ChatNT

ChatNT is a multimodal conversational agent that lets researchers interrogate DNA, RNA, and protein sequences using plain English. Genomics foundation models are typically specialized: each downstream task requires a separately fine-tuned head and a user comfortable with machine-learning tooling. ChatNT collapses this fragmentation into a single generalist system that reframes genomics prediction as a text-to-text problem, so a biologist can paste a sequence, ask a question, and receive a free-form answer without writing code.

The design borrows directly from vision-language models such as GPT-4o, which fuse an image encoder with a language model. ChatNT swaps the image encoder for a Nucleotide Transformer sequence encoder and connects it to a large language model through a learned projection, letting a single model interpret biological sequences and natural language jointly. It handles 27 distinct tasks across DNA, RNA, and proteins in one conversational interface, generalizes to questions phrased in ways it has not seen during training, and reports calibrated confidence for its answers.

ChatNT was developed by InstaDeep together with BioNTech as part of the Nucleotide Transformer family. It first appeared as a bioRxiv preprint in April 2024 (de Almeida, Richard, Dalla-Torre, Pierrot, and colleagues) and was published in Nature Machine Intelligence in June 2025.

#Key Features

• Sequence-to-language architecture: A Nucleotide Transformer v2 encoder, an English-aware projection layer, and a frozen Vicuna-7B decoder combine into a roughly 8-billion-parameter model that reads sequences and writes text.
• Generalist across modalities: A single set of weights solves 27 tasks spanning DNA regulatory elements, RNA processing, and protein properties, rather than one fine-tuned model per task.
• Natural-language interface: Users pose questions and receive answers in English, making genomic prediction accessible to researchers without machine-learning or coding backgrounds.
• Generalization to unseen questions: Because tasks are cast as English instructions, ChatNT answers rephrased and previously unseen questions rather than only the exact templates it was trained on.
• Calibrated confidence: The model derives a confidence score from the perplexity of its candidate answers and converts it to a probability via Platt scaling, giving users a signal to flag potential hallucinations.

#Technical Details

ChatNT couples three components. The encoder is a 500M-parameter Nucleotide Transformer v2, pre-trained on genomes from 850 species, which tokenizes DNA as overlapping 6-mers into up to 2,048 tokens (about 12 kb). An English-aware Perceiver Resampler compresses those 2,048 embeddings into 64 task-conditioned vectors that are inserted into the token stream of a frozen 7B-parameter Vicuna-7B decoder, which generates the response. Instruction tuning used a curated corpus of roughly 5.6 million examples built by converting established benchmarks — including the Nucleotide Transformer suite, BEND, AgroNT, DeepSTARR, and APARENT2 — into question-answer form, with dozens of question templates per task for linguistic diversity, totaling about 605 million DNA tokens and 273 million English tokens. On the Nucleotide Transformer benchmark ChatNT reaches a mean Matthews correlation coefficient of 0.77 while solving all tasks with one model, including 0.95 on human promoter prediction and 0.98 on splice-site detection, and it attains a Pearson correlation of 0.89 on protein melting-point prediction, exceeding an ESM2 baseline at 0.85.

#Applications

ChatNT is aimed at wet-lab biologists and genomics researchers who need predictions across many sequence-analysis tasks but lack the infrastructure to fine-tune and serve a separate model for each. Through a conversational prompt it can classify regulatory elements, score splice sites and methylation, predict RNA polyadenylation and degradation, and estimate protein stability and fluorescence, all in one interface. Its calibrated confidence scores help users judge when to trust an answer and when to seek independent validation.

#Impact

ChatNT demonstrates that the vision-language recipe transfers to genomics, unifying a family of specialized predictors into a single instruction-following agent and lowering the technical barrier to using foundation models on biological sequences. Its code, weights, and the instruction dataset are released on GitHub and HuggingFace, though under a non-commercial license that restricts commercial use and derivative commercial models. The model is bounded by its 27 training tasks, can hallucinate on out-of-distribution inputs, and is not a clinical diagnostic tool, so its predictions require domain judgment and experimental confirmation.