File size: 8,061 Bytes
0ed74db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""PhenoLoRAModel β€” LoRA-fine-tuned ESM-2 + per-category mean-pool + multi-task heads.

Architecture:
    For each genome, accept up to K proteins per category Γ— 8 categories.
    For each protein: tokenize β†’ ESM-2(+LoRA) β†’ masked mean-pool over residues β†’ 1 vector.
    For each category: mean-pool over its proteins β†’ 1 category vector.
    Concatenate the 8 category vectors β†’ genome vector.
    Predict the 4 phenotype targets via 4 small heads.

Trainable parameters:
    - LoRA adapters on ESM-2 attention (~1.5M params for t30, r=8 on q+v).
    - Per-protein projection (optional, default skipped β€” heads operate on raw 8Γ—D vector).
    - 4 prediction heads (~10-50K params combined).

Multi-task loss handles per-target missing labels via a binary mask.
"""
from __future__ import annotations

from dataclasses import dataclass

import torch
from torch import nn

CATEGORIES = ["temperature", "ph", "oxygen", "salt", "vitamin", "nitrogen", "carbon", "special"]
OXYGEN_CLASSES = ["aerobe", "anaerobe", "facultative_anaerobe", "microaerobe"]
N_OXYGEN_CLASSES = len(OXYGEN_CLASSES)


@dataclass
class LoraModelConfig:
    esm_model_name: str = "facebook/esm2_t12_35M_UR50D"
    lora_r: int = 8
    lora_alpha: int = 16
    lora_dropout: float = 0.05
    lora_target: tuple[str, ...] = ("query", "value")
    head_hidden_dim: int = 128
    head_dropout: float = 0.1
    max_seq_len: int = 512               # truncate long proteins to fit memory
    max_proteins_per_cat: int = 6        # cap protein count per category at train time
    gradient_checkpointing: bool = True  # trade compute for memory


class PhenoLoRAModel(nn.Module):
    def __init__(self, cfg: LoraModelConfig):
        super().__init__()
        from peft import LoraConfig, get_peft_model
        from transformers import AutoModel, AutoTokenizer

        self.cfg = cfg
        self.tokenizer = AutoTokenizer.from_pretrained(cfg.esm_model_name)
        base = AutoModel.from_pretrained(cfg.esm_model_name)
        self.embed_dim = base.config.hidden_size

        lora_cfg = LoraConfig(
            r=cfg.lora_r,
            lora_alpha=cfg.lora_alpha,
            lora_dropout=cfg.lora_dropout,
            target_modules=list(cfg.lora_target),
            bias="none",
        )
        self.esm = get_peft_model(base, lora_cfg)
        if cfg.gradient_checkpointing:
            base.gradient_checkpointing_enable()
            base.enable_input_require_grads()

        # 8 Γ— embed_dim β†’ per-target heads.
        genome_dim = self.embed_dim * len(CATEGORIES)
        hidden = cfg.head_hidden_dim

        def regression_head() -> nn.Module:
            return nn.Sequential(
                nn.Linear(genome_dim, hidden),
                nn.GELU(),
                nn.Dropout(cfg.head_dropout),
                nn.Linear(hidden, 1),
            )

        self.heads = nn.ModuleDict({
            "temp": regression_head(),
            "ph": regression_head(),
            "salt": regression_head(),
            "oxy": nn.Sequential(
                nn.Linear(genome_dim, hidden),
                nn.GELU(),
                nn.Dropout(cfg.head_dropout),
                nn.Linear(hidden, N_OXYGEN_CLASSES),
            ),
        })

    def encode_proteins(self, proteins: list[str], device: torch.device) -> torch.Tensor:
        """Tokenize and ESM-encode a list of proteins β†’ tensor of shape (n_proteins, embed_dim).

