biotite 1.5.0__cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl

biotite/structure/pseudoknots.py

@@ -0,0 +1,622 @@
+# This source code is part of the Biotite package and is distributed
+# under the 3-Clause BSD License. Please see 'LICENSE.rst' for further
+# information.
+
+"""
+This module provides functionality for pseudoknot detection.
+"""
+
+__name__ = "biotite.structure"
+__author__ = "Tom David Müller"
+__all__ = ["pseudoknots"]
+
+from itertools import chain, product
+import networkx as nx
+import numpy as np
+
+
+def pseudoknots(base_pairs, scores=None, max_pseudoknot_order=None):
+    """
+    Identify the pseudoknot order for each base pair in a given set of
+    base pairs.
+
+    By default the algorithm removes base pairs until the remaining
+    base pairs are completely nested i.e. no pseudoknots appear.
+    The pseudoknot order of the removed base pairs is incremented and
+    the procedure is repeated with these base pairs.
+    Base pairs are removed in a way that maximizes the number of
+    remaining base pairs.
+    However, an optional score for each individual base pair can be
+    provided.
+
+    Parameters
+    ----------
+    base_pairs : ndarray, dtype=int, shape=(n,2)
+        The base pairs to determine the pseudoknot order of. Each row
+        represents indices form two paired bases. The structure of
+        the :class:`ndarray` is equal to the structure of the output of
+        :func:`base_pairs()`, where the indices represent the
+        beginning of the residues.
+    scores : ndarray, dtype=int, shape=(n,), optional
+        The score for each base pair.
+        By default, the score of each base pair is ``1``.
+    max_pseudoknot_order : int, optional
+        The maximum pseudoknot order to be found. If a base pair would
+        be of a higher order, its order is specified as ``-1``.
+        By default, the algorithm is run until all base pairs
+        have an assigned pseudoknot order.
+
+    Returns
+    -------
+    pseudoknot_order : ndarray, dtype=int, shape=(m,n)
+        The pseudoknot order of the input `base_pairs`.
+        Multiple solutions that maximize the number of basepairs or
+        the given score, respectively, may be possible.
+        Therefore all *m* individual solutions are returned.
+
+    Notes
+    -----
+    The dynamic programming approach by Smit *et al*
+    :footcite:`Smit2008` is applied to detect pseudoknots.
+    The algorithm was originally developed to remove pseudoknots from a
+    structure.
+    However, if it is run iteratively on removed knotted pairs it can be
+    used to identify the pseudoknot order.
+
+    The pseudoknot order is defined as the minimum number of base pair
+    set decompositions resulting in a nested structure
+    :footcite:`Antczak2018`.
+    Therefore, there are no pseudoknots between base pairs with the same
+    pseudoknot order.
+
+    References
+    ----------
+
+    .. footbibliography::
+
+    Examples
+    --------
+    Remove the pseudoknotted base pair for the sequence *ABCbac*, where
+    the corresponding big and small letters each represent a base pair:
+
+    Define the base pairs as :class:`ndarray`:
+
+    >>> basepairs = np.array([[0, 4],
+    ...                       [1, 3],
+    ...                       [2, 5]])
+
+    Find the unknotted base pairs, optimizing for the maximum number of
+    base pairs:
+
+    >>> print(pseudoknots(basepairs, max_pseudoknot_order=0))
+    [[ 0  0 -1]]
+
+    This indicates that the base pair *Cc* is a pseudoknot.
+
+    Given the length of the sequence (6 bases), we can also represent
+    the unknotted structure in dot bracket notation:
+
+    >>> print(dot_bracket(basepairs, 6, max_pseudoknot_order=0)[0])
+    ((.)).
