Docs — RNAV8.ai

RNAV8.ai Documentation

RNAV8.ai is an open-platform mRNA design agent. It takes an ORF (coding sequence), a target cell type, and design preferences — and returns fully annotated, ranked mRNA construct candidates with scientific rationale.

The agent uses Claude's tool-use API to orchestrate six deterministic scientific tools in sequence, reason over their outputs, resolve conflicts, and synthesise candidates. All tools are built on open-source databases and algorithms.

ℹ This documentation covers the web app, Python SDK, and REST API. The core pipeline is identical across all interfaces.

Quick start

The fastest way to start is the web app — no setup required.

# Web app
# → app.rnav8.ai — paste your ORF, select cell type, click Design mRNA

# Python SDK
pip install rnav8

from rnav8 import MRNADesignAgent

agent = MRNADesignAgent()
result = agent.design(
    orf_sequence="ATGCTTCTC...TAA",
    cell_types=["t_cell"],
    nucleotide_mod="m1psi",
    cap="CleanCap",
    polya_length=120,
    optimization_goal="cpg_depleted",
)

print(result.recommended.cai)           # → 0.912
print(result.recommended.mfe_5prime)    # → -2.1 kcal/mol
print(result.recommended.full_sequence) # → full RNA sequence

Input parameters

Parameter	Type	Description
orf_sequence	string	DNA or RNA coding sequence from ATG to stop codon (inclusive). Accepts both T and U.
cell_types	string[]	Target delivery cell types. Drives codon table, miRNA panel, RBP profile, and UTR selection. See cell-type profiles.
nucleotide_mod	enum	`m1psi` · `pseudo` · `5mC` · `unmod`. Defaults to `m1psi`.
cap	enum	`CleanCap` · `ARCA` · `m7G`. Defaults to `CleanCap`.
polya_length	int	Poly-A tail length in nucleotides. Recommended: 100–150 nt.
optimization_goal	enum	`balanced` · `max_cai` · `cpg_depleted`
utr5_preference	string	UTR library ID or `"auto"` (agent selects). See UTR library.
utr3_preference	string	UTR library ID or `"auto"` (agent selects).
notes	string	Free-text constraints: e.g. "avoid BamHI site", "bicistronic with GFP", "maximize half-life".

Pipeline architecture

The orchestrator LLM receives your input and calls the six tools in sequence. Each tool returns a structured JSON result. The LLM reasons over all results — including conflicts — and calls synthesize_candidates to produce the final output.

The agent loop continues until stop_reason === "end_turn" or a maximum of 20 iterations. Typical runs complete in 6–8 tool calls with 2–3 reasoning steps.

Tool reference

validate_orf step 1

Validates the CDS. Checks ATG start, stop codon (TAA/TAG/TGA), in-frame internal stops, cryptic GT-AG splice sites, and uORF risk. Always called first.

Input	Type	Notes
orf_sequence	string	Raw DNA or RNA CDS

optimize_codons step 2

Codon-optimizes the ORF using a tissue-matched HIVE-CUT table (Athey et al. NAR 2017). Applies 5′ ramp effect (first 16 codons target ~0.55× max frequency) and CpG depletion. Returns CAI, GC%, and CpG count.

Input	Type	Notes
orf_sequence	string
cell_type	enum	Selects tissue codon table
goal	enum	max_cai · balanced · cpg_depleted

select_utrs step 3

Scores the curated UTR library by cell type match, expression boost, Kozak score, and uORF status. Returns the top-ranked 5′ and 3′ UTR sequences with full metadata.

predict_structure step 4

Predicts RNA secondary structure MFE of the 5′ region (cap + UTR + first 45 nt of ORF) using ViennaRNA RNAfold. The threshold for efficient 43S ribosome scanning is MFE > −4.0 kcal/mol. Hotspot windows below −6 kcal/mol are flagged.

scan_mirna_rbp step 5

Scans the 3′ UTR for (1) 7-mer miRNA seed matches from the cell-type–specific top-miRNA panel (FANTOM5 / miRNATissueAtlas2025), and (2) RBP binding motifs (HuR, TTP, CPEB) from ATtRACT. Returns risk score, resolved sequence, and stability score.

synthesize_candidates step 6

Assembles three full mRNA construct candidates: (A) balanced, (B) immune-evasion / max CpG depletion, (C) stability-focus with full HuR sites. Scores each with a composite function and returns ranked candidates with rationale.

