A transformer zoo for DNA, RNA, and the chatbot curious

What it does The repository hosts a growing family of pre-trained transformers and downstream tools for genomic and transcriptomic analysis. It bundles inference code and model weights for tasks ranging from whole-genome functional prediction and single-nucleotide segmentation to bulk and single-cell RNA-seq embedding. Rather than a single architecture, it functions as a shared catalog for InstaDeep’s bio-AI research, with most models available through Hugging Face.

The interesting bit NTv3 stretches context windows to one million base pairs at single-nucleotide resolution using a U-Net-like backbone, while ChatNT lets you interrogate DNA through natural language—an odd couple of brute-force sequence modeling and conversational UI sharing the same repo. The breadth is the point: plant genomics (AgroNT), multi-omics fusion (MOJO), and codon-level tokenization experiments all live here.

Key highlights

• NTv3 handles contexts up to 1 Mb and supports controllable sequence generation for enhancer design, validated by STARR-seq.
• SegmentNT (and variants using Enformer or Borzoi backbones) predicts 14 classes of human genomic elements at single-nucleotide resolution.
• The collection spans specialized domains: AgroNT for crops, Isoformer and sCT for transcriptomics, and BulkRNABert for bulk RNA-seq.
• ChatNT offers a multimodal conversational interface for interactive DNA analysis.
• All models carry a CC BY-NC-SA 4.0 license and are published in venues including Nature Methods and Nature Machine Intelligence.

Caveats

• The non-commercial CC BY-NC-SA 4.0 license means these foundation models are not freely usable for commercial applications.
• Because the repo aggregates disparate projects—each with its own docs, dependencies, and backbones—expect a catalog, not a unified toolkit.

Verdict Computational biologists and genomics ML researchers looking for off-the-shelf, research-grade transformers should bookmark this. If you need permissively licensed weights for a commercial product, look elsewhere.