LucaOne

Biological sequence models have overwhelmingly specialized: protein language models read amino acids, genomic models read nucleotides, and the two rarely share weights. That split is convenient for engineering and awkward for biology, because DNA, RNA, and protein are three encodings of the same information flow. LucaOne was built to test whether a single network trained on all three at once learns something the specialists cannot — and reports that it does, acquiring an implicit grasp of the central dogma without ever being shown a paired DNA-to-protein example during pretraining.

LucaOne is a 1.8-billion-parameter transformer encoder developed by the Apsara Lab at Alibaba Cloud Intelligence with collaborators at Sun Yat-sen University, Fudan University, Zhejiang University, and the University of Sydney, where virologist Edward C. Holmes contributed. Released as a preprint in May 2024 and published in Nature Machine Intelligence in June 2025, it uses one 39-token vocabulary covering both nucleotides and amino acids, distinguishing the two with a token-type embedding rather than separate models.

LucaOne is the foundational model of the LucaGroup family and the ancestor of several downstream systems in this catalog, including LucaVirus, a viral-specific nucleotide-and-protein model, and LucaPhylo, a hyperbolic embedding model for alignment-free phylogenetics. Where those are specialized descendants, LucaOne is the general-purpose backbone: a single set of embeddings intended to serve as input to bioinformatics tasks that would otherwise each require a bespoke model.

#Key Features

• One model for three molecule types: A shared vocabulary and token-type encoding let a single encoder process DNA, RNA, and protein sequences, so tasks with heterogeneous inputs no longer need two separate backbones stitched together.
• Emergent central dogma understanding: Nucleic acid and protein sequences from the same gene converge in embedding space despite no paired data or explicit correspondence being supplied during pretraining.
• Semi-supervised pretraining: Ten pretraining objectives span masked-token prediction, span-level region and domain recognition, sequence-level taxonomy and keyword prediction, and Cα-coordinate structure learning.
• Broad taxonomic coverage: Training sequences span 169,861 species across viruses, archaea, eukaryotes, and bacteria, drawn from RefSeq, UniProt, UniRef50, ColabFoldDB, InterPro, the PDB, and AlphaFold DB.
• Open release with multiple checkpoints: Code is Apache-2.0 and the pretrained weights are published at several training steps on HuggingFace, Zenodo, and the project FTP server. The pretraining corpus itself is too large for the Zenodo deposit and is distributed through the FTP server and CNGB.

#Technical Details

LucaOne consists of 20 transformer-encoder blocks with 40 attention heads, an embedding dimension of 2,560, and a maximum sequence length of 1,280, totaling 1.8 billion parameters. It replaces post-layer normalization with pre-layer normalization and absolute positional encoding with rotary position embeddings. Pretraining ran for 120 days on 8 NVIDIA A100 GPUs; the checkpoint used for the paper's downstream evaluations was taken at 5.6 million steps, corresponding to 36.95 billion training tokens, chosen to match the data volume seen by ESM2-3B for a fair comparison, with a 17.6-million-step checkpoint at 116.62 billion tokens also released. Across seven downstream tasks, LucaOne outperformed competing models on genus taxonomy, protein thermostability, ncRNA family classification, influenza A antigenic relationship prediction, and protein-protein interaction, and was competitive on protein subcellular localization and ncRNA-protein interaction. Gains were largest on nucleic acid-centric tasks, with accuracy improvements of 0.05 and 0.026 over DNABERT-2 on genus taxonomy and ncRNA family respectively, and a 0.015 gain in Spearman correlation over ESM2-3B on thermostability.

#Applications

LucaOne is intended as a drop-in embedding provider for bioinformatics pipelines that have limited labeled data. Because a single model handles nucleotides and amino acids, it is particularly useful for tasks whose inputs are heterogeneous — predicting interactions between an ncRNA and a protein, or relating a coding sequence to its product — where the conventional approach requires running ESM-2 and a separate DNA model and concatenating the outputs. Companion repositories provide embedding inference and downstream task fine-tuning, and the pretraining corpus is deposited in the CNGB Sequence Archive.

#Impact

LucaOne is one of the clearest demonstrations that unified nucleic acid and protein pretraining yields transferable structure rather than just parameter sharing, and its central dogma result has been widely cited as evidence for that claim. Its practical influence is visible in the model family it seeded, with LucaVirus, LucaPhylo, LucaProt, and LucaPCycle all building on the same pretraining philosophy. Limitations are worth noting: the 1,280-token context window is short relative to genomic models designed for long-range regulatory reasoning, and prediction accuracy degrades as sequence length grows. The open Apache-2.0 release of code, multiple checkpoints, and the pretraining dataset makes the model straightforward to evaluate independently.