TransGenic is a transformer for DNA-to-annotation machine translation. Gene annotations specify the structure of a gene within a DNA sequence by providing the composition of each mRNA transcript based on the coordinate locations of sub-genic features, including coding sequences (CDS), introns, and unstranslated regions (UTR). TransGenic uses a HyenaDNA encoder with the Longformer decoder to predict a text-based annotation format from raw DNA sequence. This model is a 400M parameter checkpoint with a hidden size of 768, 12 layers, and 6 attention heads.
This checkpoint was trained on 9 phylogenetically diverse plant species and acheived a maximum base-level F1 score of 92% in test set sequences from Arabidopsis thaliana.
Uses
This model is intended to be used to generate de novo annotations for plant DNA sequences which contain a gene. The model can also be used to add alternatively spliced isoforms to known primary mRNA transcripts via prompt completion.
How to Get Started with the Model
Example annotation of the AT1G58150.TAIR10 gene from Arabidopsis.
import torch
from transformers import AutoModel, AutoTokenizer
# Load the model and output tokenizer
model_name = "jlomas/HyenaTransgenic-768L12A6-400M"
model = AutoModel.from_pretrained(model_name, trust_remote_code=True)
gffTokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# Load the HyenaDNA tokenizer
dnaTokenizer = AutoTokenizer.from_pretrained("LongSafari/hyenadna-large-1m-seqlen-hf", trust_remote_code=True)
# AT1G58150.TAIR10 gene sequence
seq = '''GCTTATGTTTATCTTTTGATCTGATCTATAAATATATATACAGGTTATCAAAAGGCCTCCACCAAAACCAACTCAACATCTCCGCCTCCATCTCCGCCTCCATCTCCGCCGCGAGTTCCAGACGCTCAAGAATTGGAGTACCTTAAATCCGACTCTTTTCCCGAACACGATGCGTAGAGTTGTCATTCGGACGGAGGTGTGCGTTCCGATAAAATTAGGCTACCGCCGCGGCTTTCAGACCTTCTAGAATTGGAGAAATTGTTTCCCGAACGCGAGGCGCTGAGTTGTCCTTTGGACGGAGATGAGGATTCCAATGAACTTAGGCTACGGCCGCTGGTTCCAGACGCTCAAGAATGGAAGTACCCTAAATCCAAGTTATTTCCCAGACACGCGGCGTGGAGTTGTCATTCGGGCGGAGGTGGAGGTGGAGGCGGTGGCCGTGTATTTACAAATAAAGTAAATGCGGTAGAAGAATTCAACTTAGGAGGACTGAAGGACAGCGAATCCGATTCCGATTCCGAGTAGGGAACTTTTAAAACAACTTTGATTATGGATTTCGATATCCAGAATAATTTTAATTCACTGCTGTTGGACTTGATTAATTTCCTATCACATAACGTTTTGGTTTAACTTTGTACGACCACCA'''
# Tokenize the input sequences and remove the [SEP] token
seqs = dnaTokenizer.batch_encode_plus(
[seq],
return_tensors="pt")["input_ids"][:, :-1]
model.eval()
if torch.cuda.is_available():
seqs = seqs.to("cuda")
model.to("cuda")
# Prediction
outputs = model.generate(
inputs=seqs,
num_return_sequences=1,
max_length=2048,
num_beams=2,
do_sample=True,
decoder_input_ids = None
)
# Convert to GSF
prediction = gffTokenizer.batch_decode(
outputs.detach().cpu().numpy(),
skip_special_tokens=True
)
print(prediction)
# Output: ['<s>301|CDS1|523|+|A>CDS1']
Training Details
Training Data
Training was performed with nine plant genomes, including:
A. thaliana
G. max (Soybean)
O. sativa (Rice)
S. bicolor (Sorghum)
P. trichocarpa (Poplar)
B. distachyon (Grass)
V. vinifera (Grape)
S. italica (Millet)
P. patens (Moss)
Gene-containing genomic sequences were randomly padded with neighboring sequence to produce input sequence lengths that were multiples of 6,144nt less than 49,152nt. Training was performed with the following parameters:
