Skip to content

Malinois

Convolutional neural network for predicting cell-type-targeting cis-regulatory element (CRE) activity from DNA sequence.

Disclaimer

This is an UNOFFICIAL implementation of Machine-guided design of cell-type-targeting cis-regulatory elements by Sager J. Gosai, Rodrigo I. Castro, et al.

The OFFICIAL repository of Malinois is at sjgosai/boda2.

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 Malinois did not write this model card for this model so this model card has been written by the MultiMolecule team.

Model Details

Malinois is a deep convolutional neural network (a tuned Basset-style “branched” architecture) trained to quantitatively predict cell-type-informed CRE activity from ~200 bp DNA sequences measured by a massively parallel reporter assay (MPRA). The model emits three regression outputs, one per human cell line: K562, HepG2 and SK-N-SH (in that order).

The architecture consists of three convolutional blocks, one shared fully-connected block, and a branched grouped-linear tower with an independent parameter set per cell line. Please refer to the Training Details section for more information on the training process.

Model Specification

Num Layers Hidden Size Num Parameters (M) FLOPs (M) MACs (M) Max Num Tokens
8 420 4.11 332.95 165.70 600

Usage

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

Bash
1
pip install multimolecule

Direct Use

CRE Activity Prediction

You can use this model directly to predict the cell-type-informed CRE activity (K562, HepG2, SK-N-SH) of a sequence. Malinois pads each ~200 bp candidate to 600 bp with fixed MPRA plasmid flanks before inference; the example below uses a pre-padded 600 bp sequence:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
>>> import torch
>>> from multimolecule import DnaTokenizer, MalinoisForSequencePrediction

>>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/malinois")
>>> model = MalinoisForSequencePrediction.from_pretrained("multimolecule/malinois")
>>> sequence = "ACGT" * 150
>>> output = model(**tokenizer(sequence, return_tensors="pt"))

>>> output.logits.shape
torch.Size([1, 3])

Interface

  • Input length: fixed 600 bp window
  • Padding: each ~200 bp candidate CRE is centered and padded with fixed MPRA plasmid flanks (MPRA_UPSTREAM / MPRA_DOWNSTREAM); flank padding is part of the data pipeline, not the model
  • Output: 3 cell-line CRE activity values (K562, HepG2, SK-N-SH)

Training Details

Malinois was trained to predict quantitative, cell-type-informed CRE activity from DNA sequence.

Training Data

Malinois was trained on a lentiMPRA dataset measuring the regulatory activity of ~200 bp sequences across three human cell lines (K562, HepG2 and SK-N-SH). Each training example is a sequence with three continuous activity values (log2 fold-change over input), one per cell line. Genomic sequences were split by chromosome into training, validation, and test sets to avoid sequence leakage.

Training Procedure

Pre-training

The model was trained to minimize an L1 + KL-divergence mixed loss between predicted and measured cell-type CRE activities, with the architecture and training hyperparameters selected by Bayesian optimization.

  • Optimizer: Adam
  • Loss: L1 + KL-divergence mixed loss
  • Early stopping on validation loss

Citation

BibTeX
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
@article{gosai2024malinois,
  author    = {Gosai, Sager J. and Castro, Rodrigo I. and Fuentes, Natalia and Butts, John C. and Mouri, Kousuke and Alasoadura, Michael and Kales, Susan and Nguyen, Thanh Thanh L. and Noche, Ramil R. and Rao, Arya S. and Joy, Mary T. and Sabeti, Pardis C. and Reilly, Steven K. and Tewhey, Ryan},
  journal   = {Nature},
  month     = oct,
  number    = 8036,
  pages     = {1211--1220},
  publisher = {Springer Science and Business Media LLC},
  title     = {Machine-guided design of cell-type-targeting cis-regulatory elements},
  volume    = 634,
  year      = 2024
}

Note

The artifacts distributed in this repository are part of the MultiMolecule project. If you use MultiMolecule in your research, you must cite the MultiMolecule project as follows:

BibTeX
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
@software{chen_2024_12638419,
  author    = {Chen, Zhiyuan and Zhu, Sophia Y.},
  title     = {MultiMolecule},
  doi       = {10.5281/zenodo.12638419},
  publisher = {Zenodo},
  url       = {https://doi.org/10.5281/zenodo.12638419},
  year      = 2024,
  month     = may,
  day       = 4
}

Contact

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

Please contact the authors of the Malinois paper for questions or comments on the paper/model.

License

This model implementation is licensed under the GNU Affero General Public License.

