File size: 19,136 Bytes
0161e74
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
# CCFM (Cascaded Conditioned Flow Matching) 架构文档

## 一、项目概述

CCFM 是一个**级联流匹配**框架,融合三个模型的优势来做单细胞扰动预测:

- **scDFM**:基础流匹配架构(backbone、数据加载、训练范式)
- **LatentForcing**:级联双流思想(latent 流 + 表达流分阶段生成)
- **scGPT**:冻结的预训练模型,提供 per-gene 上下文化特征

### 核心创新

借鉴 LatentForcing 的双流架构,将其从图像域迁移到单细胞域:

| LatentForcing (图像) | CCFM (单细胞) |
|---|---|
| 像素值 | 基因表达值 |
| DINO-v2 特征(辅助生成目标) | scGPT 上下文特征(辅助生成目标) |
| 类别标签(条件信号) | control 表达 + perturbation_id(条件信号) |

**关键区分**:scGPT 特征是从 target(扰动细胞)提取的**辅助生成目标**,不是条件信号。推理时模型从噪声生成 scGPT 特征(Stage 1),再用生成的特征引导表达值生成(Stage 2)。真正的条件信号是 control 表达和 perturbation_id,它们在推理时始终可获取。

### 文件结构

```
CCFM/
├── _bootstrap_scdfm.py                    # Bootstrap scDFM 模块,命名空间隔离
├── config/
│   └── config_cascaded.py                 # CascadedFlowConfig dataclass (tyro CLI)
├── src/
│   ├── __init__.py
│   ├── utils.py                           # Re-exports scDFM utils
│   ├── _scdfm_imports.py                  # scDFM 模块集中导入
│   ├── denoiser.py                        # CascadedDenoiser (训练/推理逻辑)
│   ├── model/
│   │   ├── model.py                       # CascadedFlowModel 双流模型
│   │   └── layers.py                      # LatentEmbedder, LatentDecoder
│   └── data/
│       ├── data.py                        # scDFM 数据加载集成
│       └── scgpt_extractor.py             # FrozenScGPTExtractor
├── scripts/
│   ├── run_cascaded.py                    # 训练/评估入口
│   └── download_scgpt.py                  # 下载 scGPT 预训练模型
├── run.sh                                 # pjsub 模板
└── run_topk30_neg.sh                      # 完整参数化 job 脚本
```

### 默认超参数

| 参数 | 值 | 说明 |
|---|---|---|
| `B` | 48 | batch size |
| `G_full` | 5000 | HVG 总数 |
| `G` | 1000 | 训练时随机基因子集 |
| `d_model` | 128 | 隐藏维度 |
| `scgpt_dim` | 512 | scGPT 输出特征维度 |
| `bottleneck_dim` | 128 | LatentEmbedder 瓶颈维度 |
| `nhead` | 8 | 注意力头数 |
| `nlayers` | 4 | Transformer 层数 |
| `dh_depth` | 2 | LatentDecoder block 数 |
| `choose_latent_p` | 0.4 | 训练 latent 流的概率 |
| `latent_weight` | 1.0 | latent loss 权重 |
| `gamma` | 0.5 | MMD loss 权重 |
| `lr` | 5e-5 | 学习率 |
| `steps` | 200000 | 训练总步数 |
| `latent_steps` | 20 | 推理 ODE 步数(latent) |
| `expr_steps` | 20 | 推理 ODE 步数(表达) |
| `warmup_batches` | 200 | scGPT running stats 预热 |

---

## 二、训练流程 Tensor 数据流

### 2.1 数据准备

```
输入:
  source:           (B, G_full=5000)    控制组表达
  target:           (B, G_full=5000)    扰动后表达
  perturbation_id:  (B, 2)              扰动 token ID
  gene_ids:         (G_full=5000,)      全部基因的 vocab ID

随机采样 1000 个基因:
  input_gene_ids = randperm(5000)[:1000]          → (1000,)
  source_sub  = source[:, input_gene_ids]         → (48, 1000)
  target_sub  = target[:, input_gene_ids]         → (48, 1000)
  gene_input  = gene_ids[input_gene_ids].expand   → (48, 1000)
```

