VIRGE — Virtual Gel Electrophoresis

A restriction digest simulator: load a plasmid, pick enzymes, and see the fragment sizes plus a rendered virtual agarose gel.

VIRGE is a teaching tool. It exists to make the reasoning behind a digest visible — where an enzyme cuts and why, how fragment size becomes migration distance, what a band's brightness actually measures. The science is modelled rather than sketched, and the enzyme catalog is generated from REBASE rather than hand-typed, so what you see should hold up to scrutiny. But it models the DNA, not the reaction: VIRGE has no view on buffers, star activity or anything else that depends on what is in the tube besides the sequence. Treat it as a way to understand a digest before you plan one, not as the last check before you set one up.

Live at virge.fly.dev.

Run

npm install
npm run dev

To regenerate the enzyme catalog from REBASE (fetches the source file if absent), and to run the regression tests:

npm run build:enzymes
npm test

To deploy: fly deploy. VIRGE is static files behind nginx with no server component and no secrets — see DEPLOY.md for the image, the Content Security Policy and what was verified.

Features

DNA sources

44 sequences built from NCBI GenBank flat files, in 9 collapsible groups, with their annotations (844 features) rather than bare sequence:

Group	Contents
Cloning vectors	pUC19/18, pBR322, pACYC177/184, pBluescript SK(+)/KS(+), pGEM-3Zf(+), M13mp18 RF
Expression vectors	pTrc99a, pGEX-4T-1, pGEX-6P-1, pEGFP-N1, pEGFP-C1
Yeast vectors	pRS313/314/315/316, 2-micron circle
Plant vectors	pBI121, Ti plasmid
Phage genomes	λ, φX174 RF, T7, T4, M13 RF
Viral genomes	SV40, HPV16, HBV, adenovirus 5, HSV-1, VZV, EBV, vaccinia
Natural plasmids	pPCP1, RK2/RP4, F plasmid
Bacterial genomes	E. coli K-12 MG1655, E. coli O157:H7, B. subtilis 168
Organelle & chromosomes	human and yeast mitochondria, Arabidopsis chloroplast, yeast chromosome I

Sequences up to 62 kb are bundled; larger records (up to the 5.5 Mb E. coli O157:H7 genome) carry metadata only and are fetched from NCBI when selected, then cached for the session. That keeps whole genomes usable without carrying them in the bundle — digesting E. coli K-12 with NotI yields 23 megabase-scale fragments, the real pulsed-field experiment, in about 60 ms. See DEPLOY.md for what the bundle actually weighs.

Every accession was verified against NCBI before inclusion, and topology comes from NCBI rather than being assumed — with documented corrections where the deposited record disagrees with the molecule people actually handle (M13mp18 RF, pACYC177 and RK2 are circular despite linear deposits).

The catalog is restricted to what a restriction enzyme can actually cut. The build script rejects any accession NCBI reports as RNA, and requires any single-stranded genome to be named for its replicative form. SARS-CoV-2 was offered here until that check was written: NC_045512 is ss-RNA by its own LOCUS line, restriction endonucleases do not cut RNA, and VIRGE drew it a gel with the same confidence as a plasmid digest. φX174 and M13 are ss-DNA virions and are now named RF, which is both the duplex form and the one sold — φX174 RF cut with HaeIII is the classic small-fragment size marker.

Switching DNA keeps your lanes and recomputes every digest.

Not included: pET-28a, pcDNA3.1, pMAL, pFastBac, pLKO.1, psPAX2, pMD2.G, pX330, lentiCRISPRv2, pCAMBIA and similar. These are commercial/Addgene sequences with no clean GenBank deposition — searching NCBI for them returns unrelated fragments, not the vectors. Download the GenBank file from Addgene or your supplier and drop it in; annotations come along.

Finding DNA by name

A search box over the catalog matches common names, symbols and spellings — chickenpox finds VZV, λ finds lambda, GFP finds the pEGFP vectors, mtDNA finds the human mitochondrion, E. coli finds both genomes.

