رفتن به محتوا
newblender

13 — Molecular Nodes: how an agent could make it

این محتوا هنوز به زبان شما در دسترس نیست.
Original Brady Johnston (+ contributors), 2022–present, scientific visualisation add-on for structural biology · GitHub · PyPI
Agent-readiness today Today: the input is real, public, machine-readable data (the Protein Data Bank), the representation is Geometry Nodes (GN) driven by attributes, and correctness is checkable against the file itself. This is the most agent-native work in the 20. Taste remains in colour, framing and what to emphasise for a scientific story.
Difficulty 2 for renders and animations of a structure with the real add-on; 4 to rebuild the add-on itself (cartoon ribbons, surfaces, trajectories, density maps)
First slice Parse crambin (PDB 1CRN, 327 atoms) straight from RCSB, write atoms as a point cloud with element, chain and residue attributes, and build a ball-and-stick plus a CA backbone tube in one GN tree, coloured by element; checks prove every atom and both helices arrived.

1. What the original actually is

Section titled “1. What the original actually is”
  • A tool, not a picture. An add-on and Python package that imports PDB, mmCIF, molecular-dynamics (MD) trajectories and electron-microscopy (EM) density maps into Blender and styles them with GN (README, fetched 2026-10-07).
  • Distribution: Blender extension (Get Extensions, Blender ≥ 4.2) and PyPI. PyPI’s latest is 520.2.0 (18 Sep 2026), Python ~= 3.13: the version number tracks the Blender release line (5.2). Licence GPL-3.0. Core dependencies Biotite and MDAnalysis.
  • Styles: procedural representations such as cartoon, surface, ball-and-stick and spheres (index entry 13); colouring from per-atom attributes.
  • Funding and use: a 2024 NumFOCUS Small Development Grant via MDAnalysis; used in labs for publication renders and animations.
  • Outputs people make with it: stills for papers and covers, rotating turntables, MD trajectory playback, morphs between conformations.

2. How the humans made it

Section titled “2. How the humans made it”

Stages: S03 Shading, S04 Geometry Nodes, S07 Animation, S09 Lighting & rendering.

  • Fact (README/docs): data parsing is delegated to Biotite and MDAnalysis; Blender receives meshes/point clouds with attributes; GN node groups shipped with the add-on turn atoms into styles.
  • Fact (index): attribute-driven shading; keyframed and trajectory animation; Cycles/EEVEE.
  • Interpretation: the human craft was in the library of node groups (how a cartoon ribbon is derived from backbone atoms and secondary-structure records) and the import plumbing. A user then picks a style and colours. This is exactly the declarative-first artefact the ledger proposes (L-006): data in, recipe on top, nothing hand-modelled.

3. How I would make it: the agent-native plan

Section titled “3. How I would make it: the agent-native plan”

Two paths. Path A uses the real add-on (what a scientist would do). Path B rebuilds the minimum from scratch to prove the pipeline needs no hand work; the first slice is Path B because it needs no extension install and every step is inspectable.

Stage Agent approach Inputs (real data) Tooling
Data Fetch PDB/mmCIF from RCSB; parse ATOM/HETATM, HELIX/SHEET, SSBOND records (Path B: a 60-line parser; Path A: Biotite inside the add-on) RCSB PDB: https://files.rcsb.org/download/1CRN.pdb (crambin, 327 atoms, 46 residues, 2 helices, 1 sheet, 3 disulfides); 4HHB (deoxyhaemoglobin, 4 chains, 4,384 ATOM + 395 HETATM incl. 4 haem groups) Python inside Blender via tools/live/bl.py
S01 as data One mesh of loose vertices (one per atom) with attributes: element (int), b_factor, res_id, chain_id, is_backbone, sec_struct Same mesh.attributes.new(...).data.foreach_set
S04 Style: spheres Instance on Points (Ico Sphere), scale from a van der Waals radius lookup by element (Index Switch / Sample Index on a radii table) Bondi radii (C 1.70, N 1.55, O 1.52, S 1.80 Å) GN
S04 Style: sticks Bonds computed by distance (pairs closer than sum of covalent radii + 0.45 Å) written as mesh edges; Mesh to Curve → Curve to Mesh with a thin circle profile Covalent radii table Python for bonds (KD-tree), GN for geometry
S04 Style: backbone Filter is_backbone CA atoms in residue order → Points to Curves → smooth (Resample, Set Spline Type) → Curve to Mesh; profile width larger where sec_struct = helix HELIX/SHEET records GN
S03 Shading Colour by element (CPK-like: C grey, N blue, O red, S yellow) via Attribute node; alternative ramp on b_factor None Shader Attribute
S07 Animation Turntable (one driver), or interpolate positions between two PDB conformations by GN Mix on position; Path A loads MD trajectories Second conformation from PDB, or an MDAnalysis test trajectory GN / add-on
S09 Render Cycles, three-point light, soft AO-like ambient; orthographic option for figures Poly Haven studio HDRI (CC0), optional Cycles

Build order:

  1. Download 1CRN.pdb (≈ 60 KB). Write the checks (§5) from the file’s own header: 327 atoms, 46 CA, 2 helices (residues 7–19, 23–30), sheet 1–4 / 32–35, 3 SSBONDs.
  2. Parse and write the point mesh with attributes; run layer-1/2 checks.
  3. GN tree GN-mol_style with a style menu input (spheres / ball-and-stick / backbone).
  4. Bonds by KD-tree; check bond-length distribution and the 3 disulfide bonds.
  5. Element colours; camera framing the bounding sphere; turntable.
  6. Repeat with 4HHB to show scale (4 chains, haem groups as HETATM).
  7. Path A follow-up: install Molecular Nodes from extensions.blender.org in the owner’s Blender and render the same structure in its cartoon style as the layer-3 reference.

