https://osf.io/vj7ue/files/osfstorage/6907ba1bf5cdb710adc181cf

https://chatgpt.com/g/g-p-68e82d532be8819190b2ee9e17a0d832-agi-foundaton/shared/c/69078491-d7ec-8332-aa9f-aa042abf53d3

Everyday Structure AI: How AlphaFold-Style Models Quietly Upgrade Products, Health, and the Built World

1. Executive Summary — The “Structure-Before-Search” Pattern

Modern structure AI—spanning AlphaFold-style biology and GNN-style materials—reframes discovery as constrained design instead of unconstrained trial-and-error. In proteins, AlphaFold 3 predicts joint complexes of proteins, nucleic acids, ligands, ions, and modifications; complementary systems like RoseTTAFold and ESMFold broaden access and speed. In interactions and design, DiffDock (docking), RFdiffusion (backbone generation), and ProteinMPNN (sequence fit) turn candidate making into a guided, auditable pipeline. On the materials side, GNoME scales crystal discovery into the millions, while autonomous labs like A-Lab close the loop from prediction to synthesis. Collectively, these tools enable safer, faster, and cheaper improvement of the products people already use—mostly behind the scenes—with a growing class of advisory copilots that help professionals make better choices rather than enabling DIY wet-lab work. (Nature)

The core design objective we will use throughout is a single, Unicode-ready line:

“x* = argmaxₓ U(x) − λ·R(x) s.t. Γ(x) ≤ 0.” (1.1)

Here U(x) collects product-relevant benefits (performance, sustainability), R(x) captures model/decision risk (uncertainty, distribution shift), and Γ(x) ≤ 0 encodes hard safety, regulatory, and feasibility constraints. This “structure-before-search” lens explains why AlphaFold-era biology and graph-based materials models transfer so well to everyday pipelines: we lock in invariants first (e.g., known-safe ingredient classes; thermodynamic stability near the convex hull), then explore efficiently within those bounds. In the sections that follow, we map concrete outputs—pLDDT/pTM for structure reliability, docking success@k, ΔE_hull for materials stability—into reproducible, auditable gates for real-world decisions. (Nature)

Scope preview (public sources): proteins & complexes (AlphaFold 3; RoseTTAFold; ESMFold), docking (DiffDock), protein design (RFdiffusion; ProteinMPNN), materials discovery (GNoME) and self-driving labs (A-Lab). (Nature)

2. Orientation for AI Engineers — What These Models Actually Do

AlphaFold family (AF2 → AF3)

What it does. AF2 predicts single-protein structures with high accuracy; AF3 extends to joint complexes that can include proteins, DNA/RNA, ligands, ions, and chemical modifiers, using a diffusion-style architecture. AF3 is accessible via the AlphaFold Server (free for non-commercial research) and predictions for many proteomes are browseable in AlphaFold DB. (Nature)
Key reliability signals.
“conf_accept ⇐ (pLDDT ≥ 70) ∧ (pTM ≥ 0.7).” (2.1) (illustrative thresholds; tune per assay/target class)
pLDDT (per-residue confidence) and pTM (global topology) are standard summary signals used in downstream triage. (Nature)

Alternatives / complements you can actually run

RoseTTAFold. A three-track network for protein structure prediction, widely used in academia; original peer-reviewed report in Science (2021). (science.org)
ESMFold. Fast single-sequence structure prediction leveraging large protein language models (ESM2); public description and resources via Meta’s ESM Atlas and GitHub. (science.org)

Docking & binding (posing small molecules or partners)

DiffDock. Reframes docking as generative diffusion over ligand pose degrees of freedom; commonly reported with top-k success @ RMSD < 2 Å on PDBBind.
“Succ@k = #(RMSD ≤ 2 Å in top-k) ÷ N.” (2.2)
The original report shows large gains over search-based baselines and provides a calibrated confidence score for ranking. (arXiv)
Metric notes. “Success rate” in docking is conventionally defined via RMSD thresholds between predicted and crystal poses; papers often report SR at 2 Å or 2.5 Å. (PMC)

Sequence & generative design (making or fitting proteins)

RFdiffusion. A diffusion model (fine-tuned from RoseTTAFold) that generates backbones and functional scaffolds; demonstrated on binders, symmetric oligomers, enzymes, and motif scaffolding. (Nature)
ProteinMPNN. Learns sequences that fit a target backbone; strong sequence-recovery and experimental validations across many backbones (Science, 2022; open versions also available). (science.org)
LigandMPNN. Extends sequence design to non-protein context (explicit small molecules, nucleotides, metals) for enzyme/binder design (Nat. Methods, 2025). (Nature)

Materials discovery at scale (the “GNN side” of Structure AI)

GNoME. Scales graph networks to propose ~2.2M crystal structures, with ~381k predicted stable (on/near the convex hull) — an order-of-magnitude jump in candidate stable materials (Nature, 2023).
“ΔE_hull(x) ≤ ε ⇒ predicted-stable material.” (2.3)
Results are actively linked to downstream autonomous synthesis efforts such as A-Lab, which plans and executes inorganic syntheses with robotics and active learning. (Nature)
Context & caveats. Popular-press and commentary pieces summarize the scale and ambitions; as always, synthesis/measurement is the arbiter, and some critiques examine claims around autonomy and novelty — useful reading for risk planning. (TIME)

Where to get data / run things today

AlphaFold DB (precomputed structures), AF3 Server (complex prediction), PDB for experimental truth, PDBBind for docking benchmarks, and Materials Project for convex-hull stability references are standard entry points for reproducible workflows. (alphafold.ebi.ac.uk)

Takeaway for engineers. Treat these models as hypothesis generators with calibrated signals. Use AF-style confidence (pLDDT/pTM), docking success metrics, and ΔE_hull filters as gates before you spend wet-lab or fab-lab cycles — and document every decision threshold you set off the back of those signals. (Nature)

3. From Lab to Laundry: The Everyday Impact Arc

What changes for ordinary people? Structure AI mostly works behind the scenes: it helps suppliers and formulators shortlist better ingredients and materials before anyone touches a bench, then routes promising picks to validated labs. The result is invisible upgrades in detergents, food, packaging, skincare, batteries, and electronics.

3.1 Detergents & surface cleaners — “cold wash” that actually works

Enzyme-enhanced formulations can clean effectively at lower temperatures, cutting energy use without sacrificing hygiene. This is moving from marketing to evidence: reviews note high-performance detergents achieving strong results at reduced temperatures, and industry programs are explicitly targeting enzymes that remain active in cold cycles. Structure AI contributes by (i) predicting protein structures/complexes (AF3) to preserve catalytic geometry in surfactant systems, and (ii) screening sequence variants for stability and activity—so only the best candidates reach pilot runs. (PMC)

3.2 Food processing — smarter enzymes for texture & fermentation

Enzymes already drive fermentation, flavor, and texture; AI accelerates that pipeline. End-to-end reviews describe AI-assisted precision fermentation and broader food enzyme use, where model-proposed variants are triaged before bioreactor time. The practical consumer impact is steadier quality and shelf-life with lower inputs—again, achieved through enterprise workflows rather than DIY. (PMC)

3.3 Packaging — barrier polymers and coatings chosen by prediction

Barrier performance (oxygen and water-vapor transmission) can now be estimated with ML models trained on polymer chemistry and process data, helping teams rank candidates before costly multilayer trials. Recent work shows multi-task learning for gas permeability in polymers and related membrane-design surveys; paired with conventional packaging knowledge (e.g., humidity sensitivity of EVOH or multilayer strategies), this yields faster, safer formulation updates. (Nature)

3.4 Skincare actives — peptide advisory workflows, not DIY claims

Peptides are common cosmetic actives, but safety and efficacy still depend on lab and regulatory context. AF-style modeling and peptide-design AI help triage candidate binders or carriers (e.g., peptide–protein docking and stability heuristics), then documented assays decide. Reviews in cosmetology and peptide AI emphasize potential alongside regulatory caution—precisely where an advisory copilot belongs. (PMC)

3.5 Devices & the built world — batteries, heat spreaders, coatings

On the materials side, graph-network discovery (GNoME) proposes millions of crystals with hundreds of thousands predicted stable, feeding autonomous labs (A-Lab) that plan and execute synthesis. In parallel, ML is being used to accelerate solid-state electrolyte research and to design thermal interface materials and heat-management coatings—ingredients you never see, but that make phones cooler and batteries longer-lasting. (Berkeley Lab News Center)

The “Invisible Upgrades” pipeline (how it actually lands in products)

Procurement. Suppliers pre-screen enzyme or polymer candidates with structure/materials models and historical data; only short-listed options move forward. (PMC)
Formulation. R&D blends and stabilizes candidates, aided by docking/compatibility predictions and permeability/property forecasts; lab assays confirm. (PMC)
QA & scale-up. Autonomous and conventional labs validate performance windows (e.g., cold-wash activity, OTR/MVTR, ionic conductivity) before scale. (Nature)
Sustainability. Lower wash temperatures, less solvent, and better durability improve lifecycle metrics without changing consumer behavior. (PMC)

Bottom line. Structure AI doesn’t hand consumers a lab bench; it quietly improves everyday goods by making upstream choices smarter and more auditable.

4. Use Case Pattern A — Ingredient Screening & Reformulation (Consumer Goods)

Goal. Rapidly refresh a detergent/cleaner, food-process aid, or skincare formulation by ranking already permitted ingredients and close analogs—then hand only the best candidates to wet-lab and regulatory teams.

4.1 End-to-end workflow (reproducible & audit-ready)

Shortlist from known-safe space.
- Pull label-permitted actives/excipients from internal catalogs + UniProt (sequence & function) and vendor safety sheets; map to structures in AlphaFold DB and RCSB PDB where available. (UniProt)
Structure sanity check (AF/ESM).
- If no experimental structure: use AF2/ESMFold; if interaction context matters, use AF3 (complexes with DNA/RNA/ligands/ions).
- Gate by standard confidence signals (illustrative):
  “conf_accept ⇐ (pLDDT ≥ 70) ∧ (pTM ≥ 0.7).” (2.1)
- Store model/version, seeds, and AF DB IDs for traceability. (PubMed)
Interaction triage (docking as hypothesis, not proof).
- Generate binding hypotheses to known targets (e.g., stain components, skin receptors) using DiffDock; report top-k success @ RMSD < 2 Å on PDBBind-like panels to calibrate settings for your chemical space.
  “Succ@k = #(RMSD ≤ 2 Å in top-k) ÷ N.” (2.2)
- Keep calibrated confidence scores; do not convert docking scores to efficacy claims. (arXiv)
Batch selection + uncertainty.
- Rank candidates with risk-adjusted scoring and diversity (see §9):
  “Ŝ(x) = E[S(x)] − κ·σ̂[S(x)].” (2.1′)
  “Bₜ = argmax_{|B|=m} ∑_{x∈Cand} ( Ŝ(x) + μ·Div(x|B) ).” (5.2)
Wet-lab confirmation.
- Send only candidates passing domain gates for assays (activity window, stability, compatibility). Record failures to refine thresholds.
Regulatory dossier update.
- Version the formulation record with links to PDB IDs, AF DB entries, vendor CoAs, and assay reports; keep a one-page rationale tracing every gate and threshold used.

