FOLDIFF F6 — OWN TRAJECTORIES / WT AND SUBSTITUTED VARIANTS Local implementation 0.2.0 · 14 October 2026 PURPOSE AND SCOPE Compare the observed phi/psi distribution of ONE corresponding internal residue in WT and up to six substituted variants (one pair displayed at a time). This is an analysis of existing unbiased MD, not an MD engine, mutation generator, full protein/domain landscape, activity predictor, or absolute mutant ΔΔG. A domain can be the input construct, but removing it from a complex changes the scientific question. No independent-domain simulation is implied by the small 3D fragment shown in the interface. Troponin is not validated by the DHFR example. LOCAL WORKFLOW Open /foldiff/landscapes/compare/ using serve_local.py on 127.0.0.1:8033. 1. Add WT: identify protein, construct, force field, solvent/ions, ensemble, temperature (or leave unknown), ligands/modifications/pH/partners/pressure, source and simulation preparation. 2. Select a matching topology (PDB, mmCIF/CIF or PSF), chain, internal residue; add XTC or DCD files. File order is initially natural filename order and can be changed using arrows. Explicitly distinguish continuation segments from independent runs; a file is not automatically a replicate. 3. For DCD, declare recording interval and its evidence. For XTC, either use validated header times or explicitly declare the interval and evidence. A declared interval cannot independently validate chronology or missing frames. 4. Optionally retain every Nth frame. This does not establish independence. Confirm unbiased sampling, atom order, conditions and continuation order. 5. Add variants similarly. An individual prepared dataset can be saved as JSON. 6. Choose variant, shared grid (18/36/72 bins per angle), count threshold, and initial discard duration applied relative to each run's first retained frame. The JSON project preserves all datasets/options and recomputes on reopening. Export JSON, map CSV, three-map SVG, standalone HTML, or actual selected PDBs. No autosave. Downloaded JSON contains extracted coordinates and metadata, not the complete original trajectories; retain original files separately. NATIVE ADAPTER AND LIMITS MDTraj 1.11.2 reads coordinates in nm; angles converted to degrees, clocks to ps. Inputs: 1–12 trajectory files, <=180 MiB total HTTP upload, <=50,000 atoms, <=500,000 original frames and <=250,000 retained frames. CLI permits <=180 MiB per trajectory file but the same atom/frame limits. At least 40 retained frames per run, also after discarding. Stride integer 1..1000. Topologies/files are copied to an ASCII-named temporary directory because native readers may not support Unicode paths. Copies are deleted on completion/error. The adapter makes no remote calls. Browser uploads only to the loopback preparer. Host/Origin checks and per-process token prevent cross-origin preparation; loopback binding, one active preparation and request limits bound local exposure. This server is NOT a production upload service and must not be deployed publicly. Validate atom count, finite positions, nonsingular positive-volume periodic box, selected backbone connectivity and bond lengths 0.08..0.22 nm. Count agreement alone cannot prove atom ordering: user must supply a matching topology. Unknown residues inside the chosen protein chain and ambiguous numbering are rejected. Compute phi = C(previous), N, CA, C and psi = N, CA, C, N(next); terminal residues and chain breaks are refused. Periodic boundary conditions are used where a box is present. Actual local fragments are reconstructed whole and checked again. Header times must increase uniformly. Overlapping/reset clocks and unequal cadence are refused in header mode, including continuation boundaries. A declared clock is reconstructed from input order and explicit dt, not from DCD indices. Exactly equal boundary coordinates are removed ONLY with explicit opt-in and recorded in provenance. No approximate deduplication or inferred independent runs. 3D AND CORRESPONDENCE Keep real heavy-atom context up to five neighboring residues on either side. One representative closest to the center of each 10-degree bin is stored per run, up to 6,000 representatives. Fine 5-degree maps, or post-discard samples, may lack a saved frame even when a bin has observations: the interface says so. No frame is synthesized. Serial mappings tie each PDB to its exact sample and two torsions; browser validation recomputes them, rejecting >1 degree mismatch. PDB coordinates round to 0.001 Angstrom; timestamps remain sample timestamps. WT and variant fragments are superposed on matching numbered CA atoms using least-squares rigid alignment. This is a visual frame comparison, not a transition path, energy-minimum structure or ligand/membrane rendering. Downloaded PDBs retain source coordinates, not display alignment. Sequence comparison requires equal length, same residue numbers, same selected sequence position and at least one substitution. Indels/renumbering unsupported; >20% substituted positions is an explicit product guard, not a homology theorem. Metadata strings must match exactly after outer-space trimming, including known T. Two unknown temperatures remain