bioX Models Get API key
bioX

atlas-1

Answering

Human variants. It reads the DNA around a single-letter change — once as the reference has it, once with the change — and scores how likely that change is to cause disease.

variant effectGRCh38 cancermuscular dystrophy

What it does

score a changeone base swapped for another, anywhere in the genome, with 128 bases of context either side
rank a lista score per variant, so a list of candidates from a sequencing run comes back in order of concern
work on new genesmeasured with whole genes held out, so the number describes a gene it has never seen

Training data

variants1,370,524ClinVar, GRCh38, single-letter changes with a reviewed verdict
genes5,053at least twenty variants each
pathogenic170,958the rest, 1,199,566, are benign — so the model is weighted, not fed a class it can ignore
context257 basescut from the human reference held locally, not fetched per variant
Only records where a submitter stated the criteria they used. A verdict nobody stands behind teaches the model noise, and then flatters it on a test set carrying the same noise. Every window is checked against ClinVar's own reference base before it is kept: eight mismatches in 1.37 million, and a coordinate that disagrees is dropped rather than guessed at.

Measured

505 genes held out whole — 136,137 variants in genes the model never trained on. Reported as AUC, because seven variants in eight are benign and a model that always says benign is right 88% of the time while being worth nothing.

AUC, genes held out
hand-written features0.632
atlas-10.815

The features are the standard cheap ones: which base became which, transition or transversion, CpG context, local GC and repetitiveness. They are the bar, and atlas-1 clears it by 0.183 on genes it has never seen — 133,009 parameters, six passes over the data, seventy-eight minutes on a laptop.

An earlier version trained on three genes came out level with those features. Nothing about the network changed; it was given 127 times as much data. That is worth knowing before anyone reaches for a bigger model.

Per kind of variant

One number over all of ClinVar is mostly a statement about synonymous variants: they are 45% of the reviewed set and almost none of them are harmful. The kinds a clinical geneticist is actually stuck on are rarer and harder, so they are reported separately. Measured inside each class, on the same held-out genes.

kindatlas-1featuresvariants
splice_region0.7730.61711,044within eight bases of an intron edge
intronic0.6670.59333,123deeper inside an intron
utr30.6630.4601,746after the coding stretch
missense0.5950.55415,846one amino acid replaced
synonymous0.4240.51860,22427 harmful ones in the whole set
Read that table before the headline. Nonsense and splice-site changes are left out of it because 6,652 of 6,688 and 3,953 of 3,969 of them are harmful — there is almost nothing to tell apart, and a number there would be noise dressed as a result. For missense variants, dedicated protein-based predictors are far ahead of this model and worth using instead. Where this model earns its place is the rows at the top: the variants inside introns and beside splice sites, which those predictors do not score at all.

Per gene

The headline is an average over 505 genes. For a particular gene, held out whole and fitted on everything else:

geneatlas-1featuresvariants
DMD0.8930.6844,763Duchenne muscular dystrophy
BRCA10.8620.7584,403breast and ovarian cancer
BRCA20.8530.7686,877breast and ovarian cancer
TP530.8130.6321,473many cancers

Using it

A variant is written the way a VCF writes it: the chromosome, the position counting from one, the base the reference has there, and the base you are asking about. Send one or a list of up to 512.

curl -s https://<host>/v1/atlas/score \
  -H "Authorization: Bearer $BIOX_KEY" \
  -H "Content-Type: application/json" \
  -d '{"variants": [{"chrom": "17", "pos": 7675088, "ref": "C", "alt": "T"}]}'
field
score0 to 1. Higher means the change looks more like the ones labs have called pathogenic.
errorinstead of a score, when the position is not in GRCh38 or the reference base does not match what the genome has there. A coordinate from the wrong build is rejected rather than scored.
modelthe weights that produced it, so a result can be traced back.
A score is a ranking, not a verdict. It puts a list of candidate variants in order of concern so somebody can look at the top of it. It does not say what is wrong with a person, and it is not a substitute for a laboratory that is allowed to.
bioX Terms