+
+    If the maximum pseudoknot order is not restricted, the order of the
+    knotted pairs is determined and can be represented using dot bracket
+    letter notation:
+
+    >>> print(pseudoknots(basepairs))
+    [[0 0 1]]
+    >>> print(dot_bracket(basepairs, 6)[0])
+    (([))]
+    """
+    if len(base_pairs) == 0:
+        # No base pairs -> empty pseudoknot order array
+        return np.array([[]], dtype=np.int32)
+
+    # List containing the results
+    results = [np.full(len(base_pairs), -1, dtype="int32")]
+
+    # if no score array is given, each base pairs' score is one
+    if scores is None:
+        scores = np.ones(len(base_pairs))
+
+    # Make sure `base_pairs` has the same length as the score array
+    if len(base_pairs) != len(scores):
+        raise ValueError("'base_pair' and 'scores' must have the same shape")
+
+    # Split the base pairs in regions
+    regions = _find_regions(base_pairs, scores)
+
+    # Compute results
+    results = _get_results(regions, results, max_pseudoknot_order)
+
+    return np.vstack(results)
+
+
+class _Region:
+    """
+    This class represents a paired region.
+
+    A region is a set of base pairs. This class provides methods to
+    access the minimum and maximum index of the bases that are part of
+    the region, handles score calculation, and backtracing to the
+    original base pair array.
+
+    Parameters
+    ----------
+    base_pairs : ndarray, shape=(n,2), dtype=int
+        All base pairs of the structure the region is a subset for.
+    region_pairs : ndarray, dtype=int
+        The indices of the base pairs in ``base_pairs`` that are part of
+        the region.
+    scores : ndarray, dtype=int, shape=(n,)
+        The score for each base pair.
+    """
+
+    def __init__(self, base_pairs, region_pairs, scores):
+        # The Start and Stop indices for each Region
+        self.start = np.min(base_pairs[region_pairs])
+        self.stop = np.max(base_pairs[region_pairs])
+
+        self.region_pairs = region_pairs
+        self.score = np.sum(scores[region_pairs])
+
+    def get_index_array(self):
+        """
+        Return an index array with the positions of the region`s bases
+        in the original base pair array.
+
+        Returns
+        -------
+        region_pairs : ndarray
+            The indices of the bases in the original base pair array.
+        """
+        return self.region_pairs
+
+    def __lt__(self, other):
+        """
+        This comparison operator is required for :func:`np.unique()`. As
+        only the difference between the regions is relevant and not any
+        particular order, a distinction is made by the objects unique
+        ids.
+
+        Parameters
+        ----------
+        other : _region
+            The other region.
+
+        Returns
+        -------
+        comparision : bool
+            The evaluated comparison.
+        """
+        return id(self) < id(other)
+
+
+def _find_regions(base_pairs, scores):
+    """
+    Find regions in a base pair array. A region is defined as a set of
+    consecutively nested base pairs.
+
+    Parameters
+    ----------
+    base_pairs : ndarray, dtype=int, shape=(n, 2)
+        Each row is equivalent to one base pair and contains the first
+        indices of the residues corresponding to each base.
+    scores : ndarray, dtype=int, shape=(n,)
+        The score for each base pair.
+
+    Returns
+    -------
+    regions : Graph
+        The ``_Region`` objects as graph, where the edges represent
+        conflicts.
+    """
+    # Make sure the lower residue is on the left for each row
+    sorted_base_pairs = np.sort(base_pairs, axis=1)
+
+    # Sort the first column in ascending order
+    original_indices = np.argsort(sorted_base_pairs[:, 0])
+    sorted_base_pairs = sorted_base_pairs[original_indices]
+
+    # Rank each base
+    # E.g.: [[3, 5]  --> [[0, 1]
+    #        [9, 7]]      [3, 2]]
+    order = np.argsort(sorted_base_pairs.flatten())
+    rank = np.argsort(order).reshape(base_pairs.shape)
+
+    # The base pairs belonging to the current region
+    region_pairs = []
+    # The individual regions
+    regions = set()
+
+    # Find separate regions
+    for i in range(len(sorted_base_pairs)):
+        # if a new region is to be started append the current base pair
+        if len(region_pairs) == 0:
+            region_pairs.append(original_indices[i])
+            continue
+
+        # Check if the current base pair belongs to the region that is
+        # currently being defined
+        previous_upstream_rank = rank[i - 1, 0]
+        this_upstream_rank = rank[i, 0]
+        previous_downstream_rank = rank[i - 1, 1]
+        this_downstream_rank = rank[i, 1]
+
+        # if the current base pair belongs to a new region, save the
+        # current region and start a new region
+        if (previous_downstream_rank - this_downstream_rank) != 1 or (
+            this_upstream_rank - previous_upstream_rank
+        ) != 1:
+            regions.add(_Region(base_pairs, np.array(region_pairs), scores))
+            region_pairs = []
+
+        # Append the current base pair to the region
+        region_pairs.append(original_indices[i])
+
+    # The last region has no endpoint defined by the beginning of a
+    # new region.
