HybriDNA: A Hybrid Transformer-Mamba2 Long-Range DNA Language Model

TL;DR

HybriDNA introduces a decoder-only DNA foundation model that fuses Hybrid Transformer-Mamba2 architecture to enable single-nucleotide resolution over long contexts (up to 131k nucleotides). It is pretrained on a large multi-species corpus with base-level tokenization using next-token prediction, and finetuned with echo embeddings for understanding and prompt-tokens for generation, achieving state-of-the-art performance across multiple DNA benchmarks and enabling design of synthetic CREs. The model scales across 300M, 3B, and 7B parameters, showing improved performance with scale and longer context, while delivering superior computational efficiency relative to standard Transformers at long contexts. These capabilities position HybriDNA as a versatile tool for both fundamental DNA understanding and engineered genome design, with potential impact on biology, medicine, and biotechnology.

Abstract

Advances in natural language processing and large language models have sparked growing interest in modeling DNA, often referred to as the "language of life". However, DNA modeling poses unique challenges. First, it requires the ability to process ultra-long DNA sequences while preserving single-nucleotide resolution, as individual nucleotides play a critical role in DNA function. Second, success in this domain requires excelling at both generative and understanding tasks: generative tasks hold potential for therapeutic and industrial applications, while understanding tasks provide crucial insights into biological mechanisms and diseases. To address these challenges, we propose HybriDNA, a decoder-only DNA language model that incorporates a hybrid Transformer-Mamba2 architecture, seamlessly integrating the strengths of attention mechanisms with selective state-space models. This hybrid design enables HybriDNA to efficiently process DNA sequences up to 131kb in length with single-nucleotide resolution. HybriDNA achieves state-of-the-art performance across 33 DNA understanding datasets curated from the BEND, GUE, and LRB benchmarks, and demonstrates exceptional capability in generating synthetic cis-regulatory elements (CREs) with desired properties. Furthermore, we show that HybriDNA adheres to expected scaling laws, with performance improving consistently as the model scales from 300M to 3B and 7B parameters. These findings underscore HybriDNA's versatility and its potential to advance DNA research and applications, paving the way for innovations in understanding and engineering the "language of life".

Figures (6)

• Figure 1: Overview of HybriDNA: A Language Model for DNA Sequences. HybriDNA builds upon an efficient hybrid Transformer and Mamba2 architecture. It is initially pretrained on large-scale, multi-species genomic data at single-nucleotide resolution using a next-token prediction objective. Subsequently, HybriDNA employs an echo embedding fine-tuning approach for DNA understanding tasks and a generative fine-tuning approach for DNA generation tasks.
• Figure 2: Model Architecture of HybriDNA
• Figure 3: Pretraining loss curves for HybriDNA-300M, 3B, and 7B models
• Figure 4: Comparison of training throughput and GPU memory consumption between HybriDNA and a pure Transformer model with comparable parameters (e.g., 300M)
• Figure 5: HybriDNA Mamba2 Block