### 2.2 冻结 scGPT 提取辅助生成目标(类似 LatentForcing 的 DINO 特征提取)

```
scgpt_extractor.extract(target_sub, gene_indices=input_gene_ids):

  target_sub:        (48, 1000)         输入表达
  hvg_ids:           (1000,)            HVG → scGPT vocab ID 映射
  valid_mask:        (1000,)            bool, 过滤在 scGPT vocab 中的基因
  expr_valid:        (48, G_valid)      有效基因的表达值

  若 G_valid+1 > max_seq_len(1200):
    随机采样 1199 个基因
    seq_len = 1200
  否则:
    seq_len = G_valid + 1

  拼接 CLS token:
    src    = [cls_id | gene_ids]        → (48, seq_len)       long
    values = [0      | expr_valid]      → (48, seq_len)       float

  scGPT frozen forward:
    encoder_out = scgpt._encode(src, values, mask)  → (48, seq_len, 512)

  去掉 CLS, scatter 回固定位置:
    gene_features = encoder_out[:, 1:, :]           → (48, seq_len-1, 512)
    output = zeros(48, 1000, 512)
    output.scatter_(1, idx, gene_features)          → (48, 1000, 512)

  归一化 (running mean/var, warmup 200 batches 后冻结):
    output = (output - running_mean) / sqrt(running_var) * target_std

  z_target:  (48, 1000, 512)
```

### 2.3 级联时间采样

```
t_latent = rand(48)                         → (48,)
t_expr   = rand(48)                         → (48,)
choose_latent_mask = rand(48) < 0.4         → (48,)  bool

对每个样本二选一:
  若 mask=True  (40%概率): t_latent 保留, t_expr=0, w_expr=0, w_latent=1  → 训练 latent 流
  若 mask=False (60%概率): t_expr 保留, t_latent=1, w_expr=1, w_latent=0  → 训练表达流

输出:
  t_expr:    (48,)    表达流时间步
  t_latent:  (48,)    潜变量流时间步
  w_expr:    (48,)    表达 loss 权重 (0 或 1)
  w_latent:  (48,)    潜变量 loss 权重 (0 或 1)
```

### 2.4 Flow Path 采样(线性插值 + 速度)

```
表达流:
  noise_expr = randn_like(source_sub)                    → (48, 1000)
  path_expr  = AffineProbPath.sample(t_expr, noise_expr, target_sub)
    path_expr.x_t  = (1-t)*noise + t*target              → (48, 1000)    插值后的噪声表达
    path_expr.dx_t = target - noise                       → (48, 1000)    目标速度 (ground truth)

潜变量流:
  noise_latent = randn_like(z_target)                     → (48, 1000, 512)
  展平:
    z_target_flat      = z_target.reshape(48, 512000)     → (48, 512000)
    noise_latent_flat  = noise_latent.reshape(48, 512000) → (48, 512000)
  path_latent_flat = AffineProbPath.sample(t_latent, noise_latent_flat, z_target_flat)
  还原:
    path_latent.x_t  = reshape → (48, 1000, 512)          插值后的噪声 latent
    path_latent.dx_t = reshape → (48, 1000, 512)          目标速度
```

### 2.5 模型前向传播 (`CascadedFlowModel.forward`)

```
输入:
  gene_input:       (48, 1000)          基因 token ID
  source_sub:       (48, 1000)          源表达
  path_expr.x_t:    (48, 1000)          噪声表达
  path_latent.x_t:  (48, 1000, 512)     噪声 latent
  t_expr:           (48,)
  t_latent:         (48,)
  perturbation_id:  (48, 2)
```

