MolView

molview.fly.dev

_{Four structures at once — 8,579 atoms across four independent cameras at 0.10 ms/frame. The definition panel carries the wwPDB validation summary and says why a 1984 model scores badly rather than only showing it three red dots, and the assistant compares all four in Markdown, reporting what the turn cost in tokens.}

A WebGPU molecular viewer for the RCSB Protein Data Bank. Browse and search the PDB, and display up to four structures side by side — including macromolecular assemblies of millions of atoms.

Everything runs in the browser. There is no server component: the deployment is static files, and the app talks directly to RCSB's public endpoints, doing all decoding, geometry generation and rendering on the client. Nothing you load or save leaves your machine except the requests to RCSB itself.

Requires a WebGPU-capable browser (Chrome/Edge 113+, Safari 18+).

Tested on Chrome only — desktop Chrome on macOS, and Chrome on iPadOS for the touch gestures. That is a statement about what has been exercised, not a claim that anything else is broken: Safari, Edge and Firefox are untested rather than unsupported, and the versions above are what each browser's WebGPU support requires rather than what has been confirmed working. Wider testing is outstanding.

Two areas are the most likely to differ and the least likely to fail loudly. Touch gesture handling refuses Safari's own non-standard gesturestart events to stop a pinch zooming the page instead of the molecule, and that path has only been reasoned about, never run in Safari. And WebGPU itself is young enough that limits, texture formats and shader compilation differ between implementations, so a rendering fault on another browser is more plausible than a logic one. A structure that renders wrong somewhere is worth an issue.

npm install
npm run dev

npm run build produces a static dist/, which is all a deployment is — see Deploying below.

What it does

Browse and search. Full-text search across the PDB with filters for experimental method, resolution, polymer content, source organism and release date. Results are ranked by relevance, recency or resolution. Typing a four-character PDB ID anywhere — the search box or the ⌘K palette — loads it directly.

A first view worth looking at. A deposited coordinate frame is an accident of the crystal, so MolView turns a structure to its own principal axes as it loads — longest extent across the pane, thinnest towards the camera — and fits it to the pane's real aspect rather than to a bounding sphere. 1BNA arrives as a double helix rather than as a circle seen down its axis, GroEL shows its two stacked rings. Shapes too round for their axes to mean anything, capsids in particular, are fitted but not turned. Reset view never re-orients: once you have moved a structure, the camera is yours.

O re-orients a pane to those axes at any time, which is the way back after turning a structure by hand — distinct from R, which returns to the deposited frame and its arbitrary orientation. The ⌘K palette also offers views straight down the X, Y and Z axes, for molecules with an axis of their own.

Up to four structures at once. Single, split (horizontal or vertical) or quad layouts. Each pane has an independent camera, representation, colour scheme and shading, or you can link the cameras so rotating one rotates all of them — the usual way to compare two structures.

Macromolecules. Coordinates arrive as BinaryCIF, decoded in a worker so the interface never blocks. The 2.44M-atom HIV-1 capsid (PDB 3J3Q, a 113 MB download) loads and renders in about ten seconds.

Biological assemblies. What a depositor deposits is the asymmetric unit, which is frequently not the biological molecule. MolView reads pdbx_struct_assembly_gen and applies the symmetry operators on the GPU, so every assembly the file declares is one dropdown away and costs a list of matrices rather than a copy of the atoms:

Entry	Asymmetric unit	Assembly 1	Cost
1HHO	2 chains, 2.4k atoms	tetramer, 4 chains	2 matrices
1EI7	2 chains, 2.8k atoms	68-chain disk aggregate	34 matrices
2BTV	15 chains, 49k atoms	900-chain capsid, 2.9M atoms	60 matrices

Above roughly 70,000 atoms in view the default representation becomes spacefill rather than cartoon: a ribbon is thinner than a pixel at that scale, and what survives is dark noise costing millions of triangles. Compared side by side — Photosystem I, the 26S proteasome, a CCMV capsid, the human ribosome — spheres won every time.

2BTV's complete icosahedral capsid builds in ~1.6 s from a 2.4 MB download and encodes a frame in 0.17 ms. Assembly 1 is selected automatically unless the expansion would be extreme (icosahedral entries with thousands of copies stay opt-in), and the asymmetric unit is always available as an explicit choice.

Representations are composable. A pane is an ordered list of components — each one a selection plus how to draw it — applied like layers, so the last component covering an atom wins. "Cartoon everywhere, spheres on chain A, sticks at the haem" is three rows, not a special case. Styles are cartoon ribbons with real secondary-structure cross-sections and β-strand arrowheads, backbone trace, ball-and-stick, licorice and spacefill. Colour by chain, element, secondary structure, residue type, B-factor/pLDDT, hydrophobicity, entity, nucleotide base, or an N→C rainbow — per component or inherited from the pane.

Lone ions are drawn at half their van der Waals radius. At full size a 2.05 Å manganese beside a 0.62 Å ribbon is the largest object in the scene, and in a tRNA the magnesiums swamp the molecule they sit in.

Nucleic acid bases come in four styles. Slab fits a flat box in the base's own ring plane, so a double helix reads as one rather than as two bare tubes. Ladder joins Watson-Crick partners into a single rung — pairing is geometric, a purine N1 within hydrogen-bonding distance of a pyrimidine N3 with the ring planes side by side rather than stacked, which finds 12 of 12 pairs in 1BNA and 146 of the nucleosome's 147 in 1KX5. Stubs reduces each base to a rod, which is what stays readable on a ribosome. None leaves the bare backbone. Anything unpaired keeps a stub under Ladder, so a single strand never silently disappears. Colouring by nucleotide base puts purines warm and pyrimidines cool, so a strand reads as its purine/pyrimidine pattern and the two halves of a pair always contrast.

Assembly copies can be tinted by symmetry operator, which is what makes an icosahedral capsid's facets and 5-fold vertices visible instead of undifferentiated mush. Shading comes as named presets — Studio, Soft, Flat, Plain — with the ambient-occlusion, shadow, outline and fog sliders still there underneath.

