跳转至

RNA-MSM

Pre-trained model on non-coding RNA (ncRNA) with multi (homologous) sequence alignment using a masked language modeling (MLM) objective.

Disclaimer

This is an UNOFFICIAL implementation of the Multiple sequence alignment-based RNA language model and its application to structural inference by Yikun Zhang, Mei Lang, Jiuhong Jiang, Zhiqiang Gao, et al.

The OFFICIAL repository of RNA-MSM is at yikunpku/RNA-MSM.

Caution

The MultiMolecule team is aware of a potential risk in reproducing the results of RNA-MSM.

The original implementation of RNA-MSM used a custom tokenizer that does not append <eos> token to the end of the input sequence in consistent to MSA Transformer. This should not affect the performance of the model in most cases, but it can lead to unexpected behavior in some cases.

Please set eos_token=None explicitly in the tokenizer if you want the exact behavior of the original implementation.

See more at issue #10

Tip

The MultiMolecule team has confirmed that the provided model and checkpoints are producing the same intermediate representations as the original implementation.

The team releasing RNA-MSM did not write this model card for this model so this model card has been written by the MultiMolecule team.

Model Details

RNA-MSM is a bert-style model pre-trained on a large corpus of non-coding RNA sequences in a self-supervised fashion. This means that the model was trained on the raw nucleotides of RNA sequences only, with an automatic process to generate inputs and labels from those texts. Please refer to the Training Details section for more information on the training process.

Model Specification

Num Layers Hidden Size Num Heads Intermediate Size Num Parameters (M) FLOPs (G) MACs (G) Max Num Tokens
10 768 12 3072 95.92 21.66 10.57 1024

Usage

The model file depends on the multimolecule library. You can install it using pip:

Bash
1
pip install multimolecule

Direct Use

You can use this model directly with a pipeline for masked language modeling:

Python
 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
>>> import multimolecule  # you must import multimolecule to register models
>>> from transformers import pipeline
>>> unmasker = pipeline("fill-mask", model="multimolecule/rnamsm")
>>> unmasker("gguc<mask>cucugguuagaccagaucugagccu")

[{'score': 0.25111356377601624,
  'token': 9,
  'token_str': 'U',
  'sequence': 'G G U C U C U C U G G U U A G A C C A G A U C U G A G C C U'},
 {'score': 0.1200353354215622,
  'token': 14,
  'token_str': 'W',
  'sequence': 'G G U C W C U C U G G U U A G A C C A G A U C U G A G C C U'},
 {'score': 0.10132723301649094,
  'token': 15,
  'token_str': 'K',
  'sequence': 'G G U C K C U C U G G U U A G A C C A G A U C U G A G C C U'},
 {'score': 0.08383019268512726,
  'token': 18,
  'token_str': 'D',
  'sequence': 'G G U C D C U C U G G U U A G A C C A G A U C U G A G C C U'},
 {'score': 0.05737845227122307,
  'token': 6,
  'token_str': 'A',
  'sequence': 'G G U C A C U C U G G U U A G A C C A G A U C U G A G C C U'}]

Downstream Use

Extract Features

Here is how to use this model to get the features of a given sequence in PyTorch:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
from multimolecule import RnaTokenizer, RnaMsmModel


tokenizer = RnaTokenizer.from_pretrained("multimolecule/rnamsm")
model = RnaMsmModel.from_pretrained("multimolecule/rnamsm")

text = "UAGCUUAUCAGACUGAUGUUGA"
input = tokenizer(text, return_tensors="pt")

output = model(**input)

Sequence Classification / Regression

Note: This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for sequence classification or regression.

Here is how to use this model as backbone to fine-tune for a sequence-level task in PyTorch:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
import torch
from multimolecule import RnaTokenizer, RnaMsmForSequencePrediction


tokenizer = RnaTokenizer.from_pretrained("multimolecule/rnamsm")
model = RnaMsmForSequencePrediction.from_pretrained("multimolecule/rnamsm")

text = "UAGCUUAUCAGACUGAUGUUGA"
input = tokenizer(text, return_tensors="pt")
label = torch.tensor([1])

output = model(**input, labels=label)

Token Classification / Regression

Note: This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for nucleotide classification or regression.