#### Step 5a: 表达流 Embedding

```
gene_emb   = GeneEncoder(gene_input)                        → (48, 1000, 128)
val_emb_1  = ContinuousValueEncoder(x_t)                    → (48, 1000, 128)   encoder_1 = 噪声 target (同 scDFM)
val_emb_2  = ContinuousValueEncoder(source_sub)             → (48, 1000, 128)   encoder_2 = control (同 scDFM)
expr_tokens = fusion_layer(cat[val_emb_1, val_emb_2]) + gene_emb
            = Linear(256→128) → GELU → Linear(128→128) → LN + gene_emb → (48, 1000, 128)
```

> **设计说明**(与 scDFM 对齐):
> - `value_encoder_1` 编码噪声 target(`x_t`),`value_encoder_2` 编码 control(`source`),与 scDFM 角色一致
> - `gene_emb` 只在融合后加一次(去冗余),而非分别加到两个 encoder 的输出上
> - `val_emb_2`(control 嵌入)传入 backbone 的 DiffPerceiverBlock 作为交叉注意力 KV(稳定参考基线)

#### Step 5b: 潜变量流 Embedding

```
latent_tokens = LatentEmbedder(path_latent.x_t)
              = Linear(512→128) → GELU → Linear(128→128)    → (48, 1000, 128)
```

#### Step 5c: 双流融合

```
x = expr_tokens + latent_tokens                              → (48, 1000, 128)
```

#### Step 5d: 条件向量

```
t_expr_emb   = TimestepEmbedder(t_expr)                      → (48, 128)
t_latent_emb = TimestepEmbedder(t_latent)                     → (48, 128)
pert_emb     = GeneEncoder(perturbation_id).mean(dim=1)       → (48, 128)
               # perturbation_id: (48,2) → encoder → (48,2,128) → mean → (48,128)
c = t_expr_emb + t_latent_emb + pert_emb                     → (48, 128)
```

#### Step 5e: 共享 Backbone (4 层 DiffPerceiverBlock)

```
for i in range(4):
  # GeneadaLN: 用 gene_emb 调制 x
  x = gene_adaLN[i](gene_emb, x)                             → (48, 1000, 128)

  # Adapter: 拼接 pert_emb 后降维
  pert_exp = pert_emb[:, None, :].expand(-1, 1000, -1)        → (48, 1000, 128)
  x = cat[x, pert_exp]                                        → (48, 1000, 256)
  x = adapter_layer[i](x)
    = Linear(256→128) → LeakyReLU → Linear(128→128) → LeakyReLU
                                                               → (48, 1000, 128)

  # DiffPerceiverBlock: attention + MLP, 用 c 做 AdaLN 条件
  x = DiffPerceiverBlock(x, val_emb_2, c)                     → (48, 1000, 128)
```

#### Step 5f: 表达 Decoder Head

```
x_with_pert = cat[x, pert_exp]                                → (48, 1000, 256)
pred_v_expr = ExprDecoder(x_with_pert)["pred"]                → (48, 1000)
```

#### Step 5g: Latent Decoder Head

```
h = dh_proj(x)              # Linear(128→128)                 → (48, 1000, 128)

for j in range(2):           # dh_depth=2
  # AdaLN: c → SiLU → Linear(128→768) → chunk 6
  shift_msa, scale_msa, gate_msa,
  shift_mlp, scale_mlp, gate_mlp = adaLN_modulation(c)        各 (48, 128)

  # Self-Attention with AdaLN
  h_norm = LayerNorm(h) * (1+scale_msa[:,None,:]) + shift_msa[:,None,:]
                                                               → (48, 1000, 128)
  h_attn = MultiheadAttention(h_norm, h_norm, h_norm)          → (48, 1000, 128)
  h = h + gate_msa[:,None,:] * h_attn                          → (48, 1000, 128)

  # MLP with AdaLN
  h_norm = LayerNorm(h) * (1+scale_mlp[:,None,:]) + shift_mlp[:,None,:]
                                                               → (48, 1000, 128)
  h_mlp  = Linear(128→512) → GELU → Linear(512→128)           → (48, 1000, 128)
  h = h + gate_mlp[:,None,:] * h_mlp                           → (48, 1000, 128)

pred_v_latent = final(h)
              = LN → Linear(128→128) → GELU → Linear(128→512) → (48, 1000, 512)
```