4. What works today vs. what needs newblender

Section titled “4. What works today vs. what needs newblender”
  • Works today with the live bridge:
    • Network fetch and text parsing inside Blender’s Python; writing named attributes in bulk with foreach_set.
    • Every GN node in the plan exists in 5.2.2 (Instance on Points, Points to Curves, Curve to Mesh, Sample Index, Index Switch, Menu Switch).
    • mathutils.kdtree for bonding; evaluated-geometry checks as in tools/live/checks/chair_checks.py.
    • The real add-on tracks Blender versions closely (520.x for 5.2), so Path A is available today.
  • Painful today:
    • Building a style library as nodes.new/links.new calls; no text artefact to diff (L-006).
    • GN returns no “what did I build”: instance counts and curve lengths must be re-derived from the depsgraph (L-003).
    • Large MD trajectories in an interactive session: memory and frame-change cost; a headless batch profile would be the right home (L-011).
  • Not feasible today (for an agent alone):
    • Reproducing the add-on’s full cartoon algorithm (guide points, peptide-plane orientation, arrow heads for sheets) to its quality in a day. It is a known, specifiable algorithm, just long.
    • Molecular surfaces (solvent-excluded) at research accuracy without a library.

5. Verification plan (RFC 0001 R6, six layers)

Section titled “5. Verification plan (RFC 0001 R6, six layers)”
Layer Check Threshold
1 Validity Point mesh has no NaN coordinates and the declared attributes on the POINT domain 0 NaN; attributes {element, res_id, chain_id, is_backbone, sec_struct, b_factor} all present
1 Validity Realised style meshes clean 0 degenerate faces; 0 non-manifold edges on sphere and tube meshes
2 Spec Atom count equals the file 1CRN: exactly 327 points; 4HHB: 4,384 + 395 = 4,779 (or 4,384 + 174 if waters are excluded by spec: 172 HEM + 2 PO4)
2 Spec Residues and chains 1CRN: 46 CA atoms, chain set {A}; 4HHB: chains {A, B, C, D}
2 Spec Secondary structure transferred Points flagged helix exactly residues 7–19 and 23–30 in 1CRN
2 Spec Bonds plausible All bond lengths 0.9–2.1 Å; the 3 SSBOND pairs (3–40, 4–32, 16–26) bonded at 2.00–2.05 Å ± 0.05
2 Spec Coordinates preserved Max abs difference between file coordinates and point positions (Å → scene units at declared scale) < 1e-4
3 Reference Same structure through Molecular Nodes (Path A) Atom count equal; bounding boxes equal within 0.01 Å; centroid within 0.01 Å
4 Downstream Render readable and colour-correct Pixel sample on a known oxygen atom maps to the red element colour (hue within ±10°); no fireflies (< 0.01% > 20× local median)
4 Downstream Turntable loops Frame 0 and frame N+1 identical (PSNR ≥ 45 dB)
5 Appearance (warning) Vision model: “is this a small protein with two helices shown as ball-and-stick?” Warn on “no”
6 Taste (owner) Colour scheme, which residues to highlight, framing for a figure Owner sign-off

6. Where the human is still needed

Section titled “6. Where the human is still needed”
  • The scientific story. Which residues matter (an active site, a binding pocket, a mutation) and what to hide is a researcher’s judgement, not a rendering choice.
  • Honesty of representation. Smoothing, exaggerated radii and dramatic depth-of-field can mislead; a domain expert decides what is acceptable for a paper.
  • Style. Journal figure conventions, colour-blind-safe palettes and cover-art drama are taste.
  • Trajectories. Choosing which frames of a microsecond simulation tell the story.

7. Effort

Section titled “7. Effort”
  • First slice (live demo, ~1 day): the owner’s Blender shows an empty scene, then a cloud of 327 dots appears in the shape of crambin. Switching the GN style input turns the dots into grey/blue/red/yellow balls joined by sticks, then into a smooth backbone tube that visibly thickens along the two helices. Three yellow sulfur bridges are visible. A turntable plays; a Cycles still and a check report (327/327 atoms, 46/46 residues, helices 7–19 and 23–30, 3/3 disulfides) are saved in ~/newblender-data/demos/. A second run loads haemoglobin’s four chains with the same tree.
  • Full reproduction: of renders using the add-on: a few agent-hours per figure. Of the add-on itself (cartoon, surfaces, MD, density, UI): several hundred agent-hours plus a domain expert’s review time; render compute is small (minutes per still).

8. Risks and unknowns

Section titled “8. Risks and unknowns”
  • The add-on’s exact Python scripting API was not verified for this note (docs point to API pages; signatures unconfirmed).
  • mmCIF is the PDB’s primary format; very large structures are mmCIF-only, so Path B’s PDB parser does not scale to ribosomes.
  • Extension install needs Blender’s online access enabled; that is a one-time owner action.
  • Unit scale (Å vs metres) must be declared, or lighting and depth of field behave oddly.

9. Sources

Section titled “9. Sources”