Everything EH14 X can do, with a request you can paste into a terminal and the answer it gives back. No client library to install: it is JSON over HTTPS.
Confirm an e-mail address at /key and the key appears in the tab you asked from. Send it as a bearer token on every call.
export BIOX_KEY="your-key-here"
curl -s https://bioxcompute.com/v1/models \
-H "Authorization: Bearer $BIOX_KEY"Every answer carries the release that produced it, such as
EH14 X/1. Quote that in a paper and the same request returns the
same answer a year from now.
These routes returned a judgement on human single-letter changes. They have been
withdrawn and now answer 410.
Two different models carried the same name: the one serving this route reported 0.815, while EH14 X itself scores 0.3687 on variant effect — below the 0.5 of chance. Leaving both under one name made the number meaningless, and human variant interpretation is a different field from what this model is for.
Everything EH14 X is measured on, including what it fails at, is at eh14x.
How expected is every base of a stretch of marker-gene DNA? Low-surprise stretches are the ones evolution has refused to change, which is where a primer binds.
curl -s https://bioxcompute.com/v1/sequence/score \
-H "Authorization: Bearer $BIOX_KEY" \
-H "Content-Type: application/json" \
-d '{"sequence":"ATGCTTGTCTCAAAGATTAAGCCATGCATGTCTCAGTATAAGCTTTTACATGGCGAAACT"}'{
"length": 60,
"nll": 1.2225,
"perplexity": 3.3956,
"gc": 0.3833,
"probs": [[0.25, 0.25, 0.25, 0.25], [0.6615, 0.0646, 0.11, 0.1639], ...],
"conservation": [0.0, 0.5724, 0.3981, 0.4997, 0.5684, 0.912, ...],
"entropy": [1.432, 1.602, 1.501, 2.048, 1.093, ...],
"entropy_mean": 1.3375,
"placement": [{"cosine": 0.4283, "clazz": "Globothalamea",
"domain": "eukaryota", "phylum": "..."}, ...],
"placement_method": "4-mer profile against 60 reference groups",
"model_version": "EH14 X/1"
}conservation is the field to reach for. One number per base, high
where the model finds that position hard to vary. Those are the stretches that have
stayed the same across species, which is exactly where you put a primer.
placement_method says 4-mer profile and not a neural network, and
that is not an oversight. We trained a network for this and it lost to counting on the
same held-out taxa, so we serve the method that won. Both numbers are on
the EH14 X measurements.
| name | what |
|---|---|
| sequence | ACGT, up to 320 bases |
| sequences | a list instead, up to 200 per call |
| place | also return the reference groups it sits closest to (default true) |
| per_base | return the per-base tables (default true; set false for just the totals) |
| reading_frames | also score the six reading frames |
This is the one for unnamed sequence. Placement does not look anything up in a database, so a read that matches nothing still comes back with the company it keeps.
Continue a sequence, or write one from scratch for a lineage.
curl -s https://bioxcompute.com/v1/sequence/generate \
-H "Authorization: Bearer $BIOX_KEY" \
-H "Content-Type: application/json" \
-d '{"sequence":"ATGCTTGTCTCAAAGATTAAGCCATGCATG",
"num_tokens":200, "temperature":0.9, "top_k":4}'| name | what |
|---|---|
| sequence | optional prompt to continue |
| organism | optional lineage to write for, instead of a prompt |
| num_tokens | how many bases to write |
| temperature | higher is more varied (default 0.9) |
| top_k, top_p | the usual sampling controls |
How this works, precisely. The motif is found by counting k-mers across the target genus and the background, then improved one letter at a time. The model does not write it. Twenty-two bases that stick to that organism and to nothing else in the sample around it.