Here is how to use this model as backbone to fine-tune for a nucleotide-level task in PyTorch:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
import torch
from multimolecule import RnaTokenizer, RnaMsmForTokenPrediction


tokenizer = RnaTokenizer.from_pretrained("multimolecule/rnamsm")
model = RnaMsmForNucleotidPrediction.from_pretrained("multimolecule/rnamsm")

text = "UAGCUUAUCAGACUGAUGUUGA"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), ))

output = model(**input, labels=label)

Contact Classification / Regression

Note: This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for contact classification or regression.

Here is how to use this model as backbone to fine-tune for a contact-level task in PyTorch:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
import torch
from multimolecule import RnaTokenizer, RnaMsmForContactPrediction


tokenizer = RnaTokenizer.from_pretrained("multimolecule/rnamsm")
model = RnaMsmForContactPrediction.from_pretrained("multimolecule/rnamsm")

text = "UAGCUUAUCAGACUGAUGUUGA"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), len(text)))

output = model(**input, labels=label)

Training Details

RNA-MSM used Masked Language Modeling (MLM) as the pre-training objective: taking a sequence, the model randomly masks 15% of the tokens in the input then runs the entire masked sentence through the model and has to predict the masked tokens. This is comparable to the Cloze task in language modeling.

Training Data

The RNA-MSM model was pre-trained on Rfam. The Rfam database is a collection of RNA sequence families of structural RNAs including non-coding RNA genes as well as cis-regulatory elements. RNA-MSM used Rfam 14.7 which contains 4,069 RNA families.

To avoid potential overfitting in structural inference, RNA-MSM excluded families with experimentally determined structures, such as ribosomal RNAs, transfer RNAs, and small nuclear RNAs. The final dataset contains 3,932 RNA families. The median value for the number of MSA sequences for these families by RNAcmap3 is 2,184.

To increase the number of homologous sequences, RNA-MSM used an automatic pipeline, RNAcmap3, for homolog search and sequence alignment. RNAcmap3 is a pipeline that combines the BLAST-N, INFERNAL, Easel, RNAfold and evolutionary coupling tools to generate homologous sequences.

RNA-MSM preprocessed all tokens by replacing “T”s with “U”s and substituting “R”, “Y”, “K”, “M”, “S”, “W”, “B”, “D”, “H”, “V”, “N” with “X”.

Note that RnaTokenizer will convert “T”s to “U”s for you, you may disable this behaviour by passing replace_T_with_U=False. RnaTokenizer does not perform other substitutions.

Training Procedure

Preprocessing

RNA-MSM used masked language modeling (MLM) as the pre-training objective. The masking procedure is similar to the one used in BERT:

  • 15% of the tokens are masked.
  • In 80% of the cases, the masked tokens are replaced by <mask>.
  • In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.
  • In the 10% remaining cases, the masked tokens are left as is.

PreTraining

The model was trained on 8 NVIDIA V100 GPUs with 32GiB memories.

  • Learning rate: 3e-4
  • Weight decay: 3e-4
  • Optimizer: Adam
  • Learning rate warm-up: 16,000 steps
  • Epochs: 300
  • Batch Size: 1
  • Dropout: 0.1

Citation

BibTeX:

BibTeX
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
@article{zhang2023multiple,
    author = {Zhang, Yikun and Lang, Mei and Jiang, Jiuhong and Gao, Zhiqiang and Xu, Fan and Litfin, Thomas and Chen, Ke and Singh, Jaswinder and Huang, Xiansong and Song, Guoli and Tian, Yonghong and Zhan, Jian and Chen, Jie and Zhou, Yaoqi},
    title = "{Multiple sequence alignment-based RNA language model and its application to structural inference}",
    journal = {Nucleic Acids Research},
    volume = {52},
    number = {1},
    pages = {e3-e3},
    year = {2023},
    month = {11},
    abstract = "{Compared with proteins, DNA and RNA are more difficult languages to interpret because four-letter coded DNA/RNA sequences have less information content than 20-letter coded protein sequences. While BERT (Bidirectional Encoder Representations from Transformers)-like language models have been developed for RNA, they are ineffective at capturing the evolutionary information from homologous sequences because unlike proteins, RNA sequences are less conserved. Here, we have developed an unsupervised multiple sequence alignment-based RNA language model (RNA-MSM) by utilizing homologous sequences from an automatic pipeline, RNAcmap, as it can provide significantly more homologous sequences than manually annotated Rfam. We demonstrate that the resulting unsupervised, two-dimensional attention maps and one-dimensional embeddings from RNA-MSM contain structural information. In fact, they can be directly mapped with high accuracy to 2D base pairing probabilities and 1D solvent accessibilities, respectively. Further fine-tuning led to significantly improved performance on these two downstream tasks compared with existing state-of-the-art techniques including SPOT-RNA2 and RNAsnap2. By comparison, RNA-FM, a BERT-based RNA language model, performs worse than one-hot encoding with its embedding in base pair and solvent-accessible surface area prediction. We anticipate that the pre-trained RNA-MSM model can be fine-tuned on many other tasks related to RNA structure and function.}",
    issn = {0305-1048},
    doi = {10.1093/nar/gkad1031},
    url = {https://doi.org/10.1093/nar/gkad1031},
    eprint = {https://academic.oup.com/nar/article-pdf/52/1/e3/55443207/gkad1031.pdf},
}

Contact

Please use GitHub issues of MultiMolecule for any questions or comments on the model card.

Please contact the authors of the RNA-MSM paper for questions or comments on the paper/model.

License

This model is licensed under the AGPL-3.0 License.

Text Only
1
SPDX-License-Identifier: AGPL-3.0-or-later

multimolecule.models.rnamsm

RnaTokenizer

Bases: Tokenizer

Tokenizer for RNA sequences.

Parameters:

Name Type Description Default

alphabet

Alphabet | str | List[str] | None

alphabet to use for tokenization.

  • If is None, the standard RNA alphabet will be used.
  • If is a string, it should correspond to the name of a predefined alphabet. The options include
    • standard
    • extended
    • streamline
    • nucleobase
  • If is an alphabet or a list of characters, that specific alphabet will be used.
 None

nmers

int

Size of kmer to tokenize.

 1

codon

bool

Whether to tokenize into codons.

 False

replace_T_with_U

bool

Whether to replace T with U.

 True

do_upper_case

bool

Whether to convert input to uppercase.

 True

Examples:

Python Console Session
>>> from multimolecule import RnaTokenizer
>>> tokenizer = RnaTokenizer()
>>> tokenizer('<pad><cls><eos><unk><mask><null>ACGUNRYSWKMBDHV.X*-I')["input_ids"]
[1, 0, 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, 2]
>>> tokenizer('acgu')["input_ids"]
[1, 6, 7, 8, 9, 2]
>>> tokenizer('acgt')["input_ids"]
[1, 6, 7, 8, 9, 2]
>>> tokenizer = RnaTokenizer(replace_T_with_U=False)
>>> tokenizer('acgt')["input_ids"]
[1, 6, 7, 8, 3, 2]
>>> tokenizer = RnaTokenizer(nmers=3)
>>> tokenizer('uagcuuauc')["input_ids"]
[1, 83, 17, 64, 49, 96, 84, 22, 2]
>>> tokenizer = RnaTokenizer(codon=True)
>>> tokenizer('uagcuuauc')["input_ids"]
[1, 83, 49, 22, 2]
>>> tokenizer('uagcuuauca')["input_ids"]
Traceback (most recent call last):
ValueError: length of input sequence must be a multiple of 3 for codon tokenization, but got 10
Source code in multimolecule/tokenisers/rna/tokenization_rna.py
Python
 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
class RnaTokenizer(Tokenizer):
    """
    Tokenizer for RNA sequences.

    Args:
        alphabet: alphabet to use for tokenization.

            - If is `None`, the standard RNA alphabet will be used.
            - If is a `string`, it should correspond to the name of a predefined alphabet. The options include
                + `standard`
                + `extended`
                + `streamline`
                + `nucleobase`
            - If is an alphabet or a list of characters, that specific alphabet will be used.
        nmers: Size of kmer to tokenize.
        codon: Whether to tokenize into codons.
        replace_T_with_U: Whether to replace T with U.
        do_upper_case: Whether to convert input to uppercase.

