Source: French to English Tester Published on: 2026-04-22
Source: The Conversation – in French– By Julien Mozziconacci, Professor of Computational Biology, National Museum of Natural History (MNHN)
Rather than producing words, the Evo 2 artificial intelligence model is capable of predicting a DNA base based on a given sequence. Launched a little over a year ago, the model is improving and allows scientists to better understand the language of DNA. Its computing power, however, raises questions about energy resources.
If you have ever used a language model like ChatGPT or Mistral, you probably remember the first impression: impeccable spelling, fluent grammar, sentences that make sense. Yet, under the hood, these systems do only one very simple thing: predict the next word in a sentence. They use statistics learned from a huge corpus of texts, and that is how they “speak” French, English, and many other languages.
A fruitful idea then germinated among geneticists: what if we trained the same class of models to learn the language of life, the sequence of letters A, T, G, C, inscribed in our genomes? This is the bet of genomic language models: they learn the hidden grammar of DNA and offer researchers a valuable ally to explore, propose, and test scientific hypotheses more quickly.
What does an AI model do?
An artificial intelligence (AI) algorithm is, fundamentally, a machine for transforming numbers. The input data, which can be images, sounds, or text, are first encoded into numbers. Then the algorithm applies simple operations (additions and multiplications by parameters internal to the network and thresholding) and returns the results (other numbers) as output. On a large scale, this very simple mechanism is sufficient toto play Go, to drive a car… or to understand genomes.
The trick is not just encoding: it is especially learning. The model adjusts its internal parameters with each example (association between an input and a target output), somewhat like tuning an instrument: with each note played, you tighten or loosen the string until the melody sounds right.
The applications of this simple principle are multiple and varied. In the game ofgo, the AI looks at the position of the stones (a board of numbers) and proposes the next move; in a sentence, the model suggests the next word. In genomics, it reads A T G C… and predicts the next base. If its predictions are good, it means it has learned something about the hidden structure of the problem it is solving.
The first genomic language models
It is by following this principle that the first genomic language models were trained using genomes instead of text corpora. One of the most recent versions,Evo 2, was developed by a large team around the Arc Institute research center, in Silicon Valley. This model was trained on numerous genomes, totalingnearly 10,000 billion bases(the famous letters A, C, G, T) which represents 3,000 times the size of our genome.
The model reads one million bases at each step and the calculation always comes back to the same very simple question: among the four possible letters (A, C, G, or T), which one is the most likely just after the ones that have just been read? The gigantic size of its “reading window” allows it to grasp both local rules and long-distance dependencies (gene regulations at a distance). This leap in scale is not just a technical feat: it changes the way questions can be posed in biology, especially in these non-coding regions (those that are not translated into proteins) which often remain misunderstood and constitute the “dark matter” of the genome.
In practice, learning resembles a guessing game: each time the model correctly guesses a hidden letter within a sequence, it strengthens the internal paths that led to it; when it makes a mistake, it adjusts those paths. Over time, it identifies recurring patterns: some motifs often precede the start of a gene, others signal the end, and some sequence motifs betray how the cell cuts RNA (splicing) or the machinery for translating RNAs into proteins is assembled.
Learning is first done on a global scale. The model reads a wide variety of genomes and learns a general grammar of life. Then, it can eventually be adapted to a family of organisms or a specific question (for example, by specializing it on a group of viruses or bacteria).
AI learns the hidden grammar of DNA
This is where the research gets excited: by merely learning to complete sequences, the models recognize biological signatures without being explicitly pointed to them.
They find the three-letter periodicity of the genetic code: the text of life is read in triplets (codons), and the models “hear” this rhythm, like a measure in music. They also identify the start and stop points of genes, with strong constraints on the most important letters, where errors are expected to be rare. They detect signals useful to the cellular machinery: in bacteria, the ribosome binding sites; in eukaryotes, the boundaries between exons (conserved) and introns (sequences to be removed), as if the model distinguished paragraphs and spaces in a text.
More surprisingly, they also reveal mobile elements (for example, viruses integrated into the genome during evolution) and even fingerprints linked to the 3D shapes of proteins (alpha helices, beta sheets) and RNAs. The model then draws the contours of the final sculpture. Because it is indeed sculpture that is involved.
The genome does not only contain instructions – it encodes shapes. A protein, an RNA, are not just simple strings of letters: they fold, twist, and knot in space to adopt a precise architecture, on which their function depends. It is this shape that allows one molecule to recognize another, to attach to it, to trigger a reaction. The contacts that stabilize this shape sometimes occur between regions very far apart in the sequence – and yet, the models seem able to capture them, as if they guessed, through reading the text extensively, which letters correspond despite the distance separating them.
What may be surprising is that these discoveries have not been taught: they emerge spontaneously from learning. And sometimes, paradoxically, when one tries to refine the model by showing it well-known examples, it loses part of what it had found on its own. As if guiding the student too much made it forget what it had intuitively understood.
To make this “black box” more understandable, researchers use “sparse autoencoders” that break down the model’s internal representations into comprehensible features. Each feature lights up like a lamp above a sequence element (exon, motif, mobile element). These features serve as a guide. They indicate where the model saw a signal, its type, and how it varies from one organism to another. These features can even be transferred to poorly studied genomes, paving the way for multi-species functional atlases built more quickly and less expensively than with traditional approaches.
In our own research, Evo 2 is primarily a point of comparison: it shows how far a very large model can go when given enormous amounts of data and computing power. It must also be noted that this demonstration serves as a showcase for Nvidia, the largest manufacturer of processors for AI, which provided its computing power to the Arc Institute to design Evo 2. The underlying idea is to show that gigantic models and extraordinary computing infrastructures are necessary to decipher the secret of life. The result is impressive, but it is not necessarily the only possible path to advance biology.
We have just launched the projectPLANETOID, funded as part of France 2030, to explore a complementary strategy: building much smaller, faster models that are easier to train and deploy in academic laboratories. The goal is to leverage rich biodiversity data produced by our partners—particularly at the National Museum of Natural History and in marine stations—to annotate genomes and metagenomes (sets of genomes) at the scale of the tree of life, including for so-called “non-model” species, which represent the vast majority of life but often remain poorly understood.
PLANETOID also aims to produce reusable resources and tools, so that these approaches are not limited to a few actors capable of mobilizing industrial means, but can permeate public research, and eventually health and the environment.
The Future: Estimating the effect of a mutation or writing new genomes
Because a language model assigns a likelihood to each sequence, it becomes possible to compare the reference version and a mutated version. If the mutation causes the likelihood to drop, it becomes suspect. This score acts like a map to guide researchers: it shows areas where a variation is likely to disrupt a function and directs which experiments to prioritize.
Another application is gaining momentum: the generation of “functional” sequencesin silico. Researchers have shown that one canto compose genetic textwhich has all the characteristics of natural genomes. However, this practice raises important ethical questions (eugenic risks, possibility of synthetic viruses, etc.) and must remain strictly regulated – it is more a societal issue than an immediate research challenge.
Julien Mozziconacci is a professor at the National Museum of Natural History and a junior member of the University Institute of France. He has received funding from the National Research Agency (ANR, France 2030, PostGenAI@Paris). The views and opinions expressed are solely those of the authors and do not necessarily reflect those of the institutions that funded them.
Élodie Laine does not work for, advise, hold shares in, or receive funds from any organization that could benefit from this article, and has declared no other affiliation than her research institute.
–ref. How AI learns the secret language of DNA, and what research gains from it –https://theconversation.com/how-ai-learns-the-secret-language-of-dna-and-what-research-gains-from-it-278320