A query that matches nothing explains why, because that is usually the useful answer. Searching COVID 19 says SARS-CoV-2 is an RNA virus and restriction enzymes cut double-stranded DNA only; searching pET-28a says it has no clean GenBank deposition and points at file import. Those two cases cover most of what people look for and do not find.

The alias and exclusion lists are hand-written on purpose. Handing the query straight to NCBI is worse than useless here, measurably: nuccore's first five hits for "COVID 19" are Klebsiella pneumoniae plasmids (the pandemic is named in their isolation-source metadata), and its fifteen hits for "pET-28a" include a patent fragment, a clam mRNA and an uncultured-bacterium glucanase — but not the vector. A regression test enforces that no exclusion term names something the catalog actually carries.

Searching NCBI

Search NCBI queries nuccore and lists every candidate with its accession, length, topology and molecule type, and you pick. It never auto-loads a hit, for the reason above — the top result is frequently unrelated, and NCBI ranks by recency rather than relevance, which the results header says out loud.

Records a restriction enzyme cannot cut are disabled rather than offered: RNA is refused outright, and single-stranded genomes are marked as cuttable only as their replicative form. This is the same rule the sample build enforces, so a search cannot introduce what the catalog is not allowed to contain.

Bring your own sequence four ways:

• Drop a file — GenBank (.gb, .gbk) or FASTA, including multi-record files. This is what SnapGene, Benchling and NCBI export.
• Fetch by accession straight from NCBI (NC_005816, J01749.1, …). Uses gbwithparts, so CON/contig records resolve to real sequence.
• Paste GenBank, FASTA, or a bare sequence.
• Reload a saved one — imports are kept in a "Your sequences" group.

Topology is read from the GenBank LOCUS line rather than assumed, and can be overridden. Every import reports length, GC content, feature count and ambiguous bases, so a bad paste is obvious immediately.

Unreadable input is rejected rather than guessed at: a GenBank record with no sequence, or a file that isn't DNA, produces an explanatory error. (An early version scraped letters out of a contig record's prose and produced a confident, entirely wrong 7,195 bp "sequence" — there is now a regression test for exactly that.)

Annotated fragments

When the loaded DNA has annotations, every fragment is labelled with the genes and origins it carries — so the gel answers which band holds my insert? rather than just listing sizes. Digesting pBR322 with EcoRV + PvuII shows a 2,482 bp fragment carrying bla and the origin, and a 1,879 bp fragment carrying tet.

Enzymes — 203, from REBASE

Recognition sites, cut coordinates and commercial availability are generated from REBASE's Bairoch-format release, so they are authoritative rather than hand-typed. Every enzyme reports which suppliers carry it (200 of 203 are stocked by NEB). Includes:

• 37 Type IIS enzymes that cut outside their recognition site — BsaI, BsmBI, BbsI, SapI, AarI, BtgZI, FokI, MmeI — the basis of Golden Gate and MoClo assembly.
• Frequent cutters for RFLP, Hi-C and methylation work: MspI, HpaII, TaqI, RsaI, MboI, Sau3AI, DpnI, HinfI, HhaI, MseI, NlaIII.
• Rare and 8-cutters: AscI, PacI, PmeI, SwaI, SbfI, FseI, NotI, AsiSI, plus the homing endonucleases I-SceI, I-CeuI, PI-SceI and PI-PspI.
• Isoschizomer aliases, so searching Esp3I finds BsmBI and LguI finds SapI.

Filter by tier (Everyday / Common / All), search by name or alias, and filter by how often an enzyme cuts the loaded DNA:

• Any — the whole catalog. Leave it here when the absence of a cut is the answer you want: confirming BsaI has no internal sites before Golden Gate assembly, checking an enzyme won't cut your insert, or verifying a site you are about to introduce by primer isn't already present.
• Cutters — only enzymes that cut. Useful on small plasmids, where much of the catalog is dead weight: 84 of the 203 enzymes don't touch pUC19 at all.
• Unique (1×) — enzymes cutting exactly once, the usual cloning filter. Far scarcer than "cutters" suggests: pBR322 has 138 cutters but only 43 that cut once.

An enzyme you have already ticked stays visible under every filter, so a selection can't silently disappear.