    Examples:
        >>> from multimolecule import RnaTokenizer
        >>> tokenizer = RnaTokenizer()
        >>> tokenizer('<pad><cls><eos><unk><mask><null>ACGUNRYSWKMBDHV.X*-I')["input_ids"]
        [1, 0, 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, 2]
        >>> tokenizer('acgu')["input_ids"]
        [1, 6, 7, 8, 9, 2]
        >>> tokenizer('acgt')["input_ids"]
        [1, 6, 7, 8, 9, 2]
        >>> tokenizer = RnaTokenizer(replace_T_with_U=False)
        >>> tokenizer('acgt')["input_ids"]
        [1, 6, 7, 8, 3, 2]
        >>> tokenizer = RnaTokenizer(nmers=3)
        >>> tokenizer('uagcuuauc')["input_ids"]
        [1, 83, 17, 64, 49, 96, 84, 22, 2]
        >>> tokenizer = RnaTokenizer(codon=True)
        >>> tokenizer('uagcuuauc')["input_ids"]
        [1, 83, 49, 22, 2]
        >>> tokenizer('uagcuuauca')["input_ids"]
        Traceback (most recent call last):
        ValueError: length of input sequence must be a multiple of 3 for codon tokenization, but got 10
    """

    model_input_names = ["input_ids", "attention_mask"]

    def __init__(
        self,
        alphabet: Alphabet | str | List[str] | None = None,
        nmers: int = 1,
        codon: bool = False,
        replace_T_with_U: bool = True,
        do_upper_case: bool = True,
        additional_special_tokens: List | Tuple | None = None,
        **kwargs,
    ):
        if codon and (nmers > 1 and nmers != 3):
            raise ValueError("Codon and nmers cannot be used together.")
        if codon:
            nmers = 3  # set to 3 to get correct vocab
        if not isinstance(alphabet, Alphabet):
            alphabet = get_alphabet(alphabet, nmers=nmers)
        super().__init__(
            alphabet=alphabet,
            nmers=nmers,
            codon=codon,
            replace_T_with_U=replace_T_with_U,
            do_upper_case=do_upper_case,
            additional_special_tokens=additional_special_tokens,
            **kwargs,
        )
        self.replace_T_with_U = replace_T_with_U
        self.nmers = nmers
        self.condon = codon

    def _tokenize(self, text: str, **kwargs):
        if self.do_upper_case:
            text = text.upper()
        if self.replace_T_with_U:
            text = text.replace("T", "U")
        if self.condon:
            if len(text) % 3 != 0:
                raise ValueError(
                    f"length of input sequence must be a multiple of 3 for codon tokenization, but got {len(text)}"
                )
            return [text[i : i + 3] for i in range(0, len(text), 3)]
        if self.nmers > 1:
            return [text[i : i + self.nmers] for i in range(len(text) - self.nmers + 1)]  # noqa: E203
        return list(text)

RnaMsmConfig

Bases: PreTrainedConfig

This is the configuration class to store the configuration of a RnaMsmModel. It is used to instantiate a RnaMsm model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the RnaMsm yikunpku/RNA-MSM architecture.

Configuration objects inherit from PreTrainedConfig and can be used to control the model outputs. Read the documentation from PreTrainedConfig for more information.

Parameters:

Name Type Description Default

vocab_size

int

Vocabulary size of the RnaMsm model. Defines the number of different tokens that can be represented by the inputs_ids passed when calling [RnaMsmModel].

 26

hidden_size

int

Dimensionality of the encoder layers and the pooler layer.

 768

num_hidden_layers

int

Number of hidden layers in the Transformer encoder.

 10

num_attention_heads

int

Number of attention heads for each attention layer in the Transformer encoder.

 12

intermediate_size

int

Dimensionality of the “intermediate” (often named feed-forward) layer in the Transformer encoder.

 3072

hidden_dropout

float

The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.

 0.1

attention_dropout

float

The dropout ratio for the attention probabilities.