        Returns a zero tensor of shape (0, embed_dim) for an empty list.
        """
        if not proteins:
            return torch.zeros((0, self.embed_dim), device=device)
        enc = self.tokenizer(
            proteins,
            return_tensors="pt",
            padding=True,
            truncation=True,
            max_length=self.cfg.max_seq_len,
        )
        enc = {k: v.to(device) for k, v in enc.items()}
        outputs = self.esm(**enc)
        last = outputs.last_hidden_state  # (B, L, D)
        mask = enc["attention_mask"].unsqueeze(-1).to(last.dtype)
        pooled = (last * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1.0)
        return pooled

    def encode_genome(self, by_category: dict[str, list[str]], device: torch.device) -> torch.Tensor:
        """Build one genome vector of shape (8 Γ— embed_dim,) by category-pooling proteins."""
        cat_vectors: list[torch.Tensor] = []
        cap = self.cfg.max_proteins_per_cat
        for cat in CATEGORIES:
            proteins = by_category.get(cat) or []
            if proteins:
                # Cap per-category protein count at train time to control memory.
                # Already sorted shortest-first by the extractor.
                proteins = proteins[:cap]
                per_protein = self.encode_proteins(proteins, device)  # (n, D)
                cat_vec = per_protein.mean(dim=0)
            else:
                cat_vec = torch.zeros(self.embed_dim, device=device)
            cat_vectors.append(cat_vec)
        return torch.cat(cat_vectors, dim=0)

    def forward(self, genomes: list[dict[str, list[str]]], device: torch.device) -> dict[str, torch.Tensor]:
        """Batched forward over a list of genomes (variable protein counts per genome).

        Returns dict of per-target predictions:
            temp, ph, salt: (B,) regression predictions
            oxy:            (B, N_OXYGEN_CLASSES) logits
        """
        genome_vecs = torch.stack(
            [self.encode_genome(g, device) for g in genomes],
            dim=0,
        )  # (B, 8*D)
        return {
            "temp": self.heads["temp"](genome_vecs).squeeze(-1),
            "ph": self.heads["ph"](genome_vecs).squeeze(-1),
            "salt": self.heads["salt"](genome_vecs).squeeze(-1),
            "oxy": self.heads["oxy"](genome_vecs),
        }

    def trainable_param_count(self) -> tuple[int, int]:
        """Return (trainable, total) parameter counts."""
        trainable = sum(p.numel() for p in self.parameters() if p.requires_grad)
        total = sum(p.numel() for p in self.parameters())
        return trainable, total


def masked_multitask_loss(
    preds: dict[str, torch.Tensor],
    labels: dict[str, torch.Tensor],
    label_mask: dict[str, torch.Tensor],
    target_weights: dict[str, float] | None = None,
    oxy_class_weights: tuple[float, ...] | None = None,
) -> tuple[torch.Tensor, dict[str, float]]:
    """Sum of per-target losses, weighted by the binary label mask.

    Regression targets: MSE.
    Oxygen: cross-entropy.
    Each per-target loss is averaged over rows with mask=1; if mask is all-zero for a
    target in this batch, that target contributes 0 to the total.
    """
    if target_weights is None:
        target_weights = {"temp": 1.0, "ph": 1.0, "salt": 1.0, "oxy": 1.0}
    per_target_loss: dict[str, float] = {}
    # Graph-connected zero β€” keeps backward() safe when every mask is empty.
    total = (preds["temp"] * 0.0).sum()

    for tgt in ("temp", "ph", "salt"):
        mask = label_mask[tgt].float()
        sq = (preds[tgt] - labels[tgt]) ** 2
        # clamp(min=1) β†’ loss=0 (graph-connected via `sq`) when mask is all-zero.
        loss = (sq * mask).sum() / mask.sum().clamp(min=1.0)
        total = total + target_weights[tgt] * loss
        per_target_loss[tgt] = float(loss.detach().cpu())

    mask_oxy = label_mask["oxy"].float()
    logits = preds["oxy"]
    labels_oxy = labels["oxy"].long()
    class_weight = None
    if oxy_class_weights is not None:
        if len(oxy_class_weights) != logits.shape[-1]:
            raise ValueError("oxy_class_weights must match the number of oxygen classes")
        class_weight = torch.tensor(oxy_class_weights, dtype=logits.dtype, device=logits.device)
    per_row_loss = nn.functional.cross_entropy(
        logits,
        labels_oxy,
        weight=class_weight,
        reduction="none",
    )
    loss_oxy = (per_row_loss * mask_oxy).sum() / mask_oxy.sum().clamp(min=1.0)
    total = total + target_weights["oxy"] * loss_oxy
    per_target_loss["oxy"] = float(loss_oxy.detach().cpu())

    return total, per_target_loss