Shadows are marched through the depth buffer along the key light rather than rendered from a second camera: the pipeline lights once after rasterisation, and a shadow map would mean drawing every pane's geometry again. Sixteen texture loads per pixel buy the thing shadows are for here — telling a crevice from a shallow dip, and telling which of two strands passes in front. The limitation is inherent to the method: a caster that is off-screen or hidden behind something else cannot cast, because the depth buffer has never seen it.

NMR ensembles. Every model at once as backbone traces, which is the usual way to read the spread, or one at a time with the camera held still as you step through them. Each model becomes its own chain, so ribbons never run from one into the next, bond perception cannot join superposed models to each other, and model 3 selects a single member.

Selections use a compact atom-specification grammar:

/A,B	chains, by auth id
:1-140,200	residues by sequence number
:HEM	residues by component name
@CA,N,C,O	atoms by name
/A:1-140@CA	all three, intersected
model 3	one member of an NMR ensemble

Combined with and / or / not, parentheses, and category keywords — protein, nucleic, polymer, ligand, ion, water, hetero, helix, sheet, coil, backbone, sidechain, hydrogen, heavy. Juxtaposition means intersection, so protein /A is protein and /A. A malformed expression is reported inline and that layer is skipped; the rest of the scene still renders.

Measurements and contacts. Distances, angles and torsions by clicking atoms, with the value drawn in the viewport. Hydrogen bonds from a geometric heavy-atom criterion, restricted to what is actually visible and recomputed when the layers change. The haem Fe to proximal histidine NE2 in 4HHB measures 2.14 Å.

Interfaces. What touches what, which is the obvious question about an assembly and the one a viewer usually leaves you to answer by eye. Heavy-atom contacts within 4 Å are grouped by chain pair and ranked, and each pair can be framed or drawn as its own component in one click. On 4HHB it recovers the architecture unprompted: the α1β1 pairs strongest, the α1β2 sliding interface weaker, the α1α2 contact barely there.

Symmetry copies count too, which for an assembly is usually the whole point. Ferritin's asymmetric unit is one chain and touches nothing at all; build the 24-mer and five interfaces appear, the strongest at 203 contacts. Only the copies neighbouring the deposited coordinates are tested — in a symmetric assembly every distinct interface already appears there, so a 60-copy capsid costs 60 comparisons rather than 3,600. Bluetongue virus takes about 370 ms and finds 33 lattice contacts on top of the 28 within its asymmetric unit. The far side of such an interface is a matrix rather than a chain in the file, so only the near side can be selected, and the panel says which side it is reporting. It is still measured, though: the copy's spheres block the probe like any other neighbour, and the far half of the area is taken on the congruent arrangement the same operator's inverse produces, where both sides are real coordinates. HIV-1 protease, whose biological dimer is made by a crystallographic 2-fold and which therefore has no interface at all inside its asymmetric unit, comes out at 1,614 Ų.

Crystal packing, and telling it from biology. The assembly answers what the depositor believes; the lattice answers what the crystal actually did. They are different sets, and only one of them used to be reachable — ubiquitin generates a single assembly copy, itself, and reported no contacts at all while sitting packed against eight neighbours.

RCSB ships the cell and the space group but not the operators, so they are rebuilt from two or three generators per group, closed by composition, keyed by Hermann-Mauguin symbol. The expected multiplicity is stored beside each group as a checksum: a mistyped generator produces the wrong number of operators and the group is refused, instead of yielding a lattice that is plausible and wrong. It caught one — P 42 3 2 is absent for exactly that reason.

Keying by symbol rather than by International Tables number is not pedantry. The number does not fix the setting, and the checksum cannot catch a wrong one because the order is identical: 1AKE is P 21 2 21, whose 2-fold runs along b, and a number-keyed lookup returned the standard P 21 21 2 with its 2-fold along c. Ten of 300 sampled X-ray entries are in a setting the number alone gets wrong. The check that does catch it is geometric — a lattice neighbour must touch the model, never interpenetrate it — and it now runs across every entry tested: with the settings right, no entry has more than a handful of atoms inside 2.4 Å of a copy, and those sit on symmetry axes, where an atom is legitimately its own image.

The proof it works is 1HHP, whose biological dimer is a crystallographic 2-fold. Built from the deposited assembly operators it buries 1,614 Ų. Built independently from the space group and cell, with no assembly at all, it buries 1,614 Ų.

Neighbours are then tagged, and the tag plus the area is the evidence:

	biological interface	largest crystal contact
HIV-1 protease 1HHP	1,614 Ų	258 Ų
Haemoglobin 4HHB	816 Ų (α1β1)	353 Ų
Ubiquitin 1UBQ	none — it is a monomer	292 Ų
Myoglobin 101M	none — it is a monomer	399 Ų
Lysozyme 1LYZ	none — it is a monomer	552 Ų

MolView shows those numbers rather than a verdict badge. The separation is real but it is a tendency, not a threshold — large crystal contacts and small biological ones both exist — and a viewer that prints "biological" in green would be believed more than it deserves.

Each pair also carries the area it buries — Shrake-Rupley SASA of the two chains apart and together, reported as interface area (half the total change, the convention PISA uses). Contacts rank the pairs; the area is what a paper quotes, and the two disagree more often than you would expect, because a flat patch and a knob in a socket can touch equally and bury very differently. Haemoglobin comes out at 816–831 Ų for α1β1, 663–668 for α1β2 and 266 for α1α2, against PISA's 830–870, and the whole tetramer's surface at 24,713 Ų against a literature ~25,000. It is still a geometric criterion, not an energetic one. A 237k-atom ribosome finds 244 pairs in about 300 ms; areas are computed only for the pairs actually reported, since each is four SASA passes.

Each of those passes measures only the atoms the other side can reach. An atom whose probe sphere no partner sphere can touch has the same area apart and together and contributes exactly nothing to the difference, which is the whole answer — so dropping it is exact rather than approximate, and the occluders stay complete regardless. Haemoglobin's α1β1 is 816.4 Ų with or without the narrowing and takes 16 ms instead of 72. Bluetongue virus measures twelve pairs in 538 ms rather than 2,900.

Ligand contacts. What each bound ligand touches, grouped by the residue touching it, with the kind of each contact: hydrogen bonds, ionic pairs, aromatic stacking and van der Waals. Rings are found from the bond graph and tested for planarity rather than read from a chemical dictionary, since MolView has none — a flat six-ring stacks whatever a dictionary would call it.

