Two models. Neither is general and neither tries to be — each was built because one question kept going unanswered. Here is what those questions are, what the models can do about them today, and what they are still reaching for.
Most life on Earth has never been given a name, and the tools we built for biology can only answer with names they already have.
Take a litre of seawater and sequence what is in it. Below 500 metres, 93.8% of the sequences in public ocean data match no named species. The usual method is to look each read up in a reference database, which means everything unnamed comes back as nothing at all. The largest part of the living world is invisible to the instrument we use to look at it.
noctua-1 exists to answer where there is no name. It reads a stretch of ribosomal RNA gene and places it among sixty reference lineages by the way the sequence is written — which four-letter words it uses and how often — rather than by what it matches. A read that is in no database still comes back with the groups it sits closest to. It also scores how expected each base is, which is how you find the stretches that stay the same across species: the places where a primer will bind.
Forty groups held out of training entirely, with 400 known and 400 novel queries, read without a taxonomy prefix — the way a pasted sequence actually arrives. Give it a 150-base fragment instead of a full-length read and placement falls to 0.347: length matters more than anything else you can change.
The placement you get today is four-letter-word counting, not a neural network. We trained one — ten million parameters, 1.04 billion bases of ribosomal RNA, nearly half of it environmental reads carrying no name — and on the same held-out split it placed 0.423 against counting's 0.820. It lost, and we serve the method that won. What the model is reaching for is the thing counting cannot do: reading a sequence as a sequence, so it can say something about a gene nobody has studied in an organism nobody has described. We are not there. The number above is what we have to beat, and it is ours.
A diagnosis stalls on a list of single-letter changes that nobody can rank.
Sequence a person and you get back hundreds of places where their genome differs from the reference by one base. For most of them no laboratory has ever ruled on whether they matter — variants of unknown significance is the phrase, and it is where a great many searches for an answer simply stop. Not because the work is impossible, but because there are more candidates than there are human hours to spend on them.
atlas-1 exists to put that list in order. Give it a position and one letter changed; it gives back how likely that change is to cause disease, so the hours go to the top of the list. A second call tells you what it saw: whether the change makes or destroys the GT and AG signals that splicing uses to find the edge of an intron, how far it sits from that edge, and which neighbouring bases it reacts to most. That last part matters more than the score — a number nobody can check is worth very little.
Whole genes held out, never random rows: variants inside one gene are not independent of each other, and a random split flatters any model. The baseline is nine hand-written features — GC content, CpG, transition or transversion, distance to a splice edge — the quick method anyone reaches for first. Also measured: BRCA2 0.853, TP53 0.813, MLH1 0.807 for Lynch syndrome, CFTR 0.799 for cystic fibrosis. Research use. It orders candidates; it does not diagnose anyone.
atlas-1 learned from 1.37 million variants that clinical laboratories had already ruled on. That is why it works, and it is exactly where it stops: no verdict, no answer. Most of the genome has no verdict, most genes have almost none, and no other species has any. What it is reaching for is to stop needing them — to judge a change by reading the sequence around it, the way you can tell a sentence is wrong without being told. We are measuring that directly now, on the same variants, and so far the honest answer is that reading alone is not enough yet.
One key, both models, and a person reads every request.
Everything runs through the API — there is no console and nothing to configure. One click and the key is on your screen: no account, no password, no e-mail to wait for. The free tier is enough to try both models and to run a small study; if you need more, ask and say what for. bioX is not for profit, and access is granted rather than sold.