_ipAE (ligand binder).

Step 7: Emit manifest

Drop a JSON manifest beside the results so the run is replayable. The shared helper captures the command, config, git SHA, and pointers to the result CSVs.

bashCopy code
python3 .claude/skills/_shared/scripts/write_manifest.py \
    --output-dir ./evaluation_results/${RUN_NAME} \
    --command "complexa design configs/search_binder_local_pipeline.yaml ++run_name=${RUN_NAME} ++generation.task_name=${TASK}" \
    --skill complexa-design \
    --out ./run_manifest.json

Surface the manifest path and the result CSV to the user.

Most-common overrides

The 10 overrides that cover ~90% of runs. Full reference (every key, type, default) is in reference/overrides.md.

Hardware requirements

Typical wall-clock for 100 designs, beam_width=8, default nsteps=400:

• Protein binder + colabdesign refold: ~60–120 min on 1x A100/H100.
• Ligand binder + RF3 refold: ~90–180 min (RF3 dominates).
• AME + RF3 refold: ~120–240 min.
• Any pipeline + ESMFold refold: ~30–60 min (fast iteration).

Bumping gen_njobs=2 and eval_njobs=2 halves wall-clock on a 2-GPU host. See .claude/skills/_shared/reference/hardware.md for per-pipeline VRAM tables.

Troubleshooting (common cases)

For the full list (chain-ID mismatches, hotspot residues, ligand residue renaming for RF3, missing inverse-folding models, etc.) see reference/troubleshooting.md.

For per-pipeline details (which model, reward, inverse folding, evaluation backend, result_type, LoRA settings, USE_V2_COMPLEXA_ARCH toggle), see reference/pipelines.md.

For the full override reference (every generation.*, metric.*, aggregation.* key with type, default, example, and effect), see reference/overrides.md.

For all troubleshooting cases (cause + fix + source), see reference/troubleshooting.md.