For additional terms and clarifications, please refer to our License FAQ.

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

multimolecule.models.malinois

DnaTokenizer

Bases: Tokenizer

Tokenizer for DNA 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
    • iupac
    • 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_U_with_T

bool

Whether to replace U with T.

 True

do_upper_case

bool

Whether to convert input to uppercase.

 True

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
>>> from multimolecule import DnaTokenizer
>>> tokenizer = DnaTokenizer()
>>> tokenizer('<pad><cls><eos><unk><mask><null>ACGTNRYSWKMBDHVX|.*-?')["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, 26, 2]
>>> tokenizer('acgt')["input_ids"]
[1, 6, 7, 8, 9, 2]
>>> tokenizer('acgu')["input_ids"]
[1, 6, 7, 8, 9, 2]
>>> tokenizer = DnaTokenizer(replace_U_with_T=False)
>>> tokenizer('acgu')["input_ids"]
[1, 6, 7, 8, 3, 2]
>>> tokenizer = DnaTokenizer(nmers=3)
>>> tokenizer('tataaagta')["input_ids"]
[1, 84, 21, 81, 6, 8, 19, 71, 2]
>>> tokenizer = DnaTokenizer(codon=True)
>>> tokenizer('tataaagta')["input_ids"]
[1, 84, 6, 71, 2]
>>> tokenizer('tataaagtaa')["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/dna/tokenization_dna.py
Python
 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
class DnaTokenizer(Tokenizer):
    """
    Tokenizer for DNA 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`
                + `iupac`
                + `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_U_with_T: Whether to replace U with T.
        do_upper_case: Whether to convert input to uppercase.

    Examples:
        >>> from multimolecule import DnaTokenizer
        >>> tokenizer = DnaTokenizer()
        >>> tokenizer('<pad><cls><eos><unk><mask><null>ACGTNRYSWKMBDHVX|.*-?')["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, 26, 2]
        >>> tokenizer('acgt')["input_ids"]
        [1, 6, 7, 8, 9, 2]
        >>> tokenizer('acgu')["input_ids"]
        [1, 6, 7, 8, 9, 2]
        >>> tokenizer = DnaTokenizer(replace_U_with_T=False)
        >>> tokenizer('acgu')["input_ids"]
        [1, 6, 7, 8, 3, 2]
        >>> tokenizer = DnaTokenizer(nmers=3)
        >>> tokenizer('tataaagta')["input_ids"]
        [1, 84, 21, 81, 6, 8, 19, 71, 2]
        >>> tokenizer = DnaTokenizer(codon=True)
        >>> tokenizer('tataaagta')["input_ids"]
        [1, 84, 6, 71, 2]
        >>> tokenizer('tataaagtaa')["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_U_with_T: 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_U_with_T=replace_U_with_T,
            do_upper_case=do_upper_case,
            additional_special_tokens=additional_special_tokens,
            **kwargs,
        )
        self.replace_U_with_T = replace_U_with_T
        self.nmers = nmers
        self.codon = codon

    def _tokenize(self, text: str, **kwargs):
        if self.do_upper_case:
            text = text.upper()
        if self.replace_U_with_T:
            text = text.replace("U", "T")
        if self.codon:
            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)

MalinoisConfig

Bases: PreTrainedConfig

This is the configuration class to store the configuration of a MalinoisModel. It is used to instantiate a Malinois 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 Malinois sjgosai/boda2 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 Malinois model. Defines the number of feature channels in the one-hot encoded input fed to the first convolution. Defaults to 5.

 5

input_length

int

The fixed length (in base pairs) of the input fed to the first convolution. Upstream Malinois pads each 200 bp candidate sequence with fixed MPRA plasmid flanks up to this length before the convolution stack. Defaults to 600.

 600

conv_channels

list[int] | None

Number of output channels for each convolutional block.

 None

conv_kernel_sizes

list[int] | None

Convolution kernel size for each convolutional block.

 None

num_linear_layers

int

Number of fully-connected layers between the convolutional stack and the branched tower.

 1

linear_channels

int

Hidden size for each fully-connected layer.

 1000

linear_act

str

The non-linear activation function (function or string) applied after the convolutional and linear layers. If string, "gelu", "relu", "silu" and "gelu_new" are supported.

 'relu'

linear_dropout

float

The dropout probability for the fully-connected layers.

 0.11625456877954289

num_branched_layers

int

Number of grouped (branched) layers, one independent tower per output cell line.

 3

branched_channels

int

Hidden size for each branch in the branched tower.

 140

branched_act

str

The non-linear activation function applied between branched layers.