Checked against structures whose sites are textbook. 1CBS puts retinoic acid on Tyr134 at 2.57 Å and Arg132 at 2.73 Å, the two residues UniProt annotates as its binding site. Benzamidine in trypsin lands its ionic contact on Asp189, the S1 aspartate that makes trypsin specific for arginine — which is why benzamidine, an arginine mimic, binds there at all. Myoglobin's haem reports His93 at 2.20 Å, the proximal histidine.

Two things it deliberately does not claim. Histidine is not counted as charged, because its imidazole is cationic only when protonated and the file does not say — 101M's His93 was being reported as an ionic contact when what it makes is a coordination bond to the iron. And a residue bonded to the ligand is reported as covalently linked rather than as a partner: GFP's chromophore is built from residues 65–67 and stays in the chain, so Phe64 and Val68 flank it through peptide bonds and are not binding it. Every criterion is geometric; a long list describes proximity, not affinity.

Fit to density, computed here. Real-space correlation per residue: density calculated from the model, correlated against the map already loaded, over each residue's own envelope. It exists because MolView holds both halves in the same client, and it covers the case wwPDB's per-residue report does not — those scores are polymer only, so a ligand, the part anyone actually doubts, has no published number. 1CBS's retinoic acid comes out at 0.91 over 1,441 grid points; the worst-fitting residues are Glu103 and Glu46, surface glutamates in solvent. Ubiquitin's worst three are Arg74, Gly75 and Gly76 — its flexible C-terminal tail, which is the answer a crystallographer would predict.

Read it within one structure rather than against a published threshold. Measured against wwPDB's own per-residue RSCC on four entries it agrees at r = 0.62 and runs about 0.2 lower: the calculated density is one isotropic Gaussian per atom rather than real scattering factors, occupancy is not yet carried on the model, and the map is what the volume server sampled rather than one computed to match. The envelope radius was swept against that agreement rather than assumed, and peaks at 1.8 Å. Negative correlation is the one absolute statement — the density is somewhere the model is not. A 137-residue entry takes about 100 ms.

Morphing between conformations. With two structures superposed, the mobile pane can be slid or played between its own conformation and the reference's, along the alignment the fit already computed. Adenylate kinase open to closed says in two seconds what its 7.13 Å RMSD cannot say at all. Identical chains move together, so a homodimer does not appear to break in half. It is a straight line through space and not a physical path — bonds stretch on the way through and no barrier is respected — and the panel says so, because an animation implies a mechanism whether or not one is claimed.

Pockets. Enclosed cavities, ranked by volume, by the LIGSITE buriedness scan: a grid point counts as buried when protein lies on both sides of it along at least four of seven axes, and connected buried points are a pocket. Each one comes with the residues lining it, a ready-made selection, and any ligand sitting in it.

That last column is the check on the method rather than a restatement of the input, because ligands and waters are excluded from the grid before the scan. Myoglobin's top cavity comes back at 410 ų containing HEM, lined by Phe43, His64, His97, Arg45 — the distal pocket, residue for residue. 1CBS puts retinoic acid in its largest, 1EMA puts the GFP chromophore in the barrel. It measures concavity, not affinity: a large pocket is not a druggable one.

Overlay. A pane can draw other panes' structures inside it, each keeping its own superposition, so two folds land on top of each other in one viewport rather than side by side.

Superposition. Align one pane onto another: sequence alignment decides which residues correspond, then the paired Cα atoms are fitted and the outliers pruned. Validated against known comparisons — 4HHB's identical alpha chains 0.30 Å, alpha vs beta 1.09 Å, alpha vs sperm whale myoglobin 1.07 Å.

The alignment it computes is shown rather than discarded: one cell per aligned residue, coloured by how far the pair ended up apart after fitting, in sequence order. An RMSD is an average, and an average of a bimodal distribution lies about both halves — adenylate kinase open against closed reports 7.13 Å, and the strip shows why: a rigid core in blue and a LID domain that has swung 18 Å, which is the entire finding. Clicking a cell flies both panes to that pair.

Shareable links. The whole session compresses into the URL fragment — a four-pane project including a 2.9M-atom capsid is about 1 kB of link. A fragment never reaches a server, so the link discloses nothing in transit. Opening one works both cold and pasted into a tab already running the app.

Projects. A session has a name, shown in the centre of the title bar and renamed by clicking it; New project clears every pane back to defaults. Save a session in the browser and reopen it later, or export it as .molview.json to move between machines. Saving an already-saved project updates it in place; Save as new branches it instead, so a session built on an existing project does not overwrite its parent. Saving never downloads a file; export is the only operation that touches the file system. Coordinates are not stored — entries are referenced by PDB id and refetched, so a two-pane project is about 2 KB.

Assistant. A panel beneath the canvas talks to any model on OpenRouter. It answers as a structural biologist would — Markdown with GitHub tables and LaTeX equations — and it can drive the viewer: loading entries, building representations from selections, measuring, colouring, switching assemblies, superposing panes. Replies come back as one JSON object carrying prose plus a list of actions; each action is re-validated against live state before it runs, so a wrong chain id becomes a stated rejection rather than a silent no-op. Thirty-three action types cover loading and layout; components, colour and palette; assemblies, ensembles, overlay and superposition; camera views, focus, reset, spin, clipping, lighting and background; measurement, hydrogen bonds, interfaces, ligand contacts and pockets; density maps, real-space fit and surfaces; wwPDB validation, a computed Ramachandran analysis and UniProt annotations; similarity search and predicted structures; and nucleotide styles. Every one is a reversible local view change, so they run without a confirmation step. That count is checked against the vocabulary at build time by npm run check:actions, because this sentence has already gone stale once while nothing failed. Clear discards the rolling history as well as the visible transcript, and aborts a request in flight.

What the actions did comes back as a RESULTS message, and a turn is re-entered once when it needs it. An action that asks a question — an interface search, a measurement — returns its answer only after the model has already spoken, so without a second pass the reply is "let me check" and the findings are left for you to interpret. A rejected action gets the same treatment, told to correct it or to say plainly that it cannot be done rather than describe it as done. Both happen once, never twice: it is a correction, not a loop.

