7448 AI
Your gut is a chemical factory.
This predicts what it makes. Roughly thirty trillion bacteria live in your large intestine, manufacturing hundreds of chemicals that reach your bloodstream and organs. 7448 AI predicts which chemicals a community of bacteria will produce, from the bacteria alone.
Start from zero
Three words, one chain.
Microbiome: the community of bacteria living in your gut. Everyone has one, and they differ from person to person more than human genomes do. Two people might share only a third of the same bacterial species.
Metabolite: a small molecule produced by metabolism. When gut bacteria break down food, the leftovers are metabolites. Some are waste, some are fuel for your own cells, and some act as signals that influence your immune system, your metabolism and the nerves lining your gut.
Metabolome: all the metabolites present, taken together. It is the chemical output of the whole system.
A concrete example: when the bacterium Faecalibacterium ferments fibre, it produces butyrate. The cells lining your colon take most of their energy directly from butyrate rather than from blood sugar. Fewer of those bacteria generally means less butyrate, and low butyrate is one of the most consistently reported findings in inflammatory bowel disease.
Why predict it
Instead of measuring it.
Both can be measured directly. Sequencing tells you which bacteria are present; mass spectrometry tells you which chemicals are present. Sequencing is cheap and routine, while metabolomics is expensive, slow, and run by relatively few labs.
So most microbiome data in the world has the bacteria and not the chemistry. If the chemistry can be predicted from the bacteria with known accuracy, every one of those existing datasets gains a chemical readout it never paid for. That is the point of this project.
How it works
From bacteria to chemistry.
Step 01
Bacteria
A stool sample is sequenced, giving the relative abundance of each bacterial genus present.
Step 02
Model
A separate regression model per metabolite weighs each bacterium’s contribution. The weights are readable, so every prediction can be traced back to specific microbes.
Step 03
Chemistry
Out comes a predicted level for each of 1,098 metabolites, each with the accuracy it achieved on patients the model never saw.
Step 04
Context
Where taxa deviate from a healthy reference, published studies reporting the same shift in a condition are surfaced, with citations, not conclusions.
What it was built on
Fourteen independent cohorts.
Published studies where the same stool samples were both sequenced and run through mass spectrometry: the rare case where both halves of the answer exist for the same person.
- Paired samples
- 2,900
- Individual people
- 1,759
- Independent studies
- 14
- Metabolites predicted
- 1,098
- Median accuracy (r)
- 0.46
That median of r = 0.46 is a real but moderate correlation, and worth being plain about.418 metabolites are predicted well (r ≥ 0.5), while 185 are predicted poorly (r < 0.2). The interface marks which is which on every single value rather than presenting them as equivalent. It also does not yet transfer to labs it has not seen; see Inside the model.
Collaborate
Bring a cohort we have not seen.
The model does not yet transfer to new labs. Paired sequencing and metabolomics data from an independent cohort is the most useful thing a research group could share with us, and we will credit it in the work.
Simulator
Change the bacteria. Watch the chemistry move.
Adjust 24 key bacteria or start from a real, de-identified research sample. The model runs in your browser; nothing you enter is sent anywhere.
Open the simulator