 0.1

max_position_embeddings

int

The maximum sequence length that this model might ever be used with. Typically set this to something large just in case (e.g., 512 or 1024 or 2048).

 1024

initializer_range

float

The standard deviation of the truncated_normal_initializer for initializing all weight matrices.

 0.02

layer_norm_eps

float

The epsilon used by the layer normalization layers.

 1e-12

Examples:

Python Console Session
>>> from multimolecule import RnaMsmModel, RnaMsmConfig
>>> # Initializing a RNA-MSM multimolecule/rnamsm style configuration
>>> configuration = RnaMsmConfig()
>>> # Initializing a model (with random weights) from the multimolecule/rnamsm style configuration
>>> model = RnaMsmModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
Source code in multimolecule/models/rnamsm/configuration_rnamsm.py
Python
 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
class RnaMsmConfig(PreTrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`RnaMsmModel`][multimolecule.models.RnaMsmModel].
    It is used to instantiate a RnaMsm model according to the specified arguments, defining the model architecture.
    Instantiating a configuration with the defaults will yield a similar configuration to that of the RnaMsm
    [yikunpku/RNA-MSM](https://github.com/yikunpku/RNA-MSM) architecture.

    Configuration objects inherit from [`PreTrainedConfig`][multimolecule.models.PreTrainedConfig] and can be used to
    control the model outputs. Read the documentation from [`PreTrainedConfig`][multimolecule.models.PreTrainedConfig]
    for more information.

    Args:
        vocab_size:
            Vocabulary size of the RnaMsm model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`RnaMsmModel`].
        hidden_size:
            Dimensionality of the encoder layers and the pooler layer.
        num_hidden_layers:
            Number of hidden layers in the Transformer encoder.
        num_attention_heads:
            Number of attention heads for each attention layer in the Transformer encoder.
        intermediate_size:
            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
        hidden_dropout:
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        attention_dropout:
            The dropout ratio for the attention probabilities.
        max_position_embeddings:
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        initializer_range:
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        layer_norm_eps:
            The epsilon used by the layer normalization layers.

    Examples:
        >>> from multimolecule import RnaMsmModel, RnaMsmConfig
        >>> # Initializing a RNA-MSM multimolecule/rnamsm style configuration
        >>> configuration = RnaMsmConfig()
        >>> # Initializing a model (with random weights) from the multimolecule/rnamsm style configuration
        >>> model = RnaMsmModel(configuration)
        >>> # Accessing the model configuration
        >>> configuration = model.config
    """

    model_type = "rnamsm"

    def __init__(
        self,
        vocab_size: int = 26,
        hidden_size: int = 768,
        num_hidden_layers: int = 10,
        num_attention_heads: int = 12,
        intermediate_size: int = 3072,
        hidden_act: str = "gelu",
        hidden_dropout: float = 0.1,
        attention_dropout: float = 0.1,
        max_position_embeddings: int = 1024,
        initializer_range: float = 0.02,
        layer_norm_eps: float = 1e-12,
        position_embedding_type: str = "absolute",
        is_decoder: bool = False,
        use_cache: bool = True,
        max_tokens_per_msa: int = 2**14,
        layer_type: str = "standard",
        attention_type: str = "standard",
        embed_positions_msa: bool = True,
        attention_bias: bool = True,
        head: HeadConfig | None = None,
        lm_head: MaskedLMHeadConfig | None = None,
        **kwargs,
    ):
        super().__init__(**kwargs)

        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        self.hidden_dropout = hidden_dropout
        self.attention_dropout = attention_dropout
        self.max_position_embeddings = max_position_embeddings
        self.initializer_range = initializer_range
        self.layer_norm_eps = layer_norm_eps
        self.position_embedding_type = position_embedding_type
        self.is_decoder = is_decoder
        self.use_cache = use_cache
        self.max_tokens_per_msa = max_tokens_per_msa
        self.layer_type = layer_type
        self.attention_type = attention_type
        self.embed_positions_msa = embed_positions_msa
        self.attention_bias = attention_bias
        self.head = HeadConfig(**head if head is not None else {})
        self.lm_head = MaskedLMHeadConfig(**lm_head if lm_head is not None else {})