4.2 Guardrails (non-negotiable)

AF signals are filters, not product claims. pLDDT/pTM indicate model confidence, not real-world performance; use them to exclude low-confidence structures, never to assert effects. (PubMed)
Docking is hypothesis generation. Report Succ@k and confidence; require laboratory verification before any consumer-facing change. (arXiv)
Stay inside known-safe classes. This pattern is for reformulation and ingredient ranking among permitted classes, not novel actives.

4.3 Datasets & infrastructure you can cite/use today

AlphaFold DB (EMBL-EBI): 200M+ predicted structures; use accession IDs in your records. (alphafold.ebi.ac.uk)
UniProt / UniProtKB: canonical sequences, functions, cross-refs (PDB, EC, GO); BLAST for homology checks. (UniProt)
RCSB PDB: experimental structures for truth checking and target prep. (rcsb.org)
PDBBind: standardized protein–ligand complexes with affinities; useful for docking parameter calibration and reporting. (OUP Academic)
AlphaFold 3 (Nature, 2024) & AF Server: for complex predictions when ligands/ions/nucleic acids matter. (Nature)
DiffDock (arXiv, 2022): diffusion-based docking with calibrated confidence. (arXiv)

4.4 Minimal acceptance rule (drop-in for SOP)

“Accept_bio(x) ⇐ [conf_accept in (2.1)] ∧ [P_bind(x) ≥ τ_bind] ∧ [assay_pass = True] ∧ [regulatory_class ∈ Allowed].” (3.6′)

Practical tip. Keep a single CSV per campaign: {candidate_id, data_ids (UniProt/PDB/AFDB), model_versions, thresholds_used, Succ@k, lab_result, decision, reviewer} — plus a checksum for the whole run. This makes procurement → formulation → QA → sustainability fully auditable, while staying within public, regulator-friendly practice.

5. Use Case Pattern B — Food & Nutrition Operations (Factory-Side)

Goal. Speed up factory-side enzyme refresh cycles (fermentation, stabilization, and QA) using a design→fit→dock→stability→pilot loop, while keeping claims grounded in assays and regulatory scope.

5.1 Factory pipeline (from ideation to pilot fermenters)

Enzyme candidate ideation (backbones).
Use RFdiffusion to generate backbone scaffolds for desired functions (e.g., specific cleavage or transglycosylation motifs). RFdiffusion fine-tunes a RoseTTAFold network to denoise structures and has been validated across binders, symmetric oligomers, enzymes, and motif scaffolds. (Nature)
Sequence fit to scaffolds.
For each backbone, fit sequences with ProteinMPNN (high sequence-recovery and broad experimental validation), producing a ranked panel of variants for each target activity window. (science.org)
Docking screens for plausibility (hypothesis stage).
Run DiffDock against known substrates/inhibitors to prioritize poses; report top-k success @ RMSD ≤ 2 Å on an internal or PDBBind-like panel to calibrate parameters for your chemistry.
“Succ@k = #(RMSD ≤ 2 Å in top-k) ÷ N.” (2.2)
DiffDock’s diffusion framing improves blind-docking success and provides a calibrated confidence score—use it as a ranking signal, not evidence of efficacy. (arXiv)
Stability screens (ΔΔG proxies).
Apply established ΔΔG prediction tools (multiple methods recommended) to flag destabilizing mutations before expression; recent evaluations benchmark dozens of stability predictors and highlight strengths/limits.
“ΔΔG_pred(x) ≤ τ_stab ⇒ passes stability screen.” (3.4) (PMC)
Scale to pilot fermenters.
Move only gate-passing variants to precision fermentation pilots for activity vs. temperature/pH, yield, and downstream processing; the sector has mature playbooks from 1990s chymosin to today’s high-value ingredients. (The Good Food Institute)

Minimal gate (drop-in):
“Accept_food(x) ⇐ [RFdiffusion_backbone = ok] ∧ [ProteinMPNN fit ≥ τ_fit] ∧ [Succ@k ≥ τ_dock] ∧ [ΔΔG_pred(x) ≤ τ_stab] ∧ [pilot_assay_pass = True].” (5.1)

5.2 KPIs you can measure and report

Temperature profile (factory relevance).
“T₅₀(x) = argmin_T { activity(T) ≥ 0.5·activity_max }.” (5.2)
Track cold-active performance for energy savings and thermal tolerance for process stability; tie to actual pilot assay curves.
Shelf life uplift (finished goods).
“SL_gain = SL_new − SL_base (days or %).” (5.3)
Use predictive microbiology models alongside real-time challenge tests to quantify effects on spoilage kinetics under target storage conditions. (PMC)
Contamination resistance / robustness.
“ΔlogCFU = log₁₀(CFU_end) − log₁₀(CFU_start) under SOP challenge.” (5.4)
Combine hazard-analysis (HACCP) controls with model-informed scenario testing from the predictive microbiology literature. (PMC)
Yield & cost.
“Yield_gL, titer_gL·h⁻¹, downstream_recovery_%.” (5.5)

5.3 Documentation & governance (versioned, regulator-friendly)

Model cards + assay bundles.
For every RFdiffusion/ProteinMPNN/DiffDock run, ship a model card (intended use, data/IDs, metrics, limitations) co-packaged with pilot assay PDFs and batch records. Model cards are a well-accepted transparency pattern you can cite. (dl.acm.org)
Provenance snapshot.
Record external IDs (UniProt, PDB/PDBBind), software versions, seeds, and thresholds; keep a signed checksum for the candidate table so procurement→QA is auditable.

Bottom line. This loop turns enzyme refresh from broad trial-and-error into constrained, auditable design. Generative backbones (RFdiffusion) + sequence fit (ProteinMPNN) + docking plausibility (DiffDock) + ΔΔG filters focus fermenter time on the most promising variants—while shelf-life and contamination KPIs ensure factory-side value is measured in days saved, energy reduced, and batches kept in spec. (Nature)

6. Use Case Pattern C — Skincare & Topicals (Advisory-Only)

Scope. Provide advice and triage for over-the-counter cosmetics using only label-existing (INCI) ingredients with documented safety. No novel-active design for consumers; no therapeutic claims. Complex hypotheses (e.g., peptide–receptor binding) are for internal plausibility checks that always route to qualified assessors and validated assays. (Public Health)

6.1 Advisory workflow (what to do, step by step)

Define the allowed space.
- Pull ingredients from CosIng and brand INCI lists; check any restrictions and conditions of use; attach CIR summaries if available. (single-market-economy.ec.europa.eu)
Generate hypotheses, not claims.
- Use AlphaFold 3 to propose peptide–receptor complex geometries; gate by model confidence (illustrative):
  “conf_accept ⇐ (pLDDT ≥ 70) ∧ (pTM ≥ 0.7).” (6.1) (Nature)
- Optionally dock with a calibrated workflow for ranking only (e.g., DiffDock); never equate pose scores with efficacy.
Safety triage & exposure math.
- Compute Margin of Safety per SCCS Notes of Guidance:
  “MoS = PoD_sys ÷ SED.” (6.2)
  Use a reference MoS ≥ 100 when applicable (10× inter-species × 10× intra-species).
- For irritation/sensitisation screens, rely on OECD RhE and sensitisation NAMs in the dossier (e.g., TG 431/439/442D). (OECD)
Claim and label check (cosmetic vs drug).
- Ensure claims comply with EU 655/2013 Common Criteria (legal, truthful, evidence-based, honest, fair, clear) and do not cross into therapeutic territory; align with FDA’s “intended use” test. (eur-lex.europa.eu)
Package for the Responsible Person / assessor.
- Update the Product Information File (PIF) with ingredient IDs, model versions, MoS calculation, NAM results, and human data (if any), per EU 1223/2009. (Public Health)

6.2 Guardrails (non-negotiable)

Advisory-only. Output is ranked options and caveats for permitted INCI ingredients—not new actives or DIY protocols. (Public Health)
Models are filters, not evidence. AF3 complexes and docking poses motivate which literature/assays to run; they do not substantiate performance claims. (Nature)
NAMs first. Irritation/corrosion/sensitisation rely on validated OECD methods included in the safety dossier (RhE TG 431/439; sensitisation TG 442D). (OECD)
Claims stay cosmetic. Screen draft copy against EU 655/2013 and FDA’s cosmetic/drug boundary. (eur-lex.europa.eu)

6.3 Minimal acceptance rule (drop-in to SOP)

“Accept_skin(x) ⇐ [conf_accept in (6.1)] ∧ [INCI_verified(x) = True] ∧ [MoS(x) ≥ 100] ∧ [OECD_431/439/442D = pass] ∧ [Claim ∈ Cosmetic (EU 655/2013) ∧ not Drug (FDA)] ∧ [assay_pass = True].” (6.3)

Why this matters. AF3 helps prioritize peptide candidates within already-approved cosmetic classes; SCCS exposure math and OECD NAMs keep safety evaluation on standard rails; EU/FDA rules fence the claims. That’s how you keep peptide-forward skincare strictly advisory-only and regulator-ready. (Nature)

7. Use Case Pattern D — Batteries, Chips & the Built World (Materials)

Goal. Turn vast candidate spaces into make-and-measure loops: generate structures with GNoME, filter by thermodynamic stability and synthesizability, hand off to an autonomous lab (A-Lab) or a vendor for execution, and feed results back for active learning. (Nature)

7.1 Workflow (generation → filters → autonomous synthesis → feedback)

Candidate generation. Use GNoME to propose crystals at scale (≈2.2 M; ≈381k predicted stable on/near the convex hull). (Nature)
Thermodynamic screen. Compute distance to convex hull with Materials Project conventions; keep near-hull materials.
“ΔE_hull(x) = E_form(x) − E_hull(phase diagram).” (7.1) Keep if “ΔE_hull(x) ≤ ε.” (7.2) (docs.materialsproject.org)
Synthesizability & feasibility. Score route availability, precursor abundance, and hazards (see 7.2).
Autonomous synthesis. Ship shortlisted targets to A-Lab or a vendor: plan recipes from literature/computation, execute with robotics, use active learning to refine. (Nature)
Active learning loop. Push measured outcomes back to re-rank candidates and retrain property predictors.

