GeneChat

Computational gene function prediction is almost always posed as classification over a fixed label set — Gene Ontology terms, pathway memberships, enzyme classes. That framing is convenient for benchmarking but narrow in practice. It cannot express the mechanism, tissue context, or qualifications that a human curator writes into a gene summary, and it has nothing to say about a function that falls outside the predefined vocabulary. Since most genes in sequenced genomes carry little or no experimental annotation, the gap between what a classifier can emit and what a biologist actually wants to read is wide.

GeneChat, from Pengtao Xie's group in the Department of Electrical and Computer Engineering at UC San Diego, reframes the task as text generation. Given a nucleotide sequence and a textual prompt, the model writes a free-form natural-language description of what the gene does. It follows the adaptor pattern that has become the standard recipe for multimodal language models: a pretrained domain encoder embeds the non-text input, a learned adaptor projects those embeddings into the token-embedding space of a pretrained large language model, and the language model generates conditioned on both the projected representation and the user's prompt.

The architecture places GeneChat next to ChatNT, which couples a genomics encoder to an English decoder to answer questions about DNA, RNA, and protein sequences. Where ChatNT spans many genomics tasks recast as short-answer questions, GeneChat targets a single task — descriptive functional annotation — and optimizes for long-form narrative output.

#Key Features

• Free-form functional descriptions: Output is generated prose rather than a label drawn from a closed ontology, so the model can describe mechanism and context, including functions with no matching category in a fixed vocabulary.
• Genomic encoder built for long-range context: The gene encoder is based on DNABERT-2, whose byte-pair-encoded tokenization compresses genomic input and lets a 117M-parameter model cover long stretches of sequence efficiently.
• Cross-modal adaptor: A learned projection aligns gene representations with the language model's input embedding space, so the text backbone stays a general-purpose LLM rather than being retrained on nucleotides.
• Prompt conditioning: Descriptions are produced in response to a text prompt supplied alongside the sequence, which lets a user steer what aspect of function the output emphasizes.
• Curated training corpus: Training pairs more than 50,000 genes drawn from NCBI with their functional descriptions, giving the model supervision in the register that curators actually write in.

#Technical Details

GeneChat has three components: a DNABERT-2-based gene encoder chosen for its handling of long-range genomic context, an adaptor module that maps gene representations into the LLM input space, and Vicuna-13B — a 13-billion-parameter instruction-tuned derivative of LLaMA-2 — as the text decoder. Training uses over 50,000 genes from the NCBI database paired with their function descriptions. Evaluation is against GPT-4o on BLEU and METEOR, the standard n-gram-overlap metrics for generated text; GeneChat scores higher on both, indicating that a sequence-grounded model produces descriptions closer to curated reference text than a general-purpose LLM prompted with the same task.

#Applications

The intended use is descriptive annotation of genes that lack experimental characterization — newly assembled genomes, understudied organisms, and the long tail of predicted open reading frames that ontology-based classifiers cover poorly. A generated summary is a starting hypothesis for a curator or bench biologist rather than a verified assignment, but it can prioritize which uncharacterized genes are worth following up and it can supply context that a GO term alone does not carry.

#Impact

GeneChat is a preprint and has not been peer reviewed. Training, fine-tuning, evaluation and inference code are on GitHub, with a stage-1 checkpoint and the 51,411-gene dataset on Google Drive, but the repository carries no license and the hosted artifacts state no terms, so reuse rests on unstated permission. Its evaluation is also narrow: BLEU and METEOR measure surface overlap with reference text and can reward fluent paraphrase over factual accuracy, and the reported comparison is against a single general-purpose baseline rather than against specialized function-prediction methods. What the work does establish is that the encoder-adaptor-LLM recipe transfers cleanly to gene-level annotation, and that grounding a chat model in a genomic encoder measurably improves the descriptions it writes.