RnaMsmForContactPrediction

Bases: RnaMsmPreTrainedModel

Examples:

Python Console Session
>>> from multimolecule import RnaMsmConfig, RnaMsmForContactPrediction, RnaTokenizer
>>> config = RnaMsmConfig()
>>> model = RnaMsmForContactPrediction(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=torch.randint(2, (1, 5, 5)))
>>> output["logits"].shape
torch.Size([1, 5, 5, 2])
>>> output["loss"]
tensor(..., grad_fn=<NllLossBackward0>)
Source code in multimolecule/models/rnamsm/modeling_rnamsm.py
Python
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
class RnaMsmForContactPrediction(RnaMsmPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RnaMsmConfig, RnaMsmForContactPrediction, RnaTokenizer
        >>> config = RnaMsmConfig()
        >>> model = RnaMsmForContactPrediction(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels=torch.randint(2, (1, 5, 5)))
        >>> output["logits"].shape
        torch.Size([1, 5, 5, 2])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<NllLossBackward0>)
    """

    def __init__(self, config: RnaMsmConfig):
        super().__init__(config)
        self.rnamsm = RnaMsmModel(config, add_pooling_layer=True)
        head_config = HeadConfig(output_name="row_attentions")
        self.contact_head = ContactPredictionHead(config, head_config)
        self.head_config = self.contact_head.config

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        position_ids: Tensor | None = None,
        head_mask: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RnaMsmContactPredictorOutput:
        if output_attentions is False:
            warn("output_attentions must be True for contact classification and will be ignored.")
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.rnamsm(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=True,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            **kwargs,
        )
        output = self.contact_head(outputs, attention_mask, input_ids, labels)
        logits, loss = output.logits, output.loss

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return RnaMsmContactPredictorOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            col_attentions=outputs.col_attentions,
            row_attentions=outputs.row_attentions,
        )

RnaMsmForMaskedLM

Bases: RnaMsmPreTrainedModel

Examples:

Python Console Session
>>> from multimolecule import RnaMsmConfig, RnaMsmForMaskedLM, RnaTokenizer
>>> config = RnaMsmConfig()
>>> model = RnaMsmForMaskedLM(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=input["input_ids"])
>>> output["logits"].shape
torch.Size([1, 7, 26])
>>> output["loss"]
tensor(..., grad_fn=<NllLossBackward0>)
Source code in multimolecule/models/rnamsm/modeling_rnamsm.py
Python
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
class RnaMsmForMaskedLM(RnaMsmPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RnaMsmConfig, RnaMsmForMaskedLM, RnaTokenizer
        >>> config = RnaMsmConfig()
        >>> model = RnaMsmForMaskedLM(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels=input["input_ids"])
        >>> output["logits"].shape
        torch.Size([1, 7, 26])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<NllLossBackward0>)
    """

    _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]

    def __init__(self, config: RnaMsmConfig):
        super().__init__(config)
        self.rnamsm = RnaMsmModel(config, add_pooling_layer=False)
        self.lm_head = MaskedLMHead(config, weight=self.rnamsm.embeddings.word_embeddings.weight)

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        position_ids: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RnaMsmForMaskedLMOutput:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.rnamsm(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            **kwargs,
        )
        output = self.lm_head(outputs, labels)
        logits, loss = output.logits, output.loss

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return RnaMsmForMaskedLMOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            col_attentions=outputs.col_attentions,
            row_attentions=outputs.row_attentions,
        )

RnaMsmForPreTraining

Bases: RnaMsmPreTrainedModel

Examples:

Python Console Session
>>> from multimolecule import RnaMsmConfig, RnaMsmForPreTraining, RnaTokenizer
>>> config = RnaMsmConfig()
>>> model = RnaMsmForPreTraining(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels_mlm=input["input_ids"])
>>> output["loss"]
tensor(..., grad_fn=<AddBackward0>)
>>> output["logits"].shape
torch.Size([1, 7, 26])
>>> output["contact_map"].shape
torch.Size([1, 5, 5, 2])
Source code in multimolecule/models/rnamsm/modeling_rnamsm.py
Python
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
class RnaMsmForPreTraining(RnaMsmPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RnaMsmConfig, RnaMsmForPreTraining, RnaTokenizer
        >>> config = RnaMsmConfig()
        >>> model = RnaMsmForPreTraining(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels_mlm=input["input_ids"])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<AddBackward0>)
        >>> output["logits"].shape
        torch.Size([1, 7, 26])
        >>> output["contact_map"].shape
        torch.Size([1, 5, 5, 2])
    """