+    regions.add(_Region(base_pairs, np.array(region_pairs), scores))
+
+    # Return the graphical representation of the conflicting regions
+    return _generate_graphical_representation(regions)
+
+
+def _generate_graphical_representation(regions):
+    """
+    Find the conflicting regions and represent them graphically using
+    the ``Graph`` class from ``Networkx``.
+
+    Parameters
+    ----------
+    regions : set {_region, ...}
+        The regions representing the consecutively nested base pairs.
+
+    Returns
+    -------
+    regions : Graph
+        The ``_Region`` objects as graph, where the edges represent
+        conflicts.
+    """
+
+    # Create a graph
+    region_graph = nx.Graph()
+
+    # Add the regions to the graph as nodes
+    region_graph.add_nodes_from(regions)
+
+    # Get the region array and a boolean array, where the start of each
+    # region is ``True``.
+    region_array, (start_stops,) = _get_region_array_for(
+        regions, content=[lambda a: [True, False]], dtype=["bool"]
+    )
+
+    # Check each region for conflicts with other regions
+    for start, region in enumerate(region_array):
+        # Check each region only once
+        if not start_stops[start]:
+            continue
+
+        # Find the index of the stopping of the region in the region
+        # array
+        stop = _get_first_occurrence_for(region_array[start + 1 :], region)
+        stop += start + 1
+
+        # Store regions the current region conflicts with
+        conflicts = set()
+
+        # Iterate over the regions between the starting and stopping
+        # point of the current region
+        for other_region in region_array[start + 1 : stop]:
+            # If the other region is not already a conflict, add it to
+            # the conflict set
+            if other_region not in conflicts:
+                conflicts.add(other_region)
+            # If the other region is twice between the starting and
+            # stopping point of the current region, its starting and
+            # stopping point lie between the current region and it is
+            # thus non-conflicting
+            else:
+                conflicts.remove(other_region)
+
+        # Conflicts between regions are represented as graph edges
+        edges = []
+
+        # Convert the edges in a ``NetworkX`` compatible format
+        for conflict in conflicts:
+            edges.append((region, conflict))
+
+        # Add the edges to the graph
+        region_graph.add_edges_from(edges)
+    return region_graph
+
+
+def _get_first_occurrence_for(iterable, wanted_object):
+    """
+    Get the first occurrence of an object in an iterable.
+
+    Parameters
+    ----------
+    iterable : iterable
+        The iterable containing the object.
+    wanted_object : object
+        The object to be found.
+
+    Returns
+    -------
+    index : int
+        The index of the first occurrence of the object.
+    """
+    for i, value in enumerate(iterable):
+        if value is wanted_object:
+            return i
+
+
+def _get_region_array_for(regions, content=(), dtype=()):
+    """
+    Get a :class:`ndarray` of region objects. Each object occurs twice,
+    representing its start and end point. The regions positions in the
+    array reflect their relative positions.
+
+    Furthermore, a list of functions can be provided enabling custom
+    outputs for each objects` start and end point.
+
+    Parameters
+    ----------
+    regions : set {_region, ...}
+        The regions to be considered
+    content : list [function, ...]
+        The functions to be considered for custom outputs. For a given
+        region they must return a tuple of which the first value is
+        placed at the start position and the second value at the end
+        position of the region relative to the other regions.
+    dtype : list [str, ...]
+        The data type of the output of the custom functions.
+
+    Returns
+    -------
+    region_array : ndarray, dtype=object
+        The array of ordered region objects.
+    custom_content : list [ndarray, ...]
+        The custom output.
