GitHub - dhruvkjain/argus-ligand: finds features/motifs in DNA sequences

(i know that term ligand is used in checmical compounds but using rdkit on cloudflare workers :sigh: )

You can paste a sequence and either build a scan by hand or ask for what you want in plain English, and Workers AI turns that request into a scan.

One Cloudflare Worker, One Workers AI (Llama 3.1 8B, JSON mode), One KV Cache, written in Rust (via workers-rs). The same Rust program serves the web page and does the DNA scanning.

This is educational only. It is not a medical device, it works on example teaching sequences, and nothing here should inform any real health decision. Real clinical testing compares a person's DNA against a reference genome and databases of known variants, which is a different kind of pipeline.

Example: seeing sickle cell anemia

Sickle cell anemia is a disease which is identified with only one letter. In the beta-globin gene the codon GAG (glutamate) becomes GTG (valine): a single A to T change.

Two of the scans here show that change on an example sequence.

Healthy allele:

ATGGTGCACCTGACTCCTGAGGAGAAGTCTGCC

Sickle allele (the single base flipped, GAG to GTG):

ATGGTGCACCTGACTCCTGTGGAGAAGTCTGCC

Way 1, motif search : Search the motif CCTGAGG. It is present in the healthy allele and absent in the sickle allele. Searching CCTGTGG does the opposite. So one motif tells the two apart.
Way 2, restriction site (how it is really diagnosed) : The healthy sequence CCTGAGG is a cut site for the enzyme MstII, whose recognition sequence is CCTNAGG. The sickle mutation turns it into CCTGTGG, which MstII no longer cuts. So a restriction scan for MstII finds one site in the healthy allele and none in the sickle allele. "The enzyme stops cutting" is a genuine historical diagnostic, called an RFLP (restriction fragment length polymorphism).

Other conditions map onto the same ideas: many traits and disorders are single base changes a motif can spot, while repeat expansion disorders (for example Huntington's disease, caused by too many CAG repeats) are about counting a repeated unit.

A tini-tiny biology (just asked AI to explained in simpler terms)

DNA is a long string over four letters called bases: A, C, G, and T. It is double stranded. Each base pairs with a complement (A with T, C with G), so the second strand is the reverse complement of the first.

Soooo, a feature can sit on either strand, which is why most scans can look at both.

A motif is a short recurring pattern in the sequence that carries meaning, for example a spot where a protein binds.

Motifs are written in IUPAC codes, which add letters that stand for "any of these bases".

For example W means A or T, R means A or G, and N means any base. So the pattern TATAWAWR describes a whole family of related sequences rather than one fixed string. That is why a real promoter search uses TATAWAWR and not just TATAAA: the plain string would miss the common variants TATATAA, TATAAAA, TATAAAG, and so on.

The scanner supports these analyses:

Motif (literal or IUPAC). Where a pattern occurs, on either strand. The classic example is the TATA box, a signal found near the start of many genes.
ORF (open reading frame). A stretch from a start codon (ATG) to a stop codon in the same reading frame. ORFs are candidate genes; translating one gives the protein it would code for.
GC content and CpG islands. Regions unusually rich in G and C, measured with a sliding window. CpG islands (stretches with many CG dinucleotides) often sit at gene promoters.
GC skew. The running imbalance between G and C along the sequence. Where the cumulative skew turns around, it marks the likely replication origin in many bacterial genomes.
Restriction sites. Short specific sequences that a restriction enzyme cuts, for example EcoRI cuts GAATTC. Mapping these is a staple of cloning.
PWM (position weight matrix). A scored, probabilistic motif built from example sites. It suits fuzzy patterns like transcription factor binding sites, where no single fixed string fits every real occurrence.

A site that reads the same on both strands (a palindrome, like the EcoRI site GAATTC) is reported once with strand ± instead of being double counted.

Architecture

Project layout

📁 argus-ligand/
│
├── 📁 src/
│   ├── lib.rs            crate root: module wiring and re-exports
│   ├── error.rs          EngineError type
│   ├── types.rs          request and response JSON types (serde)
│   ├── sequence.rs       cleaning, reverse complement, Strand enum
│   ├── orf.rs            ORF finding and protein translation
│   ├── engine.rs         the scan orchestrator
│   ├── ai.rs             plain-English decomposition via Workers AI
│   ├── samples.rs        bundled real genomes, served from /samples
│   ├── worker_app.rs     HTTP handler + rate limit (wasm32 only)
│   │
│   ├── 📁 pattern/       pattern-matching scans
│   │   ├── mod.rs
│   │   ├── iupac.rs      IUPAC code to bitmask
│   │   ├── motif.rs      motif search over compiled masks
│   │   ├── restriction.rs  enzyme table + site mapping
│   │   └── pwm.rs        position weight matrix scoring
│   │
│   └── 📁 window/        sliding-window composition scans
│       ├── mod.rs
│       ├── gc.rs         GC-rich regions and CpG islands
│       └── skew.rs       GC skew (replication origin)
│
└── 📁 public/
    ├── index.html        UI markup (bento layout)
    ├── index.css         UI styles (terminal-brutalist theme)
    ├── index.js          UI logic
    │
    └── 📁 samples/
        ├── puc19.fasta   pUC19 cloning vector (2,686 bp)
        ├── pbr322.fasta  pBR322 plasmid (4,361 bp)
        └── lambda.fasta  phage lambda genome (48,502 bp)

Where does the data come from?

A sequence enters three ways:

You paste it into the box (raw nucleotides or FASTA; > header and whitespace lines are stripped automatically).
A built-in real genome from the picker. Three genuine records are bundled from NCBI GenBank and embedded in the binary: pUC19, pBR322, and phage lambda.
A tiny synthetic demo (the TATA-box and ORF buttons) for showing a specific pattern quickly.

Develop

# One-time toolchain
rustup target add wasm32-unknown-unknown
cargo install worker-build

# Test the pure Rust engine (native, no WASM, instant)
cargo test

# Run locally
npx wrangler dev

# Deploy to the internet (asks you to log in to Cloudflare)
npx wrangler deploy