Two ways to find a molecule in a genome.
Take a function you understand; search the genome for sequences that resemble it.
CPU-cheap — runs on a laptop, effectively $0 of compute.
Transparent — you can always say why a hit showed up.
A genome language model surfaces candidates from scratch — no known relative.
Can find what resembles nothing we already know.
Nothing to anchor to — the weakest-confidence tier.
Anchored to known biology — you can trust more of it up front.
Exciting by construction, uncertain by construction.
Receptor-targeted regulatory peptide
The honest starting label: a prediction, not a result.
"Computational predictions, pending wet-lab; structure-function framing only; not experimental evidence and not a disease, treatment, or efficacy claim."
Every candidate from genome mining is a starting hypothesis — not a result, however elegant the method.