    _tied_weights_keys = [
        "lm_head.decoder.weight",
        "lm_head.decoder.bias",
        "pretrain.predictions.decoder.weight",
        "pretrain.predictions.decoder.bias",
        "pretrain.predictions_ss.decoder.weight",
        "pretrain.predictions_ss.decoder.bias",
    ]

    def __init__(self, config: RnaMsmConfig):
        super().__init__(config)
        self.rnamsm = RnaMsmModel(config, add_pooling_layer=False)
        self.pretrain = RnaMsmPreTrainingHeads(config, weight=self.rnamsm.embeddings.word_embeddings.weight)

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        position_ids: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels_mlm: Tensor | None = None,
        labels_contact: Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RnaMsmForPreTrainingOutput:
        if output_attentions is False:
            warn("output_attentions must be True for contact classification and will be ignored.")
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.rnamsm(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            output_attentions=True,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            **kwargs,
        )
        total_loss, logits, contact_map = self.pretrain(
            outputs, attention_mask, input_ids, labels_mlm=labels_mlm, labels_contact=labels_contact
        )

        if not return_dict:
            output = (logits, contact_map) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return RnaMsmForPreTrainingOutput(
            loss=total_loss,
            logits=logits,
            contact_map=contact_map,
            hidden_states=outputs.hidden_states,
            col_attentions=outputs.col_attentions,
            row_attentions=outputs.row_attentions,
        )

RnaMsmForSequencePrediction

Bases: RnaMsmPreTrainedModel

Examples:

Python Console Session
>>> from multimolecule import RnaMsmConfig, RnaMsmForSequencePrediction, RnaTokenizer
>>> config = RnaMsmConfig()
>>> model = RnaMsmForSequencePrediction(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=torch.tensor([[1]]))
>>> output["logits"].shape
torch.Size([1, 2])
>>> output["loss"]
tensor(..., grad_fn=<NllLossBackward0>)
Source code in multimolecule/models/rnamsm/modeling_rnamsm.py
Python
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
class RnaMsmForSequencePrediction(RnaMsmPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RnaMsmConfig, RnaMsmForSequencePrediction, RnaTokenizer
        >>> config = RnaMsmConfig()
        >>> model = RnaMsmForSequencePrediction(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels=torch.tensor([[1]]))
        >>> output["logits"].shape
        torch.Size([1, 2])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<NllLossBackward0>)
    """

    def __init__(self, config: RnaMsmConfig):
        super().__init__(config)
        self.rnamsm = RnaMsmModel(config, add_pooling_layer=True)
        self.sequence_head = SequencePredictionHead(config)
        self.head_config = self.sequence_head.config

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        position_ids: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RnaMsmSequencePredictorOutput:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.rnamsm(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            **kwargs,
        )
        output = self.sequence_head(outputs, labels)
        logits, loss = output.logits, output.loss

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return RnaMsmSequencePredictorOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            col_attentions=outputs.col_attentions,
            row_attentions=outputs.row_attentions,
        )

RnaMsmForTokenPrediction

Bases: RnaMsmPreTrainedModel

Examples:

Python Console Session
>>> from multimolecule import RnaMsmConfig, RnaMsmForTokenPrediction, RnaTokenizer
>>> config = RnaMsmConfig()
>>> model = RnaMsmForTokenPrediction(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=torch.randint(2, (1, 5)))
>>> output["logits"].shape
torch.Size([1, 5, 2])
>>> output["loss"]
tensor(..., grad_fn=<NllLossBackward0>)
Source code in multimolecule/models/rnamsm/modeling_rnamsm.py
Python
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
class RnaMsmForTokenPrediction(RnaMsmPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RnaMsmConfig, RnaMsmForTokenPrediction, RnaTokenizer
        >>> config = RnaMsmConfig()
        >>> model = RnaMsmForTokenPrediction(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels=torch.randint(2, (1, 5)))
        >>> output["logits"].shape
        torch.Size([1, 5, 2])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<NllLossBackward0>)
    """

    def __init__(self, config: RnaMsmConfig):
        super().__init__(config)
        self.rnamsm = RnaMsmModel(config, add_pooling_layer=True)
        self.token_head = TokenPredictionHead(config)
        self.head_config = self.token_head.config

