/propose-designs — Analyse Results & Propose Next Round

• Pass rate: % of designs with filter_rmsd < 2.5 A
• Best composite score and what made it good
• ipTM distribution: median, best, worst
• Refolding RMSD distribution: how many designs fold back well
• Cross-validation survival: % that maintain ipTM > 0.5 in Boltz-2
• ipSAE/pDockQ scores if available

Step 3: Analyse by tool

Compare BoltzGen vs RFdiffusion3:

• Which tool produces better self-consistency (filter_rmsd)?
• Which survives cross-validation better?
• Are there systematic differences in binder length, secondary structure?

Step 4: Propose adjustments

Based on the analysis, propose specific changes for the next round:

Hotspot adjustments

• Which hotspot residues appear in the best designs' interfaces?
• Should we add/remove hotspots? Try different atoms?
• Example: "Add A148:OE1 to active site strategy — top designs show Gln148 contacts"

Binder length

• Are shorter or longer binders performing better?
• Current ranges: active_site 80-120, dimer 80-140, surface 60-140
• Propose narrower ranges based on what's working
• Example: "Narrow active_site to 90-110 — 85% of good designs are in this range"

BoltzGen parameters

• Should we increase --num-designs for successful strategies?
• Any strategy getting zero good designs? Consider dropping or changing it.

RFD3 parameters

• Adjust step_scale (higher = more exploration, lower = more conservative)
• Adjust gamma_0 (lower = more diverse, higher = more focused)
• Adjust num_timesteps (200 standard, 500 high quality)
• Example: "Reduce gamma_0 from 0.6 to 0.3 for dimer strategy — designs too similar"

Inpainting parameters

• Adjust partial diffusion t values (higher = more noise = more redesign)
• Try different segment lengths for segment replacement
• Example: "Try t=15 for helix_hairpin — t=10 designs are too conservative"

New strategies

• Propose entirely new configs if existing ones plateau
• Allosteric sites, cofactor-adjacent binding, epitope-targeted
• New inpainting targets based on structural analysis

Step 5: Write proposal

Output a concrete proposal with:

1. Round N+1 plan — which strategies to run, how many designs each. Limit: at most 3 design jobs per round. Pick the most promising strategies.
2. Config changes — specific YAML/JSON edits with rationale
3. New configs — any new strategy files to create
4. Expected improvement — what metrics should improve and why

Write two files:

• pgdh_campaign/rounds/r{N}_analysis.md — the full analysis of round N results (metrics breakdown by strategy/tool, what worked, what didn't, statistical summaries)
• pgdh_campaign/rounds/r{N+1}_plan.md — the concrete proposal for round N+1 (which 3 jobs to run, config changes, rationale)

Metrics Reference

Config Locations

• BoltzGen YAMLs: pgdh_campaign/configs/strategy*.yaml
• RFD3 binder JSON: pgdh_campaign/configs/rfd3_pgdh_binder.json
• RFD3 inpainting JSON: pgdh_campaign/configs/rfd3_helix_hairpin_inpaint.json
• Target structures: pgdh_campaign/structures/
• Round summaries: pgdh_campaign/rounds/r*_summary.md