Where the model supports structured outputs — Claude and GPT do — the reply shape is enforced by the API and the prompt does not restate the schema, which is both cheaper and stricter. A toggle in Settings falls back to describing the schema in the prompt, which is also what happens automatically for models that cannot be constrained.

The API key is typed into Settings, held in that tab's sessionStorage, and sent only to openrouter.ai — MolView has no server to send it to. It is never written into a project or a shareable link.

Each reply is followed by its cost — model · 1,340 in · 382 out, with reasoning tokens called out when the model bills for thinking it does not show. A turn starts at about 1,300 input tokens; what grows it is the rolling history of three exchanges, not the size of the structure, since the scene is counts and chain names rather than coordinates. Clear is what resets it.

Search results carry the same judgement in one chip: the entry's weakest metric, named and coloured together — density 7.0% in red, clash 5.5 in amber. Two cryo-EM models at the same resolution can differ fivefold in clashscore, and resolution is the tiebreaker everyone reaches for, so it is worth seeing which of a dozen entries for one protein is actually the better model before opening it. Computed models have no wwPDB validation and get no chip.

Validation. The wwPDB validates the whole archive, and the definition panel shows what it found: steric clashes, Ramachandran and rotamer outliers, fit to density where structure factors were deposited, backbone coverage for cryo-EM. Numbers rather than a grade — a green tick would invite the uncritical reading this exists to prevent — with a dot for the judgement and a sentence only when something is genuinely poor. A metric that could not be measured says why: 4HHB has no density fit because structure factors were not deposited in 1984. It is worth knowing that the same entry, the classic haemoglobin, has a clashscore of 142 where a modern structure at 1.74 Å sits near 2. The assistant sees the summary too, so it will mention a weak model unprompted.

Inspection. Hover or click any atom for its residue, chain and atom name. A click also cages the residue in magenta, drawn from its own bonds, so double-clicking to fly the camera in leaves you looking at something marked rather than at an anonymous part of the structure — which matters most on nucleic acids, where adjacent base slabs are identical. A single-atom residue, an ion or a water, gets a small cross instead of a cage. The sequence track is built from the loaded coordinates, so every residue you click actually exists in the scene. Chains can be hidden or focused individually.

Absence, drawn rather than skipped over. Where a depositor left residues unmodelled the ribbon stops and a dashed Catmull-Rom curve carries on to the next observed residue — the chain is continuous there, the coordinates are not, and neither a solid ribbon nor a bare hole says that. The same break is marked on the sequence track with the number of residues missing, and the chain header counts them, because a track that simply runs 34, 37, 69 reads as a continuous chain. The disordered histone tails of 1KX5 are the case to look at.

Which breaks are real is decided from numbering and geometry together, and neither alone would do. Author numbering is a convention that sometimes skips: trypsin is numbered by chymotrypsinogen alignment and jumps 34→37, 67→69, 125→127, 130→132, 204→209 and 217→219 with nothing missing at any of them, each across a flat 3.8 Å — so 3PTB used to be drawn with six dashed gaps that do not exist. The threshold separating the two comes from the files: across 49,852 consecutively numbered residues the CA–CA distance has median 3.80 Å and maximum 4.27 Å, so a numbering jump across less than 4.2 Å is a renumbering and gets a hairline on the track rather than a gap marker. A jump across more than that is a hole. Geometry alone catches the opposite case, a file whose numbering is consecutive across ends that cannot be bonded.

A Ramachandran plot, computed here. Phi and psi for every residue, drawn against reference contours, with one point per residue and a ranked list of outliers that selects and focuses on click. The entry panel reported wwPDB's outlier percentage for a long time without ever drawing the thing the percentage counts — a number says how many residues are wrong, the plot says which and where, and a cluster just outside a contour is a different problem from one point stranded in empty space.

Residues are judged against separate distributions for glycine, proline, pre-proline, Ile/Val and everything else, because they are genuinely different and scoring them all against one is the standard way to manufacture outliers that are not outliers. Because it is computed from coordinates rather than fetched, it works on a local file and on a predicted model, neither of which has a wwPDB report at all.

The contours are the part that makes this a validation tool rather than a scatter chart, and they are measured rather than transcribed: scripts/build-ramachandran.ts derives them from 900 X-ray structures at 1.5 Å or better, keeping residues with mean B-factor at or below 30 and no alternate conformation — 156,015 residues in all. npm run build:rama rebuilds them and rewrites the data module, which records what it was built from; coordinates are cached on disk, so changing a filter or the smoothing costs no downloads the second time. It agrees with wwPDB's own outlier percentage to a mean of 0.05 percentage points across fifteen entries, including 7CGO's 44,376 residues at 0.16% against 0.19%. The shapes were checked as well as the counts, since a contour can produce the right total from the wrong region: proline comes out with no permitted region at positive phi anywhere in its 32,400 cells, which its ring forbids and which nothing in the pipeline was told.

Alternate conformations. A residue modelled in two or three positions because the density supported more than one is underlined on the sequence track, and the pane can be switched between the conformers the file carries. Crambin at 0.54 Å (1EJG) is the case to look at: it is a mixture of two natural isoforms, so position 22 is Pro in conformer A and Ser in B and C, and 25 is Leu against Ile — switching changes the sequence, not just a side chain. A residue with no copy of the selected conformer keeps its own first one rather than disappearing, which is not a nicety: only 3 of that entry's 17 alternate positions have a C at all.

Figures that explain themselves. A colour key for the pane's scheme, drawn over the canvas and painted into screenshots and recordings — a legend that vanishes from the exported image is worse than none, since the export is the one that gets shared. Measurement labels can be dragged out of the way of what they measure. The camera's pivot can be moved onto whatever you are looking at without flying there.

Placing structures by hand. Superposition computes a placement; a pane can also be put into move mode, where the same drags turn and shift its structure instead of its camera. That is for judging a placement by eye — pushing two structures together to see whether a proposed contact is geometrically possible at all.