 'relu'

branched_dropout

float

The dropout probability for the branched tower.

 0.5757068086404574

batch_norm_eps

float

The epsilon used by the batch normalization layers.

 1e-05

batch_norm_momentum

float

The momentum used by the batch normalization layers.

 0.1

num_labels

int

Number of regression outputs. Malinois predicts cell-type-informed cis-regulatory activity for three human cell lines: K562, HepG2 and SK-N-SH (in that order).

 3

head

HeadConfig | None

The configuration of the prediction head. Defaults to a regression head (problem_type="regression"), matching Malinois’s CRE activity prediction task.

 None

Examples:

Python Console Session
1
2
3
4
5
6
7
>>> from multimolecule import MalinoisConfig, MalinoisModel
>>> # Initializing a Malinois multimolecule/malinois style configuration
>>> configuration = MalinoisConfig()
>>> # Initializing a model (with random weights) from the multimolecule/malinois style configuration
>>> model = MalinoisModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
Source code in multimolecule/models/malinois/configuration_malinois.py
Python
 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
class MalinoisConfig(PreTrainedConfig):
    r"""
    This is the configuration class to store the configuration of a
    [`MalinoisModel`][multimolecule.models.MalinoisModel]. It is used to instantiate a Malinois 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 Malinois [sjgosai/boda2](https://github.com/sjgosai/boda2) 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 Malinois model. Defines the number of feature channels in the one-hot encoded
            input fed to the first convolution.
            Defaults to 5.
        input_length:
            The fixed length (in base pairs) of the input fed to the first convolution. Upstream Malinois pads each
            200 bp candidate sequence with fixed MPRA plasmid flanks up to this length before the convolution stack.
            Defaults to 600.
        conv_channels:
            Number of output channels for each convolutional block.
        conv_kernel_sizes:
            Convolution kernel size for each convolutional block.
        num_linear_layers:
            Number of fully-connected layers between the convolutional stack and the branched tower.
        linear_channels:
            Hidden size for each fully-connected layer.
        linear_act:
            The non-linear activation function (function or string) applied after the convolutional and linear
            layers. If string, `"gelu"`, `"relu"`, `"silu"` and `"gelu_new"` are supported.
        linear_dropout:
            The dropout probability for the fully-connected layers.
        num_branched_layers:
            Number of grouped (branched) layers, one independent tower per output cell line.
        branched_channels:
            Hidden size for each branch in the branched tower.
        branched_act:
            The non-linear activation function applied between branched layers.
        branched_dropout:
            The dropout probability for the branched tower.
        batch_norm_eps:
            The epsilon used by the batch normalization layers.
        batch_norm_momentum:
            The momentum used by the batch normalization layers.
        num_labels:
            Number of regression outputs. Malinois predicts cell-type-informed cis-regulatory activity for three
            human cell lines: K562, HepG2 and SK-N-SH (in that order).
        head:
            The configuration of the prediction head. Defaults to a regression head
            (`problem_type="regression"`), matching Malinois's CRE activity prediction task.

    Examples:
        >>> from multimolecule import MalinoisConfig, MalinoisModel
        >>> # Initializing a Malinois multimolecule/malinois style configuration
        >>> configuration = MalinoisConfig()
        >>> # Initializing a model (with random weights) from the multimolecule/malinois style configuration
        >>> model = MalinoisModel(configuration)
        >>> # Accessing the model configuration
        >>> configuration = model.config
    """

    model_type = "malinois"

