Carbon-A: Unified Genome Annotation across Eukaryotic Diversity
The paper introduces Carbon-A, a unified 1.2-billion-parameter encoder model that achieves state-of-the-art structural genome annotation across diverse eukaryotic lineages using only genomic sequences, and leverages this capability to build the Carbon Annotation Database (CADB), a comprehensive resource containing predicted gene models for over 48,000 assemblies that significantly expands upon existing reference databases.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
To understand life, scientists must first find the instructions hidden within the genetic code. Every living thing, from a human to a mushroom, carries a massive library of DNA, a long string of chemical letters that holds the blueprints for building and running an organism. For decades, the challenge was simply to read this string. Now that sequencing machines can produce these strings quickly and cheaply, the bottleneck has shifted. Having the raw text is not enough; researchers must identify which parts of the text are actual instructions for making proteins, the workhorse molecules that build cells and drive biological processes. This process, known as gene annotation, is like trying to find the sentences in a book that has no spaces between words and no punctuation. In complex organisms like animals and plants, these instructions are often broken into small pieces scattered across long stretches of non-coding DNA, making them incredibly difficult to locate without a guide.
Traditionally, scientists have relied on two main strategies to find these genes. One method looks for evidence from the organism itself, such as RNA molecules that are known to be active in specific tissues. The other compares the new DNA to genes already discovered in closely related species. Both approaches work well when the organism is well-studied or when a close relative has been mapped out. However, they fail miserably when applied to the vast majority of life on Earth, which consists of species that have never been studied in detail and have no close relatives with known genetic maps. For these organisms, the instructions remain hidden, leaving huge gaps in our understanding of the tree of life.
A team of researchers has now introduced a new tool designed to solve this problem across the entire spectrum of complex life. They developed a computer model called Carbon-A, which acts as a universal translator for genetic code. Unlike previous tools that needed to be trained separately for mammals, plants, or fungi, this single model learned to recognize the patterns of protein-coding genes in all of them at once. By analyzing the raw DNA sequence directly, without needing prior knowledge of the specific species, Carbon-A can scan a genome and pinpoint exactly where the genes begin and end, even in organisms that have never been seen before.
The researchers tested this model on 42 different genomes representing six major groups of complex life: mammals, other vertebrates, invertebrates, plants, fungi, and single-celled organisms called protozoa. In every category, Carbon-A outperformed existing methods, correctly identifying the location of genes with higher accuracy than any other tool. It proved particularly robust when faced with unusual genetic rules, such as in a type of single-celled organism where the standard stop signals for genes work differently. While other models failed completely in these scenarios, Carbon-A adapted and maintained high accuracy. It also provides a confidence score for each prediction, telling researchers how reliable a specific gene finding is, which helps them decide which results to trust most.
To demonstrate the model's ability to find genes that others missed, the team compared its predictions against experimental data from four species: a cat, a hamster, a chicken, and a plant. They used a technique that reads the full length of RNA molecules to see which genes are actually being used. Carbon-A identified hundreds of gene candidates that were not present in the standard reference databases but were supported by this experimental evidence. Furthermore, computer simulations of the protein structures these genes would produce showed that they fold into stable, plausible shapes, suggesting they are real, functional parts of the organism. This indicates that the model is not just guessing but is uncovering genuine biological instructions that have been overlooked.
Building on this success, the researchers used Carbon-A to annotate every eukaryotic genome assembly available in a major public database. This massive effort resulted in the creation of the Carbon Annotation Database, a new resource containing predictions for over 566 million protein-coding genes across 22,617 different species. This database represents a ninefold increase in the amount of genetic sequence and an elevenfold increase in the number of species covered compared to the previous gold standard. For scientists studying poorly understood organisms, this provides a consistent and reliable starting point for discovery. It allows researchers to ask new questions about how genes evolve, which genes are unique to certain lineages, and how different species are related, turning a vast collection of raw genetic data into a searchable library of biological knowledge.
The work suggests that the rules for building proteins are more universal than previously thought. By learning from a diverse mix of life forms simultaneously, the model discovered shared patterns that apply across deep evolutionary divides. This approach moves beyond the need for close relatives to guide the search, offering a way to explore the genetic instructions of any complex organism, no matter how obscure. As the database grows and improves, it will serve as a foundation for future discoveries, helping to fill in the missing chapters of the story of life on Earth.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.