+    """
+    # region_array and index array
+    region_array = np.empty(len(regions) * 2, dtype=_Region)
+    index_array = np.empty(len(regions) * 2, dtype="int32")
+
+    # Content array for custom return arrays
+    content_list = [None] * len(content)
+    for i in range(len(content)):
+        content_list[i] = np.empty(len(regions) * 2, dtype=dtype[i])
+
+    # Fill the arrays
+    for i, reg in enumerate(regions):
+        indices = [2 * i, 2 * i + 1]
+        region_array[indices] = reg
+        for c in range(len(content_list)):
+            content_list[c][indices] = content[c](reg)
+        index_array[indices] = [reg.start, reg.stop]
+
+    # Order the arrays by the base indices
+    sort_mask = np.argsort(index_array)
+    region_array = region_array[sort_mask]
+
+    # if no custom array content is given only return the ordered array
+    # containing the regions
+    if content == []:
+        return region_array
+
+    # if custom content is given also return the ordered content
+    for i in range(len(content_list)):
+        content_list[i] = content_list[i][sort_mask]
+    return region_array, content_list
+
+
+def _remove_pseudoknots(regions):
+    """
+    Get the optimal solutions according to the algorithm referenced in
+    :func:`pseudoknots()`.
+
+    The algorithm uses a dynamic programming matrix in order to find
+    the optimal solutions with the highest combined region scores.
+
+    Parameters
+    ----------
+    regions : set {_region, ...}
+        The conflicting regions for whích optimal solutions are to be
+        found.
+    scores : ndarray
+        The score array.
+
+    Returns
+    -------
+    solutions : ndarray, dtype=object
+        The optimal solutions. Each solution in the ``ndarray`` is
+        represented as ``set`` of unknotted regions.
+    """
+    # Create dynamic programming matrix
+    dp_matrix_shape = len(regions) * 2, len(regions) * 2
+    dp_matrix = np.empty(dp_matrix_shape, dtype="object")
+    dp_matrix_solutions_starts = np.zeros_like(dp_matrix)
+    dp_matrix_solutions_stops = np.zeros_like(dp_matrix)
+
+    # Each index corresponds to the position in the dp matrix.
+    # ``region_array`` contains the region objects and ``start_stops``
+    # contains the lowest and highest positions of the regions
+    region_array, (start_stops,) = _get_region_array_for(
+        regions, [lambda a: (a.start, a.stop)], ["int32"]
+    )
+    # Initialise the matrix diagonal with ndarrays of empty frozensets
+    for i in range(len(dp_matrix)):
+        dp_matrix[i, i] = np.array([frozenset()])
+
+    # Iterate through the top right half of the dynamic programming
+    # matrix
+    for j in range(len(regions) * 2):
+        for i in range(j - 1, -1, -1):
+            solution_candidates = set()
+            left = dp_matrix[i, j - 1]
+            bottom = dp_matrix[i + 1, j]
+
+            # Add all solutions of the cell to the left
+            for solution in left:
+                solution_candidates.add(solution)
+
+            # Add all solutions of the cell to the bottom
+            for solution in bottom:
+                solution_candidates.add(solution)
+
+            # Check if i and j are start/end-points of the same region
+            if region_array[i] is region_array[j]:
+                # Add all solutions from the cell to the bottom left
+                # plus this region
+                bottom_left = dp_matrix[i + 1, j - 1]
+                for solution in bottom_left:
+                    solution_candidates.add(solution | set([region_array[i]]))
+
+            # Perform additional tests if solution in the left cell and
+            # bottom cell both differ from an empty solution
+            if np.any(left != [frozenset()]) and np.any(bottom != [frozenset()]):
+                left_highest = dp_matrix_solutions_stops[i, j - 1]
+                bottom_lowest = dp_matrix_solutions_starts[i + 1, j]
+
+                # For each pair of solutions check if solutions are
+                # disjoint
+                for solution1, highest in zip(left, left_highest):
+                    for solution2, lowest in zip(bottom, bottom_lowest):
+                        if highest < lowest:
+                            # Both solutions are disjoint
+                            solution_candidates.add(solution1 | solution2)
+                        else:
+                            # Both solutions are not disjoint
+                            # Add subsolutions
+                            for k in range(
+                                np.where(start_stops == lowest)[0][0] - 1,
+                                np.where(start_stops == highest)[0][0] + 1,
+                            ):
+                                cell1 = dp_matrix[i, k]
