Nature varies one base at a time. So must the data.
Human variation is overwhelmingly single-nucleotide, and WGS and GWAS return it that way. A model trained on gene knockouts is one full resolution coarser than the question it is asked.
Gene-level, in cell lines
- Deletes whole genes — patients carry variants, not knockouts
- Runs in immortalized lines standing in for patient biology
- A trillion cells at the wrong resolution is still the wrong resolution
Variant-level, in primary human cells
- The same unit a WGS or GWAS hands you, and the same unit the model must predict
- Tested in the disease-relevant primary cell, not a stand-in line
- Breaks linkage — variants tested one at a time, at scale
Not one dataset. An Atlas at a time.
Each Atlas is a finished, licensable asset the day it completes — measured data and the models trained on it. Every one deepens CODEX.
Our AI reads everything
Models trained in-house mine the world's genetic evidence and nominate every variant worth testing. For most of the field, in-silico prediction is the finished product. For us it is step one.
Map the variants
Multiplexed CRISPR in the disease-relevant primary human cell. Runs with or without a partner.
Partner with the KOL
Where a cohort investigator is involved, they bring direction, credibility and clinical reach.
Test in parallel
10K variants — every program sharing a cell type runs at once.
Ship the Atlas
The reference causal variant-to-function dataset for that indication.
Build / train the AI
Models trained on that Atlas. License the insights in addition to the data, or instead of it.
Which variants would you test if you could?
A GWAS never functionally resolved. A platform that needs a causal layer. A model that needs ground truth.