File size: 5,472 Bytes
6fa9282
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""inference procedures, algorithms 2-6.

Alg 2  forward_predict       predict endpoint, aggregate population
Alg 3  endpoint_ranking      score = reward of predicted endpoint, return topk
Alg 4  reward_guidance       jacobian pull-back ascent in embedding space
Alg 5  project_and_rerank    project e* to admissible candidates, rerank k-nn
Alg 6  greedy_combinatorial  sequentially add the best gene (mean-pool embed)
"""

from __future__ import annotations

import numpy as np
import torch

from pivot.models.encoders import build_pert_tensors
from pivot.models.pivot import PIVOT, candidate_embeddings


def encode_label(model: PIVOT, data, label: str, device) -> torch.Tensor:
    g, o, mask, pid = build_pert_tensors(data, [label], device=device)
    return model.encode(g, o, mask, pid)  # (1, m)


def encode_gene_set(model: PIVOT, data, genes: list[str], device) -> torch.Tensor:
    """encode an arbitrary gene set via the dataset operation (mean-pool)."""
    label = data.sep.join(genes) if genes else data.control_label
    return encode_label(model, data, label, device)


@torch.no_grad()
def forward_predict(model: PIVOT, c0: torch.Tensor, e_u: torch.Tensor) -> torch.Tensor:
    """Alg 2. c0: (N,d) control embeddings; e_u: (1,m) or (N,m). returns endpoint (N,d)."""
    if e_u.shape[0] == 1 and c0.shape[0] > 1:
        e_u = e_u.expand(c0.shape[0], -1)
    return model.endpoint_from_e(c0, e_u)


@torch.no_grad()
def endpoint_ranking(
    model: PIVOT,
    data,
    candidate_labels,
    c0: torch.Tensor,
    reward,
    topk: int | None = None,
    device="cpu",
    chunk: int = 4,
):
    """Alg 3. score each candidate by reward of its predicted endpoint population.

    batched: all candidates x control cells go through the flow map in one
    (chunked) forward pass, then per-row rewards are averaged per candidate. works
    for any per-row reward (centroid/cosine). returns list of (label, score) desc.
    """
    E = torch.as_tensor(
        candidate_embeddings(model, data, candidate_labels, device), device=device
    )
    C, N = E.shape[0], c0.shape[0]
    scores = np.empty(C, dtype=np.float64)
    # chunk over candidates to bound memory (c*n rows per chunk)
    cands_per_chunk = max(1, chunk * 1024 // max(N, 1))
    for s in range(0, C, cands_per_chunk):
        e = E[s : s + cands_per_chunk]
        c = e.shape[0]
        c0_rep = c0.unsqueeze(0).expand(c, N, -1).reshape(c * N, -1)
        e_rep = e.unsqueeze(1).expand(c, N, -1).reshape(c * N, -1)
        chat = model.endpoint_from_e(c0_rep, e_rep)
        if reward.kind in ("mmd", "wasserstein"):
            pops = chat.view(c, N, -1)
            r = torch.stack([reward(pop).mean() for pop in pops])
        else:
            r = reward(chat).view(c, N).mean(1)
        scores[s : s + c] = r.detach().cpu().numpy()
    order = np.argsort(-scores)
    ranked = [(candidate_labels[i], float(scores[i])) for i in order]
    return ranked if topk is None else ranked[:topk]


def reward_guidance(
    model: PIVOT,
    c0: torch.Tensor,
    reward,
    e_init: torch.Tensor,
    steps: int = 25,
    step_size: float = 0.5,
    eps: float = 1e-8,
    normalize: bool = True,
) -> torch.Tensor:
    """Alg 4. gradient ascent on reward in embedding space via the flow-map jacobian.

    e_{l+1} = e_l + gamma * g_e/(norm(g_e)+eps), g_e from autograd."""
    e = e_init.detach().clone()
    for _ in range(steps):
        e.requires_grad_(True)
        chat = model.endpoint_from_e(c0, e.expand(c0.shape[0], -1))
        r = reward(chat).mean()
        (g_e,) = torch.autograd.grad(r, e)
        with torch.no_grad():
            step = g_e / (g_e.norm() + eps) if normalize else g_e
            e = e + step_size * step
        e = e.detach()
    return e


@torch.no_grad()
def project_and_rerank(
    model: PIVOT,
    data,
    candidate_labels,
    e_star: torch.Tensor,
    c0: torch.Tensor,
    reward,
    k_nearest: int = 10,
    topk: int = 5,
    device="cpu",
):
    """Alg 5. project e* to the k nearest admissible candidates, rerank by endpoint reward."""
    E = torch.as_tensor(
        candidate_embeddings(model, data, candidate_labels, device), device=device
    )
    d = torch.cdist(e_star.view(1, -1), E).squeeze(0)
    knn = torch.argsort(d)[:k_nearest].cpu().numpy()
    sub = [candidate_labels[i] for i in knn]
    ranked = endpoint_ranking(model, data, sub, c0, reward, device=device)
    return ranked[:topk]


@torch.no_grad()
def greedy_combinatorial(
    model: PIVOT,
    data,
    gene_pool,
    c0: torch.Tensor,
    reward,
    max_size: int = 2,
    min_gain: float = 1e-4,
    device="cpu",
):
    """Alg 6. greedily add the gene that most improves the reward of the mean-pooled endpoint."""
    chosen: list[str] = []
    best_score = -np.inf
    history = []
    while len(chosen) < max_size:
        remaining = [g for g in gene_pool if g not in chosen]
        if not remaining:
            break
        # score all candidate gene-sets {chosen + g} in one batched pass
        labels = [data.sep.join(chosen + [g]) for g in remaining]
        ranked = endpoint_ranking(model, data, labels, c0, reward, device=device)
        best_label, best_local = ranked[0]
        best_gene = remaining[labels.index(best_label)]
        if best_local - best_score < min_gain:
            break
        chosen.append(best_gene)
        best_score = best_local
        history.append((list(chosen), best_score))
    return chosen, best_score, history