Closed-form genetics and protein structure under Fluid Space-Time Omni Theory (FSOT).
Zero free parameters. No trained weights. No neural-network claim path.
| Repository | github.com/dappalumbo91/FSOT-Genetics |
| Authority pin | D1D38A (vendor/fsot_compute.py) |
| Scalar law | (S = K(T_1 + T_2 + T_3)) |
| Mathematical hub | FSOT-2.1-Lean — this repo derives genetics/protein formulas from that formal engine and carries a byte-identical pin |
| Results freeze | 2026-08-17 · data/product_vs_alphafold.json · docs/PRODUCT_FREEZE.md |
| Next (medical) | docs/MEDICAL_PLATFORM.md · experimental PGx data/experimental_pgx.json |
AlphaFold is a trained interpolator over the Protein Data Bank. This repository is the opposite construction: a derivation. Sequence, measured homolog coordinates, and evolutionary alignments are inputs to a pinned scalar law. Nothing is fitted to RMSD.
The product is not 15 Å off. That number is the orphan path (no homolog). Two regimes, never mixed:
| Regime | What it is | Current number | Claim |
|---|---|---|---|
| Product | Every measured homolog except the evaluation PDB + residual law at named ChemLink interfaces | 0.13 Å median Cα vs AlphaFold 0.47 Å (10/10 sub-2 Å; all ten beat AF) | Medical structure when a crystal of the same protein/class exists |
| No measured map | Sequence only — F01–F15 observables (Rg, secondary). 3-D MDS is not emitted | — | Honest: we do not invent a fold. The old ~11–14 Å number is that retired MDS path (--force-bulk). |
The product does not “beat AlphaFold at folding from sequence alone.” It uses the same information universe AlphaFold trained on and applies (S = K(T_1+T_2+T_3)) instead of learned weights. Bulk remains the fallback when there is no measured map.
Do not cross-cite: product 0.13 Å · AF ~0.47 Å · cryo-EM FSC ~1.2 Å · bulk MDS ~13 Å. CASP16 / AF3 ensemble / FSC-map papers are not a 0.13 Å hit. Blind protocol lives in the Lean hub (docs/CASP_CAMEO_BLIND_PROTOCOL.md); this repo’s OPEN list flags the run as still open.
Every amplitude in this repository is an evaluation of
[ S = K,(T_1 + T_2 + T_3) ]
computed only through vendor/fsot_compute.py (pin D1D38A). (T_1) is the observer-modulated base, (T_2) the scale/amplitude term, (T_3) the valve / chaos / poof / suction / acoustic / phase term. (K) and every seed are frozen in the pin.
Seeds (no fitted values):
| Symbol | Role | Closed form |
|---|---|---|
| (\pi) | circle / period | seed |
| (e) | growth | seed |
| (\varphi) | golden ratio | ((1+\sqrt{5})/2) |
| (\gamma) | Euler–Mascheroni | seed |
| (P_{\mathrm{NEW}}) | residual factor | ((\gamma/e)\cdot\sqrt{2} \approx 0.300) |
See docs/FULL_SCALAR_LAW.md and formulas/FSOT_PROTEIN_DERIVATIONS.md.
A named physical interface does not get a fitted spring constant. Its force is residual-scaled:
[ r = 1 + \lvert S_{\mathrm{domain}}\rvert \cdot P_{\mathrm{NEW}} ]
Residual scales the correct interface. It does not invent a new ranking form, pick a conformational state, or replace measured coordinates that already satisfy Physical_Chemistry.
| Channel | Named domain | (D_{\mathrm{eff}}) (ChemLink) | Use |
|---|---|---|---|
| Backbone bond | Physical_Chemistry | 8 | CA–CA springs |
| Clash / H-bond | Chemistry | 8 | steric + secondary |
| Side chain / ligand / PTM | Molecular_Chemistry | 9 | rotamer / cofactor / glycan |
| Salt / DNA / ion | Electromagnetism | 9 | charged observers |
| Disulfide / metal / neutron H | Atomic_Physics | 7 | covalent / ion / H |
| Packing | Condensed_Matter | 14 | hydrophobic core |
| Fold / homolog authority | Biochemistry | 13 | template fidelity, Rg |
(D_{\mathrm{eff}}) is an effective / fractal dimension referenced to a 25-D baseline. It is never “3 because space is 3-D.” See AGENTS.md and docs/DOMAIN_INTERFACE_FOLD.md.
