`EpistasisPowerTransform`¶

EpistasisPowerTransform is a ready-made nonlinear epistasis model: it fits a Box-Cox-style power transform as the scale between the additive genetic score and the observed phenotype. It is the Sailer and Harms (2017) global-epistasis model, and a concrete subclass of EpistasisNonlinearRegression with the nonlinear function already supplied, so you do not pass a function of your own.

Observed phenotype against the additive score, with the fitted power-transform scale

The fitted red curve is the learned scale: it maps the stage-1 additive prediction (x-axis) to the observed measurement scale (y-axis), absorbing the global nonlinearity so the additive coefficients underneath stay interpretable.

When to use this model¶

Use EpistasisPowerTransform when you believe the measurement scale is a smooth, monotonic power-law-like distortion of an underlying additive landscape, and you want a parametric, identifiable transform rather than hand-writing one for EpistasisNonlinearRegression. For non-monotonic or otherwise arbitrary scales, use EpistasisSpline; for a monotone neural-style scale, use EpistasisMonotonicGE.

How it works¶

It runs the same two-stage fit as EpistasisNonlinearRegression: an order-1 additive model produces a per-genotype additive score, then the power transform f(x; lmbda, A, B) is fit so that f(additive.predict()) matches the observed phenotypes. The training-set geometric-mean reference is locked at fit time, so predictions on new genotypes reuse the same scale.

Constructor parameters¶

model_type (str, default "global"): Encoding for the additive design matrix: "global" (Hadamard) or "local".
lmbda, A, B (float | None, default None): Optional fixed values for the power-transform parameters. Leave as None to fit them; set one to hold it constant.

Workflow¶

from epistasis.models.nonlinear import EpistasisPowerTransform

model = EpistasisPowerTransform(model_type="global")
model.add_gpm(gpm)
model.fit()

y_pred = model.predict()         # phenotype on the observed (nonlinear) scale
y_linear = model.transform()     # phenotypes linearized onto the additive scale
r2 = model.score()
additive_coefs = model.additive.epistasis.values

Key methods and attributes¶

Member	Description
`fit(X=None, y=None)`	Two-stage fit: additive model, then the power transform.
`predict(X=None)`	Predicted phenotype on the observed scale.
`transform(X=None, y=None)`	Observed phenotypes projected back onto the additive scale.
`score(X=None, y=None)`	Pearson R^2 between observed and predicted.
`model.additive`	The internal order-1 `EpistasisLinearRegression`.
`model.minimizer`	`PowerTransformMinimizer`; holds the fitted scale parameters.

See the two-stage fitting guide for the shared mechanics, EpistasisSpline, and EpistasisMonotonicGE.

EpistasisPowerTransform¶