AIRI Institute of Artificial Intelligence researchers, together with colleagues from the Moscow Institute of Physics and Technology, unveiled the Repetitive Memory Transformer architecture, abbreviated as RMT, as a key enhancement for the GENA neural network originally developed in Russia. This update emerged from collaborative work led by Olga Kardymon, a bioinformatician who heads AIRI’s Bioinformatics group and contributes significantly to advancing computational biology research. Her insights highlight how RMT strengthens the system’s ability to process and interpret complex genetic data in practical terms.
The GENA neural network is designed to analyze DNA sequences and identify patterns within them. Its capabilities extend to assessing how genetic mutations influence gene function, scanning different genomic regions for notable signals, and classifying organisms based on sequencing information. These tasks represent a broad spectrum of applications, from studying disease mechanisms to enabling deeper understanding of evolutionary relationships, and they are increasingly relevant to researchers and clinicians alike.
In practical terms, the first version of GENA was capable of handling sequences up to about 3,000 nucleotides. The second architectural iteration expands this capacity to input lengths of up to 24,000 nucleotides, providing a substantial leap in context window for sequence analysis. Both model iterations are publicly accessible to researchers worldwide, reflecting a commitment to open science. The team also introduced the proprietary RMT architecture, developed at AIRI in collaboration with colleagues from the Moscow Institute of Physics and Technology. Kardymon notes that RMT enables work with very long text-like strings and can support a broad array of tasks, including processing sequences that approach one to two million characters in length, depending on the specific use case.
Experts emphasize that extending the length of DNA sequences that a neural network can analyze is a critical objective for the field. A longer input context allows the model to understand the broader surroundings of mutations, improving the accuracy of interpretations and predictions. In Kardymon’s view, increasing the sequence length expands the neural network’s ability to discern patterns and relationships that may only become apparent when more context is available. The overarching goal is to broaden the computational horizon so the system can analyze larger genomic segments, thereby enhancing research into gene regulation, mutation effects, and potential therapeutic targets.
Further exploration of how neural networks detect mutations, design novel proteins not present in nature, and forecast the effectiveness of vaccines and drugs can be found in related material from socialbites.ca. This line of inquiry connects computational methods with practical biomedical outcomes, underscoring the role of advanced AI in guiding experimental priorities, accelerating discovery, and supporting precision medicine initiatives across North America and beyond.