### 2.6 Loss 计算

```
loss_expr   = MSE(pred_v_expr - path_expr.dx_t) * w_expr[:, None]
            = ((48,1000) - (48,1000))^2 * (48,1)  → mean → scalar
              只对 w_expr=1 的样本有贡献 (约60%)

loss_latent = MSE(pred_v_latent - path_latent.dx_t) * w_latent[:, None, None]
            = ((48,1000,512) - (48,1000,512))^2 * (48,1,1) → mean → scalar
              只对 w_latent=1 的样本有贡献 (约40%)

loss = loss_expr + latent_weight * loss_latent

可选 MMD loss (对 w_expr>0 的样本):
  x1_hat = x_t + pred_v_expr * (1-t_expr)    # 单步重建 → (N_expr, 1000)
  mmd_loss = mmd2_unbiased_multi_sigma(x1_hat, target_sub)  → scalar
  loss += gamma * mmd_loss
```

### 2.7 训练循环

```
for iteration in range(200000):
    batch → source(48, 5000), target(48, 5000), pert_id(48, 2)
    loss = denoiser.train_step(source, target, pert_id, gene_ids, infer_top_gene=1000)
    optimizer.zero_grad()
    accelerator.backward(loss)
    optimizer.step()
    scheduler.step()  # CosineAnnealing

    if iteration % 5000 == 0:
        save checkpoint + run evaluation
```

---

## 三、推理流程 Tensor 数据流

推理时**不调用 scGPT 编码器**,模型从噪声自行生成 latent(辅助语义表征)。用**全部 5000 个基因**,不做子集采样。
条件信号(control 表达 + perturbation_id)在每个 ODE 步中持续提供。

```
输入:
  source:           (B, 5000)       控制组表达
  perturbation_id:  (B, 2)          扰动 ID
  gene_ids:         (B, 5000)       基因 vocab ID
```

### 3.1 Stage 1: 生成 Latent(t_latent: 0→1, t_expr 固定=0)

```
z_noise = randn(B, 5000, 512)              初始噪声

ODE 求解 (RK4, 20步, t ∈ [0,1]):
  每一步调用 latent_vf(t, z_t):
    t_latent = t.expand(B)                  → (B,)     当前时间步
    t_expr   = zeros(B)                     → (B,)     固定为 0

    model.forward(
      gene_ids,                             (B, 5000)
      source,                               (B, 5000)   源表达
      source,                               (B, 5000)   ← 注意: x_t 用 source (因为 t_expr=0)
      z_t,                                  (B, 5000, 512)  当前 latent 状态
      t_expr=0,                             (B,)
      t_latent=t,                           (B,)
      perturbation_id,                      (B, 2)
    )
    → _, v_latent                           (B, 5000, 512)  latent 速度场

    return v_latent

  z_traj = odeint(latent_vf, z_noise, linspace(0,1,20))
                                            → (20, B, 5000, 512)
  z_generated = z_traj[-1]                  → (B, 5000, 512)  ★ 生成的 latent
```

### 3.2 Stage 2: 用生成的 Latent 引导表达生成(t_expr: 0→1, t_latent 固定=1)

```
expr_noise = randn_like(source)             → (B, 5000)  高斯噪声 (或 Poisson)

ODE 求解 (RK4, 20步, t ∈ [0,1]):
  每一步调用 expr_vf(t, x_t):
    t_expr   = t.expand(B)                  → (B,)     当前时间步
    t_latent = ones(B)                      → (B,)     固定为 1 (latent 已完成)

    model.forward(
      gene_ids,                             (B, 5000)
      source,                               (B, 5000)
      x_t,                                  (B, 5000)   ← 当前噪声表达
      z_generated,                          (B, 5000, 512)  ★ Stage 1 的结果 (固定)
      t_expr=t,                             (B,)
      t_latent=1,                           (B,)
      perturbation_id,                      (B, 2)
    )
    → v_expr, _                             (B, 5000)  表达速度场

    return v_expr

  x_traj = odeint(expr_vf, expr_noise, linspace(0,1,20))
                                            → (20, B, 5000)
  x_generated = clamp(x_traj[-1], min=0)   → (B, 5000)  ★ 最终预测表达
```

