File size: 6,017 Bytes
9651517
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
SACFMDenoiser — Source-Anchored Conditional Flow Matching.

Training: x_0 = source + sigma_aug * eps (noisy source, not pure noise).
          Standard affine path: x_t = (1-t)*x_0 + t*target.
          Velocity target: dx_t = target - x_0.
          Gene-weighted MSE loss.

Inference: ODE from clean source (no noise, no SDE).
"""

import torch
import torch.nn as nn
import torchdiffeq

from ._scdfm_imports import AffineProbPath, CondOTScheduler


def pairwise_sq_dists(X, Y):
    return torch.cdist(X, Y, p=2) ** 2


@torch.no_grad()
def median_sigmas(X, scales=(0.5, 1.0, 2.0, 4.0)):
    D2 = pairwise_sq_dists(X, X)
    tri = D2[~torch.eye(D2.size(0), dtype=bool, device=D2.device)]
    m = torch.median(tri).clamp_min(1e-12)
    s2 = torch.tensor(scales, device=X.device) * m
    return [float(s.item()) for s in torch.sqrt(s2)]


def mmd2_unbiased_multi_sigma(X, Y, sigmas):
    m, n = X.size(0), Y.size(0)
    Dxx = pairwise_sq_dists(X, X)
    Dyy = pairwise_sq_dists(Y, Y)
    Dxy = pairwise_sq_dists(X, Y)
    vals = []
    for sigma in sigmas:
        beta = 1.0 / (2.0 * (sigma ** 2) + 1e-12)
        Kxx = torch.exp(-beta * Dxx)
        Kyy = torch.exp(-beta * Dyy)
        Kxy = torch.exp(-beta * Dxy)
        term_xx = (Kxx.sum() - Kxx.diag().sum()) / (m * (m - 1) + 1e-12)
        term_yy = (Kyy.sum() - Kyy.diag().sum()) / (n * (n - 1) + 1e-12)
        term_xy = Kxy.mean()
        vals.append(term_xx + term_yy - 2.0 * term_xy)
    return torch.stack(vals).mean()


class SACFMDenoiser(nn.Module):
    """
    Source-Anchored Conditional Flow Matching Denoiser.

    Key differences from scDFM baseline:
    - x_0 = source + sigma_aug * eps (not pure noise)
    - Gene-weighted velocity loss
    - Inference starts from clean source

    Key differences from SB:
    - No SigmaNet (sigma_aug is fixed, data-driven)
    - No ScoreDecoder
    - No SDE inference
    - No bridge formulation (standard affine path)
    """

    def __init__(
        self,
        model: nn.Module,
        sigma_aug: torch.Tensor,       # (G,) per-gene augmentation noise
        gene_weight: torch.Tensor,      # (G,) per-gene loss weight
        noise_type: str = "Gaussian",
        use_mmd_loss: bool = True,
        gamma: float = 0.5,
    ):
        super().__init__()
        self.model = model
        self.noise_type = noise_type
        self.use_mmd_loss = use_mmd_loss
        self.gamma = gamma

        # Fixed per-gene tensors (not learned)
        self.register_buffer("sigma_aug", sigma_aug)
        self.register_buffer("gene_weight", gene_weight)

        # Standard affine flow matching path (same as scDFM)
        self.flow_path = AffineProbPath(scheduler=CondOTScheduler())

    def train_step(
        self,
        source: torch.Tensor,          # (B, G_sub) control expression
        target: torch.Tensor,          # (B, G_sub) perturbed expression
        perturbation_id: torch.Tensor, # (B, n_pert)
        gene_input: torch.Tensor,      # (B, G_sub) vocab-encoded gene IDs
        input_gene_ids: torch.Tensor,  # (G_sub,) indices into full gene set
    ) -> dict:
        B = source.shape[0]
        device = source.device

        # 1. Sample time (uniform, clamped away from boundaries)
        t = torch.rand(B, device=device).clamp(1e-5, 1 - 1e-5)

        # 2. Look up per-gene sigma and weight for current gene subset
        sigma_sub = self.sigma_aug[input_gene_ids]      # (G_sub,)
        weight_sub = self.gene_weight[input_gene_ids]    # (G_sub,)

        # 3. Source-anchored x_0 with data-driven augmentation noise
        eps = torch.randn_like(source)
        x_0 = source + sigma_sub.unsqueeze(0) * eps     # (B, G_sub)

        # 4. Standard affine flow matching (reuses scDFM AffineProbPath)
        #    x_t = (1-t)*x_0 + t*target
        #    dx_t = target - x_0 = (target - source) - sigma * eps
        path_sample = self.flow_path.sample(t=t, x_0=x_0, x_1=target)

        # 5. Model forward (scDFM convention: cell_1=x_t, cell_2=source)
        pred_v = self.model(
            gene_input, path_sample.x_t, path_sample.t, source,
            perturbation_id, gene_input, mode="predict_y",
        )

        # 6. Gene-weighted velocity loss
        loss_v = (weight_sub.unsqueeze(0) * (pred_v - path_sample.dx_t) ** 2).mean()

        # 7. MMD loss (optional, same as scDFM baseline)
        loss_mmd = torch.tensor(0.0, device=device)
        if self.use_mmd_loss:
            t_col = t.unsqueeze(-1)
            x1_hat = path_sample.x_t + pred_v * (1 - t_col)
            sigmas_mmd = median_sigmas(target, scales=(0.5, 1.0, 2.0, 4.0))
            loss_mmd = mmd2_unbiased_multi_sigma(x1_hat, target, sigmas_mmd)

        loss = loss_v + self.gamma * loss_mmd

        return {
            "loss": loss,
            "loss_v": loss_v.detach(),
            "loss_mmd": loss_mmd.detach(),
        }

    @torch.no_grad()
    def generate(
        self,
        source: torch.Tensor,          # (B, G)
        perturbation_id: torch.Tensor,  # (B, n_pert)
        gene_ids: torch.Tensor,         # (B, G) or (G,)
        steps: int = 20,
        method: str = "rk4",
    ) -> torch.Tensor:
        """Generate perturbed expression via PF-ODE starting from clean source."""
        B, G = source.shape
        device = source.device

        if gene_ids.dim() == 1:
            gene_ids = gene_ids.unsqueeze(0).expand(B, -1)

        # Start from clean source (no augmentation noise at inference)
        x_0 = source.clone()

        def ode_func(t_scalar, x):
            t_batch = torch.full((B,), t_scalar.item(), device=device)
            pred_v = self.model(
                gene_ids, x, t_batch, source,
                perturbation_id, gene_ids, mode="predict_y",
            )
            return pred_v

        t_span = torch.linspace(0, 1, steps, device=device)
        trajectory = torchdiffeq.odeint(
            ode_func, x_0, t_span,
            method=method, atol=1e-4, rtol=1e-4,
        )
        return torch.clamp(trajectory[-1], min=0)