7.2 Practical gates (drop-in)

Stability gate (thermo).
“G_stab(x) = 𝟙{ ΔE_hull(x) ≤ ε }.” (7.3) (ε often 0–0.05 eV/atom, adjust per domain.) (docs.materialsproject.org)
Property fit (task-specific).
Example (solid-state electrolyte):
“Req_SSE(x) ⇐ [σ_ion(x) ≥ σ_min] ∧ [E_window(x) ≥ V_req] ∧ [T_decomp(x) ≥ T_req].” (7.4) (sciencedirect.com)
Synthesizability proxy (route/abundance/hazard).
“S_syn(x) = sigm( u·score_route(x) + v·abundance(x) − w·hazard(x) ).” (7.5)
Decision rule (materials).
“Accept_mat(x) ⇐ G_stab(x) ∧ Req_task(x) ∧ [S_syn(x) ≥ τ_syn].” (7.6)

7.3 KPIs to report (battery, thermal, cost, ESG)

Ionic conduction window (batteries).
“σ_ion(x) @ T_use, E_window(x) = V_ox − V_red vs Li/Li⁺.” (7.7) (Target ranges from SSE reviews; report along with interfacial notes.) (sciencedirect.com)
Thermal stability / heat management (devices).
“T_decomp, k_through-plane(x), CTE_mismatch(x, stack).” (7.8) (TIMs and coatings: emphasize through-plane k and reliability under cycling.) (onlinelibrary.wiley.com)
Abundance & cost.
“Cost̂(x) = ∑ᵢ μᵢ·price(elemᵢ) + process_cost(x).” (7.9)
ESG footprint.
“GHG_intensity(x) ≤ τ_GHG, Criticality_index(x) ≤ τ_crit.” (7.10)

Reporting tip. Pair every accepted candidate with its Materials Project ID, ΔE_hull, predicted properties, synthesis plan version, and measured results (mean ± SD). Use the A-Lab run IDs to trace experiments end-to-end. (next-gen.materialsproject.org)

7.4 Why this works now

Scale on the front end. GNoME expands the stable-materials frontier by nearly an order of magnitude, making downstream filters essential. (Nature)
Throughput on the back end. A-Lab closes the compute-→-synthesis gap with robotics and active learning; results roll back into ranking. (Nature)

Bottom line. Treat ΔE_hull as the first gate, add task-specific property thresholds (e.g., σ_ion and stability window for SSEs), enforce a synthesizability score, and let autonomous labs iterate. That pipeline reliably converts millions of predictions into a small, testable set of materials that matter for batteries, chips, and coatings. (docs.materialsproject.org)

8. Data & Model Access — What to Use Where

Biology (structures, truth, annotations)

AlphaFold DB (predictions). Use for fast coverage and confidence scores (pLDDT, pTM). Record AFDB accession IDs in your tables; snapshot the database date in your methods. (alphafold.ebi.ac.uk)
RCSB PDB (experimental truth). Fetch crystal/NMR/cryo-EM structures for ground truth, target prep, and benchmarking. Always log the PDB ID and revision. (rcsb.org)
UniProt / UniProtKB (annotations). Canonical sequences, isoforms, function, cross-refs (PDB, GO, EC). Store the UniProt accession and release version used. (UniProt)

Tip. When an AFDB model exists but a PDB structure does too, prefer PDB for benchmarking and use AFDB as a supplemental candidate generator with explicit confidence filters.

Docking benchmark (pose & affinity panels)

PDBBind. Standardized protein–ligand complexes with curated binding data; cite the release year/split, and report success@k with RMSD thresholds. (OUP Academic)

Materials (energetics, convex hull, IDs)

Materials Project. Pull formation energies and compute distance to the convex hull for stability screening; always keep the MP material ID and API snapshot date. (docs.materialsproject.org)

Open models you can run today

ESMFold (fast single-sequence). Good for rapid triage when you need many structures quickly; browse large-scale predictions in the ESM Metagenomic Atlas. (esmatlas.com)
RoseTTAFold (academic stack). Peer-reviewed three-track model; use base RF for structure and RF-family variants for design tasks in academic pipelines. (science.org)
Open Catalyst Project (surface chemistry). Datasets and code for adsorbate–catalyst systems; include dataset commit/tag when reporting results. (GitHub)

Minimal reproducibility checklist (drop-in)

IDs. AFDB accession, PDB ID, UniProt accession, MP material ID.
Versions. Database release/snapshot date; model commit/tag; software environment.
Provenance hash. Keep a run-level checksum of inputs, params, code refs, and outputs (ties back to §9 pipeline).
Metrics to report. pLDDT/pTM distributions (biology), Succ@k with RMSD thresholds (docking), ΔE_hull and property deltas (materials).

This combination gives engineers a clear “where to get it / how to cite it” map that regulators and reviewers recognize.

9. Engineering the Workflow — A Reproducible, Auditable Pipeline

This section turns Sections 2–8 into a step-by-step SOP you can actually run. It uses only public tools and standards (AF2/AF3, RoseTTAFold, ESMFold, DiffDock, RFdiffusion, ProteinMPNN, Materials Project), and it anchors documentation to Model Cards and the NIST AI Risk Management Framework so QA/Regulatory can audit every decision. (Nature)

9.1 Planning — define invariants before search

Scope & guardrails (write these first):

Allowed design space. e.g., “only INCI-listed actives,” “only vendor-approved excipients,” or “only elements without supply-risk flags.”
Truth sets. For biology, use RCSB PDB as ground truth; for materials, use Materials Project formation energies/convex hull.
Reporting units & IDs. AFDB accession, UniProt accession, PDB ID, MP material ID, dataset snapshot dates. (rcsb.org)

Batch record fields (mandatory): {candidate_id, data_IDs, model_name@version, parameters, thresholds, seeds, metrics, decision, reviewer} — plus a provenance hash (see 9.6).

9.2 Stage 1 — Structure (predict or retrieve)

What to run (choose the lightest tool that answers the question):

AF2 / ESMFold for single proteins; AF3 when complexes (protein–protein, protein–DNA/RNA, ligands, ions) matter. RoseTTAFold remains a well-cited open academic option. (alphafoldserver.com)

Gate:
“conf_accept ⇐ (pLDDT ≥ 70) ∧ (pTM ≥ 0.7).” (9.0) (illustrative; tune by target class)

Recording: Store PDB IDs when available; otherwise store AFDB/ESM Atlas IDs and the AlphaFold Server job reference for AF3 runs. (alphafoldserver.com)

9.3 Stage 2 — Interact (generate binding/pose hypotheses)

Use DiffDock to rank plausible poses; report top-k success @ RMSD ≤ 2 Å on a PDBBind-style calibration panel to choose k and thresholds for your chemistry.
“Succ@k = #(RMSD ≤ 2 Å in top-k) ÷ N.” (9.1a) (arXiv)

For triage to the lab, use a simple logistic rule that mixes pose score, model confidence, and strain penalties:
“P(bind) ≈ σ(α·score + β·conf − γ·strain).” (9.1) (arXiv)

Important: Docking is hypothesis generation, not evidence. Keep the score, confidence, and calibration plot in the run folder; never convert a docking score into a product claim. (arXiv)

9.4 Stage 3 — Design (make candidates on purpose)

Proteins (food, cleaners, binders):

Backbones: RFdiffusion for scaffolds, symmetry, binder motifs.
Sequences: ProteinMPNN to fit sequences to backbones; optionally add LigandMPNN when non-protein context (metals, ligands) is essential. (Nature)

Materials (batteries, coatings):

Crystals: GNoME proposals → Materials Project energy evaluations → ΔE_hull screen.
“ΔE_hull(x) = E_form(x) − E_hull(phase diagram).” (9.3) Keep if “ΔE_hull(x) ≤ ε.” (9.3′) (Nature)

9.5 Stage 4 — Decide (thresholds + human-in-the-loop)

Make domain gates explicit, then combine with a global accept rule.

Biology gate (example):
“Accept_bio(x) ⇐ [conf_accept in (9.0)] ∧ [P(bind) ≥ τ_bind] ∧ [ΔΔG_pred(x) ≤ τ_stab] ∧ [toxicity_screen = pass].” (9.4)

Materials gate (example):
“Accept_mat(x) ⇐ [ΔE_hull(x) ≤ ε] ∧ [Req_task(x) = pass] ∧ [S_syn(x) ≥ τ_syn].” (9.5)

Global decision rule:
“Accept ⇐ [Γ_hard(x) ≤ 0] ∧ [Γ_soft(x) ≤ δ] ∧ [Gate_domain(x) = True] ∧ [Reviewer_signoff = True].” (9.2)

Human-in-the-loop: Any metric < τ or any rule violation ⇒ escalate with a comment; do not silently override gates.

9.6 Stage 5 — Validate (wet-lab or autonomous lab) and log

Wet-lab: send only Accepted candidates to assays; report mean ± SD and SOP identifiers.
Autonomous lab (materials): hand off to A-Lab or vendor; record synthesis plan, robot run IDs, and results; push measurements back for active learning. (Nature)

Provenance hash (store with every candidate):
“h(x) = H( data_IDs ⊕ model_IDs ⊕ seeds ⊕ params ⊕ code_refs ⊕ outputs ).” (9.6)

9.7 Minimal file/folder convention (per campaign)

/campaign_YYMMDD/
  planning.yaml           # scope, invariants, allowed space
  data_snapshot.txt       # AFDB/PDB/MP release IDs and dates
  stage1_structure.csv    # pLDDT, pTM, IDs, conf_accept (9.0)
  stage2_interact.csv     # docking scores, conf, Succ@k (9.1a)
  stage3_design.csv       # RFdiffusion/ProteinMPNN or GNoME/ΔE_hull
  decisions.csv           # 9.4/9.5 gates + global rule (9.2)
  assays/                 # lab PDFs, run IDs, mean±SD
  model_cards/            # one per model run (see 9.8)
  audit/                  # NIST AI RMF trace, risk notes

9.8 Documentation & governance (what auditors expect)

Model Cards for each model/run: intended use, data sources/IDs, metrics (pLDDT/pTM dists; Succ@k; ΔE_hull), limits, ethical/use constraints. Cite the Model Cards paper in your methods. (arXiv)
NIST AI RMF 1.0 alignment: keep a short checklist in /audit/ mapping risks to mitigations (e.g., distribution shift → ensemble uncertainty; bio claims → advisory-only; materials novelty → ΔE_hull + synthesis confirmation). Include the RMF reference. (nvlpubs.nist.gov)

