File size: 7,739 Bytes
9627ce0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from abc import ABC, abstractmethod

import attr
import torch
from tqdm import tqdm

from src.data.esm.sdk.api import (
    ESM3InferenceClient,
    ESMProtein,
    ESMProteinError,
    ESMProteinTensor,
    SamplingConfig,
    SamplingTrackConfig,
)
from src.data.esm.sdk.forge import ESM3ForgeInferenceClient
from src.data.esm.tokenization import get_esm3_model_tokenizers


class GuidedDecodingScoringFunction(ABC):
    @abstractmethod
    def __call__(self, protein: ESMProtein) -> float:
        pass


class ESM3GuidedDecoding:
    """This class can be used to perform derivative-free guided decoding, based on
    the method described in "Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-Based Decoding"
    https://arxiv.org/abs/2408.08252
    """

    def __init__(
        self,
        client: ESM3InferenceClient,
        scoring_function: GuidedDecodingScoringFunction,
    ):
        if isinstance(client, ESM3):
            self.tokenizers = client.tokenizers
        elif isinstance(client, ESM3ForgeInferenceClient):
            self.tokenizers = get_esm3_model_tokenizers(client.model)
        else:
            raise ValueError(
                "client must be an instance of ESM3 or ESM3ForgeInferenceClient"
            )

        self.client = client
        self.scoring_function = scoring_function

    def guided_generate(
        self,
        protein: ESMProtein,
        num_decoding_steps: int,
        num_samples_per_step: int,
        denoised_prediction_temperature: float = 0.0,
        track: str = "sequence",
        verbose: bool = True,
    ) -> ESMProtein:
        protein_tensor = self.client.encode(protein)

        assert not isinstance(protein_tensor, ESMProteinError)

        if track == "structure":
            protein_tensor = self.maybe_add_default_structure_tokens(protein_tensor)

        num_masked_positions = self.get_number_of_masked_positions(
            protein_tensor, track=track
        )
        num_positions_to_unmask = num_masked_positions // num_decoding_steps

        current_score = -1

        if verbose:
            pbar = tqdm(range(num_decoding_steps), desc="Current score: -1")
        else:
            pbar = range(num_decoding_steps)

        for step in pbar:
            if step == num_decoding_steps - 1:
                # At the last step, unmask all remaining positions
                num_positions_to_unmask = self.get_number_of_masked_positions(
                    protein_tensor, track=track
                )

            samples = []
            scores = []
            for _ in range(num_samples_per_step):
                sample = self.randomly_unmask_positions(
                    protein_tensor, num_positions_to_unmask, track=track
                )
                scores.append(
                    self.reward_function(
                        sample,
                        denoised_prediction_temperature=denoised_prediction_temperature,
                    )
                )
                samples.append(sample)

            # Select best scoring sample
            best_sample = samples[scores.index(max(scores))]
            current_score = max(scores)
            protein_tensor = best_sample

            if verbose:
                pbar.set_description(f"Current score: {current_score:.2f}")  # type: ignore

        # Fully predict and decode final protein
        protein_tensor_output = self.client.forward_and_sample(
            protein_tensor,
            SamplingConfig(
                sequence=SamplingTrackConfig(temperature=0.0),
                structure=SamplingTrackConfig(temperature=0.0),
            ),
        )

        assert not isinstance(protein_tensor_output, ESMProteinError)
        protein_tensor = protein_tensor_output.protein_tensor

        decoded_protein = self.client.decode(protein_tensor)
        assert not isinstance(decoded_protein, ESMProteinError)
        return decoded_protein

    def reward_function(
        self,
        protein_tensor: ESMProteinTensor,
        denoised_prediction_temperature: float = 0.0,
    ) -> float:
        denoised_protein = self.predict_denoised(
            protein_tensor, temperature=denoised_prediction_temperature
        )
        return self.scoring_function(denoised_protein)

    def get_number_of_masked_positions(
        self, protein_tensor: ESMProteinTensor, track: str = "sequence"
    ) -> int:
        assert isinstance(protein_tensor, ESMProteinTensor)
        track_tensor = getattr(protein_tensor, track)
        track_tokenizer = getattr(self.tokenizers, track)
        is_mask = track_tensor == track_tokenizer.mask_token_id
        return is_mask.sum().item()  # type: ignore

    def randomly_unmask_positions(
        self,
        protein_tensor: ESMProteinTensor,
        num_positions_to_unmask: int,
        temperature: float = 1.0,
        track: str = "sequence",
    ) -> ESMProteinTensor:
        track_tensor = getattr(protein_tensor, track)
        assert track_tensor is not None
        protein_tensor = attr.evolve(protein_tensor)
        setattr(protein_tensor, track, track_tensor.clone())

        track_tensor = getattr(protein_tensor, track)
        track_tokenizer = getattr(self.tokenizers, track)

        is_mask = track_tensor == track_tokenizer.mask_token_id
        num_masked_positions = is_mask.sum().item()

        if num_positions_to_unmask > num_masked_positions:
            num_positions_to_unmask = num_masked_positions  # type: ignore

        mask_indices = is_mask.nonzero(as_tuple=False)
        mask_indices = mask_indices[torch.randperm(mask_indices.size(0))]
        mask_indices = mask_indices[:num_positions_to_unmask]

        sampling_config = SamplingConfig()
        setattr(sampling_config, track, SamplingTrackConfig(temperature=temperature))

        denoised_protein_tensor_output = self.client.forward_and_sample(
            protein_tensor, sampling_configuration=sampling_config
        )
        assert not isinstance(denoised_protein_tensor_output, ESMProteinError)
        denoised_protein_tensor = denoised_protein_tensor_output.protein_tensor
        output_track_tensor = getattr(denoised_protein_tensor, track)
        assert output_track_tensor is not None
        track_tensor[mask_indices] = output_track_tensor[mask_indices]
        setattr(protein_tensor, track, track_tensor)

        return protein_tensor

    def predict_denoised(
        self, protein_tensor: ESMProteinTensor, temperature: float = 0.0
    ) -> ESMProtein:
        denoised_protein_tensor_output = self.client.forward_and_sample(
            protein_tensor,
            sampling_configuration=SamplingConfig(
                sequence=SamplingTrackConfig(temperature=temperature),
                structure=SamplingTrackConfig(temperature=temperature),
            ),
        )
        assert not isinstance(denoised_protein_tensor_output, ESMProteinError)
        denoised_protein_tensor = denoised_protein_tensor_output.protein_tensor
        denoised_protein = self.client.decode(denoised_protein_tensor)
        assert not isinstance(denoised_protein, ESMProteinError)
        return denoised_protein

    def maybe_add_default_structure_tokens(
        self, protein_tensor: ESMProteinTensor
    ) -> ESMProteinTensor:
        empty_protein_tensor = ESMProteinTensor.empty(
            len(protein_tensor) - 2,
            tokenizers=self.tokenizers,
            device=protein_tensor.device,
        )
        if protein_tensor.structure is None:
            setattr(protein_tensor, "structure", empty_protein_tensor.structure)
        else:
            print("Warning: structure already exists in protein_tensor")
        return protein_tensor