    def __init__(
        self,
        vocab_size: int = 5,
        input_length: int = 600,
        conv_channels: list[int] | None = None,
        conv_kernel_sizes: list[int] | None = None,
        num_linear_layers: int = 1,
        linear_channels: int = 1000,
        linear_act: str = "relu",
        linear_dropout: float = 0.11625456877954289,
        num_branched_layers: int = 3,
        branched_channels: int = 140,
        branched_act: str = "relu",
        branched_dropout: float = 0.5757068086404574,
        batch_norm_eps: float = 1e-5,
        batch_norm_momentum: float = 0.1,
        num_labels: int = 3,
        head: HeadConfig | None = None,
        **kwargs,
    ):
        super().__init__(num_labels=num_labels, **kwargs)
        if conv_channels is None:
            conv_channels = [300, 200, 200]
        if conv_kernel_sizes is None:
            conv_kernel_sizes = [19, 11, 7]
        if len(conv_channels) != len(conv_kernel_sizes):
            raise ValueError(
                f"conv_channels and conv_kernel_sizes must have the same length, "
                f"got {len(conv_channels)} and {len(conv_kernel_sizes)}."
            )
        if len(conv_channels) != 3:
            raise ValueError(f"Malinois uses exactly 3 convolutional blocks, got {len(conv_channels)}.")
        if input_length <= 0:
            raise ValueError(f"input_length must be positive, got {input_length}.")
        if num_linear_layers <= 0:
            raise ValueError(f"num_linear_layers must be positive, got {num_linear_layers}.")
        if num_branched_layers <= 0:
            raise ValueError(f"num_branched_layers must be positive, got {num_branched_layers}.")
        self.vocab_size = vocab_size
        self.input_length = input_length
        self.conv_channels = conv_channels
        self.conv_kernel_sizes = conv_kernel_sizes
        self.num_conv_layers = len(conv_channels)
        self.num_linear_layers = num_linear_layers
        self.linear_channels = linear_channels
        self.linear_act = linear_act
        self.linear_dropout = linear_dropout
        self.num_branched_layers = num_branched_layers
        self.branched_channels = branched_channels
        self.branched_act = branched_act
        self.branched_dropout = branched_dropout
        self.batch_norm_eps = batch_norm_eps
        self.batch_norm_momentum = batch_norm_momentum
        if head is None:
            head = HeadConfig(problem_type="regression")
        else:
            head = HeadConfig(head)
            if head.problem_type is None:
                head.problem_type = "regression"
        self.head = head

    @property
    def flatten_factor(self) -> int:
        hook = self.input_length // 3 // 4
        return (hook + 2) // 4

    @property
    def hidden_size(self) -> int:
        return self.num_labels * self.branched_channels

MalinoisForSequencePrediction

Bases: MalinoisPreTrainedModel

Examples:

Python Console Session
1
2
3
4
5
6
7
8
9
>>> import torch
>>> from multimolecule import MalinoisConfig, MalinoisForSequencePrediction, DnaTokenizer
>>> config = MalinoisConfig()
>>> model = MalinoisForSequencePrediction(config)
>>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/malinois")
>>> input = tokenizer(["ACGT" * 150, "TGCA" * 150], return_tensors="pt")
>>> output = model(**input, labels=torch.randn(2, 3))
>>> output["logits"].shape
torch.Size([2, 3])
Source code in multimolecule/models/malinois/modeling_malinois.py
Python
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
class MalinoisForSequencePrediction(MalinoisPreTrainedModel):
    """
    Examples:
        >>> import torch
        >>> from multimolecule import MalinoisConfig, MalinoisForSequencePrediction, DnaTokenizer
        >>> config = MalinoisConfig()
        >>> model = MalinoisForSequencePrediction(config)
        >>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/malinois")
        >>> input = tokenizer(["ACGT" * 150, "TGCA" * 150], return_tensors="pt")
        >>> output = model(**input, labels=torch.randn(2, 3))
        >>> output["logits"].shape
        torch.Size([2, 3])
    """

    def __init__(self, config: MalinoisConfig):
        super().__init__(config)
        self.model = MalinoisModel(config)
        self.sequence_head = SequencePredictionHead(config)
        self.head_config = self.sequence_head.config

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

    @property
    def output_channels(self) -> list[str]:
        if self.config.num_labels == 3:
            return ["K562", "HepG2", "SK-N-SH"]
        return [f"cell_{index}" for index in range(self.config.num_labels)]

    @can_return_tuple
    def forward(
        self,
        input_ids: Tensor | NestedTensor | None = None,
        attention_mask: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Tuple[Tensor, ...] | SequencePredictorOutput:
        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            return_dict=True,
            **kwargs,
        )

        output = self.sequence_head(outputs, labels)
        logits, loss = output.logits, output.loss

        return SequencePredictorOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

MalinoisModel

Bases: MalinoisPreTrainedModel

Examples:

Python Console Session
1
2
3
4
5
6
7
8
>>> from multimolecule import MalinoisConfig, MalinoisModel, DnaTokenizer
>>> config = MalinoisConfig()
>>> model = MalinoisModel(config)
>>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/malinois")
>>> input = tokenizer(["ACGT" * 150, "TGCA" * 150], return_tensors="pt")
>>> output = model(**input)
>>> output["pooler_output"].shape
torch.Size([2, 420])
Source code in multimolecule/models/malinois/modeling_malinois.py
Python
 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
class MalinoisModel(MalinoisPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import MalinoisConfig, MalinoisModel, DnaTokenizer
        >>> config = MalinoisConfig()
        >>> model = MalinoisModel(config)
        >>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/malinois")
        >>> input = tokenizer(["ACGT" * 150, "TGCA" * 150], return_tensors="pt")
        >>> output = model(**input)
        >>> output["pooler_output"].shape
        torch.Size([2, 420])
    """

    def __init__(self, config: MalinoisConfig):
        super().__init__(config)
        self.embeddings = MalinoisEmbedding(config)
        self.encoder = MalinoisEncoder(config)
        self.pooler = MalinoisPooler(config)

