File size: 7,061 Bytes
ce7bf5b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
# Copyright Generate Biomedicines, Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Layers for computing sequence complexities.
"""

import numpy as np
import torch
import torch.nn.functional as F

from chroma.constants import AA20
from chroma.layers.graph import collect_neighbors


def compositions(S: torch.Tensor, C: torch.LongTensor, w: int = 30):
    """Compute local compositions per residue.

    Args:
        S (torch.Tensor): Sequence tensor with shape `(num_batch, num_residues)`
            (long) or `(num_batch, num_residues, num_alphabet)` (float).
        C (torch.LongTensor): Chain map with shape `(num_batch, num_residues)`.
        w (int, optional): Window size.

    Returns:
        P (torch.Tensor): Local compositions with shape
            `(num_batch, num_residues - w + 1, num_alphabet)`.
        N (torch.Tensor): Local counts with shape
            `(num_batch, num_residues - w + 1, num_alphabet)`.
        mask_P (torch.Tensor): Mask with shape
            `(num_batch, num_residues - w + 1)`.
    """
    device = S.device
    Q = len(AA20)
    mask_i = (C > 0).float()
    if len(S.shape) == 2:
        S = F.one_hot(S, Q)

    # Build neighborhoods and masks
    S_onehot = mask_i[..., None] * S
    kx = torch.arange(w, device=S.device) - w // 2
    edge_idx = (
        torch.arange(S.shape[1], device=S.device)[None, :, None] + kx[None, None, :]
    )
    mask_ij = (edge_idx > 0) & (edge_idx < S.shape[1])
    edge_idx = edge_idx.clamp(min=0, max=S.shape[1] - 1)
    C_i = C[..., None]
    C_j = collect_neighbors(C_i, edge_idx)[..., 0]
    mask_ij = (mask_ij & C_j.eq(C_i) & (C_i > 0) & (C_j > 0)).float()

    # Sum neighborhood composition
    S_j = mask_ij[..., None] * collect_neighbors(S_onehot, edge_idx)
    N = S_j.sum(2)

    num_N = N.sum(-1, keepdims=True)
    P = N / (num_N + 1e-5)
    mask_i = ((num_N[..., 0] > 0) & (C > 0)).float()
    mask_ij = mask_i[..., None] * mask_ij
    return P, N, edge_idx, mask_i, mask_ij


def complexity_lcp(
    S: torch.LongTensor,
    C: torch.LongTensor,
    w: int = 30,
    entropy_min: float = 2.32,
    method: str = "naive",
    differentiable=True,
    eps: float = 1e-5,
    min_coverage=0.9,
    # entropy_min: float = 2.52,
    # method = "chao-shen"
) -> torch.Tensor:
    """Compute the Local Composition Perplexity metric.

    Args:
        S (torch.Tensor): Sequence tensor with shape `(num_batch, num_residues)`
            (index tensor) or `(num_batch, num_residues, num_alphabet)`.
        C (torch.LongTensor): Chain map with shape `(num_batch, num_residues)`.
        w (int): Window size.
        grad_pseudocount (float): Pseudocount for stabilizing entropy gradients
            on backwards pass.
        eps (float): Small number for numerical stability in division and logarithms.

    Returns:
        U (torch.Tensor): Complexities with shape `(num_batch)`.
    """

    # adjust window size based on sequence length
    if S.shape[1] < w:
        w = S.shape[1]

    P, N, edge_idx, mask_i, mask_ij = compositions(S, C, w)

    # Only count windows with `min_coverage`
    min_N = int(min_coverage * w)
    mask_coverage = N.sum(-1) > int(min_coverage * w)

    H = estimate_entropy(N, method=method)
    U = mask_coverage * (torch.exp(H) - np.exp(entropy_min)).clamp(max=0).square()

    # Compute entropy as a function of perturbed counts
    if differentiable and len(S.shape) == 3:
        # Compute how a mutation changes entropy for each neighbor
        N_neighbors = collect_neighbors(N, edge_idx)
        mask_coverage_j = collect_neighbors(mask_coverage[..., None], edge_idx)
        N_ij = (N_neighbors - S[:, :, None, :])[..., None, :] + torch.eye(
            N.shape[-1], device=N.device
        )[None, None, None, ...]
        N_ij = N_ij.clamp(min=0)
        H_ij = estimate_entropy(N_ij, method=method)
        U_ij = (torch.exp(H_ij) - np.exp(entropy_min)).clamp(max=0).square()
        U_ij = mask_ij[..., None] * mask_coverage_j * U_ij
        U_differentiable = (U_ij.detach() * S[:, :, None, :]).sum([-1, -2])
        U = U.detach() + U_differentiable - U_differentiable.detach()

    U = (mask_i * U).sum(1)
    return U


def complexity_scores_lcp_t(
    t,
    S: torch.LongTensor,
    C: torch.LongTensor,
    idx: torch.LongTensor,
    edge_idx_t: torch.LongTensor,
    mask_ij_t: torch.Tensor,
    w: int = 30,
    entropy_min: float = 2.515,
    eps: float = 1e-5,
    method: str = "chao-shen",
) -> torch.Tensor:
    """Compute local LCP scores for autoregressive decoding."""
    Q = len(AA20)
    O = F.one_hot(S, Q)
    O_j = collect_neighbors(O, edge_idx_t)
    idx_i = idx[:, t, None]
    C_i = C[:, t, None]
    idx_j = collect_neighbors(idx[..., None], edge_idx_t)[..., 0]
    C_j = collect_neighbors(C[..., None], edge_idx_t)[..., 0]

    # Sum valid neighbor counts
    is_near = (idx_i - idx_j).abs() <= w / 2
    same_chain = C_i == C_j
    valid_ij_t = (is_near * same_chain * (mask_ij_t > 0)).float()[..., None]
    N_k = (valid_ij_t * O_j).sum(-2)

    # Compute counts under all possible extensions
    N_k = N_k[:, :, None, :] + torch.eye(Q, device=N_k.device)[None, None, ...]

    H = estimate_entropy(N_k, method=method)
    U = -(torch.exp(H) - np.exp(entropy_min)).clamp(max=0).square()
    return U


def estimate_entropy(
    N: torch.Tensor, method: str = "chao-shen", eps: float = 1e-11
) -> torch.Tensor:
    """Estimate entropy from counts.

        See Chao, A., & Shen, T. J. (2003) for more details.

    Args:
        N (torch.Tensor): Tensor of counts with shape `(..., num_bins)`.

    Returns:
        H (torch.Tensor): Estimated entropy with shape `(...)`.
    """
    N = N.float()
    N_total = N.sum(-1, keepdims=True)
    P = N / (N_total + eps)

    if method == "chao-shen":
        # Estimate coverage and adjusted frequencies
        singletons = N.long().eq(1).sum(-1, keepdims=True).float()
        C = 1.0 - singletons / (N_total + eps)
        P_adjust = C * P
        P_inclusion = (1.0 - (1.0 - P_adjust) ** N_total).clamp(min=eps)
        H = -(P_adjust * torch.log(P_adjust.clamp(min=eps)) / P_inclusion).sum(-1)
    elif method == "miller-maddow":
        bins = (N > 0).float().sum(-1)
        bias = (bins - 1) / (2 * N_total[..., 0] + eps)
        H = -(P * torch.log(P + eps)).sum(-1) + bias
    elif method == "laplace":
        N = N.float() + 1 / N.shape[-1]
        N_total = N.sum(-1, keepdims=True)
        P = N / (N_total + eps)
        H = -(P * torch.log(P)).sum(-1)
    else:
        H = -(P * torch.log(P + eps)).sum(-1)
    return H