# Reconstruction notes — real methodology extracted from EoE project artifacts

Source of truth for the 5 skills + 1 specialist. Every threshold/schema below is
lifted from actual project artifacts (not invented).

## 1. binder-design-campaign
Pipeline (EoE_binder_design_report.md, binder_validation_protocol.md):
  Target PDB → RFdiffusion backbones → (Soluble)ProteinMPNN sequences → co-fold fold-back → on-target selection
- Backbones: RFdiffusion Complex_base checkpoint, 40/target (spec says 1k–5k in scale-up),
  binder len 65–90 aa, condition on mapped epitope hotspots.
- Sequences: SolubleMPNN v_48_020, 8 seqs × 3 temperatures × N backbones; sequence-complexity filter.
- Fold-back: ESMFold2-Fast triage → Boltz-2 / Chai-1 / AF-Multimer final ranking.
- Composite on-target score = ipTM × epitope_coverage × pLDDT (rewards confident AND on-epitope).
- Acceptance thresholds (Accept / Strong):
    ipTM > 0.6 / > 0.8 ; pAE_interface < 10 / < 5 (Å) ; predicted DockQ > 0.23 / > 0.49 ;
    binder pLDDT > 80 / > 90 ; epitope recapitulation ≥ 50% / ≥ 75% ; ddG < -30 / < -45 REU (if run).
- Filter: drop ipTM<0.6 OR epitope-recap<50%; rank by (ipTM + epitope-recap - normalized pAE); top 10-20/target.
- OUTPUT SCHEMAS:
  binder_design_targets.csv: target, modality, framework, epitope, key_hotspots, target_KD_nM, selectivity_screen, gpu_design
  S1_binder_designs.csv: design, target, binder_len, mpnn_score, iptm, complex_plddt, confidence, pass, seq
  binder_validation.csv: per-design full metrics.
- Composes catalog skills: proteinmpnn, solublempnn, ligandmpnn, alphafold2, boltz, chai1, esmfold2, remote-compute-modal.
- KNOWN GAP: no RFdiffusion skill in catalog — backbone step must be wrapped/installed; note this explicitly.
- Honesty: ipTM/pLDDT are model confidences, NOT measured affinity. Wet-lab: SEC+SPR/BLI, functional assay.

## 2. binder-developability-hardening
(celiac_gut_restricted_binder_report.md, celiac_binder_*.csv/.fasta/.json)
- Liability scorecard cols: length, net_charge, n_cys, n_aromatic, pct_charged, gravy,
  trypsin_sites, chymo_sites, pepsin_sites, elastase_sites, harden_target_sites, total_sites.
- 4 GI protease cleavage rules (PeptideCutter-style):
    pepsin; trypsin (after K/R, NOT before P); chymotrypsin (after F/Y/W/L/M); elastase (after A/V/G/S/I/L).
- Cross-ref every cut site vs fixed binding-interface residues (in_interface flag; unremovable if on interface).
- Interface-preserving hardening: re-run SolubleMPNN with interface residues FIXED and
    omit_AAs="KRFWYM" (kills trypsin K/R + chymo/pepsin aromatics F/W/Y/M; keeps aliphatics + Cys for disulfides).
- Honest trade-off: elastase sites RISE (composition pushed to small aliphatics); managed at FORMAT level (compact disulfide-locked VHH), not sequence.
- Result exemplars: trypsin 32→5 (MICA-decoy), 18→1 (IL-15); 46-49% total reduction on hardenable proteases; interface preserved exactly.
- Delivery-format spec (6 steps): minimize→VHH graft (~13kDa, ref oral anti-TNF V565)→harden framework→disulfide-lock (Cys22-Cys92)→mucin-anchor. Target-depth: apical surface=favorable, submucosal pool=caveat.
- OUTPUT SCHEMAS:
  protease_map.csv: binder, position, residue, n_proteases, proteases, in_interface, removable
  developability.csv (cols above); hardened.fasta; delivery_spec.json (6-step path + target-depth).
- Honesty: CPU sequence/structure design; protease rules are predictions not digestion; needs SGF/SIF digest + co-fold.