Suggested digests

✨ Suggest proposes digests for a stated purpose, picked with the selector beside it:

Purpose	Aims for	Cares about
Diagnosis	~5 well-separated bands	resolvable, readable, mostly gene-sparing
Cloning	~2 bands	not cutting inside annotated genes, common enzymes
Fingerprinting	~12 bands	resolution and spread; gene-cutting irrelevant

It replaces the lanes rather than appending to them — the picks are chosen as a set that reads well together, so mixing them into whatever was already loaded gives a gel that is neither. A purpose with no suitable digest says so and leaves the existing lanes alone.

It rejects pairs that can't share a tube (the same temperature check the manual lane warning uses), collapses isoschizomers and no-op pairs by comparing the cuts they actually produce rather than enzyme names, and returns fewer than three options rather than padding the list with a bad one. On pBR322 the cloning setting proposes NdeI + PciI, EcoRI + SspI and ClaI + SspI — none of which touch bla or tet.

This replaced a version that scored band spread alone, and an audit showed why that was a toy: every suggestion on annotated DNA cut through a gene, one proposed an un-performable enzyme pair, every result sat pinned at 13–14 bands because an unbounded band-count reward swamped the band-count penalty, and isoschizomers were offered as if they were different experiments. Those properties are now pinned by tests.

Cutting a feature out

✂ Cut out picks a named gene or origin from the loaded annotations and finds enzyme sets that cut on both sides of it and nowhere inside — the bench task of lifting a gene off the backbone intact so it can be gel-purified.

It reports the excised fragment and the flanking DNA that comes with it: on pBR322, AvaI + HindIII takes tet off in a 1,396 bp fragment with 56 bp upstream and 149 bp downstream, and BamHI + BspHI takes bla off in 1,008 bp. Excess flank is what the scoring minimises, since that is what makes the band hard to tell from the backbone; fewer cuts and commoner enzymes are secondary. Pairs that cannot share a tube are rejected, as in Suggest.

Selecting enzymes by hand also warns when a cut lands inside an annotated feature — the "will this destroy my insert?" check. Both paths go through featuresCutBy(), so the suggestion engine and the manual panel cannot disagree about what counts as cutting a gene.

Methylation

Digests are computed in a methylation context, because a plasmid from a standard miniprep is not naked DNA:

• E. coli dam⁺/dcm⁺ (default) — Dam methylates GATC, Dcm methylates CCWGG. Enzymes whose site overlaps a methylated target are blocked, and the lane reports how many sites were lost.
• None — PCR product or synthetic DNA.
• Eukaryotic — CpG methylated.

Blocking is context-dependent, not blanket: XbaI is blocked only where its site forms TCTAGATC, and cuts normally elsewhere. The classic teaching case works out of the box — MboI, Sau3AI and DpnI all recognize GATC, but on a dam⁺ prep MboI is blocked, Sau3AI is indifferent, and DpnI requires the methylation.

Digest engine

Searches both strands (needed for non-palindromic Type IIS sites), handles circular topology including sites spanning the origin, IUPAC degenerate bases, overlapping sites, and multi-enzyme digests. Cut coordinates are tracked separately for each strand, so overhangs are derived rather than assumed.

Verified against published maps: pUC19 PvuII 2364 + 322, the 43 bp EcoRI–PstI polylinker fragment, λ/HindIII and λ/EcoRI fragment sets, and Type IIS coordinates (BsaI GGTCTC(1/5), FokI GGATG(9/13), MmeI's 2 nt 3′ overhang).

Gel

Log-linear migration calibrated to agarose percentage (0.7–2 %), with compressed rather than clamped migration outside the resolving range so oversized fragments still separate. The ladder prints dimmer than the sample lanes and is captioned as a size standard, and its choice sets the gel's zoom window. PNG export.

Pulsed field (CHEF)

A constant field runs out above ~20 kb: every larger fragment reaches the same limiting mobility and piles into one band. The Field selector switches to a CHEF pulsed field, where the field periodically changes direction and a molecule must reorient before it can move again — the time that takes scales with length, which restores size dependence out to megabases.

