FishMamba-1: A Linear-Complexity Foundation Model for Deciphering Polyploid Cyprinid Genomes

Abstract

The Cypriniformes order, comprising essential aquaculture species like carps and minnows, presents unique genomic challenges due to complex whole-genome duplication (WGD) events and extensive repetitive elements. Conventional annotation tools and Transformer-based foundation models often struggle to capture long-range dependencies in these expanded genomes due to quadratic computational complexity. Here, we introduce FishMamba-1, the first genomic foundation model tailored for the aquatic clade, built upon the selective state-space model (SSM) architecture. By leveraging Mamba-2s linear scaling efficiency, FishMamba-1 processes context windows of 32,768 base pairs (32k)--significantly surpassing the 4-6k limit of standard DNA Transformers--enabling the modeling of distal regulatory patterns on a single GPU. We curated Cypri-24, a comprehensive dataset comprising 28.8 Gb of high-quality genome assemblies from 24 representative species, to pre-train FishMamba-1 on 15 billion tokens. Subsequent fine-tuning for genome segmentation (FishSegmenter) demonstrates the models capability to annotate gene structures at single-nucleotide resolution with remarkable precision. Evaluation on a held-out test set reveals that FishMamba-1 achieves a precision of 64.6% in exon identification, effectively distinguishing coding regions from the vast non-coding background without relying on RNA-seq evidence. Furthermore, interpretability analysis confirms that the model captures biological syntax such as splice acceptor motifs. FishMamba-1 provides a scalable, open-source framework for decoding the complex genomes of non-model organisms, providing a scalable computational resource to support downstream applications in molecular breeding and ecological monitoring. The complete source code, pre-trained model weights, and datasets are freely available at https://github.com/lu1000001/FishMamba. Additionally, the FishMamba Hub, a web-based inference platform, is accessible at https://huggingface.co/spaces/lu1000001/FishMamba-Hub to facilitate real-time genomic segmentation for the aquatic research community. Graphic Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=112 SRC="FIGDIR/small/710409v1_ufig1.gif" ALT="Figure 1"> View larger version (71K): org.highwire.dtl.DTLVardef@f5ee02org.highwire.dtl.DTLVardef@1dcfd31org.highwire.dtl.DTLVardef@1729ef8org.highwire.dtl.DTLVardef@22e4f1_HPS_FORMAT_FIGEXP M_FIG C_FIG