Turntable video. Any pane records to a WebM clip of one or more full revolutions. The camera is driven frame by frame rather than the screen being captured in real time, so a structure that renders slowly costs a slow export and still produces a smooth turn — and the clip is the length you asked for rather than however long the export took.

Clipping, set by eye. A front and a rear plane, perpendicular to the view, dragged on a small side-on schematic that shows where they sit relative to the structure — which is the thing a pair of sliders labelled in Ångströms cannot tell you. The pair defines a slab, so a section through a capsid stays a section instead of becoming a silhouette against everything behind it. The planes follow the camera, so rotating rotates the cut.

The planes belong to the structure rather than to the pane showing it. Overlay one pane into another and each keeps its own: section pane 1 and its structure opens up while the guest from pane 2 stays whole, which is the usual reason to put them together. Cutting both is a matter of setting clipping on both.

Saturation and intensity. Two sliders per pane adjust how strong the colours are without changing the scheme itself — 1 is the scheme as authored, 0 saturation is greyscale, above 1 pushes further from grey. Saturation defaults to 2 and intensity to 1: the schemes were authored to be distinguishable, and against a near-black canvas with fog and ambient pulling everything towards the background, "distinguishable" arrives looking washed out.

The asymmetry is measured, not tasteful. Saturation moves a colour away from grey, which increases the distance between colours; intensity is an exposure multiplier into a ceiling the composite clamps at 1. Minimum separation across the element palette on ordinarily lit geometry is 36 at 1/1, 41 at 2/1, and 0 at 2/2 with all eleven colours clipped. So 2/1 is better than the authored 1/1 on every palette measured, and 2/2 is worse than either. Vividness was never what cost anything; the exposure was. They are applied to the material colour as the geometry pass writes it, before lighting, so they retune the palette rather than the exposure and cost nothing: no geometry rebuild, and the frame stays at 0.10 ms while you drag.

Applying them there rather than in the resolve is what lets each structure keep its own. The resolve sees one pane-wide uniform per pixel, so a guest overlaid from another pane was given the host's palette — restyled, and stripped of the pin that keeps its ramp readable. The geometry pass is the only place each structure is drawn under its own settings. Overlay a B-factor-coloured pane into a chain-coloured one and drag the host to greyscale: the host goes grey, the guest keeps its ramp.

They are held at 1 wherever colour carries meaning rather than decoration, and the sliders say which case applies. The composite clamps to 0..1 with only a 1/1.05 gamma, so there is no rolloff and the bright end simply stops.

With the colour-blind-safe palette, those eight colours are spaced so the closest pair survives deuteranopia; at 2, every fragment facing the key light clips and #e69f00 and #f0e442 come out identical under simulation. Measured separation across the lit range: 24–60 units at 1, 0.0 at 2.

With a quantitative scheme — B-factor, pLDDT, RSRZ, geometry outliers, hydrophobicity, AlphaMissense, the N→C rainbow, and a Coulombic surface — the ramp flattens into a plateau. At 2, on ordinarily lit geometry, 47 of 100 steps become indistinguishable and values 13% and 33% along the scale render as the same colour; two of the four pLDDT bands collapse outright on the brightest fragments. A key that no longer maps to the picture is worse than a dull picture.

Strengthening a colour and preserving a distinction turn out to be opposite operations once there is a ceiling. Everything categorical — chain, element, residue, secondary structure, entity, base — is unaffected and still responds.

They reach a molecular surface, because a chain-coloured envelope has to match the cartoon inside it, and the colour key, because a legend that no longer matches the picture is worse in an export than no legend. They deliberately do not reach a density map: its blue, green and red identify which map and which sign of the difference, and desaturating those would erase the distinction the colours exist to make.

Colour-blind-safe palette. The default chain palette runs cyan, orange, purple, green, pink — and the green/pink pair is the classic deuteranopia collision, which makes a per-chain figure read as one colour to about eight per cent of men. One toggle in Settings swaps in Okabe and Ito's eight colours, which stay distinguishable. Eight rather than fourteen is the honest cost: a structure with many chains repeats sooner.

Take the view elsewhere. The active pane exports as a PyMOL .pml or ChimeraX .cxc script. Selections are recompiled from their parsed form rather than string-substituted, so not water and nested parentheses survive into a different grammar instead of quietly producing the wrong picture. What does not transfer says so in a comment: the camera orientation, and MolView's per-residue colour tables, which become the target's own nearest scheme rather than a thousand colour commands.

Finding structures from a structure. The search box answers questions you can phrase; two more services answer one you can only point at. By shape compares the assembly on screen against every assembly in the archive — 4HHB returns 1COH and 2HHB at the top, which is the right answer. By sequence takes the longest chain and reports the identity of each hit, so 1CBS finds its own family of retinoic-acid-binding proteins, including versions solved at better resolution. The two disagree usefully: a hit in one and not the other is generally the interesting one.

Predicted structures. AlphaFold DB alongside the PDB, searched by protein name, gene or UniProt accession. A prediction gets its own Definition panel, because the fields the experimental one is built from — method, resolution, citation, validation — do not exist for it, and showing them empty would suggest the model merely lacks metadata rather than lacking an experiment. What it has instead is confidence, of two kinds:

• pLDDT, per residue, read in AlphaFold's own four bands at 90/70/50 rather than as a smooth ramp. The bands are how the number is actually used, and a gradient hides exactly the two lines that matter.
• PAE, the predicted error between every pair of residues, as a matrix beside the structure. This is the one people skip, and it answers what pLDDT cannot: two domains can each be confidently folded and still be placed relative to each other with no confidence at all. Calmodulin is the case to look at — two dark blocks on the diagonal, a pale everything else. Drag across a block to focus those residues in the pane.

AlphaMissense rides along where it exists: a predicted pathogenicity for every possible substitution, averaged per residue, as a colour scheme. It says where a mutation would matter, which is a different question from where the model is confident — p53 shows it plainly, with the DNA-binding domain scored pathogenic and the disordered tails benign.

The ESM Metagenomic Atlas, reached by MGnify accession in the same field. AlphaFold covers UniProt — proteins from organisms somebody has named. The Atlas covers sequence read straight out of soil, seawater and gut samples, where the source organism was frequently never cultured and often is not known at all. That coverage is the point, not accuracy: ESMFold folds from a single sequence with no alignment, which is what made hundreds of millions of predictions tractable and also what puts it below AlphaFold2 on the same protein.

