Repository of Codon-based Foundation Models (EnCodon & DeCodon)

This repository contains the code for the EnCodon and DeCodon models, codon-resolution large language models pre-trained on the NCBI Genomes database described in the paper "A Suite of Foundation Models Captures the Contextual Interplay Between Codons".

Get started 🚀

Installation

From source

Currently, this is the only way to install the package but will push a pip installable version soon. To install the package from source, run the following command:

pip install git+https://github.com/goodarzilab/cdsFM.git

Applications

Now that you have cdsFM installed, you can use AutoEnCodon and AutoDeCodon classes which serve as wrappers around the pre-trained models. Here are some examples on how to use them:

Sequence Embedding Extraction with EnCodon

Following is an example of how to use the EnCodon model to extract sequence embeddings:

from cdsFM import AutoEnCodon

# Load your dataframe containing sequences
seqs = ...

# Load a pre-trained EnCodon model
model = AutoEnCodon.from_pretrained("goodarzilab/encodon-620M")

# Extract embeddings
embeddings = model.get_embeddings(seqs, batch_size=32)

Sequence Generation with DeCodon

You can generate organism-specific coding sequences with DeCodon simply by:

from cdsFM import AutoDeCodon

# Load a pre-trained DeCodon model
model = AutoDeCodon.from_pretrained("goodarzilab/DeCodon-200M")

# Generate!
gen_seqs = model.generate(
    taxid=9606, # NCBI Taxonomy ID for Homo sapiens
    num_return_sequences=32, # Number of sequences to return
    max_length=1024, # Maximum length of the generated sequence
    batch_size=8, # Batch size for generation
)

---

Tokenization

EnCodon and DeCodon are pre-trained on coding sequences of length up to 2048 codons (i.e. 6144 nucleotides), including the <CLS> token prepended automatically to the beginning of the sequence and the <SEP> token appended at the end. The tokenizer's vocabulary consists of 64 codons and 5 special tokens namely <CLS>, <SEP>, <PAD>, <MASK> and <UNK>.

---

HuggingFace 🤗

A collection of pre-trained checkpoints of EnCodon & DeCodon models are available on HuggingFace 🤗. Following table contains the list of available models:

Model	name	num. params	description	weights
EnCodon	encodon-80M	80M	Pre-trained checkpoint	🤗
EnCodon	encodon-80M-euk	80M	Eukaryotic-expert	🤗
EnCodon	encodon-620M	620M	Pre-trained checkpoint	🤗
EnCodon	encodon-620M-euk	620M	Eukaryotic-expert	🤗
DeCodon	decodon-200M	200M	Pre-trained checkpoint	🤗
DeCodon	decodon-200M-euk	200M	Eukaryotic-expert	🤗

---

Citation

@article{Naghipourfar2024,
  title = {A Suite of Foundation Models Captures the Contextual Interplay Between Codons},
  url = {http://dx.doi.org/10.1101/2024.10.10.617568},
  DOI = {10.1101/2024.10.10.617568},
  publisher = {Cold Spring Harbor Laboratory},
  author = {Naghipourfar,  Mohsen and Chen,  Siyu and Howard,  Mathew and Macdonald,  Christian and Saberi,  Ali and Hagen,  Timo and Mofrad,  Mohammad and Coyote-Maestas,  Willow and Goodarzi,  Hani},
  year = {2024},
  month = oct 
}