## 3. multivalent-display-design
(eoe_pmhc_nanoparticle_*.csv, table3_nanoparticle_specs.csv)
- Payload (pMHC / antigen / binder) → scaffold + fusion architecture + valency scenarios + assembly model.
- NP core exemplar: 20 nm Fe3O4/γ-Fe2O3, PEG2k-maleimide ~50-100/NP (Navacim/Santamaria precedent), thiol-maleimide Michael addition pH 7.0-7.5.
- OUTPUT SCHEMAS:
  fusion_architecture.csv: component, spec, function, notes
  assembly_scenarios.csv: Scenario, Valency, Surface_occupancy_%, Inter_epitope_nm, Cross_link_feasible, Estimated_avidity_fold, Advantages, Disadvantages
  specs.csv: Parameter, Value, Rationale/Reference
- Scenario sweep: conservative(3/NP)/standard(5/NP)/high — avidity fold ~12/24/…, inter-epitope nm for TCR clustering (~5.8 nm optimal).
- Mechanism framing: multivalency → TCR cross-linking → Tr1 (CD39+CD73+ IL-10) tolerance.

## 4. biotech-pitch-and-brand
(company_brand_brief.md, BIOTECH_PITCH_DECK_GUIDE.txt, business_plan.md)
- Brand brief sections: (1) name (mechanism-forward; alternatives table w/ "why not #1"); (2) mission;
  (3) founding thesis = 4 claims; (4) positioning statement (For X / <Company> is Y / Unlike competitors / <Company> does Z);
  (5) positioning axes = the whitespace quadrant (2 axes, competitors cluster in one corner); (6) visual identity
  (motif, palette with hex + role, type, tone of voice); (7) "What we are NOT".
- VC deck = 10 slides w/ speaker notes: title; disease burden/unmet need; approach/discovery; structural validation & confidence;
  platform architecture; mechanism; clinical dev roadmap (phases, $ , go-gates); regulatory & commercial; competitive landscape & IP; milestones & go/no-go.
- Pulls numbers from an existing market report + program dossier (prevalence, TAM, ipTM/pLDDT, budget, timeline).
- Audience adaptation: therapeutic biotech / NP-tech / VC / academic emphasis switches.
- Honesty rule: explicitly separate tetramer-validated vs computational-prior claims; never overclaim.

## 5. compute-cost-audit
(gpu_stalled_tasks_report.md, gpu_stalled_tasks_costs.csv)
- Method: audit every work session transcript for GPU/deferral language; cross-check deferred items
  against the executable protocol specs written in lieu of running them; tabulate.
- Cost model: realistic = single_pass_GPU_hr × iteration_factor (2-2.5× on generative steps) × provider $/hr.
  Modal rates used: A100-40 ~$2.10, A100-80 ~$2.50, L40S ~$1.95, L4 ~$0.80 (approx 2026 on-demand; spot 30-60% cheaper).
- Report BOTH single clean-pass and realistic (with iteration) totals.
- Dependency chain matters more than dollars: e.g. E1→E2→E3→E4 (can't cherry-pick); flag decoupled tasks; name cheapest-high-value first move.
- Also flag SKILL gaps (tool not wrapped) distinct from pure compute gaps.
- OUTPUT SCHEMAS:
  costs.csv: id, campaign, task, tool, gpu, gpu_hr_low, gpu_hr_high, iter_factor, source, rate_hr, clean_low, clean_high, real_hr_low, real_hr_high, real_low, real_high
  report.md (headline $ + task table + dependency graph + practical notes); cost bar chart (figure-style).

## 6. PROTEIN_ENGINEER specialist
- Structure-design bench counterpart to INFLAM_TARGET_ANALYST (target/epitope) and DRUG_PROGRAM_ARCHITECT (program altitude).
- Owns: target+epitope → designed, validated, developable binders/scaffolds.
- Bundles skills 1-3 + model skills (proteinmpnn/solublempnn/ligandmpnn/alphafold2/boltz/chai1/esmfold2) + remote-compute-modal.
- Discipline: fold-back thresholds, developability gates, "in-silico ≠ experimentally validated" honesty.