It needed the PDB reader, because the Atlas serves legacy PDB, and it needed its own way in: the existing route is UniProt search then accession, and searching UniProt can never produce an MGnify accession. The panel names ESMFold as the predictor and drops AlphaMissense entirely rather than reporting it missing — AlphaMissense is computed over UniProt, so a sequence nobody has named could not have one, and "no annotation for this entry" would read as a gap in the data instead of a category that does not apply.

One trap worth recording: ESMFold writes pLDDT on a 0-1 scale where AlphaFold writes 0-100. Everything downstream assumes the hundred, so a confident model reported itself as "0.9, very low" and painted itself in the worst band. It is normalised where the scale is known.

A command line. A line beginning with / in the assistant's composer runs as commands rather than going to a model — the same twenty-seven verbs the assistant has, executed by the same function, so a second vocabulary cannot drift from the first. Several at once, separated by newlines or semicolons:

/load 4HHB; /color chain; /view orient; /clip slab 25

It needs no API key and no account, which is half the point: the app is drivable by typing whether or not you have an assistant. /help prints the vocabulary, and an unrecognised verb suggests the nearest one.

What the residues are for. Everything else here is geometry; this is function. UniProt's positioned annotations — active and binding sites, modified residues, motifs, domains, mutagenesis results, natural variants — landed on the residues of the loaded entry, each with a selection so it can be framed or drawn rather than merely read.

The difficulty is entirely in the numbering, and it is a three-step chain: UniProt position → entity sequence position via rcsb_polymer_entity_align → auth_seq_id via auth_to_entity_poly_seq_mapping → a residue in the pane. Skipping the first step is the classic error and it fails silently. 1CBS is offset by one, and doing it properly puts its binding site on Arg132 and Tyr134 — 2.7 and 2.6 Å from the carboxylate of the retinoic acid, which is what those residues are there for. 101M's "proximal binding residue" comes out as His93, whose NE2 sits 2.20 Å from the haem iron: the coordination bond itself.

Validation, per residue. The entry summary says "9.5% rotamer outliers"; two colour schemes say which residues. Fit to density (RSRZ) shows how badly each residue matches its own experimental density, and Geometry outliers counts clashes and bond, angle and stereochemistry faults together. Both ramps are anchored to fixed thresholds rather than to each entry's range, so the same colour means the same thing in two structures, and the legend says where the thresholds are. A ranked list of the worst residues sits under it; clicking one flies there. Paired with the density map this is the whole loop — colour by fit, find the worst residue, look at the evidence under it.

Molecular surfaces. The envelope a molecule presents to the solvent, as a Gaussian surface over the atoms currently drawn or over any selection, blended so the cartoon stays visible through it. It carries the pane's colour scheme, so a haemoglobin surface shows where its four subunits meet rather than being one undifferentiated blob. It is a Gaussian surface, not a solvent-excluded one: there is no rolling probe, so the re-entrant saddles of a true SES are absent and the overall envelope is the same.

Coulombic colouring. A molecular surface can be coloured by electrostatic potential — red negative, blue positive — from formal charges on the ionisable side chains, the termini, the nucleic acid phosphates and the ions, with a distance-dependent dielectric. The nucleosome is the check: its histone core comes out overwhelmingly blue with the H2A/H2B acidic patch showing as a red spot, and the charge model totals −292 for 1KX5's DNA, which is exactly its 292 phosphates, against about +120 for the octamer.

It is called Coulombic rather than "electrostatics" for a reason. There is no Poisson-Boltzmann solve, no solvent screening and no ionic strength; it is read for gross character — which face is acidic, which groove binds DNA — and not for a number.

Experimental density. A model is an interpretation of an experiment, and MolView will show you the experiment. The Panes tool fetches the deposited map for the pane — 2Fo-Fc and Fo-Fc for X-ray entries, the EMDB map for cryo-EM — and contours it around the model. Contours are quoted in sigma, because that is the unit the argument is conducted in.

The default presentation is chicken wire, which is what a crystallographer reads a map in and is also the only see-through surface a deferred renderer gets for free; a blended solid surface is one toggle away. The difference map is off until asked for and then appears as both lobes at once, green where the data want atoms the model does not have and red where the model has atoms the data do not support — reading only the green half is the classic way to talk yourself into a ligand that is not there.

The map is fetched as a box around the model rather than as a whole unit cell, so the server applies the spacegroup and the density arrives where the molecule is. Density is then masked to within a few Ångströms of the atoms actually drawn, because an X-ray box contains the crystal packing too. Entries with no map say why: 4HHB was deposited in 1984 without structure factors, which is the same fact the validation panel reports as a missing density fit.

Open your own structure. Drop a file onto a pane, or open one from the Panes tool: mmCIF, BinaryCIF, or legacy PDB. Nothing is uploaded — parsing and rendering happen in your browser.

Legacy PDB matters because it is what most tools still emit. A refined model, a docking pose, a predicted complex: none of them could be opened at all until the reader existed, and everything MolView could show was bounded by what RCSB serves.

Anything computed from coordinates then works on it — representations, selections, measurements, hydrogen bonds, interfaces and buried area, pockets, surfaces, superposition and morphing, ligand contacts, chain gaps, alternate conformations, and the Ramachandran plot. That last one is the useful case: it is computed rather than fetched, so it gives outliers on a model wwPDB has never seen. What cannot work is anything keyed to an entry id — the Definition panel, the wwPDB validation summary, UniProt annotations, similarity search — and /validation says so rather than failing quietly.

The reader is checked against the format it replaces rather than against examples: building the same entry from both its PDB and its mmCIF and comparing atom count, residue count, residue kinds, secondary structure, polymer chains, the element histogram, Ramachandran totals, unmodelled gaps and space group, forty-nine entries agree exactly — twenty picked for their edge cases and twenty-nine drawn at random.

