Nucleotide Transformers (NT, SegmentNT, ChatNT, AgroNT)

The Nucleotide Transformers family — comprising Nucleotide Transformer (NT), SegmentNT, ChatNT, and AgroNT — is a suite of transformer-based genomic language models developed by InstaDeep and available on the DeepChain platform. These state-of-the-art foundation models are trained on extensive datasets of DNA sequences across multiple reference genomes, enabling highly accurate predictions of molecular phenotypes and precise segmentation of genomic elements at single-nucleotide resolution. They are designed for researchers in genomics, plant biology, and life sciences who need powerful, flexible tools for understanding and interpreting DNA sequences.

Each model in the family addresses a distinct challenge in genomics — from bridging the gap between genotype and phenotype, to enabling conversational interaction with biological sequence data, to powering crop genomic improvement — making the suite broadly applicable across academic and industrial R&D contexts.

Nucleotide Transformer (NT)

• Pre-trained on DNA sequences from 3,202 diverse human genomes and 850 genomes spanning a wide range of model and non-model organisms.
• Generates transferable, context-specific representations of nucleotide sequences that enable accurate molecular phenotype prediction even in low-data settings.
• Representations alone match or outperform specialised methods on 11 of 18 prediction tasks, and up to 15 tasks after fine-tuning.
• Without explicit supervision, the model learns to focus attention on key genomic elements, including enhancers and other gene expression regulators.
• Demonstrated ability to improve the prioritisation of functional genetic variants using model representations alone.

SegmentNT

• Built on top of the Nucleotide Transformer, SegmentNT is a segmentation model that processes input DNA sequences up to 30 kb in length.
• Predicts 14 different classes of genomic elements at single-nucleotide resolution.
• Leverages pre-trained NT weights, surpassing ablation models including convolutional networks with one-hot encoded sequences and models trained from scratch.
• Supports zero-shot generalisation for sequences of up to 50 kbp across multiple sequence lengths.
• Demonstrates improved performance on splice site detection and strong nucleotide-level precision throughout the genome.
• Evaluates all gene elements simultaneously, enabling prediction of the impact of sequence variants on splice site changes, as well as exon and intron rearrangements in transcript isoforms.
• A model trained on human genomic elements can generalise to elements of different species; a multispecies SegmentNT model achieves stronger generalisation for all genic elements on unseen species.
• Easily extensible to additional genomic elements and species, representing a new paradigm for DNA analysis and interpretation.

ChatNT

• The first multimodal conversational agent with advanced understanding of biological sequences, bridging the gap between biology foundation models and conversational AI.
• Achieves state-of-the-art results on the Nucleotide Transformer benchmark while solving all tasks simultaneously, in English, and generalising to unseen questions.
• Supports tasks spanning DNA, RNA, and proteins across multiple species, tissues, and biological processes, using a curated set of biologically relevant instruction tasks.
• Reaches performance on par with state-of-the-art specialised methods on these tasks.
• Introduces a novel perplexity-based technique to help calibrate model prediction confidence.
• Accessible to users with no coding background, removing barriers to entry for non-computational biologists.
• The genomics instruction-tuning framework is easily extensible to more tasks and biological data modalities, including structure and imaging.

AgroNT

• A foundational large language model trained on genomes from 48 plant species, with a predominant focus on crop species.
• Achieves state-of-the-art predictions for regulatory annotations, promoter and terminator strength, tissue-specific gene expression, and functional variant prioritisation.
• Includes a large-scale in silico saturation mutagenesis analysis on cassava, evaluating the regulatory impact of over 10 million mutations and providing predicted effects as a resource for variant characterisation.
• Introduces the Plants Genomic Benchmark (PGB), a comprehensive benchmark for deep learning-based methods in plant genomic research, compiled from diverse datasets.
• Designed to support crop genomic improvement by enabling accurate predictions from genome-wide molecular phenotype data.

All models in the Nucleotide Transformers family are accessible through the DeepChain platform, which offers minimal setup, customisable workflows, and an AI assistant. The platform is designed to accelerate life sciences R&D across genomics, agriculture, and related fields, making advanced AI capabilities available to both computational and non-computational researchers.