Generalized Biological Foundation Model with Unified Nucleic Acid and Protein Language

Yong He, Pan Fang, Yongtao Shan, Yuanfei Pan, Yanhong Wei, Yichang Chen, Yi‐Hao Chen, Yi Liu, Zhenyu Zeng, Zhan Zhou, Feng Zhu, Edward C. Holmes, Jieping Ye, Jun Li, Yuelong Shu, Mǎng Shī, Zhaorong Li

Abstract

The language of biology, encoded in DNA, RNA, and proteins, forms the foundation of life but remains challenging to decode due to its complexity. Traditional computational methods often struggle to integrate information across these molecules, limiting a comprehensive understanding of biological systems. Advances in Natural Language Processing (NLP) with pre-trained models offer new possibilities for interpreting biological language. Here, we introduce LucaOne, a pre-trained foundation model trained on nucleic acid and protein sequences from 169,861 species. Through large-scale data integration and semi-supervised learning, LucaOne demonstrates an understanding of key biological principles, such as DNA-Protein translation. Using few-shot learning, it effectively comprehends the central dogma of molecular biology and performs competitively on tasks involving DNA, RNA, or protein inputs. Our results highlight the potential of unified foundation models to address complex biological questions, providing an adaptable framework for bioinformatics research and enhancing the interpretation of life’s complexity.

Introduction

In molecular biology, the "Central Dogma" describes the fundamental flow of information: DNA is transcribed into RNA, which is then translated into proteins. Historically, computational biology has treated these biopolymers as distinct languages. Specialized models like ESM2 focus exclusively on protein sequences, while others like DNABert are tailored for genomic data. This fragmentation creates a significant bottleneck; biological function often emerges from the intricate interplay between these molecules.

LucaOne was developed to bridge this gap. As a 1.8-billion parameter foundation model, it treats DNA, RNA, and proteins as a single, unified language. By training on a massive cross-species dataset, LucaOne aims to capture the intrinsic relationships between the genetic code and its physical manifestations, offering a holistic perspective on biological systems.

Methodology: A Unified Biological Vocabulary

LucaOne utilizes a Transformer-Encoder architecture, similar to the BERT models used in natural language processing, but adapted for the specific requirements of biological sequences.

Unified Tokenization and Encoding

To process diverse biopolymers within a single framework, the researchers implemented a unified vocabulary of 39 tokens. This set encompasses the four nucleotides of DNA (A, T, C, G), the five found in RNA (including Uracil), and the 20 standard amino acids. To resolve potential ambiguities—such as the letter "C" representing both Cytosine in DNA and Cysteine in proteins—the model employs a binary Token-Type Encoding. This indicator (0 for nucleic acids, 1 for proteins) allows the model to maintain context-specific meanings for shared characters.

Knowledge-Enhanced Pre-training

Beyond standard self-supervised masked language modeling (predicting hidden tokens), LucaOne incorporates eight additional supervised tasks during pre-training. This semi-supervised approach forces the model to learn functional and structural nuances, including:

  • Taxonomy Prediction: Identifying the species of origin.
  • Gene Region Identification: Distinguishing between introns and exons.
  • Structural Awareness: Predicting the 3D spatial coordinates of protein $C_\alpha$-atoms, teaching the model that sequence dictates three-dimensional shape.

Data and Scale

The model's robust performance is rooted in the scale and diversity of its training data, spanning 169,861 species:

  • Nucleic Acids: 1.18 billion DNA and 136 million RNA sequences from RefSeq.
  • Proteins: Hundreds of millions of sequences from UniProt, UniRef50, and ColabFold.
  • Structure: Experimentally determined structures from RCSB-PDB and predicted structures from AlphaFold2.

Key Insights and Results

Emergent Understanding of the Central Dogma

One of the most significant findings was LucaOne’s ability to perform "emergent translation." Despite never being explicitly taught the genetic code, the model achieved 84.53% accuracy in matching DNA sequences to their corresponding proteins. This suggests that by observing billions of sequences, the model independently discovered the rules of translation that govern life.

Performance Benchmarks

LucaOne demonstrated state-of-the-art (SOTA) or competitive performance across eight distinct downstream tasks:

  • Genus Taxonomy (GenusTax): Outperformed specialized models like DNABert2 by a margin of 0.05 in accuracy.
  • Protein Stability (ProtStab): Surpassed ESM2-3B, showing a 0.015 increase in Spearman’s Rank Correlation Coefficient (SRCC).
  • Viral Representation: In Influenza A (InfA) classification, LucaOne achieved a perfect 1.0 accuracy, demonstrating an exceptional ability to represent viral RNA relationships.
  • Clustering: T-SNE visualizations of the model’s internal embeddings showed that LucaOne organizes species and protein families more cleanly than specialized models, creating a more coherent "map" of the biological universe.

Discussion and Limitations

While LucaOne represents a major advancement toward a universal biological AI, several challenges remain. The model currently operates on sequence data in isolation, meaning it cannot yet account for environmental variables such as temperature or pH, which significantly influence gene expression and protein folding. Furthermore, the model's performance slightly declines when encountering species with non-standard codon usage, highlighting the difficulty of capturing the "long tail" of biological diversity.

Conclusion

LucaOne proves that a unified model is more than the sum of its parts. By integrating DNA, RNA, and proteins into a single representational space, it captures the fundamental logic of the Central Dogma. This research provides a powerful, adaptable framework for bioinformatics, moving the field closer to a comprehensive, multimodal understanding of life’s complexity.

Summary updated

Abstract

Abstract The language of biology, encoded in DNA, RNA, and proteins, forms the foundation of life but remains challenging to decode due to its complexity. Traditional computational methods often struggle to integrate information across these molecules, limiting a comprehensive understanding of biological systems. Advances in Natural Language Processing (NLP) with pre-trained models offer new possibilities for interpreting biological language. Here, we introduce LucaOne, a pre-trained foundation model trained on nucleic acid and protein sequences from 169,861 species. Through large-scale data integration and semisupervised learning, LucaOne demonstrates an understanding of key biological principles, such as DNA-Protein translation. Using few-shot learning, it effectively comprehends the central dogma of molecular biology and performs competitively on tasks involving DNA, RNA, or protein inputs. Our results highlight the potential of unified foundation models to address complex biological questions, providing an adaptable framework for bioinformatics research and enhancing the interpretation of life’s complexity.
BibTeX
@inproceedings{he2025generalized,
  author = {Yong He and Pan Fang and Yongtao Shan and Yuanfei Pan and Yanhong Wei and Yichang Chen and Yi‐Hao Chen and Yi Liu and Zhenyu Zeng and Zhan Zhou and Feng Zhu and Edward C. Holmes and Jieping Ye and Jun Li and Yuelong Shu and Mǎng Shī and Zhaorong Li},
  title = {{Generalized Biological Foundation Model with Unified Nucleic Acid and Protein Language}},
  booktitle = {Nature Machine Intelligence},
  year = {2025},
  doi = {10.1038/s42256-025-01044-4},
}
Figure