RCSB endpoints

Purpose	Endpoint
Molecular definitions and metadata	data.rcsb.org/graphql
Search	search.rcsb.org/rcsbsearch/v2/query
Coordinates (BinaryCIF)	models.rcsb.org/{id}.bcif
Coordinates (mmCIF fallback)	files.rcsb.org/download/{ID}.cif
Density maps (BinaryCIF)	maps.rcsb.org/{x-ray|em}/{id}/box/...
Shape and sequence similarity	the Search API's structure and sequence services

Predicted structures come from two more, both CORS-open:

Purpose	Endpoint
Predicted models, PAE, AlphaMissense	alphafold.ebi.ac.uk/api/prediction/{accession}
Name and gene to accession	rest.uniprot.org/uniprotkb/search
Metagenomic models (legacy PDB)	api.esmatlas.com/fetchPredictedStructure/{MGYP…}
Their PAE	api.esmatlas.com/fetchConfidencePrediction/{MGYP…}

AlphaFold is keyed by UniProt accession and nothing else, which is why the UniProt lookup exists. File URLs are taken from the API response rather than composed: the database version is in every filename and it moves — the widely copied _v4 URLs are already dead.

Per-residue validation comes from the same GraphQL endpoint, as polymer instance features rather than as a validation category — RSRZ, RSCC and OWAB as arrays along the entity sequence, faults as one position per residue. They are numbered in entity sequence, so auth_to_entity_poly_seq_mapping converts before anything is coloured.

Search returns identifiers only; a single batched GraphQL query then fetches the definition for a whole page of hits.

Architecture

src/
  rcsb/        GraphQL + Search clients, MessagePack, BinaryCIF, mmCIF text
               parser, wwPDB validation summary, density maps
  mol/         Structure model, element data, bond perception, colour schemes,
               selections, components, measurements, alignment, chain contacts,
               chain discontinuities, ligand contacts, pockets, density fit,
               space groups, loading worker
  gfx/         WebGPU engine, camera, geometry generation, WGSL shaders, text,
               first-view orientation, marching-cubes isosurfaces
  ai/          Action vocabulary and executor, prompt, OpenRouter client,
               reply parser, Markdown/KaTeX renderer
  viewer/      Controller bridging React state to GPU resources
  ui/          Application shell, panels, command palette, assistant
  state/       Zustand store, project serialisation, IndexedDB, share links

The structure model is structure-of-arrays: flat typed arrays for coordinates, elements, residues and chains, with atom and residue names interned into a shared string table. Loading happens in a worker and the whole model is transferred back zero-copy.

Rendering is deferred. Atoms are ray-traced sphere impostors — four vertices each, exact silhouettes and exact depth from the fragment shader — which is what makes millions of atoms tractable. Bonds are instanced cylinder meshes. Cartoons are a generated triangle mesh, swept along a Catmull-Rom spline through the Cα trace using Carson–Bugg peptide-plane frames, with cross-sections that vary by secondary structure.

Instance data lives in storage buffers rather than vertex buffers, which is what makes assemblies nearly free: a draw issues instanceCount x transformCount instances and the shader recovers atom = i % n, copy = i / n. A scene is a list of groups — one per assembly generator, since generators replicate different chain subsets — each holding its geometry once plus the matrices it repeats under.

All four panes render into one G-buffer via viewport and scissor rectangles, then a single resolve pass applies screen-space ambient occlusion, three-point lighting, depth-discontinuity outlines and fog. Because shading happens once, after rasterisation, an impostor sphere and a ribbon are lit identically and read as parts of the same object.

Occlusion radius, outline thresholds and fog density are normalised against scene size, so the same settings look the same on a 20 Å ligand and a 1000 Å capsid.

Labels are a separate pass after the resolve, blended straight onto the swap chain so they never touch the G-buffer. The font atlas is rasterised at runtime with Canvas 2D — no font ships, and the metrics match whatever monospace face the platform actually has. An opaque block reserved in the atlas lets the background pills go through the same pipeline as the glyphs.

Assistant replies are rendered with marked + KaTeX + DOMPurify. Two orderings matter: math is lifted out before Markdown parsing, because emphasis rules will otherwise chew through \frac{a}{b}; and KaTeX runs after sanitisation, writing into the already-cleaned DOM, since sanitising KaTeX's own output would strip the markup it depends on. Model HTML is dropped and non-http links are defanged. The whole renderer is dynamically imported, keeping KaTeX's 330 kB out of the initial bundle.

Interaction

Rotate	drag
Pan	shift-drag, or middle/right drag
Roll	alt-drag
Zoom	wheel
Focus a residue	double-click
Select	click
Command palette	⌘K
Search	⌘F
Panes 1–4	1 2 3 4
Reset view	R
Orient to the structure's axes	O
Auto-rotate	S
Link cameras	L
Toggle panels	[ ]