9.9 Tuning notes (what typically needs calibration)

AF/ESM thresholds: raise pLDDT/pTM if your domain is sensitive to topology; lower for early, high-throughput filters. (Nature)
Docking: choose k and RMSD threshold by reproducing literature success rates on your PDBBind split; store the calibration figure. (arXiv)
Materials: pick ε for ΔE_hull by comparing against MP stability labels in your chemistry (0–50 meV/atom are common ranges in screening papers). (docs.materialsproject.org)

9.10 One-page “Decision Math” (drop-in for Methods)

“P(bind) ≈ σ(α·score + β·conf − γ·strain).” (9.1)
“Accept if [metric ≥ τ] ∧ [no rule violations]; else escalate.” (9.2)
“ΔE_hull(x) = E_form(x) − E_hull(phase diagram); accept if ΔE_hull(x) ≤ ε.” (9.3–9.3′)
“Accept_bio(x) and Accept_mat(x) as in (9.4–9.5).” (9.4–9.5)
“h(x) = H( data_IDs ⊕ model_IDs ⊕ seeds ⊕ params ⊕ code_refs ⊕ outputs ).” (9.6)

Why this pipeline is credible

Each stage sits on peer-reviewed, public foundations: AF2/AF3 for structures, DiffDock for poses, RFdiffusion/ProteinMPNN for design, GNoME+Materials Project for materials, and NIST/Model Cards for governance and documentation. If you keep IDs, thresholds, and hashes as shown, a reviewer can replay your decisions end-to-end. (Nature)

10. Metrics that Matter — Model-Side and Product-Side

This section fixes what to compute, report, and threshold so results are comparable across teams and audits.

10.1 Biology (structure, interaction, hit quality)

Structure confidence (report distributions, not only means).
“pLDDTᵢ ∈ [0,100], pTM ∈ [0,1]. Define conf_accept ⇐ (pLDDT̄ ≥ τ_LDDT) ∧ (pTM ≥ τ_TM).” (10.1)
Source of signals: pLDDT (per-residue) and pTM (global topology) are standard outputs of AlphaFold-style models. Report histograms/percentiles per domain and per complex. (Nature)

Docking pose quality (calibrated on a PDBBind-style split).
“Succ@k = #(RMSD ≤ 2 Å in top-k) ÷ N; RMSD_ref = crystal pose.” (10.2)
Also log model confidence if available (e.g., DiffDock). Report Succ@{1,5,10} on your calibration panel before using scores in triage. (arXiv)

Hit-rate uplift vs. baseline screens.
“Uplift = (Hits_AI ÷ Tests_AI) − (Hits_base ÷ Tests_base).” (10.3)
Here “hit” must be a pre-declared assay threshold (activity, EC₅₀, etc.). Keep 95% CIs by bootstrap.

Logistic triage probability for lab picks (record coefficients).
“P(bind) ≈ σ(α·score + β·conf − γ·strain).” (10.4)
Publish α,β,γ learned on your calibration set and freeze them per campaign. (arXiv)

10.2 Materials (stability, synthesizability, property deltas)

Thermodynamic stability (energy above hull).
“ΔE_hull(x) = E_form(x) − E_hull(phase diagram); keep if ΔE_hull(x) ≤ ε.” (10.5)
Report the distribution of ΔE_hull for accepted vs. rejected candidates and the chosen ε (e.g., 0–50 meV/atom by domain). (docs.materialsproject.org)

Synthesizability proxy (make an explicit score).
“S_syn(x) = σ( u·score_route(x) + v·abundance(x) − w·hazard(x) ). Accept if S_syn ≥ τ_syn.” (10.6)
Back this with a literature-anchored classifier or regression trained on made vs. not-made labels (e.g., post-2019 corpora); publish AUROC/PR on a held-out set. (Nature)

Task property fit (report deltas to incumbent).
“ΔProp(x) = p̂(x) − p_incumbent; pass if |ΔProp(x)| ≥ δ_min and sign is favorable.” (10.7)
Examples: ionic conductivity at T_use, electrochemical window (V), thermal conductivity through-plane (W·m⁻¹·K⁻¹), CTE mismatch (ppm/K).

Close-the-loop metric (autonomous synthesis).
“Yield_AL = successful_synth ÷ attempts; Δt_loop = t(measure) − t(propose).” (10.8)
Use A-Lab (or vendor) run IDs to tie predictions to attempts and outcomes. (Nature)

10.3 Business & sustainability (program-level)

Time & cost efficiency.
“Time-to-candidate = t(accepted_for_assay) − t(start).” (10.9)
“Assay_cost_per_hit = total_assay_cost ÷ #hits_confirmed.” (10.10)

Portfolio focus.
“%Budget_to_top_quartile = spend on Q4(score) ÷ total spend.” (10.11)

Carbon & energy.
“ΔCO₂eq_per_unit = CO₂eq(new) − CO₂eq(baseline).” (10.12)
“ΔkWh_per_use = kWh(new workflow) − kWh(baseline).” (10.13)
Tie to LCA or metered factory data (e.g., cold-wash cycles, low-temp curing).

10.4 Reporting template (one table per campaign)

Biology: {target, PDB/AFDB/UniProt IDs, pLDDT̄, pTM, Succ@k (2 Å), P(bind) in (10.4), assay_hit (Y/N)}. (Nature)
Materials: {MP ID, ΔE_hull (10.5), S_syn (10.6), ΔProp (10.7), A-Lab run, result}. (docs.materialsproject.org)
Program: {Time-to-candidate (10.9), Assay_cost_per_hit (10.10), ΔCO₂eq_per_unit (10.12)}.

Why these metrics? pLDDT/pTM and Succ@k are peer-standard reliability and pose measures; ΔE_hull is the canonical stability filter in the Materials Project ecosystem; synthesizability scores and autonomous-lab yields connect prediction to real make-and-measure outcomes. Together with time/cost/CO₂eq, they let engineering, QA, and sustainability read the same scoreboard. (Nature)

11. Validation, Compliance, and Risk Management (Non-Negotiable)

Frameworks to anchor your SOP. Use NIST AI RMF 1.0 (four functions: Govern, Map, Measure, Manage) as the lifecycle backbone, and align any health-adjacent workflow with WHO guidance (2021 ethics principles and 2024–2025 LMM governance). These are public, regulator-recognized baselines. (nvlpubs.nist.gov)

11.1 Map NIST AI RMF → your Structure-AI pipeline

GOVERN — policies, roles, and red lines

Define accountable owners (engineering, QA, Responsible Person) and prohibited uses (e.g., consumer DIY bio, unapproved claims). Keep a written risk appetite and escalation paths. (nvlpubs.nist.gov)
Maintain Model Cards per model/run (intended use, data IDs, limits, group performance, update policy). Store alongside assay PDFs. (arXiv)

MAP — context & data constraints

Document problem context, affected users, datasets (AFDB/PDB/UniProt/MP) with snapshot dates; list foreseeable impacts if predictions are wrong. (nvlpubs.nist.gov)

MEASURE — uncertainty, performance, drift

Track pLDDT/pTM distributions (biology), Succ@k@2 Å (docking), ΔE_hull (materials), synthesizability score AUROC/PR, and calibration plots; add drift monitors on inputs. (nvlpubs.nist.gov)

MANAGE — controls, incidents, continuous improvement

Enforce gates (Section 9) and change management for thresholds; keep an incident playbook (mislabel, assay failure, claim breach) with rollback steps. Use the NIST Playbook as a living checklist. (NIST)

11.2 Health-adjacent: follow WHO rules (advisory, not diagnostic)

No diagnosis or treatment recommendations from non-regulated tools; route anything clinical to licensed providers and regulated pathways. WHO’s 2021 ethics report (six principles) and 2024/2025 LMM guidance set explicit expectations for safety, equity, transparency, and human oversight. (World Health Organization)
Keep claim boundaries clear in UI and documentation; record referral/deflection logic to clinical channels. (World Health Organization)

Policy snippet (drop-in):
“Health_mode ⇐ AdvisoryOnly ∧ NoDiagnosis ∧ RegulatedReferral.” (11.1)

11.3 Mandatory disclaimers (product & docs)

Hypothesis, not proof. “These models propose hypotheses; experiments decide.” Put in UI, spec sheets, and model cards. (nvlpubs.nist.gov)
Scope restriction. “This tool analyzes label-existing ingredients/materials only; novel actives route to regulated partners.”
Clinical boundary. “Not for diagnosis or treatment. Seek professional advice.” (World Health Organization)

Release rule (tie to Section 9):
“Release ⇐ [All domain gates pass] ∧ [Reviewer_signoff = True] ∧ [Disclaimers_present = True].” (11.2)

11.4 Evidence pack (what auditors expect to see)

Governance: risk register, roles, red-lines, exception log (NIST RMF “Govern”). (nvlpubs.nist.gov)
Data & IDs: AFDB/PDB/UniProt/MP IDs + snapshot dates; reproducible seeds; environment lockfile.
Metrics: pLDDT/pTM histograms; Succ@{1,5,10}@2 Å calibration; ΔE_hull distribution; synthesizability AUROC/PR; assay mean±SD.
Model Cards: one per model/run; update cadence and deprecation policy. (arXiv)
WHO alignment (health-adjacent): statement of non-diagnostic use, human oversight steps, equity/access notes for LMMs. (World Health Organization)

11.5 Risk controls you must implement

Human-in-the-loop gates. Any metric < τ or rule violation ⇒ escalate; write the rationale into decisions.csv. (nvlpubs.nist.gov)
Calibration before use. Reproduce literature-level docking performance on your PDBBind split and freeze α,β,γ in “P(bind)” (9.1). (NIST)
Change management. No threshold edits without JIRA-style ticket + reviewer signoff; re-run calibration if chemistry or assay changes.
Security & privacy. Segregate proprietary recipes and supplier data; log access; rotate keys. (NIST RMF “Manage”). (nvlpubs.nist.gov)

11.6 Incident triggers and responses (pre-declare them)

Trigger examples: assay reproducibility failure; regression in Succ@k or ΔE_hull pass-rate; claim language drifting beyond cosmetic/consumer boundaries.
Response: immediate rollback to last validated release; stakeholder notice; retraining plan; added guardrail or disclaimer update per NIST “Manage”. (nvlpubs.nist.gov)

11.7 Minimal math (for your Methods)

“Policy_pass ⇐ NIST{Govern,Map,Measure,Manage} ∧ WHO_Health(AdvisoryOnly).” (11.3)
“Ship ⇐ Release in (11.2) ∧ IncidentRate ≤ τ_incident over Δt.” (11.4)

Why this matters. NIST AI RMF 1.0 gives you an audit-ready lifecycle; WHO guidance draws a firm line for health-adjacent work. If you keep these artifacts and rules with every run, reviewers can replay your decisions end-to-end and see that hypotheses were tested before claims. (nvlpubs.nist.gov)

12. Case Study Templates — Fill-in-the-Blanks Blueprints

Below are three drop-in templates you can copy into a repo and fill. Each one ends with: (i) acceptance math, (ii) reporting tables, and (iii) a provenance hash so audits can replay decisions. Equations follow the same Unicode, single-line style.

