AI Foundation Models for Genomics

A hub for InstaDeep's cutting-edge deep learning models and research for genomics, originating from the Nucleotide Transformer and its evolutions.

---

🎯 Our Focus: Advancing Genomics with AI

Welcome to the InstaDeep AI for Genomics repository! This is where we feature our collection of transformer-based genomic language models and innovative downstream applications. Our work in the genomics space began with The Nucleotide Transformer, developed in collaboration with Nvidia and TUM and trained on Cambridge-1, and has expanded to include projects like the Agro Nucleotide Transformer (in collaboration with Google, trained on TPU-v4 accelerators), SegmentNT, and ChatNT.

Our mission is to provide the scientific community with powerful, reproducible, and accessible tools to unlock new insights from biological sequences. This repository serves as the central place for sharing our models, inference code, pre-trained weights, and research contributions in the genomics domain, with explorations into future areas like single-cell transcriptomics.

We are thrilled to open-source these works and provide the community with access to the code and pre-trained weights for our diverse set of genomics language models and segmentation models.

✨ Featured Models & Research Evolutions

This section highlights the key models and research directions from our team. Each entry provides a brief overview and links to detailed documentation, publications, and resources. (Detailed code examples, setup for specific models, and in-depth figures are now located in their respective documentation pages within the ./docs folder.)

---

🧬 The Nucleotide Transformer v3 (NTv3)

A unified multi-species foundation model that brings representation learning, functional-track prediction, genome annotation and controllable sequence generation into one framework. NTv3 uses a U-Net-like architecture with single-base tokenization to model contexts up to 1 Mb at nucleotide resolution. It is pretrained on 9 trillion base pairs from OpenGenome2, then post-trained on 16,262 functional tracks from 15 animal and plant species, and reaches state-of-the-art accuracy for functional-track prediction and genome annotation across species. It can be fine-tuned into a controllable generative model for designing enhancer sequences at a requested activity level, validated experimentally by STARR-seq.

One 700M-parameter backbone is trained once and then specialised per capability, so each checkpoint inherits what the others learned about sequence rather than starting over.

Checkpoint	Use it for
NTv3_pretrained	Embeddings, masked-token prediction, and your own fine-tuning
NTv3_tracks	Functional signal: accessibility, histone marks, TF binding, expression, from 16 kb to 1 Mb inputs
NTv3_tracks_1mb	The same tracks from 1 Mb inputs only: the specialist behind the NTv3 paper's main results
NTv3_annot_mammals / _plants / _insects	Calling genes from sequence alone, one model per clade

Annotate a genome end to end with NTv3, FASTA --> GFF3:

pip install "nucleotide-transformer[gpu]"
ntv3-annotate --genome /data/cow/genome.fa --model_cfg mammals --out /data/cow.gff3.gz

• Keywords: Multi-species Foundation Model, Long-context (1 Mb), Single-base Resolution, Functional-track Prediction, Genome Annotation, Controllable Sequence Generation, U-Net Architecture, Masked Diffusion
• ➡️ Model Details & Usage
• 🧬 Genome Annotation, annotate a genome from FASTA to GFF3 with ntv3-annotate
• 🤗 NTv3 collection on Hugging Face
• 📜 Paper

NTv3 locus-analysis skill

The NTv3 skill runs NTv3 inside Claude Code on a genomic locus you name. You ask in one sentence. The skill asks only for what is missing, shows one plan, and runs nothing until you approve it. The result is one self-contained report page, results/<run>/ntv3_report.html, with the predicted tracks in a genome browser.

