petab1to2: rescale noise and priors when converting log10 to log

[FFroehlich]

PEtab v2 has no log10-based noise distributions or priors, so petab1to2 has to express them on natural-log scale. Until now it did that without rescaling: log10-normal noise became log-normal with the same noise formula, which changes the likelihood, and an estimated sigma silently changed meaning by a factor of ln 10. log10-laplace noise raised NotImplementedError.

Why ln 10. ln(y) = ln(10) · log10(y). If log10(y) is normal (Laplace) with standard deviation (scale) σ, then ln(y) is normal (Laplace) with ln(10) · σ. The two densities are identical: the residual is (ln s − ln m) / (ln(10) σ) = (log10 s − log10 m) / σ, and the normalisation term gets the same ln 10 factor from the Jacobian. Priors work the same way: log10(x) ~ N(μ, σ) is ln(x) ~ N(ln(10) μ, ln(10) σ), and likewise for Laplace (location and scale).

Changes (petab/v2/petab1to2.py)

• Observables: log10 + normal/laplace now map to log-normal/log-laplace, with the noise formula log(10) * (<formula>). The parentheses keep compound formulas intact, and noise placeholders are extracted as before. The old "Using log-normal instead" warning is replaced by a single warning that lists the affected observables. It points out that parameters in their noise formulas (e.g., an estimated sigma) keep their v1 meaning, i.e., they refer to log10(y); the ln 10 factor sits in the v2 noise formula.
• Priors: parameterScaleNormal/parameterScaleLaplace on log10-scaled parameters now map to log-normal/log-laplace, with both prior parameters multiplied by ln 10. log10-laplace previously raised NotImplementedError. The prior warning is removed as well.

Tests (tests/v2/test_conversion.py)

• Table level: log10-normal and log10-laplace map to log-*; the noise formula is log(10) * (...) and placeholders are unchanged.
• End to end: v1 and v2 calculate_llh agree (rel. 1e-12) after petab1to2, for normal and laplace noise, with a numeric sigma and with noiseParameter1_obs_B + offset (an estimated sigma via the measurement table).
• Priors: the v2 prior density matches the v1 Prior.pdf for log10/log/lin parameterScaleNormal/parameterScaleLaplace (rel. 1e-12, truncated at the bounds).
• The new log10 tests fail on main and pass with this change. pytest tests/v2: 87 passed, 1 skipped.

Benchmark collection (checked locally, with a fixed simulation per observable and time): v1 and v2 log-likelihoods now agree to ≤ 4e-15 relative for all six affected models (Bachmann_MSB2011, Lucarelli_CellSystems2018, Schwen_PONE2014, Borghans_BiophysChem1997, Elowitz_Nature2000, Perelson_Science1996). The log10 parameterScaleNormal priors in Bachmann, Isensee, Raimundez and Schwen match the v1 densities to ~1e-13.

The issue's repro now prints the same value for v1 and for v2 as converted (2.667567).

Closes #529

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.69%. Comparing base (1a83733) to head (62fd87d).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #535      +/-   ##
==========================================
+ Coverage   76.63%   76.69%   +0.06%     
==========================================
  Files          67       67              
  Lines        7527     7538      +11     
  Branches     1342     1344       +2     
==========================================
+ Hits         5768     5781      +13     
  Misses       1262     1262              
+ Partials      497      495       -2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

🟢 Approval recommended

The conversions are mathematically correct, preserve placeholder behavior, and have comprehensive regression coverage.

0 open findings

What changed in this PR

Corrects PEtab v1-to-v2 conversion so log10-based noise models and priors preserve their original probability densities.

Changes:

• Rescales log10 normal/Laplace noise formulas and prior parameters by ln(10).
• Adds table-level, likelihood, placeholder, and prior-density regression tests.

File	Description
petab/​v2/​petab1to2.py	Implements exact log10 noise and prior conversion.
tests/​v2/​test_conversion.py	Verifies distributions, formulas, likelihoods, and prior densities.

🧠 Review effort: Balanced

---

Give feedback about Copilot approvals in this survey to enter a drawing for a $150 gift card.

[dilpath]

👍

In cases where the v1 noiseFormula is an estimated parameter sigma, then it's unclear whether this is an improvement, since then the v2 sigma should be interpreted with its log(10) factor too (i.e., a user of the converted v2 problem might get confused about whether the sigma is for the log10- or log- noise distribution).

A warning for this could be useful.

[FFroehlich]

👍

In cases where the v1 noiseFormula is an estimated parameter sigma, then it's unclear whether this is an improvement, since then the v2 sigma should be interpreted with its log(10) factor too (i.e., a user of the converted v2 problem might get confused about whether the sigma is for the log10- or log- noise distribution).

A warning for this could be useful.

added warning