Template A — Consumer Enzyme Refresh (Detergent / Food-Process Aid)

Objective (paste one sentence):
Improve ______ by ≥ ____% without changing restricted INCI classes or supplier constraints.

Datasets & IDs:

Sequences/annotations: UniProt IDs: ________
Experimental structures: PDB IDs: ________ (if none, note “N/A”)
Predicted structures: AlphaFold DB accessions: ________
Docking calibration: PDBBind split/version: ________

Models & versions:

AF2 / ESMFold (commit/tag): ________
AF3 for complexes (server job ref): ________
DiffDock (commit/tag): ________
RFdiffusion (commit/tag): ________
ProteinMPNN (commit/tag): ________

Filters & gates (fill thresholds):
“conf_accept ⇐ (pLDDT ≥ τ_LDDT) ∧ (pTM ≥ τ_TM). τ_LDDT = ____, τ_TM = ____.” (12.1)
“Succ@k = #(RMSD ≤ 2 Å in top-k) ÷ N; choose k = ____; pass if Succ@k ≥ ____.” (12.2)
“ΔΔG_pred(x) ≤ τ_stab ⇒ stability pass; τ_stab = ____ (kcal·mol⁻¹).” (12.3)
“Accept_bio(x) ⇐ [conf_accept] ∧ [P(bind) ≥ τ_bind] ∧ [ΔΔG_pred ≤ τ_stab] ∧ [tox = pass]. τ_bind = ____.” (12.4)

Assays (declare before running):

Primary activity assay: SOP ________; acceptance ≥ ________.
Stability/compatibility: SOP ________; acceptance ≥ ________.
Shelf-life proxy or challenge test (if applicable): SOP ________.

Release notes (one page):

What changed, why, expected benefit, any label/INCI implications, and sustainability note.

Reporting tables (CSV headings):
candidate_id, uniprot_id, pdb_id/afdb_id, pLDDT, pTM, dock_score, dock_conf, Succ@k_calib, dG_stability, assay_result, decision, reviewer, timestamp

Acceptance math (paste into Methods):
“Ŝ(x) = E[S(x)] − κ·σ̂[S(x)]. Rank by Ŝ; choose batch B maximizing ∑(Ŝ + μ·Div).” (12.5)
“Accept(x) ⇐ [Γ_hard ≤ 0] ∧ [Γ_soft ≤ δ] ∧ [Accept_bio(x)] ∧ [Reviewer_signoff].” (12.6)

Provenance hash:
“h(x) = H(data_IDs ⊕ model_IDs ⊕ seeds ⊕ params ⊕ code_refs ⊕ outputs).” (12.7)

Template B — Battery Electrolyte Explorer (Solid-State / Liquid Window)

Objective:
Identify electrolyte candidates with σ_ion ≥ ____ mS·cm⁻¹ at T_use = ____ °C and electrochemical window ≥ ____ V, manufacturable at cost ≤ ____.

Candidate generation:

GNoME shortlist IDs: ________
Optional human seed structures: ________

Energetics & filters:

Pull formation energies from Materials Project; compute distance to convex hull.
“ΔE_hull(x) = E_form(x) − E_hull(phase diagram). Keep if ΔE_hull ≤ ε; ε = ____ eV·atom⁻¹.” (12.8)

Synthesizability heuristic (declare features):
“S_syn(x) = σ(u·route_score + v·abundance − w·hazard). Accept if S_syn ≥ τ_syn; τ_syn = ____.” (12.9)

Autonomous lab plan (A-Lab or vendor):

Recipe source(s): literature IDs ________ / computed hints ________
Robot run IDs & iterations: ________
Active learning policy (brief): ________

Property tests (declare SOPs):

Ionic conductivity σ_ion @ T_use: SOP ________; pass ≥ ____.
Electrochemical window E_window: SOP ________; pass ≥ ____.
Thermal stability T_decomp: SOP ________; pass ≥ ____.
Interface notes (e.g., Li|electrolyte compatibility): SOP ________.

KPIs to log:
“Yield_AL = successful_synth ÷ attempts; Δt_loop = t(measure) − t(propose).” (12.10)
“Cost̂(x) = ∑ᵢ μᵢ·price(elemᵢ) + process_cost(x).” (12.11)
“ESG: GHG_intensity(x) ≤ τ_GHG; Criticality_index(x) ≤ τ_crit.” (12.12)

Decision rule:
“Accept_mat(x) ⇐ [ΔE_hull ≤ ε] ∧ [σ_ion ≥ σ_min] ∧ [E_window ≥ V_req] ∧ [T_decomp ≥ T_req] ∧ [S_syn ≥ τ_syn].” (12.13)

Reporting tables (CSV headings):
mp_id, formula, ΔE_hull_eV_atom, S_syn, σ_ion_mS_cm, E_window_V, T_decomp_°C, run_id, attempts, successes, Yield_AL, decision, reviewer, timestamp

Provenance hash: use (12.7).

Template C — Skincare Peptide Advisory (Label-Existing Only)

Scope & boundary statement (paste verbatim and fill):
Advisory-only tool restricted to INCI-listed peptides/actives. No diagnosis/treatment claims. Decisions route to qualified assessors. Product jurisdiction: ______ (e.g., EU cosmetic).

Allowed space & sources:

INCI list / CosIng entries: ________
CIR / SCCS safety summaries: ________

Complex hypothesis (triage only):

AF3 peptide–receptor complex runs: job refs ________
Confidence gate (illustrative; adjust by target class):
“conf_accept ⇐ (pLDDT ≥ τ_LDDT) ∧ (pTM ≥ τ_TM); τ_LDDT = ____, τ_TM = ____.” (12.14)

Exposure & safety math (declare inputs):

Product use pattern: leave-on / rinse-off, application ____ g/day, body surface ____ cm², frequency ____.
Margin of Safety:
“MoS = PoD_sys ÷ SED. Require MoS ≥ 100 unless otherwise justified.” (12.15)

Non-animal methods (NAMs) to include:

OECD RhE irritation/corrosion (TG 431/439) → pass/fail
Sensitization NAMs (e.g., TG 442D) → pass/fail

Claim compliance:

Check against EU 655/2013 common criteria and local advertising codes.
FDA “intended use” screen if selling in US.

Decision rule:
“Accept_skin(x) ⇐ [INCI_verified] ∧ [conf_accept] ∧ [MoS ≥ 100] ∧ [OECD NAMs = pass] ∧ [Claim ∈ Cosmetic] ∧ [Assessor_signoff].” (12.16)

Reporting tables (CSV headings):
inci_name, af3_job, pLDDT, pTM, literature_refs, MoS, OECD_431, OECD_439, OECD_442D, claim_check, assessor, decision, timestamp

Mandatory disclaimers (paste):
“These models propose hypotheses; experiments decide. Not for diagnosis or treatment.”

Provenance hash: use (12.7).

Shared checklists (paste under each case file)

/planning.yaml

Scope, allowed space, red-lines, jurisdictions, stakeholders, escalation path.

/data_snapshot.txt

AFDB/PDB/UniProt/Materials Project snapshot IDs + dates; PDBBind split/version.

/model_cards/

One per run (intended use, data IDs, metrics, limits, update policy).

/audit/nist_rmf.md

GOVERN/MAP/MEASURE/MANAGE bullets with links to artifacts.

One-line acceptance summary for all templates

“Release ⇐ [All domain gates pass] ∧ [Reviewer_signoff = True] ∧ [Disclaimers_present = True] ∧ [Provenance hash recorded].” (12.17)

These blueprints keep the “structure-before-search” focus while making every decision replayable: thresholds are declared up front, IDs and SOPs are logged, and acceptance math is a single line you can paste into Methods.

13. Compute, MLOps, and Costing

What to plan for. Most spend lands in structure prediction (esp. AF3 for complexes), docking inference (GPU but cheaper per-candidate), and generative design (RFdiffusion backbones; ProteinMPNN is comparatively light). Use ESMFold for fast prefiltering—its single-sequence runtime is reported in seconds on a V100—reserve AF3 for the small set that truly needs complex context. (science.org)

13.1 GPU budget anatomy (where time goes)

Structure (triage → escalate):
ESMFold for breadth (seconds per 300–400 aa on V100); AF2 for higher-fidelity singles; AF3 for complexes (proteins, nucleic acids, ligands, ions). Expect AF3 to dominate GPU time when used. (science.org)
Docking (pose hypotheses):
DiffDock uses diffusion over pose DoF; fast inference with calibrated confidence; report top-k success @ 2 Å on a PDBBind split. (arXiv)
Design (make candidates):
RFdiffusion (backbones) = heaviest of the design stack; ProteinMPNN (sequence fit) = light and fast; both are peer-reviewed. (Nature)
Materials screening:
Use precomputed Materials Project energetics for ΔE_hull gates; GNoME-style generation is large-scale but upstream of most enterprise runs; A-Lab costs are off-GPU (robotics/lab time). (docs.materialsproject.org)

Rule of thumb (planning): run ESMFold → AF2 → AF3 as a compute ladder; run DiffDock only on structures that pass confidence gates; run RFdiffusion/ProteinMPNN on few retained targets.

13.2 Cost estimators (drop-in formulas)

“GPUh_total = ∑ₜ Nₜ · t̂ₜ_gpu ÷ Utilₜ.” (13.1)
“Cost_total = GPUh_total·price_gpu + CPUh_total·price_cpu + GB_store·price_store·months + GB_egress·price_egress.” (13.2)
“t̂ₜ_gpu = median_{runs=100 warm} (runtimeₜ).” (13.3)

How to use: measure per-task warm runtime on your hardware (ESMFold, AF2/AF3, DiffDock, RFdiffusion, ProteinMPNN), then scale by candidate counts. Keep Utilₜ (0–1) realistic given queuing/IO.