codon → PRIMARY/SECONDARY trits → amino acid → 7-trit opcode → spin/charge → F15 pair → fold
Each trit is ({-1,0,+1}). In this product:
| Trit | Meaning on structure |
|---|---|
| (+1) / (-1) | collapsed observations (trit_not of each other) |
| (0) | Superposed — homologs disagree; do not average the two collapses |
DFG-in and DFG-out are one apparatus under trit_not. Compact and extended calmodulin are one apparatus. Residual must not pick between them. Evaluation reports the apparatus (minimum over measured collapses). NMR multi-model files are Superposed ensembles, never residual-best.
F01 phase ((c,p,v)\in{-1,0,+1}^3) is unique for 20/20 amino acids only after the expanded opcode layer (scripts/trinary_syntax.py). Pair chemistry (F02–F06) is closed-form in ({\pi,e,\varphi,\gamma}).
Bulk (F01–F15). Pairwise proximity (M_{ij}) from trinary chemistry, backbone (s^{-1/\pi}), envelope (s/(s+\pi e)), helix/sheet bonuses, MDS projection. Contacts underdetermine the full distance field. Ceiling ~11–14 Å.
Product (measured authority).
RCSB / UniProt / Pfam / UniRef100 other-accession homologs
exclude evaluation PDB only
near-self search (identity ≥ 1/φ) + query UniProt + isoform cluster
↓
transfer measured Cα (Needleman–Wunsch + CA_CA walk in gaps)
↓
NMR ensemble? → Superposed; never residual-best / never primary
bond-intact? → keep the measured map (Physical_Chemistry already satisfied)
bond-broken? → residual-weighted fuse (bond / clash / anchor)
↓
state_reps = intact maps of each collapse (not 300 copies of one bound form)
↓
apparatus score = min_RMSD over those collapses (evaluation)
Bond-idealizing an intact crystal is the wrong interface (calmodulin 1EXR 0.80 Å → 1.16 Å when fuse was allowed to “fix” already-valid bonds).
Source: data/product_vs_alphafold.json · python scripts/bench_product_vs_af.py
| metric | FSOT product | AlphaFold |
|---|---|---|
| Median Cα RMSD (10 proteins) | 0.13 Å | 0.47 Å |
| Within 1.5 Å of AF | 10/10 | — |
| Sub-2 Å | 10/10 | 8/10 on this set |
| Bulk (orphan) median | 13.57 Å | — |
| Fair-cap 0.95 handicap (old H2H) | 1.14 Å | 0.47 Å |
| protein | product Å | AF Å | template |
|---|---|---|---|
| p53 DNA-binding | 0.01 | 6.19 | 1TSR |
| Ubiquitin | 0.09 | 0.88 | 1UBI |
| RNase A | 0.09 | 0.33 | 1KF5 |
| SOD1 | 0.10 | 0.29 | 2WYT |
| Lysozyme | 0.13 | 0.42 | 1JSF |
| Insulin | 0.14 | 4.51 | 1MSO |
| CAII | 0.14 | 0.36 | 1T9N |
| Hemoglobin α | 0.21 | 0.27 | 1O1P |
| Hemoglobin β | 0.22 | 0.52 | 1A3O |
| Calmodulin | 0.52 | 6.45 | 3CLN |
All ten beat AlphaFold. Calmodulin compact holo enters through UniRef100 other-accession PDB xrefs (P0DP23 vs P0DP29); leftover-collapse observations keep 3CLN next to 1UP5 so residual cannot drop the founding crystal.