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        position_ids: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RnaMsmTokenPredictorOutput:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.rnamsm(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            **kwargs,
        )
        output = self.token_head(outputs, attention_mask, input_ids, labels)
        logits, loss = output.logits, output.loss

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return RnaMsmTokenPredictorOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            col_attentions=outputs.col_attentions,
            row_attentions=outputs.row_attentions,
        )

RnaMsmModel

Bases: RnaMsmPreTrainedModel

Examples:

Python Console Session
>>> from multimolecule import RnaMsmConfig, RnaMsmModel, RnaTokenizer
>>> config = RnaMsmConfig()
>>> model = RnaMsmModel(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input)
>>> output["last_hidden_state"].shape
torch.Size([1, 7, 768])
>>> output["pooler_output"].shape
torch.Size([1, 768])
Source code in multimolecule/models/rnamsm/modeling_rnamsm.py
Python
 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
class RnaMsmModel(RnaMsmPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RnaMsmConfig, RnaMsmModel, RnaTokenizer
        >>> config = RnaMsmConfig()
        >>> model = RnaMsmModel(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input)
        >>> output["last_hidden_state"].shape
        torch.Size([1, 7, 768])
        >>> output["pooler_output"].shape
        torch.Size([1, 768])
    """

    def __init__(self, config: RnaMsmConfig, add_pooling_layer: bool = True):
        super().__init__(config)
        self.pad_token_id = config.pad_token_id
        self.embeddings = RnaMsmEmbeddings(config)
        self.encoder = RnaMsmEncoder(config)
        self.pooler = RnaMsmPooler(config) if add_pooling_layer else None

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        position_ids: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RnaMsmModelOutputWithPooling:
        if kwargs:
            warn(
                f"Additional keyword arguments `{', '.join(kwargs)}` are detected in "
                f"`{self.__class__.__name__}.forward`, they will be ignored.\n"
                "This is provided for backward compatibility and may lead to unexpected behavior."
            )
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if isinstance(input_ids, NestedTensor):
            input_ids, attention_mask = input_ids.tensor, input_ids.mask
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        if input_ids is not None:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
        elif inputs_embeds is None:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if attention_mask is None:
            attention_mask = (
                input_ids.ne(self.pad_token_id) if self.pad_token_id is not None else torch.ones_like(input_ids)
            )

        unsqueeze_input = input_ids.ndim == 2
        if unsqueeze_input:
            input_ids = input_ids.unsqueeze(1)
        if attention_mask.ndim == 2:
            attention_mask = attention_mask.unsqueeze(1)

        embedding_output = self.embeddings(
            input_ids=input_ids,
            position_ids=position_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
        )
        encoder_outputs = self.encoder(
            embedding_output,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = encoder_outputs[0]
        if unsqueeze_input:
            sequence_output = sequence_output.squeeze(1)
        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None

        if not return_dict:
            return (sequence_output, pooled_output) + encoder_outputs[1:]

        return RnaMsmModelOutputWithPooling(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            col_attentions=encoder_outputs.col_attentions,
            row_attentions=encoder_outputs.row_attentions,
        )

RnaMsmPreTrainedModel

Bases: PreTrainedModel

An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models.

Source code in multimolecule/models/rnamsm/modeling_rnamsm.py
Python
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
class RnaMsmPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = RnaMsmConfig
    base_model_prefix = "rnamsm"
    supports_gradient_checkpointing = True
    _no_split_modules = ["RnaMsmLayer", "RnaMsmAxialLayer", "RnaMsmPkmLayer", "RnaMsmEmbeddings"]

    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
    def _init_weights(self, module: nn.Module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm) and module.elementwise_affine:
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)