Three switch-time programmes set the resolving window (1–6 s → 10–150 kb, 10–60 s → 0.1–1 Mb, 60–120 s → 0.2–2.5 Mb), with two pulsed-field standards: λ concatemers (48.5 kb to 1.02 Mb) and S. cerevisiae chromosomes (230 kb to 1.53 Mb). Ladders are partitioned by mode, since a 1 kb ladder on a pulsed-field gel is meaningless; asking the assistant for a pulsed-field ladder switches the mode on its own.

Inside the window, migration is linear in size rather than in log(size). That is not cosmetic — it is the property that makes a ramped-pulse gel readable, and it is why the λ ladder prints as evenly spaced rungs here and would crush together on a log axis.

The difference on the experiment this exists for — E. coli K-12 cut with NotI, 23 fragments from 4.3 kb to 1.08 Mb: a constant-field gel resolves 2 bands across 27 % of the track, and the same digest at 10–60 s resolves 10 across the full track.

Neither mode is a physical simulation. Both are resolving-window models: the window comes from published protocol ranges and behaviour outside it is the same compression used everywhere else. The run conditions are drawn onto the gel, so a saved PNG still says how it was run.

Lanes are rendered as an accumulated intensity profile, not one rectangle per fragment. Each fragment deposits a Gaussian of stain whose mass scales with its length (fragments of a digest are equimolar, so signal ∝ length) and whose width grows as fragments get smaller and diffuse more. Overlapping fragments add up the way stained DNA actually does, which means the same code produces sharp bands for a plasmid digest and a genuine smear for a genomic one — where drawing hundreds of separate rectangles just overdrew into a solid block.

Two consequences worth knowing:

• Fragments beyond the resolving range pile into compression zones at the ends of the gel instead of disappearing. A NotI digest of E. coli shows the megabase fragments burned into the top of the lane, exactly as on a real gel.
• Lane brightness is scaled to a high percentile of the occupied rows and allowed to clip, mimicking a gel doc's exposure. Without this the compression zone — which holds most of the mass — would render everything below it black.

Two imaging controls in Gel Setup work like the ones on a gel doc:

• Exposure moves that reference point in stops (powers of two, 0.25×–5.7×), the way exposure compensation works on a camera. Turn it up to pull faint low-molecular-weight bands out of the background at the cost of burning out the bright ones; turn it down to resolve structure inside a saturated smear or compression zone.
• Contrast is the gamma of the transfer curve (0.25–1.6, labelled very flat → very hard). A flat curve pushes everything toward mid-brightness so faint detail survives; a hard curve steepens it so only strong signal registers and the background crushes to black. Exposure decides how much signal reads; contrast decides how sharply faint is separated from bright.

Double-click either slider to return it to its default. Both redraw the gel only — digests are cached, so dragging stays responsive even on a 4.6 Mb genome — and both are saved with a configuration.

Sticky ends and ligation compatibility

Selecting an enzyme shows the overhang it leaves and every other enzyme in the catalog whose end will ligate to it — BamHI, BglII, BclI, BstYI, DpnII, MboI and Sau3AI all leave GATC, so an insert cut with one drops into a vector cut with another. PstI/NsiI/SbfI share a 3′ TGCA end; XbaI/NheI/SpeI/AvrII share CTAG; AgeI/XmaI/BspEI/NgoMIV and others share CCGG.

Type IIS enzymes are correctly reported as having a variable overhang, since they cut outside their recognition site and the end is set by the flanking sequence — which is precisely what makes Golden Gate assembly programmable.

Reaction-compatibility warnings

Selecting enzymes with different incubation temperatures (e.g. EcoRI at 37 °C with SmaI at 25 °C) flags a sequential-digest recommendation — usually where people first meet the idea that two enzymes may not share a tube.

Out of scope: per-enzyme buffer activity tables (the "% activity in NEBuffer r1.1/r2.1/r3.1/rCutSmart" matrix), star activity propensity and heat inactivation conditions. These are supplier catalog data rather than properties of the enzyme: absent from REBASE, particular to one vendor's product line, and revised on that vendor's schedule. VIRGE teaches that a double digest needs a compatible buffer, and deliberately declines to say which — that answer belongs to whoever sold you the enzyme, and a stale or guessed matrix in a tool people trust would be worse than none. Check your supplier's own chart. See BACKLOG.md for what implementing it would involve, should the tool ever take on a bench-planning role.

