Science and resources
Beyond the polygenic score
A polygenic score says where someone stands among other people. We build models that go further, to the chance of a disease over a stated period.
Informational and Research Use Only. Our models estimate health risk. They do not diagnose or predict whether a person will develop a condition.
Many small effects, added up
Most common diseases are not caused by a single faulty gene. The genetic part of the risk comes from many thousands of variants, each with a small effect. A polygenic risk score adds up those effects, weighting each variant by how strongly published research has associated it with the condition, and compares the total with a reference population.
Beyond a score
A polygenic score places a person among other people. It does not say how likely a diagnosis is, or when. The same score can mean very different things at 40 and at 70, in men and in women, and in people with and without other risk factors. Over any period, the chance of a diagnosis also competes with the chance of dying of something else.
Our models start from the scores and go further.
| Question | A polygenic score alone | Our models |
|---|---|---|
| What it gives you | A rank among other people | The chance of a first diagnosis |
| Over what period | None | A stated period, from a stated age |
| Death from other causes | Not considered | Allowed for |
| What it needs | A genome | Any mix of a genome, health answers, age and sex |
| Without a genome | No result | An estimate from the other inputs |
| Uncertainty for one person | None | A range that reflects the inputs supplied |
| Ancestry | Depends on the reference group chosen | A continuous position, not a category |
How an estimate is made
An estimate combines several kinds of information into one number, then turns that number into risk over time.
- Inputs
- Polygenic scores from a person’s genome, answers to a short set of health questions, age and sex.
- One index
- The scores and answers are combined into a single number. Which inputs count, and by how much, is learned separately for each condition.
- Ancestry
- Ancestry is measured from the DNA as a continuous position, not a category, and each input is read within that position.
- Risk over time
- The index, age and sex become the chance of a first diagnosis over a stated period, for someone alive and free of the condition at the start. The model allows for the chance of dying of something else first.
- Partial inputs
- When an input is left out, the model averages over what it could have been, given the inputs it has, and the range widens.
- Uncertainty
- An estimate comes with a range where the model supplies one, so a reader can see how far the number could move.
- Combined estimates
- With consent, an estimate can add health and lifestyle data to genetics.
Trained on summaries, not individuals
Our models are fit to aggregate statistics, not to individual records. Person-level data stays inside the research environment it came from. Only summaries that meet the data holder’s disclosure rules come out, and each model is checked before release so that the summary tables behind it cannot be reconstructed from it.
How the models were evaluated
Before fitting, we set aside a share of participants. The models never saw them, and every choice about a model’s design was made without them. We then asked four questions of those held-out people.
- Does it rank people better?
- We compare each model with age and sex alone, and with the strongest single published score.
- Do the numbers match what happened?
- We compare estimated risk with observed risk, condition by condition, and check whether the ranges hold.
- What does each input add?
- We repeat the test with inputs taken away, to see what a genome and the health answers each contribute.
- Who does it work for?
- We look at results within genetic ancestry groups wherever there are enough people to do so.
Results differ by condition.
These were not clinical trials, and the models are not clinically validated. Results in one research cohort may not carry over to other groups. The evidence is in the whitepaper.
Interpreting the number
An estimate of your risk of developing the condition over a stated period, with your genetic score compared with a reference population. It is an estimate, not an exact probability, and not a diagnosis. Being in a high percentile means your inherited contribution is larger than most people’s in that reference group. Your environment, your behavior and many non-genetic factors also shape most outcomes.
What to keep in mind
- Ancestry
- The large studies these models are built on were conducted mostly in people of European ancestry. Calibration varies across populations. This is shared across the field and is an active area of research. On today’s models, the percentile in a report is unadjusted for genetic ancestry.
- One part of risk
- Genetics is one part of risk, and usually the smaller part, for most of these conditions. A high score does not mean you will develop the condition, and a low score does not mean you will not.
- Single genes
- A report does not test for rare, high-risk variants in single genes such as BRCA1 and BRCA2. No score on it, however low, rules them out, and it does not account for one you already know you carry.
- One research program
- Our results come from one large research program of volunteers.
- Length of follow-up
- Longer-range estimates extend the fitted model beyond the follow-up period in the evaluation.
- Risk levels
- How closely estimated risk matches observed risk varies by condition. Some risk levels are aligned to published lifetime figures.
- Data source
- Accuracy can vary with the platform a sample is read on and the coverage of a file. A report lists the conditions it scored.
- Models improve
- Your score is based on current research. As the models are refined, your score can change.
Regulatory status
Our models are Informational and Research Use Only. They are not FDA approved and are not intended for clinical or medical decision-making. Any change to your treatment should come from a qualified healthcare provider reading your report with you.
Resources
- Whitepaper
- Request the whitepaper
- Publications
- Papers on these models will be listed here as they are published.