Source code for peptdeep.rescore.fdr

import numba
import numpy as np
import pandas as pd




@numba.njit
def fdr_to_q_values(fdr_values: np.ndarray) -> np.ndarray:
    """convert FDR values to q_values.

    Parameters
    ----------
    fdr_values : np.ndarray
        FDR values, they should be
        sorted according to the descending order of the `score`

    Returns
    -------
    np.ndarray
        q_values

    """
    q_values = np.zeros_like(fdr_values)
    min_q_value = np.max(fdr_values)
    for i in range(len(fdr_values) - 1, -1, -1):
        fdr = fdr_values[i]
        if fdr < min_q_value:
            min_q_value = fdr
        q_values[i] = min_q_value
    return q_values





def calc_fdr(
    df: pd.DataFrame, score_column: str, decoy_column: str = "decoy"
) -> pd.DataFrame:
    """Calculate FDR values (q_values in fact) for the given dataframe

    Parameters
    ----------
    df : pd.DataFrame
        PSM dataframe to calculate FDRs

    score_column : str
        score column to sort in decending order

    decoy_column : str, optional
        decoy column in the dataframe.
        1=target, 0=decoy. Defaults to 'decoy'.

    Returns
    -------
    pd.DataFrame
        PSM dataframe with 'fdr' column added

    """
    df = df.reset_index(drop=True).sort_values(
        [score_column, decoy_column], ascending=False
    )
    target_values = 1 - df[decoy_column].values
    decoy_cumsum = np.cumsum(df[decoy_column].values)
    target_cumsum = np.cumsum(target_values)
    fdr_values = decoy_cumsum / target_cumsum
    df["fdr"] = fdr_to_q_values(fdr_values)
    return df



# wrapper
calc_fdr_for_df = calc_fdr




@numba.njit
def fdr_from_ref(
    sorted_scores: np.ndarray, ref_scores: np.ndarray, ref_fdr_values: np.ndarray
) -> np.ndarray:
    """Calculate FDR values from the given reference scores and fdr_values.
    It is used to extend peptide-level or sequence-level FDR (reference)
    to each PSM, as PSMs are more useful for quantification.

    Parameters
    ----------
    sorted_scores : np.array
        the scores to calculate FDRs,
        they must be sorted in decending order.

    ref_scores : np.array
        reference scores that used to
        calculate ref_fdr_values, also sorted in decending order.

    ref_fdr_values : np.array
        fdr values corresponding to ref_scores

    Returns
    -------
    np.array
        fdr values corresponding to sorted_scores.

    """
    q_values = np.zeros_like(sorted_scores)
    i, j = 0, 0
    while i < len(sorted_scores) and j < len(ref_scores):
        if sorted_scores[i] >= ref_scores[j]:
            q_values[i] = ref_fdr_values[j]
            i += 1
        else:
            j += 1
    while i < len(sorted_scores):
        q_values[i] = ref_fdr_values[-1]
        i += 1
    return q_values





def calc_fdr_from_ref(
    df: pd.DataFrame,
    ref_scores: np.ndarray,
    ref_fdr_values: np.ndarray,
    score_column: str,
    decoy_column: str = "decoy",
) -> pd.DataFrame:
    """Calculate FDR values for a PSM dataframe from the given reference
     scores and fdr_values. It is used to extend peptide-level or
     sequence-level FDR (reference) to each PSM, as PSMs are more useful
     for quantification.
    ``

    Parameters
    ----------
    df : pd.DataFrame
        PSM dataframe

    ref_scores : np.array
        reference scores that used to
        calculate ref_fdr_values, also sorted in decending order.

    ref_fdr_values : np.array
        fdr values corresponding to ref_scores

    score_column : str
        score column in the dataframe

    decoy_column : str, optional
        decoy column in the dataframe.
        1=target, 0=decoy. Defaults to 'decoy'.

    Returns
    -------
    pd.DataFrame
        dataframe with 'fdr' column added

    """
    df = df.reset_index(drop=True).sort_values(
        [score_column, decoy_column], ascending=False
    )
    sorted_idxes = np.argsort(ref_fdr_values)
    ref_scores = ref_scores[sorted_idxes]
    ref_q_values = ref_fdr_values[sorted_idxes]
    df["fdr"] = fdr_from_ref(df.score.values, ref_scores, ref_q_values)
    return df



calc_fdr_from_ref_for_df = calc_fdr_from_ref