marked unconfirmed; kJ/mol is disabled. Equal metadata never proves equivalent physics or model correctness. CALCULATIONS Common periodic equal-area histogram bins in [-180,180) degrees, no smoothing or pseudocounts. Within each condition, counts are pooled by frames, then normalized: P_i = n_i / sum(n). Longer independent runs therefore carry more weight. Delta probability = P_variant - P_WT, displayed in percentage points. JSD uses base-2 logs: 0 identical, 1 disjoint distributions. Overlap = sum min(P,Q). These describe the observed distributions; neither is a mutation success score. F_i/(RT) = ln(n_max/n_i) for bins above threshold; with confirmed temperature, F_i = RT ln(n_max/n_i), R = 0.00831446261815324 kJ mol^-1 K^-1. Each condition has an independent additive energy zero; NO subtraction of these energy maps into mutant ΔΔG. Unobserved bins have no estimate, not infinite energy. Low-count probability differences remain shown with hatching. This is not a significance test. Change grid/threshold/discard to inspect sensitivity. QUALITY Shared F5 count engine reports first/second-half JSD, per-run and pairwise JSD, counts, cadence, four contiguous blocks per run, limited autocorrelation of sin/cos(phi/psi), and advisory ESS. Fixed-reference block SD in kJ/mol is available only with known T; unknown T produces null physical energy/SD arrays. Circular angle means use atan2(mean sin,mean cos); resultant is also exported. No confidence intervals or p-values are manufactured from correlated frames. A 256-lag diagnostic is not a convergence/independence certificate. One run per condition cannot separate mutation effects from between-run variation. Unknown T, unequal sampling, drift, constant coordinates and sparse bins are not evidence of equivalent physics. Report the limitations, not a single universal accuracy. PUBLISHED VERIFICATION CASE Ebru Cetin, Ali Rana Atilgan, Burcu Dedeoglu, Canan Atilgan (2023). Kinetic barrier to enzyme inhibition is manipulated by dynamical local interactions in E. coli DHFR. https://doi.org/10.1021/acs.jcim.3c00818 Dataset https://zenodo.org/records/7966540 · CC BY 4.0 Authors' code https://github.com/midstlab/JCIM_Cetin_etal_2023 Use protein-only Dataset 1, tmpp-wt-[1..5].dcd and tmpp-l28r-[1..5].dcd, with their corresponding -pro.psf. Original simulated context includes TMP+ and NADPH; CHARMM36, TIP3P/0.15M KCl, NPT. Ligands and solvent are absent from these deposited coordinate files. We do not reproduce the paper's FEP analysis. Each condition: 5 x 1001 frames, FOUR exactly duplicated boundary frames removed, 5001 distinct frames in ONE 1-microsecond run. 200 ps cadence is inferred from the documented duration and frame count, not trusted from DCD header clocks. Default initial discard 10 ns leaves 4951 frames per condition; source deposit start offset is not independently established. This setting is adjustable. Unbiased production-MD temperature unconfirmed. The authors' separate FEP script uses 310 K; it is not used as evidence for the selected trajectories' temperature. Selected coordinate: residue 28 backbone phi/psi. Detected substitution: L28R. Default 36x36 grid: JSD ~0.0030263, overlap ~0.970511. This local angular agreement is not evidence that the complete landscapes, function or stability are equal. Raw input checksums verified against Zenodo MD5; SHA256 recorded. Transformations, source files and conditions retained in JSON. These are derived, reduced datasets; not endorsed by the source authors. License https://creativecommons.org/licenses/by/4.0/ VERIFICATION / REPRODUCIBILITY research/fetch_dhfr.py — download and source/checksum manifest. research/prepare_dhfr.py — reproducible native preparation. tests/adapter_test.py — 6 groups: formats, time, stitching/replicates, topology, geometry, PBC and limits, using small real-coordinate fixtures (not new MD). tests/compare.test.cjs — 7 groups: known distributions, source mapping, metadata, roles, biased rejection, temperature scaling, actual PDB validation, sampling. tests/validate_comparison.py — independent NumPy dihedral formula and counts: all 10002 unique frames, 49 PDB representatives, 6 grid/discard configurations. Native-angle maximum difference <0.000016 degrees. PDB round-trip <0.112 degrees. Initial 0.1-degree PDB tolerance failed by rounding; revised to 0.2 degrees for 0.001-Angstrom PDB rounding. Native tolerance stayed 0.001 degrees; numeric probability/FRT/JSD/overlap tolerance 1e-12. No source data altered to pass a test. See samples/dhfr-validation.json for exact errors, source hashes and limits. tests/server_test.py — Host/Origin/token isolation, limits, native multipart. tests/compare-browser.cjs — actual native upload, RU/EN, 320/390/768/1440 pixels, dark mode, keyboard, real WebGL, exports, JSON roundtrip and incompatible metadata. F5 regressions remain in tests/core.test.cjs and tests/browser.cjs. This verifies implementation against published MD and independent mathematics. It does not experimentally validate the force field, equilibrium coverage, causality, disease relevance, troponin constructs or every claim in the paper. No new MD, biased reweighting, free-energy perturbation or clinical use.