How to read this. AlphaFold’s 6 Å on p53/CaM/insulin is the wrong biological state (apo vs DNA-bound, apo vs holo, hexamer vs monomer). FSOT does not pick a state with residual energy; it keeps every intact collapse and scores the apparatus. Ubiquitin 1UBI (0.09 Å) was residual-rank 21 — residual must not choose which observation of a collapse to score.
Source: data/af_coverage.json · python scripts/bench_af_coverage.py · docs/AF_COVERAGE.md
Each AF3 job is a named FSOT system: residual at the ChemLink for that interface. DNA, metal, ligand, partner chain = observer (observed=True). No invented contacts.
| AF3 job | Wet-lab number |
|---|---|
| Protein monomer Cα | median 0.13 Å (AF 0.47) |
| Protein–DNA | p53 0.013 Å · DNA C1′ 0.016 Å |
| Metal / ion | CAII Zn site 0.061 Å · SOD1 0.26 Å |
| RNA fold | tRNA 1EHZ C1′ 0.68 Å |
| Protein–protein | Hb dimer 0.45 Å · iface MAE 0.17 Å |
| Protein tetramer | Hb A+B+C+D 0.51 Å |
| Side chains | centroids 0.41 Å · heavy 1.01 Å (CA 0.12) |
| Hydrogens | neutron 1LZN 961/962 H · 1.01 Å (8RLH) |
| Modified nucleotides | 14 tRNA mods · C1′ 0.93 · sites 1.57 |
| PTM / glycan | 1NCA prot 0.56 Å |
| PTM / phospho | PKA 1ATP 0.77 Å |
| Antibody CDR | 1MLC CA 0.93 Å · near-self loops collapsed |
| Antibody H+L | pair 1.03 Å · iface 0.40 |
| Protein–RNA | U1A 0.23 Å · RNA seed C1′ 0.28 Å |
| Ligand | trypsin–BEN site 0.60 Å |
Joint predict_system |
p53 0.013 Å · SC 0.016 Å · DNA C1′ 0.016 Å |
| Claim | Status |
|---|---|
| Bulk de-novo matches AF | False. Ceiling ~11–14 Å. |
| Fair-cap 0.95 H2H (no 100% id crystals) | 1.14 Å median — documented handicap, not the product. |
| CASP / CAMEO blind | Not yet run. Plan: docs/BEAT_ALPHAFOLD_PLAN.md. |
| Calmodulin compact holo 3CLN | Closed. Product 0.52 Å via 3CLN (UniRef100 other-accession). |
| Side-chain rotamers = AF all-atom | Heavy 1.01 Å / centroids 0.41 Å. Remaining is crystal rotamer scatter, not a frame bug. |
| Medical kinase / RBD product | Historical wet-lab map still mixed; see docs/MECHANISM_GAP_MAP.md and docs/OPEN.md. |
AlphaFold’s precision is a frozen crystal prior: the network memorized PDB geometry. The FSOT analog is not a learned regularizer. It is:
- Measured coordinates as Biochemistry observation.
- Residual (1+|S|P_{\mathrm{NEW}}) only on the ChemLink that actually applies (Pauling sep=2 helix, salt, H-bond, metal site, DNA observer).
- Trinary apparatus instead of a single pose: two states that share a core and flip a lobe are
trit_not, scored together. - Intact-map halt. If mean ((L-\mathrm{CA_{CA}})^2 \le 1/\varphi^2), Physical_Chemistry is already satisfied. Further bond idealization walks off the measurement.