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

    @merge_with_config_defaults
    @capture_outputs
    def forward(
        self,
        input_ids: Tensor | NestedTensor | None = None,
        attention_mask: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> MalinoisModelOutput:
        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")
        elif input_ids is None and inputs_embeds is None:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if isinstance(input_ids, NestedTensor):
            if attention_mask is None:
                attention_mask = input_ids.mask
            input_ids = input_ids.tensor
        if isinstance(inputs_embeds, NestedTensor):
            if attention_mask is None:
                attention_mask = inputs_embeds.mask
            inputs_embeds = inputs_embeds.tensor

        embedding_output = self.embeddings(
            input_ids=input_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
        )
        sequence_output = self.encoder(embedding_output)
        pooled_output = self.pooler(sequence_output)

        return MalinoisModelOutput(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
        )

MalinoisModelOutput dataclass

Bases: ModelOutput

Base class for outputs of the Malinois model.

Parameters:

Name Type Description Default

last_hidden_state

`torch.FloatTensor` of shape `(batch_size, flattened_conv_features)`

Flattened feature map produced by the convolutional encoder.

 None

pooler_output

`torch.FloatTensor` of shape `(batch_size, num_labels * branched_channels)`

Branch-major sequence-level representation produced by the fully-connected and branched tower. The first branched_channels features belong to the K562 branch, the next to HepG2, and the last to SK-N-SH.

 None

attentions

`tuple(torch.FloatTensor)`, *optional*

Always None; Malinois is a convolutional model without attention.

 None
Source code in multimolecule/models/malinois/modeling_malinois.py
Python
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
@dataclass
class MalinoisModelOutput(ModelOutput):
    """
    Base class for outputs of the Malinois model.

    Args:
        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, flattened_conv_features)`):
            Flattened feature map produced by the convolutional encoder.
        pooler_output (`torch.FloatTensor` of shape `(batch_size, num_labels * branched_channels)`):
            Branch-major sequence-level representation produced by the fully-connected and branched tower. The first
            `branched_channels` features belong to the K562 branch, the next to HepG2, and the last to SK-N-SH.
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or
            when `config.output_hidden_states=True`):
            Hidden-states of the model at the output of each layer.
        attentions (`tuple(torch.FloatTensor)`, *optional*):
            Always `None`; Malinois is a convolutional model without attention.
    """

    last_hidden_state: torch.FloatTensor | None = None
    pooler_output: torch.FloatTensor | None = None
    hidden_states: tuple[torch.FloatTensor, ...] | None = None
    attentions: tuple[torch.FloatTensor, ...] | None = None

MalinoisPreTrainedModel

Bases: PreTrainedModel

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

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

    config_class = MalinoisConfig
    base_model_prefix = "model"
    _can_record_outputs: dict[str, Any] | None = None
    _no_split_modules = ["MalinoisConvBlock", "MalinoisBranchedLayer"]

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        # Use transformers.initialization wrappers (imported as `init`); they check the
        # `_is_hf_initialized` flag so they don't clobber tensors loaded from a checkpoint.
        if isinstance(module, nn.Conv1d):
            init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
            if module.bias is not None:
                init.zeros_(module.bias)
        # copied from the `reset_parameters` method of `class Linear(Module)` in `torch`.
        elif isinstance(module, nn.Linear):
            init.kaiming_uniform_(module.weight, a=math.sqrt(5))
            if module.bias is not None:
                fan_in, _ = nn.init._calculate_fan_in_and_fan_out(module.weight)
                bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
                init.uniform_(module.bias, -bound, bound)
        elif isinstance(module, MalinoisGroupedLinear):
            init.kaiming_uniform_(module.weight, a=math.sqrt(3))
            fan_in, _ = nn.init._calculate_fan_in_and_fan_out(module.weight)
            bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
            init.uniform_(module.bias, -bound, bound)
        elif isinstance(module, (nn.BatchNorm1d, nn.LayerNorm, nn.GroupNorm)):
            init.ones_(module.weight)
            init.zeros_(module.bias)