Genome-Predicting "Chimaera" Gives Scientists a New Way to Explore DNA's 3D Architecture
Russian bioinformaticians have developed a neural network that can analyze and predict the three-dimensional organization of genomes across organisms ranging from fungi to mammals. The work suggests that the rules governing genome architecture differ from one species to another.
For decades, geneticists interpreted DNA as a flat, linear sequence of information. Inside living cells, however, the genome folds into an extraordinarily complex three-dimensional structure in which the spatial proximity of genes plays a critical role in gene regulation. Russian bioinformaticians have now taken a major step forward by developing Chimaera, an interpretable neural network capable of predicting the genome's three-dimensional organization directly from DNA sequences. Trained on data from 22 species, including yeast, plants, fish, mammals and humans, the model opens new opportunities for comparative biology.
The study was conducted by Russian researchers and published in Nucleic Acids Research. One of the paper's lead authors, Alexey Shkolikov, is affiliated with the Institute of General Genetics of the Russian Academy of Sciences, while the project was led by bioinformatician and evolutionary biologist Mikhail S. Gelfand. The researchers analyzed both DNA sequence data and three-dimensional genome organization across 22 organisms representing a wide range of taxonomic groups, including humans, mice, frogs, zebrafish, gastropod mollusks, honeybees, ants, fruit flies, mosquitoes, ticks, nematodes, comb jellies, amoebae, yeast and the model plant Arabidopsis thaliana, among others.
The End of the "Black Box" Era
Chimaera's defining advantage is its interpretability. Many AI systems function as "black boxes," producing accurate predictions without revealing how those predictions are made. Chimaera takes a different approach. It not only predicts how chromatin folds but also identifies the specific DNA regions responsible for shaping that three-dimensional architecture. The study also uncovered an important finding: structurally similar genome features in different organisms can arise through entirely different biological mechanisms. That insight offers researchers a powerful framework for investigating genome evolution and generating new biological hypotheses.
The Chimaera neural network was trained on datasets containing DNA fragments paired with maps of three-dimensional genome organization. Through that process, the model learned to identify hidden principles governing chromatin folding and to predict genome architecture using DNA sequence information alone. Building on those patterns, the researchers were also able to construct what amounts to an evolutionary tree of three-dimensional genome organization spanning organisms from plants to mammals.
From CRISPR to Genome Architecture
Chimaera's emergence in 2026 reflects the steady evolution of Russia's bioinformatics research community rather than a single breakthrough. Four years earlier, in 2022, researchers at Skoltech developed an AI system that improved the accuracy of the CRISPR gene-editing platform. In 2023, AIRI and the Moscow Institute of Physics and Technology introduced GENA, Russia's first DNA language model. In 2024, the compact LegNet network won first place in the international DREAM challenge by predicting gene activity, and in 2025, the GENA_LM family learned to analyze ultra-long DNA sequences. Today, we are witnessing a historic transition from simply reading the letters of the genetic code to understanding its three-dimensional architecture.
A Foundation for Future Medicine
Chimaera is a tool for fundamental research that could accelerate efforts to identify the causes of inherited diseases, discover regulatory DNA regions and develop approaches for personalized medicine.
The work also has important implications for agriculture. As Russia expands technologies supporting technological sovereignty, the ability to model disease resistance in plants and animals at the level of three-dimensional genome organization could strengthen next-generation breeding programs. Domestic software platforms for genomic data analysis may also become a foundation for independent bioinformatics infrastructure.
On the global market, Russia is positioned to offer not only biological resources but also advanced computational services. Export potential could be especially strong for secure, locally deployed solutions serving international universities and biotechnology companies that prefer not to place sensitive genomic data in external cloud environments.
Building the Digital Groundwork
To move Chimaera toward practical applications, researchers say it will require what they describe as "digital ground" – large collections of high-quality experimental Hi-C maps and carefully annotated datasets. The study identifies fragmented biological data as one of the field's primary limitations. Information scattered across independent laboratories makes it difficult to train models at the scale required. Addressing that challenge calls for the systematic integration of Russian biobanks and sequencing centers into a unified ecosystem. Only by combining AI models with real clinical and agricultural research can scientists develop technologies ready for practical use.
The development of Chimaera highlights the growing role Russian AI is playing in fundamental scientific research. By linking advanced computational methods with the biology of genome organization, the work could contribute to future advances in medicine, pharmaceuticals and biotechnology while strengthening the country's position in the rapidly expanding field of computational biology.
