Skip to content

models

models provide a collection of pre-trained models.

Model Class

In the 🤗 transformers library, the names of model classes can sometimes be misleading. While these classes support both regression and classification tasks, their names often include xxxForSequenceClassification, which may imply they are only for classification.

To avoid this ambiguity, MultiMolecule provides a set of model classes with clear, intuitive names that reflect their intended use:

  • multimolecule.AutoModelForSequencePrediction: Sequence Prediction
  • multimolecule.AutoModelForTokenPrediction: Token Prediction
  • multimolecule.AutoModelForContactPrediction: Contact Prediction

Each of these models supports both regression and classification tasks, offering flexibility and precision for a wide range of applications.

Contact Prediction

Contact prediction assign a label to each pair of token in a sentence. One of the most common contact prediction tasks is protein distance map prediction. Protein distance map prediction attempts to find the distance between all possible amino acid residue pairs of a three-dimensional protein structure

Nucleotide Prediction

Similar to Token Classification, but removes the <bos> token and the <eos> token if they are defined in the model config.

<bos> and <eos> tokens

In tokenizers provided by MultiMolecule, <bos> token is pointed to <cls> token, and <sep> token is pointed to <eos> token.

Usage

Build with multimolecule.AutoModels

Python
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

# For additional terms and clarifications, please refer to our License FAQ at:
# <https://multimolecule.danling.org/about/license-faq>.


from transformers import AutoTokenizer

from multimolecule import AutoModelForSequencePrediction

model = AutoModelForSequencePrediction.from_pretrained("multimolecule/rnafm")
tokenizer = AutoTokenizer.from_pretrained("multimolecule/rnafm")

sequence = "UAGCGUAUCAGACUGAUGUUG"
output = model(**tokenizer(sequence, return_tensors="pt"))

Direct Access

All models can be directly loaded with the from_pretrained method.

Python
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

# For additional terms and clarifications, please refer to our License FAQ at:
# <https://multimolecule.danling.org/about/license-faq>.


from multimolecule.models import RnaFmForTokenPrediction, RnaTokenizer

model = RnaFmForTokenPrediction.from_pretrained("multimolecule/rnafm")
tokenizer = RnaTokenizer.from_pretrained("multimolecule/rnafm")

sequence = "UAGCGUAUCAGACUGAUGUUG"
output = model(**tokenizer(sequence, return_tensors="pt"))

Build with transformers.AutoModels

While we use a different naming convention for model classes, the models are still registered to corresponding transformers.AutoModels.

Python
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

# For additional terms and clarifications, please refer to our License FAQ at:
# <https://multimolecule.danling.org/about/license-faq>.


from transformers import AutoModelForSequenceClassification, AutoTokenizer

import multimolecule  # noqa: F401

model = AutoModelForSequenceClassification.from_pretrained("multimolecule/mrnafm")
tokenizer = AutoTokenizer.from_pretrained("multimolecule/mrnafm")

sequence = "UAGCGUAUCAGACUGAUGUUG"
output = model(**tokenizer(sequence, return_tensors="pt"))

import multimolecule before use

Note that you must import multimolecule before building the model using transformers.AutoModel. The registration of models is done in the multimolecule package, and the models are not available in the transformers package.

The following error will be raised if you do not import multimolecule before using transformers.AutoModel:

Python
ValueError: The checkpoint you are trying to load has model type `rnafm` but Transformers does not recognize this architecture. This could be because of an issue with the checkpoint, or because your version of Transformers is out of date.

Initialize a vanilla model

You can also initialize a vanilla model using the model class.

Python
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

# For additional terms and clarifications, please refer to our License FAQ at:
# <https://multimolecule.danling.org/about/license-faq>.


from multimolecule.models import RnaFmConfig, RnaFmForTokenPrediction, RnaTokenizer

config = RnaFmConfig()
model = RnaFmForTokenPrediction(config)
tokenizer = RnaTokenizer()

sequence = "UAGCGUAUCAGACUGAUGUUG"
output = model(**tokenizer(sequence, return_tensors="pt"))

Available Models

DeoxyriboNucleic Acid (DNA)

RiboNucleic Acid (RNA)