AI assistant

A chat panel in the right-hand pane, backed by the Claude API. It answers questions about the loaded DNA and operates the app through tools rather than only describing what to do — it can read the current state, search the enzyme catalog with live cut counts, preview a digest, load a sequence, add lanes, and adjust the gel and imaging controls.

Ask "set up a digest that separates the resistance genes" on pBR322 and it checks the fragment annotations, then adds an EcoRV + PvuII lane splitting bla from tet.

Setup. Paste your own Anthropic API key into the panel (get one here). It is kept in that browser's localStorage and reused on your next visit; Forget clears both the key and the conversation. Until a key is entered the panel shows the prompt and nothing else in VIRGE is affected.

There is no server component — the browser calls api.anthropic.com directly, so the assistant works in a static build exactly as it does in dev.

Model. The picker in the panel header chooses between Opus 5 (default — most capable, best at multi-step setup), Sonnet 5 (balanced) and Haiku 4.5 (fastest and cheapest, weaker at planning a digest). The choice persists in localStorage and you can switch mid-conversation: the history is plain messages and tool results that any of them can pick up, though the next turn pays a prompt-cache miss. It is locked while a reply is in flight, so a single exchange never finishes on a different model than it started on. Since you are spending your own credit, the tradeoff is yours to make — a "which enzymes cut this once?" lookup does not need Opus.

Replies render as Markdown and LaTeX — headings, lists, tables, code blocks and maths — while your own turn stays plain text.

Two decisions worth knowing. The reply is sanitised, always: the model has just read a GenBank file you dropped in, whose labels and definitions are someone else's text, so <script>, event handlers, javascript: URLs, frames, forms and style attributes are stripped before the HTML reaches the page. And maths renders to MathML rather than KaTeX's default HTML, because that output writes inline style attributes — four on x^2 alone — which VIRGE's style-src 'self' policy blocks. MathML emits none, needs no webfonts, and browsers render it natively, so the strict CSP stands and ~1 MB of KaTeX font files never enters the bundle.

If no working sanitiser is available the reply is escaped and shown verbatim rather than rendered — the failure mode is ugly, not unsafe.

Token usage sits under the composer: 1,284 in · 312 out · 1,150 cached for the last request. Cache reads are shown because they are not counted in input_tokens and are most of a long conversation's input — the system prompt and tool schemas are cached behind a breakpoint, so omitting them would misreport what a turn costs. A turn is not one request, though: the tool loop resends the whole conversation each round, so the line reports the last request and the tooltip carries the turn total and how many requests it took.

Clear (above Send) drops the conversation the assistant remembers, so the next question starts from nothing — useful when you switch topic and don't want earlier turns steering the answer, or to stop paying to resend a long history. It leaves your key, the gel and its lanes alone; it is disabled when there is no conversation to clear and while a reply is in flight.

Understand the tradeoff before using it. Holding an API key in a page means anything else running in that page — a browser extension, or any script that gets injected — can read it. That is why the SDK calls the option dangerouslyAllowBrowser, and why the API demands an explicit anthropic-dangerous-direct-browser-access header. It is a reasonable choice for a local tool with your own key, and a poor one for a key you can't rotate or a shared machine. The panel says as much, shows only the last four characters back to you, and never writes the key anywhere but localStorage.

Implementation notes: the tool-use loop runs in the browser because the tools operate the live app, and parallel tool calls return in a single user turn. Refusals are checked before reading content, server-side fallbacks are requested with graceful degradation if the account lacks the beta, and the stable system prompt plus tool schemas are cached while the volatile state snapshot rides after the cache breakpoint.

Configuration library

Save the current setup (DNA, lanes, gel settings, methylation context) by name in localStorage; alphabetical listing, one-click load, two-click delete, and JSON export/import for moving libraries between machines. Configs built on a pasted sequence embed it, so they stay portable.