+                                cell2 = dp_matrix[k + 1, j]
+                                for subsolution1 in cell1:
+                                    for subsolution2 in cell2:
+                                        solution_candidates.add(
+                                            subsolution1 | subsolution2
+                                        )
+
+            # Make solution candidates ``ndarray`` array of sets
+            solution_candidates = np.array(list(solution_candidates))
+
+            # Calculate the scores for each solution
+            solution_scores = np.zeros(len(solution_candidates))
+            for s, solution in enumerate(solution_candidates):
+                score = 0
+                for reg in solution:
+                    score += reg.score
+                solution_scores[s] = score
+            # Get the indices where the score is at a maximum
+            highest_scores = np.argwhere(
+                solution_scores == np.amax(solution_scores)
+            ).flatten()
+
+            # Get the solutions with the highest score
+            solution_candidates = solution_candidates[highest_scores]
+
+            # Add the solutions to the dynamic programming matrix
+            dp_matrix[i, j] = solution_candidates
+
+            solution_starts = np.zeros_like(solution_candidates, dtype="int32")
+            solution_stops = np.zeros_like(solution_candidates, dtype="int32")
+
+            for s, solution in enumerate(solution_candidates):
+                solution_starts[s] = min([reg.start for reg in solution], default=-1)
+                solution_stops[s] = max([reg.stop for reg in solution], default=-1)
+
+            dp_matrix_solutions_starts[i, j] = solution_starts
+            dp_matrix_solutions_stops[i, j] = solution_stops
+
+    # The top right corner contains the optimal solutions
+    return dp_matrix[0, -1]
+
+
+def _get_results(regions, results, max_pseudoknot_order, order=0):
+    """
+    Use the dynamic programming algorithm to get the pseudoknot order
+    of a given set of regions. If there are remaining conflicts their
+    results are recursively calculated and merged with the current
+    results.
+
+    Parameters
+    ----------
+    regions : set {_region, ...}
+        The regions for whích optimal solutions are to be found.
+    results : list [ndarray, ...]
+        The results
+    max_pseudoknot_order : int
+        The maximum pseudoknot order to be found. If a base pair would
+        be of a higher order, its order is specified as -1. If ``None``
+        is given, all base pairs are evaluated.
+    order : int
+        The order that is currently evaluated.
+
+    Returns
+    -------
+    results : list [ndarray, ...]
+        The results
+    """
+
+    # Remove non-conflicting regions
+    non_conflicting = [isolate for isolate in nx.isolates(regions)]
+    regions.remove_nodes_from(non_conflicting)
+
+    # Non-conflicting regions are of the current order:
+    index_list_non_conflicting = list(
+        chain(*[region.get_index_array() for region in non_conflicting])
+    )
+    for result in results:
+        result[index_list_non_conflicting] = order
+
+    # If no conflicts remain, the results are complete
+    if len(regions) == 0:
+        return results
+
+    # Get the optimal solutions for given regions. Evaluate each clique
+    # of mutually conflicting regions seperately
+    cliques = [component for component in nx.connected_components(regions)]
+    solutions = [
+        set(chain(*e))
+        for e in product(*[_remove_pseudoknots(clique) for clique in cliques])
+    ]
+
+    # Get a copy of the current results for each optimal solution
+    results_list = [
+        [result.copy() for result in results] for _ in range(len(solutions))
+    ]
+
+    # Evaluate each optimal solution
+    for i, solution in enumerate(solutions):
+        # Get the pseudoknotted regions
+        pseudoknotted_regions = regions.copy()
+        pseudoknotted_regions.remove_nodes_from(solution)
+
+        # Get an index list of the unknotted base pairs
+        index_list_unknotted = list(
+            chain(*[region.get_index_array() for region in solution])
+        )
+
+        # Write results for current solution
+        for j, result in enumerate(results_list[i]):
+            result[index_list_unknotted] = order
+
+        # If this order is the specified maximum order, stop evaluation
+        if max_pseudoknot_order == order:
+            continue
+
+        # Evaluate the pseudoknotted region
+        results_list[i] = _get_results(
+            pseudoknotted_regions,
+            results_list[i],
+            max_pseudoknot_order,
+            order=order + 1,
+        )
+
+    # Flatten the results
+    return list(chain(*results_list))