基因组语言模型能否为病毒基因组片段生成有分类学信息量的嵌入?bioRxiv 预印本评估 DNABERT-2、DNABERT-S、ViraLM 与 Evo2-7B
Assessing the Capacity of Genomic Language Models to Generate Informative Embeddings of Viral Genome Fragments

bioRxiv 预印本在基于 ICTV 病毒元数据资源构建的平衡参考数据集上,比较 DNABERT-2、DNABERT-S、ViraLM 和 Evo2-7B 四种基因组语言模型与 k-mer 谱、MinHash sketch 对短病毒基因组片段的表征能力。
原文摘要
出版方登记 · CrossrefGenomic language models (gLMs) trained on large volumes of nucleotide sequences are becoming an alternative to classical methods in metagenomic analysis, particularly for studying genomic "dark matter". In this work, we investigate the ability of four gLMs (DNABERT-2, DNABERT-S, ViraLM, and Evo2-7B) to generate taxonomically informative vector representations of short viral genome fragments. We constructed a balanced reference dataset of viruses based on the ICTV Virus Metadata Resource and compared the representations generated by gLMs with conventional k-mer profiles and MinHash sketches using clustering and neighborhood-based metrics (adjusted Rand index (ARI), precision@k, and CROSS precision@k). On the reference dataset, three of the four gLMs outperformed the 4-mer profile in terms of ARI and precision@5 under the same evaluation protocol. The best performance was generally achieved by intermediate model layers. Conventional k-mer profiles consistently outperformed MinHash sketches. We further evaluated the methods on two viral orders (egoviruses and mriyaviruses) absent from the reference dataset, with model layers and k-mer lengths selected exclusively on the reference dataset. DNABERT-S showed the most consistently strong performance across these two orders, ranking first on mriyaviruses and second on egoviruses, while k-mer profiles with Jensen-Shannon distance remained competitive. We also observed the consistency of representations between forward fragments and their reverse complements. Overall, the best-performing gLMs surpassed conventional baselines, although their advantage depended on the dataset and evaluation metric. The consistently strong performance of DNABERT-S suggests a potential benefit of contrastive fine-tuning and indicates that larger models are not necessarily more effective for generating taxonomically informative representations.
页面未能读取全文,以上为出版方登记的摘要;方法与图表见原文
来源:bioRxiv 生物信息学 · biorxiv.org