### 3.3 评估流程

```
对每个扰动条件:
  source = 随机采样 128 个控制细胞              → (128, 5000)
  按 batch_size 调用 generate:
    pred = denoiser.generate(source, pert_id, gene_ids)  → (128, 5000)

汇总:
  pred_adata = AnnData(X=所有预测表达, obs=扰动标签)
  real_adata = AnnData(X=所有真实表达, obs=扰动标签)
  MetricsEvaluator(pred_adata, real_adata) → results.csv, agg_results.csv
```

---

## 四、训练 vs 推理 对比总结

| 维度 | 训练 | 推理 |
|---|---|---|
| 基因数 | 随机采样 G=1000 | 全量 G=5000 |
| scGPT 特征(辅助生成目标) | 冻结提取 target 特征 `z_target` 作为 latent 流的生成目标 | **不使用**,模型从噪声自行生成 latent(Stage 1) |
| 时间步 | 级联采样:40% 训练 latent, 60% 训练表达 | 两阶段串行:先 latent(0→1), 再表达(0→1) |
| 流采样 | 单步:`x_t = (1-t)*noise + t*target` | ODE 积分:RK4, 20 步 |
| Loss | MSE(velocity) + MMD | 无 |
| 输出 | scalar loss | `(B, G_full)` 表达矩阵 |
| 条件信号 | control + pert_id(真条件);scGPT 特征是生成目标 | control + pert_id(每步输入);latent 由 Stage1 生成 |
| 时间步 | `t_expr``t_latent` 随机,互斥激活 | Stage1: `t_expr=0, t_latent∈[0,1]`; Stage2: `t_latent=1, t_expr∈[0,1]` |

---

## 五、模型架构总览图

```
                        ┌──────────────────────────────────────────────┐
                        │           CascadedFlowModel                  │
                        │                                              │
  gene_id (B,G) ───────→│  GeneEncoder ──→ gene_emb (B,G,d)           │
                        │       │                                      │
  x_t (B,G) ──────────→│  ValEnc1 ──→ val_emb_1 (B,G,d)  噪声target  │
  source (B,G) ────────→│  ValEnc2 ──→ val_emb_2 (B,G,d)  control     │
                        │       │                                      │
                        │  fusion_layer(cat[val_emb_1, val_emb_2])     │
                        │       + gene_emb (融合后加一次)                │
                        │       ↓                                      │
                        │  expr_tokens (B,G,d)                         │
                        │       │                                      │
  z_t (B,G,512) ──────→│  LatentEmbedder(512→128→d)                   │
                        │       ↓                                      │
                        │  latent_tokens (B,G,d)                       │
                        │       │                                      │
                        │  x = expr_tokens + latent_tokens  (B,G,d)   │
                        │                                              │
  t_expr (B,) ────────→│  TimestepEmb ──┐                             │
  t_latent (B,) ──────→│  TimestepEmb ──┼→ c = sum (B,d)             │
  pert_id (B,2) ──────→│  PertEmb ──────┘                             │
                        │                                              │
                        │  ┌─ 4x DiffPerceiverBlock ──────────────┐    │
                        │  │  gene_adaLN(gene_emb, x)             │    │
                        │  │  cat[x, pert_emb] → adapter → x      │    │
                        │  │  DiffPerceiverBlock(x, val_emb_2, c)  │    │
                        │  └──────────────────────────────────────┘    │
                        │       │                                      │
                        │       ├──→ ExprDecoder(cat[x, pert]) ──→ pred_v_expr (B,G)
                        │       │                                      │
                        │       └──→ LatentDecoder(x, c)               │
                        │              dh_proj → 2x AdaLN Block        │
                        │              → final proj ──→ pred_v_latent (B,G,512)
                        └──────────────────────────────────────────────┘
```