curl -s https://bioxcompute.com/v1/design \
-H "Authorization: Bearer $BIOX_KEY" \
-H "Content-Type: application/json" \
-d '{"genus":"Trypanosoma","rounds":6}'{
"genus": "Trypanosoma", "disease": "sleeping sickness",
"designed": {
"seq": "CACACTGCAAACGATGACACCC",
"dekking": 0.94, "achtergrond": 0.00136,
"gc": 0.545, "smeltpunt": 56.7, "bruikbaar": true
},
"found_by_counting": { ... },
"tested": false,
"use": "research; a candidate, not a validated assay"
}# without `genus`, it lists what it will answer for
{"targets": [{"genus": "Candida", "disease": "candidiasis"},
{"genus": "Trypanosoma", "disease": "sleeping sickness"}, ...],
"note": "Pass one of these as `genus`. Detection probes only: the corpus is
ribosomal RNA, which carries identity and no function."}Call it without genus to see which targets it answers for. It
answers for that fixed list and refuses everything else. The corpus it was
built on is ribosomal RNA, which carries identity and not function, so nothing about
how an organism works can be written out of it.
"tested": false is on every answer and it is not
boilerplate. What comes out is a candidate. Whether it binds in a tube is what your
laboratory is for, and we would very much like to hear what you find.Is this barcode read a PCR artefact or something real. This is the one task where EH14 X beats the standard tool in the field: AUC 0.9735 against UCHIME's 0.9442 on held out taxonomic groups, difference 0.0293 with an interval of 0.0208 to 0.0385. Counting 4-mers in the two halves sits at 0.5000 and shuffled labels at 0.4409.
Read the caveat that comes back with every answer. That comparison is at equal abundance and abskew 1.0, where both methods see only the letters. In a real sample UCHIME also has the abundance, and we do not model it.
curl -s https://bioxcompute.com/v1/sequence/chimera \
-H "Authorization: Bearer $BIOX_KEY" \
-H "Content-Type: application/json" \
-d '{"sequences":["ATGC...321 bases...","ATGC...321 bases..."]}'sequence or sequences — one read, or a list of at
most 200. A list counts as a single call.score — 0 to 1. At or above 0.5 the answer carries
"call": "chimera".250 to 332 bases. The head was trained on barcode length amplicons,
and the feature window is tied to that geometry. A read outside the range comes back
with "score": null and the reason, rather than a number we cannot stand
behind. For longer reads the head has to be retrained at that length.
Probabilities and embeddings as a base64 NPZ, the way the Evo 2 NIM returns them, so an existing pipeline can point at this instead.
curl -s https://bioxcompute.com/v1/sequence/forward \
-H "Authorization: Bearer $BIOX_KEY" \
-H "Content-Type: application/json" \
-d '{"sequence":"ATGCTTGTCTCAAAGATTAAGCC","output_layers":["probs"]}'Some data does not travel. Patient variants are the clear case: sending them to anyone else's server means a processing agreement and a legal review, and the review usually says no before the researcher has finished asking.
So the model travels instead. The part that ranks variants is 529 kB. It trains on a laptop in minutes, which means you can sharpen it on your own cases without a single record leaving the building.
What you get back is three numbers on the same held-out genes:
hand-written features the quick method anyone reaches for first
EH14 X as it is without having seen a letter of your data
EH14 X after once it hasThe middle number is there on purpose. Sharpening a model can also do nothing, or make it worse, and you want to see that rather than only the figure at the end. If it did not help, the report says so.
Your data needs five columns, and a sixth helps:
chrom,pos,ref,alt,label[,gene]
chr17,43092919,G,A,1
chr13,32339132,G,A,0label is 1 for pathogenic and 0 for benign. Without gene
the split cannot be made per gene, and the number at the end comes out more
optimistic than it should. The report warns you when the held-out set is too small to
mean anything.
| code | what it means |
|---|---|
| 401 | no key, or an unknown one. Get one at /key |
| 400 | the request is missing something; the message says what |
| 429 | over the hourly limit. Ask for more and say what for |
| 503 | that model is not loaded on this instance |
Every error is JSON with an error field in plain English. If one is ever
unclear, that is a bug and we would like to know.