Known limits

• Secondary structure comes from the file's struct_conf / struct_sheet_range records. When a file carries none, a Cα-geometry heuristic stands in; it is not DSSP and will differ at the edges of helices and strands.
• One alternate conformation is drawn at a time, and which one is a per-pane choice rather than a per-residue one — there is no way to show conformer A of one residue beside B of its neighbour, and no way to draw both at once. Occupancy is not carried on the model either, so the 21%-occupied conformer counts the same as the 57% one everywhere it is used, including the density correlation. Atoms deposited at zero occupancy are dropped outright, which is why 1KX5's histone tails read as unmodelled: the coordinates are there and the occupancies are 0.00. The assistant is told which conformers an entry carries and which is drawn, so it will not describe one as the whole story, but there is no action for switching: that is a control, not a conversation.
• A chain discontinuity is classified from numbering and geometry together, and the classification can be wrong in one direction. A component inside a polymer chain that is not in the recognised residue lists — a fused chromophore like GFP's CRO, an unusual modified residue — ends the run rather than being counted across, so a real gap adjacent to one is not reported. That is deliberate: nothing is claimed where nothing can be known.
• The Ramachandran contours are measured from 156,015 residues, which is enough for the favoured contour and thin for the allowed one: the 99.95% level is cut from 7,587 prolines and 6,464 pre-prolines, where it is decided more by the smoothing than by the data. Cis and trans proline share one distribution rather than the two MolProbity separates. Outlier counts will also differ from wwPDB's wherever an entry deposits backbone at zero occupancy, since those atoms are dropped here and validated there — 1KX5 reads 3.98% against wwPDB's 7.62% for that reason alone, and 7.41% with them restored.
• The computed real-space correlation is not on wwPDB's scale and must not be read against published thresholds. It agrees with their per-residue RSCC at r = 0.62 and runs about 0.2 lower, for reasons given where the feature is described. Rank residues within one structure with it; do not quote it.
• Ligand contacts are found only within the deposited coordinates. A ligand sitting at a crystal contact has partners in a symmetry copy that are not reported, and charge is assigned to the protein or nucleic side only — an ionic contact requires the partner to be a known charged group, so a charged group on the ligand meeting a neutral partner is reported as a plain polar contact.
• Whole-structure bond perception is skipped above 250,000 atoms, so ball-and-stick on very large structures falls back to ligand connectivity. Cartoon and spacefill are unaffected.
• Picking is a linear scan over atom centres, repeated for each assembly copy whose bounding sphere the ray crosses; the hover interval widens with scene size to keep the cost off the frame budget.
• Secondary structure, hydrogen bonds and superposition are geometric approximations, not DSSP, an energetic H-bond analysis, or a structure-based aligner. They are good enough to look at and to reason from; they are not what you would cite.
• The PDB reader covers coordinates, secondary structure and the unit cell, but not REMARK 350, so a PDB file offers no biological assemblies where the same entry's mmCIF would. Legacy PDB also cannot express a structure past 99,999 atoms or a multi-character chain id, which is why RCSB serves no .pdb for the largest entries — those still need mmCIF.
• A pane opened from a local file has no id to refetch, so a project must embed its bytes to restore it. That is an opt-in, capped at 25 MB per file, and never included in a shareable link.
• Transparency is not depth-sorted. Isosurfaces are drawn back faces first and then front faces, which is the right order for one closed surface and enough for almost everything; two separate transparent surfaces in front of each other still blend in the order they were submitted.
• Molecular surfaces are Gaussian, not solvent-excluded: no rolling probe, so no re-entrant saddles.
• Recording produces WebM, the only format browsers encode natively, and the pane has to stay visible while it runs.
• The production Content-Security-Policy pins connect-src to an explicit host list, so an endpoint added without updating security-headers.conf works in development and is blocked once deployed. npm run check:csp diffs the source against that list and runs as part of npm run build, alongside npm run check:actions. Both guard things that fail silently rather than loudly, and both have to know which build they are in: the CSP applies only in production, while the README the action count is checked against exists only outside the image, since .dockerignore excludes *.md. That half is skipped there, and says so.
• Dragged label positions last for the session. A saved project restores its measurements by atom reference and gives them new ids, so the offsets have nothing to reattach to.
• Surfaces and density contours are generated on the CPU, so building one costs a stall — a fifth of a second for a small protein, a couple of seconds for a nucleosome — and it blocks the frame while it runs.
• A density map's opening contour is chosen against a triangle budget rather than fixed, so a map never arrives half-drawn; the level it settles on is the one the panel reports, and it is never lower than the one asked for.
• Contour levels are quoted in sigma, taken from the map header, which is the conventional value. The header has twice been seen to arrive carrying another map's sigma — a block named 2Fo-Fc holding the Fo-Fc value, which would put every contour at 2.5x its intended level — so it is checked against the samples and replaced only when it is off by more than a factor of two. A ten per cent difference is normal and is left alone: the server computes over its own sampled region, not over the returned box.
• Symmetry contacts are found only around the deposited copy — enough to see every distinct interface in a symmetric assembly, not a per-copy census.
• Lattice contacts need a space group symbol the operator table knows. It is keyed by symbol rather than by International Tables number, because the number does not fix the setting — 1AKE is P 21 2 21, whose 2-fold runs along b where the standard P 21 21 2 puts it along c. It covers 299 of 300 sampled X-ray entries; the miss was P 42 3 2, whose generators failed their own checksum and were dropped rather than guessed. Anything unlisted, and anything without a real cell (NMR, cryo-EM and predictions carry a placeholder 1 Å P 1), reports assembly contacts only.
• The assistant is only as good as the model behind it. Models that cannot be held to the reply schema mostly cope, since the shape is described in the prompt and fenced JSON is unwrapped, but a small one will emit prose where an action was wanted and may describe a change it failed to make. The rejection notices below a reply are the record of what actually happened.

Deploying

The repo carries a fly.io configuration. There is no server component, so a deployment is a Vite build served by nginx: no secrets, no environment variables, no volumes, no database.

Install flyctl and sign in:

fly auth login

Claim the app name — this reserves it and starts nothing, so it costs nothing:

fly apps create molview

If the name is taken it fails immediately; pick another and change app in fly.toml to match. Then, from the repo root:

fly deploy

That builds the Dockerfile on Fly's remote builder — Docker does not need to be running locally — and starts one machine. Every later deploy is the same command. The app is then at https://<app>.fly.dev.

HTTPS is mandatory, not a preference: WebGPU requires a secure context, so the app does not start over plain http. force_https in fly.toml covers it.

The machine sleeps when idle (auto_stop_machines = 'suspend' with min_machines_running = 0), so an unvisited deployment costs nothing and the first request after an idle period pays about a second of wake-up. Set min_machines_running = 1 if that matters.

To check the image locally before deploying:

docker build -t molview:test . && docker run --rm -p 8099:80 molview:test

The files involved are Dockerfile, nginx.conf, security-headers.conf, fly.toml and .dockerignore. DEPLOY.md explains what each one assumes — caching, the Content Security Policy and the hosts it allows, and why the token cost of the assistant is the only cost that scales.

Contributions

Pull requests are closed automatically. MolView is a personal project, so this is not a judgement on any particular change — nothing gets merged, and a PR left open would only waste the time of whoever wrote it. Fork it and take it where you like.

Issues are a different matter: a bug report, or a structure that renders wrong, is genuinely useful and gets read.

Licence

MIT. The dependencies are all permissive too — MIT, ISC, and DOMPurify under your choice of MPL-2.0 or Apache-2.0 — so a fork carries no copyleft obligation. Structures come from the RCSB PDB, whose data is in the public domain; MolView neither redistributes nor caches it.