Structure

• scripts/build-enzymes.mjs — REBASE ingestion; curated catalog + metadata
• scripts/build-samples.mjs — fetches the built-in samples from NCBI as GenBank and parses their annotations
• src/genbank.js — GenBank/FASTA parsers, shared by the build script and the browser so the two import paths cannot drift apart
• scripts/test.mjs — science regression tests (152), checked against published restriction maps rather than this code's own output
• scripts/test-dom.mjs + scripts/dom-harness.mjs — UI regression tests (60). The harness parses the real index.html, installs the browser globals, and records canvas drawing instead of rasterising it, so a test can ask what was actually drawn. Every check corresponds to a bug that shipped: pulsed-field mode silently not engaging, Suggest appending instead of replacing, the import status naming the previous sequence, the Clear button wiped by an innerHTML re-render. Each was reintroduced and confirmed to fail the suite — the science tests catch none of them, and a screenshot passes for all of them
• src/data/enzymes.js — generated enzyme data (do not edit by hand)
• src/data/samples.js — GenBank sequences with group metadata
• src/enzymes.js — catalog helpers (overhangs, display, buffer checks)
• src/digest.js — both-strand site finding, methylation, fragments
• src/suggest.js — purpose-driven digest scoring, and excisionOptions() for cutting a named feature out in one piece
• src/gel.js — canvas gel renderer: constant-field and pulsed-field (CHEF) migration, ladders, compression zones, exposure and contrast
• src/main.js — UI state, config library, wiring
• src/assistant.js — AI assistant: key handling, tool loop, chat UI
• src/assistant-config.js — assistant system prompt and tool schemas
• src/markdown.js — Markdown + LaTeX rendering for replies, sanitised, with maths as MathML so the CSP needs no 'unsafe-inline'
• src/dna-search.js — curated name aliases, the "deliberately absent, and here is why" list, and the NCBI candidate search
• scripts/check-csp.mjs — fails the build if the source fetches a host the production CSP forbids (dev applies no CSP, so nothing else catches it)
• Dockerfile, nginx.conf, security-headers.conf, fly.toml — deployment

Contributing

Pull requests are not accepted — see CONTRIBUTING.md. Bug reports about incorrect biology are welcome as issues, and the MIT licence means you are free to fork.

Licence

MIT — Copyright (c) 2026 Andre Paquette. See LICENSE.

Bundled data is not covered by that licence and keeps its own terms: enzyme definitions derive from REBASE, Copyright (c) Dr. Richard J. Roberts, free for academic use; sequences come from NCBI GenBank.

Theme

Colours come from the skew-t project — near-black surfaces with a cyan accent — mapped onto this app's variables in the :root block of src/style.css (bg-void→bg, bg-panel→card, bg-raised→panel, bg-input→input, line→border, text-faint→muted). Semantic colours (--warn, --error, --ok) come from the same palette, and the previously hardcoded reds, ambers and greens now reference them — so retheming means editing that one block.

Buttons are neutral light grey (--btn-text and friends) rather than accented, which leaves cyan to mean this is data — cut counts, fragment sizes, the active sequence — instead of competing with every control on the page. The two button states that carry meaning keep their colour: disabled greys out, and the armed half of a two-click delete turns red.

The gel is monochrome, like a gel photographed on a gel doc — a single channel with nothing colour-coded. Because hue can no longer separate the lanes, brightness does: sample bands print at full strength and the ladder deliberately dimmer (which is also how a mass-adjusted ladder exposes beside a concentrated digest), backed up by the ladder's caption and the dashed divider so the cue is never carried by tone alone. An uncut plasmid keeps its italic uncut label where it used to rely on being orange.

Every tone — bands, labels, wells, the slab itself — derives from one TINT constant at the top of src/gel.js, so the gel is provably single-hue (verified by checking that all 1.66M canvas pixels have identical RGB channels). Set TINT to something like [176, 226, 255] to tint the whole gel cool instead of neutral grey.

Notes

The REBASE source file is fetched at build time rather than redistributed here.

λ/HindIII yields 7 fragments, not the 8 of commercial ladders — the 125 bp band comes from the λ cI857 strain, whose mutation creates an extra HindIII site absent from the wild-type GenBank reference.