Isoformer is a model able to accurately predict differential transcript expression, outperforming existing methods and leveraging the use of multiple modalities. Our framework achieves efficient transfer knowledge from three pre-trained encoders: Enformer for the DNA modality, Nucleotide Transformer v2 for the RNA modality and ESM2 for the protein modality.

Developed by: InstaDeep

How to use

A small snippet of code is given here in order to retrieve embeddings and gene expression predictions given a DNA, RNA and protein sequence.

python

from transformers import AutoTokenizer, AutoModelForMaskedLM
import numpy as np
import torch

# Import the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained("InstaDeepAI/isoformer", trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained("InstaDeepAI/isoformer",trust_remote_code=True)

protein_sequences = ["RSRSRSRSRSRSRSRSRSRSRL" * 9]
rna_sequences = ["ATTCCGGTTTTCA" * 9]
sequence_length = 196_608
rng = np.random.default_rng(seed=0)
dna_sequences = ["".join(rng.choice(list("ATCGN"), size=(sequence_length,)))]

torch_tokens = tokenizer(
    dna_input=dna_sequences, rna_input=rna_sequences, protein_input=protein_sequences
)
dna_torch_tokens = torch.tensor(torch_tokens[0]["input_ids"])
rna_torch_tokens = torch.tensor(torch_tokens[1]["input_ids"])
protein_torch_tokens = torch.tensor(torch_tokens[2]["input_ids"])

torch_output = model.forward(
    tensor_dna=dna_torch_tokens,
    tensor_rna=rna_torch_tokens,
    tensor_protein=protein_torch_tokens,
    attention_mask_rna=rna_torch_tokens != 1,
    attention_mask_protein=protein_torch_tokens != 1,
)

print(f"Gene expression predictions: {torch_output['gene_expression_predictions']}")
print(f"Final DNA embedding: {torch_output['final_dna_embeddings']}")

Training data

Isoformer is trained on RNA transcript expression data obtained from the GTex portal, namely Transcript TPMs measurements across 30 tissues which come from more than 5000 individuals. In total, the dataset is made of ∼170k unique transcripts, of which 90k are protein-coding and correspond to ∼20k unique genes.