Find the unusual,
in the geometry.
gAnalyze learns the geometric shape of normal across independent datasets, then ranks what does not fit. Its first run surfaced things worth a second look, before it was told what anything was.
Request accessRanked by distance from normal
Each mark is a lead the engine lifted out of millions. What they are, and why they surfaced, is what we show you on access.
Hover a lead
Genomics is one domain
The same engine runs wherever the same entities are measured across independent datasets. This is one instantiation. Others are in development.
What it has already shown
Tested against openly licensed public datasets, with no biological knowledge built in, the engine has already done three things worth saying out loud:
- It re-found textbook disease variants on its own. Several classical single-gene disease variants, established in the medical literature for decades, surfaced independently in two open datasets at once, without the engine ever being told what they were. Where one such variant did not surface, we report that too.
- Its alarms come with a guarantee, and the guarantee held. Every flag carries a mathematical ceiling on how often it may be a false alarm. Measured on millions of open records, the ceiling held.
- Its map of tissues reassembled anatomy unprompted. Given only regulatory measurements from dozens of human tissues, the engine's learned geometry grouped brain with brain and organ with organ, without being told what a brain is.
A full technical write-up with the methods, the numbers and every check behind these statements is in preparation as a preprint; the detailed claims will become public with it, not before. Results that build on consortium data stay in the private preview until each data owner has given written permission: the Psychiatric Genomics Consortium's Data Access Committee has granted ours in writing for the seven datasets concerned, and the remaining requests are with their owners.