Design an mRNA sequence by letting physics settle it into place.

Pick a protein and a host cell. Every position is a categorical choice of synonymous codon; the machine writes down an energy — rare codons cost, GC content off-target costs, repeated runs cost — and relaxes toward the lowest-energy sequence with annealed block-Gibbs sampling. This is Extropic's published codon-optimization energy, run as a live instrument. Every gauge below is a real measurement of the sampler running in your browser.

E = w_f·Σ rarity(c_p) + w_GC·(GC−target)² + w_R·Σ repeat(c_p,c_p+1)

Protein

Positions optimized 200

First 200 of the 1273-aa SARS-CoV-2 spike. Slide to the full protein.

Host cell

Real codon-usage tables (python_codon_tables). Human = the mRNA-vaccine setting; the COVID spike optimized for expression in a human cell.

Energy weights

Target GC fraction 0.50

Codon-usage weight w_f 0.10

GC weight w_GC 1

log scale · 1 → 20000. Standard=1 (soft), Hard=2×10⁴ (GC enforced).

Repeat weight w_R 0.10

Sampler

Parallel chains 8

Anneal β (inverse temp) 0.30

Auto-ramps low→high while running (simulated annealing). Drag to steer heat by hand.

Adaptive GC coefficient

Linearizes the global GC term into a per-codon bias μ that adapts each step (Lagrangian) — avoids all-to-all coupling.

Re-anneal reheats to β_min and re-ramps — the chains never wipe, they keep their state and re-settle.

Best total energy

—

GC content

—

Sweeps

Flips / sweep

—

Overview — every codon, best chain GC 0 GC 3 just flipped

Live sequence — amino acid over its chosen codon

Click any codon above to inspect its live distribution — the sampler's actual conditional right now.

Total energy—

Energy termsusage gc repeat

GC fraction → targetμ

mRNA output — lowest-energy sequence found

—

What this is. A browser simulation of the identical algorithm — simulated-annealing block-Gibbs sampling on the Potts codon energy from Extropic's codon_opt, with real E. coli / human / yeast codon-usage tables. The same program was reproduced at scale on the JAX/THRML implementation over the full 1273-aa spike: best energy 444.11 vs the paper's 444.0 (Hard weights) and 233.92 vs 234.0 (Standard) — matching to the decimal.
What this is not. The headline energy advantage (Extropic estimate: 10⁵–10⁹× less energy than a GPU) rests on their prototype-hardware measurements — cited as theirs, not reproduced here. This page proves the algorithm and the solution quality, run see-through; the substrate advantage is their measured claim.