13.3 Throughput & batching

“Throughputₜ ≈ batch_sizeₜ · GPUs ÷ t̂ₜ_gpu.” (13.4)

Structure models like ESMFold and AF2 support moderate batching; AF3 batching is constrained by memory (ligands/nucleic acids increase tokens).
For DiffDock, shard by targets; cache target embeddings.
For RFdiffusion, pre-bake conditioning (symmetry, motifs) and checkpoint intermediate steps for restart. (science.org)

13.4 Reproducibility (pin, seed, checksum)

Pin environments. Record CUDA/cuDNN, drivers, GPU type, and exact repo commits/tags for AF2/AF3, ESMFold, DiffDock, RFdiffusion, ProteinMPNN.
Seeds and determinism. Log random seeds; enable deterministic kernels where supported; record non-det flags explicitly.
Snapshot external truth. Archive PDB IDs, UniProt accessions, Materials Project IDs with snapshot dates; follow RCSB PDB citation/identifier guidance. (rcsb.org)
Content-addressable outputs. Store a run-level provenance hash:
“h(x) = H(data_IDs ⊕ model_IDs ⊕ seeds ⊕ params ⊕ code_refs ⊕ outputs).” (13.5)
Model Cards (one per run). Intended use, data/IDs, metrics, limits, update policy; this is a standard transparency artifact you can cite. (arXiv)

13.5 Storage layout (minimal but audit-ready)

/campaign_YYMMDD/
  planning.yaml
  env.lock  # CUDA/cuDNN/driver, pip/conda lock, git SHAs
  data_snapshot.txt  # PDB/UniProt/MP IDs + dates
  stage1_structure.csv  # pLDDT, pTM, IDs
  stage2_interact.csv   # DiffDock scores, Succ@k
  stage3_design.csv     # RFdiffusion/ProteinMPNN artifacts
  materials.csv         # MP IDs, ΔE_hull, properties
  assays/               # PDFs (mean ± SD), SOP IDs
  model_cards/          # Mitchell et al. format
  audit/                # NIST RMF mapping
  artifacts/            # pickles/pt/npz, named by h(x)

13.6 Cost-cutting tactics that don’t hurt science

Breadth with ESMFold, depth with AF3. Use ESMFold for first-pass culling (seconds per protein); escalate only when complexes matter. (science.org)
Calibrate docking once, then cache. Fix k, RMSD threshold, and logistic coefficients for “P(bind)” per chemistry; reuse across campaigns. (arXiv)
Design sparingly. Generate fewer, higher-value RFdiffusion backbones; enumerate sequences cheaply with ProteinMPNN. (Nature)
Exploit public energetics. Use Materials Project ΔE_hull directly; do not re-DFT unless necessary. (docs.materialsproject.org)
Asynchronous queuing. Keep AF3/RFdiffusion on the biggest GPUs; run ESMFold/ProteinMPNN/DiffDock on smaller/spot GPUs.

13.7 Minimal math you can cite

“Compute_ladder: ESMFold → AF2 → AF3; escalate only if Gate_domain requires complexes.” (13.6)
“Accept if [metric ≥ τ] ∧ [no rule violations]; else escalate.” (13.7)
“Cost_total as in (13.2); provenance hash as in (13.5).” (13.8)

Takeaway. Treat compute as a ladder (cheap breadth → expensive depth), measure per-task warm runtimes, pin everything (IDs, seeds, commits), and checksum outputs. With RCSB/UniProt/MP IDs archived and Model Cards attached, your pipeline is both cost-aware and fully replayable. (science.org)

14. Limits & Failure Modes

Modern structure AI is powerful—but it is not physics, chemistry, wet-lab, or fab. Below are the known places it breaks, plus “know-when-to-stop” gates you can paste into SOPs.

14.1 Biology: dynamics, disorder, solvent/ions

Static snapshots vs real kinetics. AF-style predictors output single conformations; many targets bind by conformational selection / induced fit, where relevant states exist off the predicted minimum. Docking to one pose can be misleading. (PMC)
Intrinsically disordered regions (IDRs). IDPs/IDRs lack stable 3D structure; activity often hinges on context. Empirical tests show high failure rates when “fold-like” heuristics are applied to IDR peptides. Gate aggressively here. (PMC)
Water and ions matter. Neglecting structured waters and metal coordination is a common source of docking error; explicit solvent/ion treatment or MD often changes the story. (PMC)

Stop rule (biology context):
“Stop_bio ⇐ [IDR_fraction ≥ τ_IDR] ∨ [site_requires{water/metal} ∧ model_omits{water/metal}] ∨ [conf_accept = False].” (14.1)

14.2 Docking: overconfidence and novel chemotypes

Pose success ≠ affinity truth. Even with DiffDock’s calibrated confidence and better top-k RMSD, pose scores are hypotheses, not effect sizes. Calibrate on a PDBBind split and report Succ@k @ 2 Å before using scores in triage. (arXiv)
OOD chemistries. Docking and ML scorers can be over-confident off-distribution (new scaffolds, protonation/tautomer states, RNA targets). Use applicability-domain (AD) checks from QSAR practice. (PMC)

Stop rule (docking/OOD):
“Stop_dock ⇐ [Succ@k_calib < τ_k] ∨ [AD_score(x) < τ_AD] ∨ [protomer/tautomer unresolved].” (14.2)

14.3 Materials: stability vs makeability

Thermodynamic stability ≠ synthesizability. ΔE_hull ≤ ε is a solid first gate, but routes, precursors, hazards, and kinetics still block many candidates. GNoME expands the frontier (≈2.2 M proposals; ≈381k predicted stable), yet A-Lab results underline the gap between prediction and synthesis—success requires iterative make-and-measure. (Google DeepMind)

Stop rule (materials):
“Stop_mat ⇐ [ΔE_hull > ε] ∨ [S_syn < τ_syn] ∨ [precursor_unavailable ∨ ESG_red_flag].” (14.3)

14.4 Covariate shift & applicability domain (cross-cutting)

Distribution shift is the norm. New organisms, new assay conditions, or new chemistries are out-of-distribution (OOD) for models trained on legacy data. Report and enforce AD boundaries (descriptor/structure/response space) per OECD QSAR guidance; track OOD explicitly in bio/materials tasks. (OECD)

Generic stop rule:
“Stop_OOD ⇐ [AD(x)=out] ∨ [calibration_gap(x) > τ_gap].” (14.4)

14.5 “Know-when-to-stop” gate (health-adjacent, consumer-facing)

Even when a model looks great, do not ship claims without evidence and jurisdictional fit.

“Stop_ship ⇐ [health_adjacent ∧ (no human assessor ∨ no NAMs/safety math)] ∨ [claim crosses cosmetic→drug boundary] ∨ [no validated assay].” (14.5)

Rationale: AF3 and docking propose peptide–receptor geometries; WHO/NIST-aligned governance requires assessor review and validated experiments before any consumer claim. (Nature)

14.6 Minimal incident triggers to pre-declare

Biology: repeated assay failures where pLDDT/pTM passed; IDR-heavy targets treated as folded; water/metal neglected in known metalloproteins. (alphafoldserver.com)
Docking: calibration drifts (Succ@k down), surge in OOD scaffolds without AD guardrails. (arXiv)
Materials: ΔE_hull good but A-Lab repeatedly fails synthesis; supply/ESG blockers found late. (Nature)

Escalation math (paste into SOP):
“Escalate ⇐ Stop_bio ∨ Stop_dock ∨ Stop_mat ∨ Stop_OOD ∨ Stop_ship.” (14.6)

14.7 One-line disclaimer (product & docs)

“These models propose hypotheses; experiments decide. Use AD checks, calibration metrics, and domain gates; stop when any red flag triggers.” (14.7)

Why this section exists. AlphaFold-era tools (AF3), diffusion docking (DiffDock), and GNN materials discovery (GNoME) are major advances—but dynamics, disorder, solvent/ions, OOD drift, and synthesis reality remain hard constraints. Writing explicit stop rules is how you avoid cargo-cult claims and keep health-adjacent work compliant. (Nature)

15. Roadmap (12–24 Months)

What’s coming next: better complex modeling on the bio side, context-aware sequence design that “sees” ligands and ions, and scaled make-and-measure loops for materials that close the gap between prediction and reality. Consumer-visible benefits will remain mostly indirect—greener detergents, tougher device materials, and cautious peptide advisory—delivered via audited suppliers, not DIY labs.

0–12 months: consolidation and hand-offs

Complex modeling becomes the default for tricky targets. AlphaFold-3–class models that jointly reason over proteins, nucleic acids, ligands, ions, and modifications move from “special case” to routine triage for targets where binding context matters. Expect wider server access, clearer confidence readouts, and more integration into downstream docking/design steps. (Nature)
Context-aware sequence design lands in production pipelines. Tools like LigandMPNN (sequence design explicitly conditioned on non-protein context such as small molecules, nucleotides, metals) reduce the hand-off friction between structure, docking, and design—particularly in enzyme engineering and binder design. (Nature)
Materials: property predictors and autonomous loops tighten. After GNoME’s large front-end expansion of candidate crystals, teams standardize ΔE_hull gating and plug shortlists into self-driving labs (A-Lab–style) with active learning; parallel work on open surface-chemistry datasets (OC20/OC22) improves electrocatalysis and interface modeling. (Nature)
Factory-facing impact (quiet but real).
• Detergents: additional cold-wash–capable enzymes arrive through enterprise channels; LCA studies and policy work keep pushing low-temperature washing where performance allows. (OUP Academic)
• Devices: ML-guided thermal materials/coatings get modest throughput gains; teams report through-plane k and reliability under cycling as standard KPIs. (pubs.acs.org)

12–24 months: integration and measured scale-up

“Structure ↔ Design” becomes one loop for selected programs. Mature stacks couple AF-style complex prediction, diffusion docking with calibration, and context-aware sequence design into a single, auditable workflow for enzyme refresh and binder projects—still hypothesis-first, but with fewer manual glue steps. (Nature)
Materials: from can exist to can make and measure at scale. Expect more labs to adopt autonomous planning/execution akin to A-Lab, with program dashboards reporting Yield_AL and Δt_loop as first-class metrics; specialized predictors for solid-state electrolytes and interfacial stability get better with new curated datasets. (Nature)
Benchmarking culture hardens. Open efforts (e.g., OC20/OC22; MP IDs; PDB/PDBBind splits) keep standardizing what gets reported (ΔE_hull distributions, Succ@k@2 Å, pLDDT/pTM histograms), improving comparability across teams and audits. (arXiv)

What consumers are likely to notice (via audited suppliers)

Greener detergents without behavior change. More products that actually clean at lower temperatures, backed by better enzymes and documented sustainability gains; policy and LCA work continue nudging cold-wash adoption where performance permits. (OUP Academic)
More durable device materials. Incremental upgrades in heat spreaders, coatings, and electrolytes translate into slightly cooler phones, longer-lasting batteries, and fewer premature failures—reported as reliability and warranty metrics rather than splashy claims. (pubs.acs.org)
Modest, compliant skincare personalization. Peptide advisory within label-existing INCI classes (triaged by complex modeling, validated by literature and safety math) yields small, targeted tweaks—kept firmly inside cosmetic-claim boundaries. (Nature)

Program-level expectations to set with stakeholders

Evidence cadence: “Model → Gate → Assay” remains non-negotiable; no diagnostic or therapeutic claims from unregulated tools. (Nature)
Reporting discipline: keep IDs (PDB/AFDB/UniProt/MP/OC), calibration plots, ΔE_hull histograms, Succ@k tables, and assay PDFs in every release pack. (Nature)
Sustainability tracking: continue to publish ΔCO₂eq_per_unit and ΔkWh_per_use when low-temperature workflows are enabled by enzyme upgrades. (susproc.jrc.ec.europa.eu)

Bottom line: Over the next two years, expect tighter loops—AF-style complexes + ligand-aware sequence design on the bio side, and GNN-driven proposals + autonomous synthesis on the materials side—yielding steady, audited improvements that show up in everyday products without turning consumers into experimentalists. (Nature)

16. Appendices

A. Datasets & IDs — How to Cite and Record Them

Use this appendix as your Methods copy-paste for IDs, snapshots, and citations. Keep one block per campaign in your repo.