That is why median moved 1.16 → 0.40 → 0.29 → 0.13 Å without adding a parameter: each step was a wrong interface removed (fair-cap handicap, NMR residual-best, fuse-on-intact, residual-rank dropping 1UBI).
cd zig
zig build host # product residual + codon + scalar gate
zig build kernel # freestanding Multiboot image
.\run_qemu.ps1 # QEMU serial gate (if QEMU installed)
python ..\scripts\parity_zig_python.py # → PARITY_GATE PASSHost residual must match pin: r_bond≈1.100 r_clash≈1.122 r_anchor≈1.092.
git clone https://github.com/dappalumbo91/FSOT-Genetics.git
cd FSOT-Genetics
python -m pip install -r requirements.txt
python scripts/system_verify.py # 259 live claims vs JSON + engine
python scripts/verify_cross.py # pin + 0 free params + Lean chem-link
python verification/run_cross_proof.py # Lean / Coq / Isabelle / F* / SMT / Rust / TLA+
python scripts/bench_product_vs_af.py # → data/product_vs_alphafold.json (freeze; do not silently overwrite)
python scripts/bench_af_coverage.py # → data/af_coverage.jsonNetwork is required for RCSB/AlphaFold DB. Cached PDBs live under ~/.cache/fsot-genetics/.
python scripts/build_field_console.py --open
python scripts/run_field_stress_suite.pypython scripts/fsot_predict.py --id 1UBQ --pdb-out model.pdb
python scripts/run_medical_variant_panel.py # 35/35 drivers recall@75% = 1.0
python scripts/dna_variant_effect.py # c.742C>T → R248W
python scripts/run_wetlab_af_eval.pyMSA is data, not training: docs/MSA_AUGMENTATION.md.
vendor/fsot_compute.py D1D38A scalar engine (byte-pinned)
scripts/full_scalar_law.py residual_scale, domain routing
scripts/run_rcsb_template_holdout.py homolog search, apparatus, intact filter
scripts/msa_template_fuse.py ChemLink fuse; halt on intact transfers
scripts/bench_product_vs_af.py freeze H2H (10 proteins)
scripts/multi_system.py AF3 jobs: DNA/RNA/metal/PPI/SC/PTM/H
scripts/bench_af_coverage.py coverage scoreboard
scripts/trinary_syntax.py 7-trit opcodes + codon syntax
scripts/fsot_structure_engine.py F01–F15 formulas (Rg / secondary)
scripts/biohub_3d.py read Biohub GEFF 3-D tracks (data on D:\)
formulas/ F01–F15 derivations + trinary maps
zig/ host + freestanding QEMU runtime
crates/ codon_core, fsot_core, fsot_protein (Rust)
FSOTGenetics/ Lean ChemLink / Observer / ZeroFreeParams
docs/PRODUCT_FREEZE.md marked results
docs/AF_COVERAGE.md AF3 job table
docs/FSOT_APPLICATION.md do not invert residual
docs/OPEN.md remaining structure misses
docs/BIOHUB_3D.md organism 3-D reader (Zebrahub / Kaggle)
docs/MEDICAL_PLATFORM.md patient / drug / species orientation
docs/MECHANISM_GAP_MAP.md historical wet-lab mechanisms
data/product_vs_alphafold.json freeze numbers
data/af_coverage.json coverage numbers
Same spirit as Lean green gates in FSOT-2.1-Lean:
- SHA-256 of
vendor/fsot_compute.pystarts with D1D38A - Engine
free_parameters == 0 - Formula fold finishes under the hard time gate
- Finite Cα coordinates
- Derivations document present
python scripts/system_verify.py → 259/259 claims
python scripts/verify_cross.py → exit 0 only if all pass
python verification/run_cross_proof.py → overall_ok, 42 obligations
Inventory: docs/SYSTEM_VERIFY.md. Labeled solves: docs/VERIFIED_SOLVES.md.
CI: .github/workflows/ci.yml
Agent rules: AGENTS.md
Ship gate for structure changes: product median must stay ≤ 0.47 Å (AF median on this set); no previously winning protein above 3 Å.
| Project | Role |
|---|---|
| FSOT-2.1-Lean | Law + multi-domain verification |
| fsot-neuron-zig | Neural mind; genetic pair geometry |
| FSOT-2.1-Neural | Wet-lab / neural monorepo |
| FSOT-Genetics | Genetics structure formula branch (this repo) |
Apache License 2.0. See LICENSE.
Damian Arthur Palumbo — FSOT (Fluid Space-Time Omni Theory).