File size: 6,536 Bytes
6019d52
 
 
 
 
 
98ed1b7
6019d52
98ed1b7
 
6019d52
 
98ed1b7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6019d52
 
 
 
98ed1b7
6019d52
 
 
 
 
98ed1b7
6019d52
 
 
 
 
 
 
 
 
 
 
98ed1b7
 
 
 
6019d52
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98ed1b7
 
 
 
 
 
 
6019d52
98ed1b7
 
 
 
 
6019d52
 
 
 
98ed1b7
 
 
 
 
 
6019d52
 
 
 
 
 
 
 
98ed1b7
 
 
 
 
 
6019d52
 
 
 
 
 
 
 
98ed1b7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6019d52
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98ed1b7
 
6019d52
 
 
 
 
98ed1b7
6019d52
 
 
 
 
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
"""Deprecated MitoInteract v1 architecture retained for artifact forensics.

The v1 checkpoint failed its evaluation gates and uses unsafe pickle
serialization. New work lives under ``recovery/``. Loading and inference are
disabled by default so this module cannot silently present an undocumented
legacy score as pKd.
"""

import torch
import torch.nn as nn
from transformers import EsmModel, AutoModel, AutoTokenizer


class MitoInteract(nn.Module):
    def __init__(
        self,
        esm_model_name="facebook/esm2_t33_650M_UR50D",
        mol_model_name="seyonec/ChemBERTa-zinc-base-v1",
        protein_dim=1280,
        mol_dim=768,
        proj_dim=256,
        n_heads=8,
        dropout=0.1,
        freeze_encoders=True,
    ):
        super().__init__()
        self.freeze_encoders = freeze_encoders
        self.esm = EsmModel.from_pretrained(esm_model_name)
        self.protein_dim = protein_dim
        self.mol_encoder = AutoModel.from_pretrained(mol_model_name)
        self.mol_dim = mol_dim
        if freeze_encoders:
            for p in self.esm.parameters():
                p.requires_grad = False
            for p in self.mol_encoder.parameters():
                p.requires_grad = False
        self.prot_proj = nn.Sequential(
            nn.Linear(protein_dim, proj_dim),
            nn.LayerNorm(proj_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
        )
        self.mol_proj = nn.Sequential(
            nn.Linear(mol_dim, proj_dim),
            nn.LayerNorm(proj_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
        )
        self.cross_attn_mol2prot = nn.MultiheadAttention(
            proj_dim, n_heads, dropout=dropout, batch_first=True
        )
        self.cross_attn_prot2mol = nn.MultiheadAttention(
            proj_dim, n_heads, dropout=dropout, batch_first=True
        )
        self.ln_mol2prot = nn.LayerNorm(proj_dim)
        self.ln_prot2mol = nn.LayerNorm(proj_dim)
        fused_dim = proj_dim * 2
        self.mlp = nn.Sequential(
            nn.Linear(fused_dim, 512),
            nn.BatchNorm1d(512),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(512, 256),
            nn.BatchNorm1d(256),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(256, 128),
            nn.BatchNorm1d(128),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(128, 1),
        )

    def encode_protein(self, input_ids, attention_mask):
        ctx = torch.no_grad() if self.freeze_encoders else torch.enable_grad()
        with ctx:
            out = self.esm(input_ids=input_ids, attention_mask=attention_mask)
        mask = attention_mask.unsqueeze(-1).float()
        pooled = (out.last_hidden_state * mask).sum(1) / mask.sum(1).clamp(min=1e-9)
        return pooled, out.last_hidden_state

    def encode_molecule(self, input_ids, attention_mask):
        ctx = torch.no_grad() if self.freeze_encoders else torch.enable_grad()
        with ctx:
            out = self.mol_encoder(input_ids=input_ids, attention_mask=attention_mask)
        return out.pooler_output, out.last_hidden_state

    def forward(
        self, prot_input_ids, prot_attention_mask, mol_input_ids, mol_attention_mask
    ):
        prot_pooled, prot_seq = self.encode_protein(prot_input_ids, prot_attention_mask)
        mol_pooled, mol_seq = self.encode_molecule(mol_input_ids, mol_attention_mask)
        prot_seq_proj = self.prot_proj(prot_seq)
        mol_seq_proj = self.mol_proj(mol_seq)
        prot_q = self.prot_proj(prot_pooled).unsqueeze(1)
        mol_q = self.mol_proj(mol_pooled).unsqueeze(1)
        prot_pad_mask = prot_attention_mask == 0
        mol_pad_mask = mol_attention_mask == 0
        h_prot2mol, _ = self.cross_attn_prot2mol(
            prot_q, mol_seq_proj, mol_seq_proj, key_padding_mask=mol_pad_mask
        )
        h_mol2prot, _ = self.cross_attn_mol2prot(
            mol_q, prot_seq_proj, prot_seq_proj, key_padding_mask=prot_pad_mask
        )
        h_prot2mol = self.ln_prot2mol(h_prot2mol.squeeze(1))
        h_mol2prot = self.ln_mol2prot(h_mol2prot.squeeze(1))
        fused = torch.cat([h_prot2mol, h_mol2prot], dim=-1)
        return self.mlp(fused).squeeze(-1)


def load_model(checkpoint_path, device="cpu", *, allow_unsafe_legacy=False):
    """Load the failed v1 pickle only after explicit risk acknowledgement."""
    if not allow_unsafe_legacy:
        raise RuntimeError(
            "MitoInteract v1 is deprecated and full_model.pt requires unsafe pickle "
            "deserialization. See AUDIT.md. Pass allow_unsafe_legacy=True only for "
            "controlled artifact forensics, never for untrusted files."
        )
    checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=False)
    config = checkpoint["config"]
    model = MitoInteract(
        esm_model_name=config["esm_model"],
        mol_model_name=config["mol_model"],
        protein_dim=config["protein_dim"],
        mol_dim=config["mol_dim"],
        proj_dim=config["proj_dim"],
        n_heads=config["n_heads"],
        dropout=config["dropout"],
        freeze_encoders=True,
    )
    model.load_state_dict(checkpoint["model_state_dict"])
    model.eval()
    return model, config


def predict_binding(
    model, protein_seq, smiles, device="cpu", *, allow_invalid_legacy_output=False
):
    """Return the undocumented v1 score only after explicit acknowledgement."""
    if not allow_invalid_legacy_output:
        raise RuntimeError(
            "v1 output has undocumented target semantics and failed validation. "
            "It cannot be interpreted as pKd or converted to Kd. See AUDIT.md."
        )
    prot_tokenizer = AutoTokenizer.from_pretrained(model.esm.name_or_path)
    mol_tokenizer = AutoTokenizer.from_pretrained(model.mol_encoder.name_or_path)

    prot_enc = prot_tokenizer(
        protein_seq, return_tensors="pt", padding=True, truncation=True, max_length=512
    )
    mol_enc = mol_tokenizer(
        smiles, return_tensors="pt", padding=True, truncation=True, max_length=200
    )

    model = model.to(device)
    with torch.no_grad():
        legacy_score = model(
            prot_enc["input_ids"].to(device),
            prot_enc["attention_mask"].to(device),
            mol_enc["input_ids"].to(device),
            mol_enc["attention_mask"].to(device),
        )

    return {
        "legacy_score": legacy_score.item(),
        "warning": "Do not interpret this value as pKd, Kd, or validated affinity.",
    }