Isoformer

Isoformer is a multimodal deep learning model that bridges DNA, RNA, and protein sequences to predict how much each RNA transcript isoform is expressed across human tissues. Most biological foundation models operate on a single sequence modality — DNA, RNA, or protein — yet many genomics problems are intrinsically cross-modal: a stretch of DNA is transcribed into multiple RNA isoforms, each of which may encode a distinct protein. Isoformer is the first model to integrate all three modalities, applying that integration to transcript isoform expression prediction, a task where single-modality DNA models struggle because different isoforms of the same gene share regulatory DNA yet are expressed at very different levels.

The model's central idea is transfer learning between existing foundation models rather than training a single large network from scratch. Isoformer reuses three pre-trained, modality-specific encoders and connects them through a cross-attention aggregation module, so knowledge captured during each encoder's separate pre-training is combined and transferred toward the shared expression task. This lets the model exploit both intra-modality pre-training and inter-modality transfer at once.

Isoformer was introduced in June 2024 by Juan Jose Garau-Luis, Patrick Bordes, Guillaume Richard, Thomas Pierrot, and colleagues at InstaDeep, and was presented at NeurIPS 2024 in the paper "Multi-modal Transfer Learning between Biological Foundation Models."

#Key Features

• Tri-modal integration: Isoformer is the first model to jointly reason over DNA, RNA, and protein sequences for a single genomics task, rather than operating on one modality in isolation.
• Cross-attention aggregation: The three encoder outputs are combined through a cross-attention module with residual connections, where each modality attends successively to the others to form a unified multimodal representation.
• Transfer between foundation models: The design reuses Enformer, Nucleotide Transformer v2, and ESM2, demonstrating that pre-trained encoders can be stitched together and fine-tuned jointly for a downstream task.
• Isoform-resolution expression: Predictions are made per transcript isoform across 30 human tissues, capturing isoform-specific differences that gene-level DNA models cannot resolve.
• Open artifacts: Model weights, an inference notebook, and the curated training dataset are publicly released on Hugging Face and GitHub.

#Technical Details

Isoformer combines Enformer for the DNA modality (accepting up to 196,608 nucleotides of genomic context), Nucleotide Transformer v2 for the RNA modality (up to 12,288 nucleotides), and ESM2-150M for the protein modality (up to 1,200 amino acids). Each encoder is initialized with its pre-trained weights and fine-tuned jointly during training, and their representations are merged by the cross-attention aggregation module before a head predicts expression. Training used RNA transcript TPM measurements from the GTEx portal spanning 30 tissues and more than 5,000 individuals, covering roughly 170,000 unique transcripts (about 90,000 protein-coding) across some 20,000 genes; the released multi_omics_transcript_expression dataset pairs each transcript with its DNA, RNA, and protein sequences. The multimodal model reaches an R2 of about 0.53 and Spearman correlation near 0.72 on transcript expression, substantially exceeding a DNA-only Enformer baseline (R2 near 0.21). Ablations show that pre-trained initialization is essential — randomly initializing the encoders collapses performance — and that removing any single modality reduces accuracy, evidence that the model benefits from both intra-modality pre-training and inter-modality transfer.

#Applications

Isoformer is aimed at computational biologists studying transcript-level regulation, tissue-specific alternative splicing, and the relationship between sequence and isoform abundance. Because it predicts expression at isoform rather than gene resolution, it is useful for interpreting how regulatory sequence gives rise to distinct transcript products and for prioritizing isoforms relevant to disease. More broadly, its architecture offers a template for combining separately trained genomics, transcriptomics, and proteomics models on tasks that no single modality can address alone.

#Impact

Isoformer establishes multimodal transfer learning as a practical strategy in genomics, showing that independently pre-trained DNA, RNA, and protein foundation models can be fused to outperform strong single-modality baselines on isoform expression prediction. As the first tri-modal DNA-RNA-protein model, it points toward a modular approach in which specialized encoders are reused rather than retrained. Its practical scope is bounded: predictions cover protein-coding genes, the released weights and dataset carry a non-commercial CC BY-NC-SA 4.0 license, and results are computational, but the open release lowers the barrier for others to build on multimodal biological modeling.