It answers five asks:

• Predict: the tracks you name, at a gene or an exact region, from NTv3_tracks at any window up to 1,048,576 bp, on a GPU when there is one (on CPU, 16,384 bp or more is slow and comes with a warning). The model predicts the central 37.5% of the window. A gene is centred on its transcription start site. Only the requested model heads are computed.
• Saliency: for each track, the gradient of its mean over the predicted span at each base, centred across A, C, G and T and kept at the reference base.
• Compare: one SNV, the tracks for each allele, and how far each moves (alt − ref).
• Annotate: a FASTA, or the 1 Mb around a Predict run, with a clade (mammals, plants, insects): the model (InstaDeepAI/NTv3_annot_<clade>), decoder (hmm_constrained by default), 1,048,576 bp windows with the central 75% kept, and the released filters. After approval it runs ntv3-annotate's pipeline on a GPU (about 25 GB per window) and writes a GFF3 the browser draws, next to the tracks for a locus.
• Explain: which model, species, window, crop, bins, and model head produced each number.

Browser lanes are only the ones you name: GENCODE (knownGene) and ClinVar (clinvarMain) on human, NCBI RefSeq (ncbiRefSeq) on mouse, or none. Another UCSC lane, such as ENCODE cCREs, DNase clusters, phastCons, or a histone mark, is added only when you name it. Naming a tissue or an assay adds no lane. Integrated Gradients, SmoothGrad, ISM, design, knockout, and attention are not offered.

Start here: ntv3_skills/README.md covers requirements, installation from zero, a first ask, the run outputs, and troubleshooting.

---

🧬 The Nucleotide Transformer (NT)

Our foundational language models leverage DNA sequences from over 3,200 diverse human genomes and 850 genomes from a wide range of species. These models provide extremely accurate molecular phenotype prediction compared to existing methods. This family includes multiple variants (e.g., 500M_human_ref, 2B5_1000G, NT-v2 series) which are detailed further in the specific documentation.

• Keywords: Foundational Model, Genomics, DNA/RNA, Pre-trained, Sequence Embeddings, Phenotype Prediction
• ➡️ Model Details, Variants & Usage
• 📜 Read the Paper (Nature Methods 2025)
• 🤗 Hugging Face Collection
• 🚀 Fine-tuning Notebooks (HF): (LoRA and regular)

---

🌾 Agro Nucleotide Transformer (AgroNT)

A novel foundational large language model trained on reference genomes from 48 plant species, with a predominant focus on crop species. AgroNT demonstrates state-of-the-art performance across several prediction tasks ranging from regulatory features, RNA processing, and gene expression in plants.

• Keywords: Plant Genomics, Foundational Model, Crop Science, Gene Expression, Agriculture AI
• ➡️ Model Details & Usage
• 📜 Read the Paper (Communications Biology 2024)
• 🤗 Hugging Face Collection

---

🧩 SegmentNT (& family: SegmentEnformer, SegmentBorzoi)

Segmentation models using transformer backbones (Nucleotide Transformers, Enformer, Borzoi) for predicting genomic elements at single-nucleotide resolution. SegmentNT, for instance, predicts 14 different classes of human genomic elements in sequences up to 30kb (generalizing to 50kbp) and demonstrates superior performance.

• Keywords: Genome Segmentation, Single-Nucleotide Resolution, Genomic Elements, U-Net, Enformer, Borzoi
• ➡️ Model Details & Usage (Covers SegmentNT, SegmentEnformer, SegmentBorzoi)
• 📜 Read the Paper (Nature Methods 2025)
• 🤗 Hugging Face Collection
• 🚀 SegmentNT Inference Notebook (HF)

---

💬 ChatNT

A multimodal conversational agent designed with a deep understanding of DNA biological sequences, enabling interactive exploration and analysis of genomic data through natural language.

• Keywords: Conversational AI, Multimodal, DNA Analysis, Genomics Chatbot, Interactive Biology
• ➡️ Model Details & Usage
• 📜 Read the Paper (Nature Machine Intelligence 2025)
• 🤗 ChatNT on Hugging Face
• 🚀 ChatNT Inference Notebook (Jax)

---

3️⃣ Codon-NT (Exploring 3-mer Tokenization)

A Nucleotide Transformer model variant trained on 3-mers (codons). This work investigates alternative tokenization strategies for genomic language models and their impact on downstream performance and interpretability.