Composite scoring

Candidates are ranked by a weighted composite score (0–1):

score = (
    0.30 × CAI
  + 0.20 × MFE_accessibility   # (MFE + 8) / 8, capped 0–1
  + 0.20 × RBP_stability_score
  + 0.20 × (1 − miRNA_risk_score)
  + 0.10 × GC_penalty_term        # penalise if GC% outside 45–55%
)

Weights are configurable in the Pro tier via API parameter scoring_weights.

Data sources

UTRdb 2.0

26M eukaryotic UTRs from 6M genes across 573 species. Annotations: uORFs, IRES, miRNA targets (TarBase v8), Rfam motifs, APA sites, m6A, RNA editing, PhastCons conservation.

utrdb.cloud.ba.infn.it · CC BY-NC · Lo Giudice et al. NAR 2023

HIVE-CUT

Human tissue-specific codon usage tables computed from GTEx mRNA-seq. 12 tissues: liver, muscle, brain, lung, blood, and more. Gold standard for cell-type matched optimization.

hive.biochemistry.gwu.edu · CC0 · Athey et al. NAR 2017

miRBase v22

Canonical miRNA sequence database. 2,654 human mature miRNA sequences. Seed regions (positions 2–8) used for 3′ UTR 7-mer scanning.

mirbase.org/ftp · CC BY 4.0

FANTOM5

492 short RNA-seq libraries across primary human cell types. Used to build cell-type–specific top-miRNA expression panels for the scanner.

fantom.gsc.riken.jp · CC BY 4.0 · De Rie et al. Nat Biotech 2017

miRNATissueAtlas2025

61,593 miRNA samples across 373 tissue types. Most comprehensive miRNA expression resource available. Supplements FANTOM5 for tissues not covered.

ccb.uni-saarland.de/tissueatlas2025 · open access

ATtRACT

370 human RBPs with 1,858 PWMs. Consensus motif scanning for HuR (stabilizing), TTP/BRF1 (destabilizing), CPEB (neuronal translation), MBNL1 (muscle).

attract.cnic.es · CC BY-NC · Giudice et al. 2016

ViennaRNA

Gold-standard RNA secondary structure prediction by free energy minimization. Python bindings via pip install ViennaRNA. Also supports LinearFold for long constructs.

github.com/ViennaRNA · academic free · Lorenz et al. AMB 2011

FAQ

Where is RNAV8 Bio located?

RNAV8 Bio is a US-based company. Our team operates in the United States, and all data processing occurs on US infrastructure. We comply with US data privacy regulations.

Is my sequence data kept private?

Yes. Sequences you submit are used only to run your design job. They are not stored, shared, or used to train any model. See our Privacy Policy for details.

Can I submit a protein (amino acid) sequence instead of DNA?

Yes. Paste a single-letter amino acid sequence (e.g. MKTLLLTLVVVTIVCLDLGYTPEEITQSPATLSVSPGETVTITCRASSSVSYMHWYQQKPGSSPKPWIYATSNLASGVPARFSGSGSGTSYSLTISRMEAEDAATYYC…) and the agent will automatically convert it to a codon-optimized ORF using tissue-matched codon tables before running the pipeline. You can also click the Protein / AA toggle in the input panel.

What does the composite score mean?

The composite score (0–1) ranks candidates across five dimensions: CAI (30%), 5′ MFE accessibility (20%), RBP stability (20%), miRNA risk (20%), and GC content proximity to 50% (10%). Higher is better. Scores above 0.70 indicate a well-balanced therapeutic construct.

How many runs do I get?

RNAV8.ai is free to use. After registering (no credit card needed), you can run as many designs as you need.

What cell types are supported?

Hepatocyte, skeletal muscle, T cell, dendritic cell, neuron, lung epithelium, and HEK293. Each uses a tissue-specific HIVE-CUT codon frequency table and a curated miRNA seed panel derived from primary human cell transcriptomics.

Can I use this for commercial research?

RNAV8.ai is free for internal research, including commercial pre-clinical work. For GMP or regulatory submissions, contact us for details.

REST API

Pro and Enterprise tiers have access to the REST API. Authentication uses Bearer tokens.

# POST /v1/design
curl -X POST https://api.rnav8.ai/v1/design \
  -H "Authorization: Bearer $RNAV8_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "orf_sequence": "ATGCTTCTC...TAA",
    "cell_types": ["t_cell"],
    "nucleotide_mod": "m1psi",
    "cap": "CleanCap",
    "polya_length": 120,
    "optimization_goal": "cpg_depleted"
  }'

⚠REST API is in private beta. Join the waitlist for early access.