Generate pvalues

Generate p-values for genes where larger effects are less likely to be false positives.

Parameters:

Name	Type	Description	Default
effects	Tensor	A tensor of effects	required
large_effect_percentile	float	The percentile of effects that are considered large effects. Defaults to 0.9	0.9
large_effect_upper_pval	float	The upper bound of the p-values for large effects. Defaults to 0.2	0.2

Returns:

Type	Description
torch.Tensor	A tensor of p-values

Raises:

Type	Description
ValueError	If effects is not a tensor or the values themselves are not numeric
ValueError	If large_effect_percentile is not between 0 and 1
ValueError	If large_effect_upper_pval is not between 0 and 1

Source code in yeastdnnexplorer/probability_models/generate_data.py

def generate_pvalues(
    effects: torch.Tensor,
    large_effect_percentile: float = 0.9,
    large_effect_upper_pval: float = 0.2,
) -> torch.Tensor:
    """
    Generate p-values for genes where larger effects are less likely to be false
    positives.

    :param effects: A tensor of effects
    :type effects: torch.Tensor
    :param large_effect_percentile: The percentile of effects that are considered large
        effects. Defaults to 0.9
    :type large_effect_percentile: float, optional
    :param large_effect_upper_pval: The upper bound of the p-values for large effects.
        Defaults to 0.2
    :return: A tensor of p-values
    :rtype: torch.Tensor
    :raises ValueError: If effects is not a tensor or the values themselves are not
        numeric
    :raises ValueError: If large_effect_percentile is not between 0 and 1
    :raises ValueError: If large_effect_upper_pval is not between 0 and 1

    """
    # check inputs
    if not isinstance(effects, torch.Tensor):
        raise ValueError("effects must be a tensor")
    if not torch.is_floating_point(effects):
        raise ValueError("effects must be numeric")
    if not 0 <= large_effect_percentile <= 1:
        raise ValueError("large_effect_percentile must be between 0 and 1")
    if not 0 <= large_effect_upper_pval <= 1:
        raise ValueError("large_effect_upper_pval must be between 0 and 1")

    # Generate p-values
    pvalues = torch.rand(effects.shape[0])

    # Draw p-values from a uniform distribution where larger abs(effects) are
    # less likely to be false positives
    large_effect_threshold = torch.quantile(torch.abs(effects), large_effect_percentile)
    large_effect_mask = torch.abs(effects) >= large_effect_threshold
    pvalues[large_effect_mask] = (
        torch.rand(torch.sum(large_effect_mask)) * large_effect_upper_pval
    )

    return pvalues