• Keywords: Genomics, Language Model, Codon, Tokenization, 3-mers, Nucleotide Transformer Variant
• ➡️ Model Details & Usage
• 📜 Read the Paper (Bioinformatics 2024)
• 🤗 Hugging Face Link

---

🧬 Isoformer

A model designed for learning isoform-aware embeddings directly from RNA-seq data, enabling a deeper understanding of transcript-specific expression and regulation.

• Keywords: RNA-seq, Transcriptomics, Isoforms, Gene Expression, Embeddings
• ➡️ Model Details & Usage
• 📜 Read the Paper (NeurIPS 2024)
• 🤗 Hugging Face Link

---

🔬 sCT (single-Cell Transformer)

Our foundational transformer model for single-cell and spatial transcriptomics data. sCT aims to learn rich representations from complex, high-dimensional single-cell datasets to improve various downstream analytical tasks.

• Keywords: Single-cell RNA-seq, Spatial Transcriptomics, Foundational Model, Transformer, Gene Expression
• ➡️ Model Details & Usage
• 📜 Read the Paper (OpenReview preprint)
• 🤗 Hugging Face Link
• 🚀 sCT Inference Notebook (HF)

---

🧪 BulkRNABert

BulkRNABert is a transformer-based, encoder-only foundation model designed for bulk RNA-seq data. It learns biologically meaningful representations from large-scale transcriptomic profiles.

• Keywords: Bulk RNA-seq, Foundational Model, Transformer, Cancer prognosis
• ➡️ Model Details & Usage
• 📜 Read the Paper (Machine Learning for Health 2024)
• 🤗 Hugging Face Link
• 🚀 BulkRNABert Inference Notebook (HF)

---

🔗 MOJO (Multi-Omics JOint representation)

MOJO is a multimodal model designed learn embeddings of multi-omics data. It integrates bulk RNA-seq and DNA methylation data to generate powerful joint representations tailored for cancer-type classification and survival analysis.

• Keywords: Bulk RNA-seq, DNA Methylation, Foundational Model, Transformer, Multimodal, Cancer prognosis
• ➡️ Model Details & Usage
• 📜 Read the Paper (ICML Workshop on Generative AI and Biology 2025)
• 🤗 Hugging Face Link
• 🚀 MOJO Inference Notebook (HF)

💡 Why Choose InstaDeep's Genomic Models?

• Built on Strong Foundations: Leveraging large-scale pre-training and diverse genomic datasets.
• Cutting-Edge Research: Incorporating the latest advancements in deep learning for biological sequence analysis.
• High Performance: Designed and validated to achieve state-of-the-art results on challenging genomic tasks.
• Open and Accessible: We provide pre-trained weights, usage examples, and aim for easy integration into research workflows.
• Collaborative Spirit: Developed with leading academic and industry partners.
• Focused Expertise: Created by a dedicated team specializing in AI for genomics at InstaDeep.

🚀 Getting Started

To begin using models from this repository:

1. Clone the repository:

   git clone https://github.com/instadeepai/nucleotide-transformer.git
   cd nucleotide-transformer
   

2. Set up your environment (virtual environment recommended):

   python -m venv .venv
   source .venv/bin/activate # On Windows use `source .venv\Scripts\activate`
   

3. Install the package and dependencies:

   pip install . # Installs the local package
   # Or, for a general requirements file if you have one:
   # pip install -r requirements.txt 
   

For detailed instructions on individual models, including specific dependencies, downloading pre-trained weights, and Python usage examples, please refer to their dedicated documentation pages linked in the "Featured Models & Research Evolutions" section above (e.g., ./docs/nucleotide_transformer.md).

⚖️ Licences

• Code: CC BY-NC-SA 4.0.
• NTv3 model weights: the InstaDeep Open Model Licence, for non-commercial use only. You can read it here before requesting access to the gated models on Hugging Face.

🤝 Community & Support

• Questions & Bug Reports: Please use the GitHub Issues page.
• Discussions: For broader discussions or questions, please use the GitHub Discussions tab (if enabled).
• Stay Updated: Follow InstaDeep's official channels for announcements on new model releases and research updates.