Foundation models (FMs) are artificial intelligence models trained on large-scale datasets and possessing generalized learning capabilities that can be used in various tasks. Decoding DNA syntax provides an indispensable infrastructure for processes such as disease diagnosis, drug discovery, and synthetic biology. The fact that DNA sequences are continuous nucleotide chains without distinct boundary markers or gaps as in natural languages, presents significant challenges in current modeling approaches. The current state of knowledge in the field shows that tokenization methods with fixed vocabularies fragment the biologically meaningful motifs. On the other hand, Nucleotide-level models require excessive computational costs in long contexts. Accordingly, how to strike a balance between computational feasibility and biological accuracy in the analysis of long DNA sequences remains as a fundamental problem awaiting an answer in the literature.
The study conducted by Arnav Shah, Junzhe Li, and colleagues, aims to develop and test the performance of “dnaHNet”, which is a scalable and hierarchical foundation model for genomic sequence learning. The goal is to overcome the shortcomings of current approaches, such as their reliance on fixed schemes to fragment biological structures or their high computational overhead. The research investigates the hypothesis of whether a dynamic chunking mechanism can discover biologically significant segmentations without external guidance and whether it can scale more efficiently for long genomic contexts compared to existing architectures.
In the study, an autoregressive model independent of chunking has been used. The model directly processes raw nucleotides, and adaptively compresses them into latent tokens via a differentiable dynamic chunking mechanism. The developed architecture is designed with a recursive hierarchy allowing for multiple compression stages to reflect the nested nature of genomic information. The model was trained on high-volume prokaryotic genome data obtained from the Genome Taxonomy Database (GTDB). The system’s performance was tested on experimental fitness data (MaveDB) and gene essentiality (Database of Essential Genes, DEG) datasets to predict protein variant effects. This architectural approach is presented as a distinctive method due to its ability to allow the model to generate its own partitions according to the sequential context of the data.
The analyses conducted demonstrate that the dnaHNet model achieves greater data-processing efficiency than the leading architectures due to its hierarchical compression mechanism. The model increased the inference speed by more than three times compared to transform models. The findings indicate that the method can automatically discover functional boundaries such as codon structures, promoters, and intergenetic regions in a context-dependent manner. The results revealed that the model achieved higher accuracy rates in predicting gene essentiality and protein variant effects compared to existing models. The study is considered to offer a new perspective in terms of computational efficiency in genomic analysis methodologies. It is anticipated that this architecture may provide a foundation for the investigation of eukaryotic genomes and the discovery of novel functional genetic elements in future research.
Yazar: Nehir Necem Ünlü
Editor: Elinsu Ak
Referans: Shah, A., Li, J., Idehpour, P., Fallahpour, A., Wang, B., Hwang, S., Wang, B., Hsu, P. D., Goodarzi, H., & Gu, A. (2026). dnaHNet: A scalable and hierarchical foundation model for genomic sequence learning (arXiv:2602.10603). arXiv. https://arxiv.org/abs/2602.10603
-Bioinfocodes Scientific News Service-
News articles prepared by our team members, reviewing and compiling scientific research
published in journals with and impact factor greater than 20 (click here for the list)