A.1 AlphaFold Protein Structure Database (AFDB)

What to record (per entry):
AFDB_accession (UniProt), AFDB_version_or_snapshot_date, model_confidence (pLDDT̄, pTM), AFDB_page_example

How to cite (choose what applies):

AlphaFold DB resource: “Fleming, J. et al. AlphaFold Protein Structure Database and 3D-Beacons: New Data and Capabilities. J. Mol. Biol. (2025).” (alphafold.ebi.ac.uk)
AlphaFold 2 method paper (if used): “Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature (2021).” (ebi.ac.uk)
AlphaFold DB overview (IDs are keyed by UniProt accessions): Varadi, M. et al. Nucleic Acids Res. 50(D1):D439–D444 (2022). (OUP Academic)

Notes: AFDB entries are keyed by UniProt accession (e.g., Q7RTU9); individual entries are downloadable and should be referenced with accession + snapshot date. (alphafold.ebi.ac.uk)

Example line (Methods):
“AFDB: accession QXXXXY (snapshot 2025-04-15); cite Fleming 2025 (resource) and Jumper 2021 (method). pLDDT̄ = 83, pTM = 0.78.” (alphafold.ebi.ac.uk)

A.2 RCSB PDB (Experimental “Ground Truth”)

What to record (per structure):
PDB_ID, experimental_method, resolution_or_map_QC, deposition_version_date, DOI, primary_publication

How to cite (canonical):

PDB ID + archive DOI; include the primary paper when available. Example DOI format: 10.2210/pdb1ema/pdb. wwPDB guidance: “A PDB structure with a corresponding publication should be referenced by PDB ID and cited using both the corresponding DOI and publication.” (wwpdb.org)
If no journal paper exists, cite the PDB archive DOI for the entry (e.g., 10.2210/pdb1B54/pdb) and the RCSB PDB resource paper. (OUP Academic)

Example line (Methods):
“PDB: 1EMA, X-ray 2.1 Å (deposited 1996-06-01), DOI 10.2210/pdb1ema/pdb; primary citation Ormö & Remington 1996.” (wwpdb.org)

Tip: Keep a link to the RCSB ‘Identifiers in PDB’ help page in your lab wiki for ligand/chain-level selectors. (rcsb.org)

A.3 Materials Project (MP)

What to record (per material):
MP_ID (mp-xxxxx), formula, structure_version_hash (if available), database_snapshot_date, formation_energy, ΔE_hull, calculation_method_tag

How to cite (resource):

Materials Project “How to Cite” page (next-gen site) and/or legacy “Citing” page; include the database version or snapshot date used for your queries. (next-gen.materialsproject.org)

Core stability field to report:
“ΔE_hull(x) = E_form(x) − E_hull(phase diagram).” (report in eV/atom with ε gate used in screening)

Example line (Methods):
“Materials Project: mp-4681 (snapshot 2025-07-10). E_form = −3.12 eV/atom; ΔE_hull = 0.018 eV/atom; cited per MP ‘How to Cite’.” (materialsproject.org)

A.4 PDBBind (Docking/Scoring Benchmarks)

What to record (per run):
PDBBind_release_year, split (general/refined/core), CASF_or_custom_split, preprocessing_rules (tautomer/protomer), calibration_metrics (Succ@k@2 Å)

How to cite (dataset + split):

PDBBind is commonly organized into general, refined, and core subsets; CASF benchmarks are typically evaluated on the core set. Always name the version (e.g., v2020) and which split you used. (arXiv)
Many recent works explicitly document “v2020 refined set for training; core set for tuning/evaluation.” Mirror this phrasing in your Methods. (pubs.acs.org)

Calibration metric to report:
“Succ@k = #(RMSD ≤ 2 Å in top-k) ÷ N (reference pose = crystal).” (report Succ@{1,5,10})

Example line (Methods):
“PDBBind v2020, refined for training (5,316 complexes), core for calibration; report Succ@{1,5,10} @ 2 Å and logistic triage coefficients.” (pubs.acs.org)

A.5 Minimal “ID & Snapshot” table (drop-in)

Resource	Must-have ID(s)	Snapshot/Version	What to cite
AlphaFold DB	UniProt accession (e.g., Q7RTU9); AFDB page ref	Date pulled or dataset release	Fleming 2025 (resource), Jumper 2021 (method), Varadi 2022 (overview) (alphafold.ebi.ac.uk)
RCSB PDB	PDB ID; archive DOI; primary paper DOI	Deposition/revision date	wwPDB “Cite Us” + RCSB NAR 2023 guidance (wwpdb.org)
Materials Project	MP ID (mp-xxxxx)	API snapshot date	MP “How to Cite” (next-gen/legacy) (next-gen.materialsproject.org)
PDBBind	Release year (e.g., v2020); split (general/refined/core)	Release date	Dataset papers/notes; CASF usage notes; recent split hygiene (LP-PDBBind) (arXiv)

A.6 One-line provenance rule (paste into every campaign)

“h(x) = H( data_IDs ⊕ model_IDs ⊕ seeds ⊕ params ⊕ code_refs ⊕ outputs ), with {AFDB_accession, PDB_ID+DOI, MP_ID, PDBBind_version/split, snapshot_dates} included in data_IDs.” (A.1)

Why this appendix matters. Stable IDs + snapshot dates make your results replayable; citing resource papers (AFDB, RCSB/wwPDB, MP) and dataset versions (PDBBind) is what reviewers and auditors now expect. (alphafold.ebi.ac.uk)

16. Appendices — B. Reference Equations (Unicode, single-line)

Core set (copy-paste):
“x* = argmaxₓ U(x) − λ·R(x) s.t. Γ(x) ≤ 0.” (B.1)
“Succ@k = #(RMSD ≤ 2 Å in top-k) ÷ N.” (B.2)
“ΔE_hull(x) = E(x) − E_hull(phase diagram).” (B.3)

Variable key (one-liners):

U(x) = utility (e.g., performance, sustainability); R(x) = risk/penalty; Γ(x) = hard constraints.
RMSD = root-mean-square deviation to crystal pose; N = number of complexes evaluated.
ΔE_hull = energy above convex hull (eV·atom⁻¹) used for stability screening.

Need to preserve your earlier tags? If your manuscript already cites (A.1)–(A.3), use the same three lines with those labels—they are identical.

16. Appendices — C. Minimal Reproducibility Checklist

Copy-paste this into your repo’s README_repro.md. It’s the smallest set that lets another team replay your decisions end-to-end.

C.1 Model card (one per run)

Name & version: model, commit/tag, container digest.
Intended use / limits: what it can’t be used for.
Data IDs: AFDB/UniProt/PDB/MP/PDBBind version & snapshot date.
Metrics: pLDDT/pTM histograms; Succ@{1,5,10}@2 Å; ΔE_hull distro; AUROC/PR for synthesizability/AD.
Calibration: plots + coefficients used (e.g., α,β,γ in 9.1).
Update policy: when it’s retrained/deprecated.

C.2 Data snapshots (frozen truth)

Biology: list of AFDB accessions, UniProt accessions, PDB IDs (+ DOIs), PDBBind release & split.
Materials: MP IDs, API date, calculation method tag.
Files: store a data_snapshot.txt with dates and counts.

C.3 Seeds & determinism

Seeds: global RNG seeds for Python/NumPy/PyTorch/JAX.
Determinism flags: cuDNN determinism on/off; any non-det ops named.
Replications: N repeat runs for runtime & variance.

C.4 Environment lock

Hardware: GPU model, driver, CUDA/cuDNN.
Software: exact repo SHAs, pip/conda lockfile, container hash.
Invocation: CLI commands with args and batch sizes.

C.5 Code & parameters

Configs: YAML files checked in (thresholds τ, ε, k, learning rates).
Feature toggles: which options were enabled/disabled.
Pre/post-processing: protonation/tautomer rules, structure cleaning, symmetry/motif settings.

C.6 Dataset hygiene

Splits: train/val/test (or refined/core for PDBBind), leakage checks.
Applicability domain (AD): descriptor ranges, OOD flags per sample.
Provenance: raw → processed mapping scripts.

C.7 Unit tests (must pass before running campaigns)

I/O sanity: load/save round-trip and schema checks.
Metrics math: RMSD calculation, Succ@k, ΔE_hull, CI/bootstraps.
Gates: acceptance rules (9.4/9.5/9.2) return expected booleans on fixtures.
Hashing: provenance hash stable across platforms.

C.8 Evaluation artifacts

Tables: stage1_structure.csv, stage2_interact.csv, stage3_design.csv, materials.csv, decisions.csv.
Figures: calibration plots, pLDDT/pTM histograms, ΔE_hull distributions.
Assays: PDFs with mean ± SD, SOP IDs, lot numbers.

C.9 Thresholds & gates (declare before use)

Biology: τ_LDDT, τ_TM, τ_bind, τ_stab, dock k & RMSD.
Materials: ε for ΔE_hull, σ_min, V_req, T_req, τ_syn.
Global: Γ_hard, Γ_soft=δ, reviewer sign-off rule.

C.10 Governance & compliance

NIST AI RMF mapping: GOVERN/MAP/MEASURE/MANAGE bullets.
WHO (health-adjacent): advisory-only statement + referral logic.
Disclaimers: “These models propose hypotheses; experiments decide.”

