The Traceability Ledger: Calibrated Genotype–Phenotype Inference (working draft v0.1.1)
Authors/Creators
Description
Predicting how a genetic variant affects a person is hard. The usual assumption is that it's hard only because our models aren't good enough yet, and that bigger, better models will eventually close the gap. This monograph argues that part of the gap is of a different kind: it can't be closed by better molecular models because the missing information isn't in the molecule. It lives further downstream, where a variant's effect depends on environment, on other genes, and on things we simply haven't measured.
To handle this, we built a tool called the Traceability Ledger. For any prediction, it separates uncertainty into two kinds: uncertainty that more evidence could resolve ("curable"), and uncertainty that no amount of better modeling can remove ("constitutional"). It also flags a third dangerous case: when a model is confident and self-consistent yet still wrong because the real cause is something it never accounted for. The practical payoff is that the tool tells you not just how uncertain a prediction is, but whether collecting more data would help — and what kind.
We test it on real data and report the result honestly. For the PTEN gene, it worked: the variants the tool flagged as "curable" were exactly the ones a lab assay could resolve. We then ran a pre-registered repeat on a second gene, TP53, and it did not work for an instructive reason. The method requires two independent predictors to disagree; on TP53, they agreed too closely, leaving nothing for the "curable" axis to measure. That is itself a finding: the method only works when your predictors genuinely diverge, and we can now estimate roughly where that boundary lies. So the central claim is supported by one gene and is undefined — not disproven — for a second.
Throughout, we are careful to mark what is mathematically proven, what is only illustrated on simulated data, what is supported by a single real gene, and what remains an open research program.
Files
traceability_ledger_monograph.pdf
Files
(852.4 kB)
| Name | Size | |
|---|---|---|
|
md5:3cff142ffe5d8a8ab29b11e3c4757789
|
4.4 kB | Preview Download |
|
md5:46d8f3affe40f448decbcad4409533cb
|
5.7 kB | Preview Download |
|
md5:06b64ef8511a20fac63d90475028443e
|
8.4 kB | Preview Download |
|
md5:ac265f2640264511f05610820ec91ec3
|
30.8 kB | Preview Download |
|
md5:1111dafeb25c9ccc5fe0d2ad80687902
|
262.4 kB | Download |
|
md5:189860abbdb045e570f78c03a0780a16
|
473.7 kB | Preview Download |
|
md5:1b7f40b65f4eed5b8a284594efed7033
|
67.0 kB | Download |
Additional details
Software
- Programming language
- Python
References
- 10.1038/s41586-021-04043-8 (EVE — Frazer et al., Nature 2021) 10.1038/s41586-025-10014-0 (AlphaGenome, Nature 2026) 10.5281/zenodo.8208688 (AlphaMissense data release) 10.1007/s10994-021-05946-3 (Hüllermeier & Waegeman, aleatoric/epistemic) 10.1016/j.cell.2017.05.038 (omnigenic model) 10.1038/s41588-018-0204-y (Giacomelli et al., TP53 mutational landscape, Nat Genet 2018)