# -----------------------------------------------------------------------------

# Copyright (C) 2008-2022 Yunshun Chen, Aaron TL Lun, Davis J McCarthy, Matthew E Ritchie, Belinda Phipson, Yifang Hu, Xiaobei Zhou, Mark D Robinson, Gordon K Smyth

# Copyright (C) 2024 Maximilien Colange

# This program is free software: you can redistribute it and/or modify

# it under the terms of the GNU General Public License as published by

# the Free Software Foundation, either version 3 of the License, or

# (at your option) any later version.

# This program is distributed in the hope that it will be useful,

# but WITHOUT ANY WARRANTY; without even the implied warranty of

# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the

# GNU General Public License for more details.

# You should have received a copy of the GNU General Public License

# along with this program.  If not, see <https://www.gnu.org/licenses/>.

# -----------------------------------------------------------------------------

# This file is based on the file 'R/exactTestByDeviance.R' of the Bioconductor edgeR package (version 3.38.4).

import numpy as np

from scipy.stats import nbinom

from .binomTest import binomTest

from .edgepy_cpp import compute_unit_nb_deviance

from .exactTestDoubleTail import exactTestDoubleTail

def exactTestByDeviance(y1, y2, dispersion=0.0):

    """

    Compute genewise *p*-values for differences in the means between two groups of negative-binomially distributed counts.

    This function uses the deviance goodness of fit statistics to define the

    rejection region, and is therefore equivalent to a conditional likelihood

    ratio test.

    See also

    --------

    exactTest

    Arguments

    ---------

    y1 : matrix

        matrix of counts for the first of the two experimental groups to be

        tested for differences. Rows correspond to genes and columns to

        libraries. Libraries are assumed to be equal in size -- *e.g.* adjusted

        pseudocounts from the output of :func:`equalizeLibSizes`.

    y2 : matrix

        matrix of counts for the second of the two experimental groups to be

        tested for differences. Rows correspond to genes and columns to

        libraries. Libraries are assumed to be equal in size -- *e.g.* adjusted

        pseudocounts from the output of :func:`equalizeLibSizes`.

    dispersion : array_like of floats

        an array of dispersions, either of length one or of length equal to the

        number of genes.

    Returns

    -------

    ndarray

        array of genewise *p*-values, one for each row of :code:`y1` and :code:`y2`

    """

    y1 = np.asarray(y1)

    y2 = np.asarray(y2)

    if y1.shape[0] != y2.shape[0]:

        raise ValueError("Number of rows of y1 not equal to number of rows of y2")

    ntags = y1.shape[0]

    if np.isnan(y1).any() or np.isnan(y2).any():

        raise ValueError("NAs not allowed")

    n1 = y1.shape[1]

    n2 = y2.shape[1]

    if n1 == n2:

        return exactTestDoubleTail(y1=y1, y2=y2, dispersion=dispersion)

    dispersion = np.asarray(dispersion)

    sum1 = np.round(y1.sum(axis=1))

    sum2 = np.round(y2.sum(axis=1))

    if (dispersion == 0).all():

        return binomTest(sum1, sum2, p=n1 / (n1 + n2))

    if (dispersion == 0).any():

        raise ValueError("dispersion must be either all zero or all positive")

    dispersion = np.broadcast_to(dispersion, (ntags,))

    pvals = np.zeros(ntags)

    if ntags == 0:

        return pvals

    # Eliminate all zero rows

    all_zeros = (sum1 == 0) & (sum2 == 0)

    if all_zeros.any():

        pvals[~all_zeros] = exactTestByDeviance(

            y1=y1[~all_zeros, :],

            y2=y2[~all_zeros, :],

            dispersion=dispersion[~all_zeros],

        )

        pvals[all_zeros] = 1

        return pvals

    # The code below was originally written in C++

    nlibs = n1 + n2

    stotal = sum1 + sum2

    mu = stotal / nlibs

    mu1 = mu * n1

    mu2 = mu * n2

    r1 = n1 / dispersion

    r2 = n2 / dispersion

    p = r1 / (r1 + mu1)

    # The aim is to sum conditional probabilities for all partitions of the

    # total sum with deviances greater than that observed for the current

    # partition. We start computing from the extremes in both cases

    phi1 = 1 / r1

    phi2 = 1 / r2

    for i in range(ntags):

        obsdev = compute_unit_nb_deviance(

            sum1[i], mu1[i], phi1[i]

        ) + compute_unit_nb_deviance(sum2[i], mu2[i], phi2[i])

        # Going from the left

        for j in range(int(stotal[i]) + 1):

            if obsdev <= compute_unit_nb_deviance(

                j, mu1[i], phi1[i]

            ) + compute_unit_nb_deviance(stotal[i] - j, mu2[i], phi2[i]):

                pvals[i] += nbinom.pmf(j, r1[i], p[i]) * nbinom.pmf(

                    stotal[i] - j, r2[i], p[i]

                )

            else:

                break

        # Going from the right, or what's left of it

        for k in range(int(stotal[i]) - j + 1):

            if obsdev <= compute_unit_nb_deviance(

                k, mu2[i], phi2[i]

            ) + compute_unit_nb_deviance(stotal[i] - k, mu1[i], phi1[i]):

                pvals[i] += nbinom.pmf(k, r2[i], p[i]) * nbinom.pmf(

                    stotal[i] - k, r1[i], p[i]

                )

            else:

                break

    totalr = r1 + r2

    pvals /= nbinom.pmf(stotal, totalr, totalr / (totalr + mu1 + mu2))

    return np.minimum(pvals, 1)