C.11 Storage layout (minimal)

/campaign_YYMMDD/
  planning.yaml
  env.lock
  data_snapshot.txt
  stage1_structure.csv
  stage2_interact.csv
  stage3_design.csv
  materials.csv
  decisions.csv
  assays/
  figures/
  model_cards/
  audit/  # NIST mapping, risk register
  artifacts/  # binaries named by hash

C.12 One-line provenance & ship rules (paste into Methods)

“h(x) = H( data_IDs ⊕ model_IDs ⊕ seeds ⊕ params ⊕ code_refs ⊕ outputs ).” (C.1)
“Release ⇐ [All domain gates pass] ∧ [Reviewer_signoff = True] ∧ [Disclaimers_present = True] ∧ [∀x: h(x) recorded].” (C.2)

That’s it—model card + snapshots + seeds + unit tests plus IDs and hashes. If these are present, an external reviewer can reconstruct your pipeline and verify that hypotheses were tested before claims.

16. Appendices — D. Glossary for AI Engineers

pLDDT (predicted Local Distance Difference Test).
Per-residue confidence from AlphaFold-style models on a 0–100 scale (higher = more reliable local geometry). Typical screen: pLDDT ≥ 70 as “usable,” ≥ 90 as “high.” Use as a filter, not evidence of function or stability.

pTM (predicted TM-score).
Global topology confidence on [0,1] (higher = more reliable overall fold/complex arrangement). A common gate is pTM ≥ 0.7 for “reasonable global topology.” Like pLDDT, it is a model confidence signal, not a claim.

RMSD (root-mean-square deviation).
Geometric deviation between a predicted structure/pose and a reference (often crystal).
“RMSD = √{ (1/N) · ∑ᵢ ‖rᵢ − rᵢ^ref‖² }.” (D.1)
In docking, heavy-atom RMSD ≤ 2 Å is a standard “near-native” pose threshold; see also Succ@k in (B.2).

ΔE_hull (energy above convex hull).
Thermodynamic stability measure for materials (eV·atom⁻¹): energy distance from the phase diagram’s lower envelope.
“ΔE_hull(x) = E(x) − E_hull(phase diagram).” (B.3)
Interpretation: 0 ⇒ on hull (stable to decomposition); 0–0.05 eV/atom is often considered “promising,” but not a guarantee of makeability.

Convex hull (materials).
The lower envelope of formation energy vs. composition; phases on the hull are stable against breaking into mixtures of others. Screening uses ΔE_hull relative to this envelope to rank candidates.

Synthesizability (makeability proxy).
A pragmatic score for “can we make it at sensible cost/risk?” combining route availability, precursor abundance, hazard/complexity, and historical success. One generic form:
“S_syn(x) = σ( u·route_score(x) + v·abundance(x) − w·hazard(x) ).” (D.2)
Use S_syn ∈ [0,1] as a gate (e.g., S_syn ≥ τ_syn) alongside ΔE_hull and target properties.

Assay (wet-lab or fab-lab test).
A validated experimental procedure that decides reality (activity, binding, conductivity, stability, etc.). Always pre-declare acceptance criteria (SOP ID, endpoints) and report mean ± SD, N, and lot/run IDs. In this article’s logic: models propose; assays decide.

Appendix E — Public Pattern Transfer to Everyday Domains (No New Theory Names)

Purpose. This appendix shows how to apply the same structure-before-search pipeline to non-bio tasks (scripts, cooking, meetings, business plans) using only public, neutral terms. Keep the same auditing habits: declare invariants, generate options, score with multiple critics, gate with thresholds, and run a small assay (real-world test) to decide.

E.0 Universal spine (drop-in anywhere)

“x* = argmaxₓ U(x) − λ·R(x) s.t. Γ(x) ≤ 0.” (E.1)
“P(pass) = σ(α·score + β·conf − γ·risk).” (E.2)
“Accept ⇐ [metric ≥ τ] ∧ [no rule violations]; else escalate.” (E.3)
“CSAₖ(x) = #(independent critics agree) ÷ k.” (E.4)
“Stop ⇐ [CSAₖ < τ_csa] ∨ [AD(x) = out].” (E.5)

U(x) bundles task utility; R(x) is uncertainty/overfit penalty; Γ are hard constraints.
CSAₖ = cross-tool agreement (e.g., model + rule-based checker + human rubric).
AD(x) = applicability domain flag from simple range checks on inputs/features.

E.1 Play Scripts (outline → scenes → table-read)

Invariants Γ (declare before generation): rating, runtime ≤ T, genre, must-hit beats B (e.g., theme stated, inciting incident, midpoint, crisis, climax), cast size, locations ≤ L.

Generation: beat sheet → multiple outline variants → scene drafts.

Critics & metrics (single-line):
“BeatCov(s) = #(required beats present) ÷ |B|.” (E.6)
“ArcCoherence(s) ∈ [0,1] from character-arc rubric.” (E.7)
“PaceVar(s) = stdev(scene_lengths).” (E.8)
“P(success) ≈ σ(α·BeatCov + β·ArcCoherence − γ·PaceVar).” (E.9)

Gate & accept:
“Accept_script(s) ⇐ [BeatCov ≥ τ_b] ∧ [ArcCoherence ≥ τ_a] ∧ [CSAₖ ≥ τ_csa].” (E.10)

Assay (experiment decides): cold table-read with target audience; report “EngageScore” (mean ± SD) and “Clarity” ≥ τ.
Log fields: {script_id, Γ, BeatCov, ArcCoherence, PaceVar, CSA_k, table_read_scores, decision, reviewer}.

E.2 Cooking (constrained recipe design)

Invariants Γ: allergens ∉ ban list, nutrition near target vector, equipment set E, budget ≤ B, total time ≤ T, servings S.

Generation: candidate recipes + substitutions from pantry.

Critics & metrics:
“NutriDist(r) = ‖macros(r) − macros_target‖₂.” (E.11)
“FlavorScore(r) ∈ [0,1] (data-driven flavor-pair/co-occurrence).” (E.12)
“Cost(r) = ∑ᵢ priceᵢ·qtyᵢ.” (E.13)
“Time(r) = prep + cook + cleanup.” (E.14)
“P(success) ≈ σ(α·FlavorScore − β·NutriDist − γ·OverTime − δ·OverBudget).” (E.15)

Gate & accept:
“Accept_recipe(r) ⇐ [AllergenSafe] ∧ [NutriDist ≤ ε] ∧ [Cost ≤ B] ∧ [Time ≤ T] ∧ [FlavorScore ≥ τ_f] ∧ [CSAₖ ≥ τ_csa].” (E.16)

Assay: quick tasting panel (N testers) with satisfaction ≥ τ_s and repeatability over two cooks.
Log fields: {recipe_id, Γ, NutriDist, FlavorScore, Cost, Time, CSA_k, panel_mean±SD, decision}.

E.3 Board Meeting (agenda that guarantees decisions)

Invariants Γ: objectives O, must-decide items D, duration ≤ T, quorum Q, required attendees, dependency constraints.

Generation: multiple agenda drafts with timeboxes and decision slots.

Critics & metrics:
“Coverage(a) = #(objectives covered) ÷ |O|.” (E.17)
“TimeBuffer(a) = (T − ∑ slots) ÷ T.” (E.18)
“ConflictLoad(a) = #(speaker/time conflicts).” (E.19)
“P(decision) ≈ σ(α·Coverage + β·TimeBuffer − γ·ConflictLoad).” (E.20)

Gate & accept:
“Accept_agenda(a) ⇐ [Coverage ≥ τ_c] ∧ [TimeBuffer ≥ τ_b] ∧ [Quorum ≥ Q] ∧ [|DecisionSlots| ≥ |D|] ∧ [CSAₖ ≥ τ_csa].” (E.21)

Assay: dry-run (calendar holds placed; pre-reads received rate ≥ τ_r) or simulate with historic meeting stats.
Log fields: {agenda_id, Γ, Coverage, TimeBuffer, ConflictLoad, quorum_ok, preread_rate, CSA_k, decision}.

E.4 Business Plan (options that clear finance & ops)

Invariants Γ: capex ≤ B, runway ≥ R months, compliance satisfied, staffing ≤ H headcount, risk tail ≤ q.

Generation: strategic options with pricing, go-to-market, ops, staffing.

Critics & metrics:
“NPV(p) = ∑ₜ CFₜ ÷ (1 + k)ᵗ.” (E.22)
“IRR(p) = arg k : NPV(p) = 0.” (E.23)
“RiskTail(p) = P(Loss ≥ L₀).” (E.24)
“OpsFeasible(p) ∈ {True, False} from capacity sim.” (E.25)
“P(choose) ≈ σ(α·NPV_norm + β·IRR_norm − γ·RiskTail).” (E.26)

Gate & accept:
“Accept_plan(p) ⇐ [NPV ≥ τ_npv] ∧ [IRR ≥ τ_irr] ∧ [RiskTail ≤ q] ∧ [Capex ≤ B] ∧ [OpsFeasible = True] ∧ [CSAₖ ≥ τ_csa].” (E.27)

Assay: pilot or staged A/B (limited geography or channel); report uplift vs baseline and variance bands.
Log fields: {plan_id, Γ, NPV, IRR, RiskTail, Capex, OpsFeasible, CSA_k, pilot_results, decision}.

E.5 Minimal reproducibility for non-bio tasks

IDs & versions: scenario IDs, dataset snapshots (e.g., pricing tables, pantry inventory date), model commits, seeds.
Thresholds: {τ_b, τ_a, τ_f, ε, τ_c, τ_b (buffer), τ_npv, τ_irr, q, τ_csa} captured in planning.yaml.
Provenance:
“h(x) = H( data_IDs ⊕ model_IDs ⊕ seeds ⊕ params ⊕ code_refs ⊕ outputs ).” (E.28)

One-line ship rule (for these domains):
“Release ⇐ [All Γ satisfied] ∧ [CSAₖ ≥ τ_csa] ∧ [Assay_pass = True] ∧ [Reviewer_signoff = True] ∧ [h(x) recorded].” (E.29)

Disclaimer

This book is the product of a collaboration between the author and OpenAI's GPT-5, Wolfram's GPTs language model. While every effort has been made to ensure accuracy, clarity, and insight, the content is generated with the assistance of artificial intelligence and may contain factual, interpretive, or mathematical errors. Readers are encouraged to approach the ideas with critical thinking and to consult primary scientific literature where appropriate.

This work is speculative, interdisciplinary, and exploratory in nature. It bridges metaphysics, physics, and organizational theory to propose a novel conceptual framework—not a definitive scientific theory. As such, it invites dialogue, challenge, and refinement.

I am merely a midwife of knowledge.

Sunday, November 2, 2025