Scholarlink

Context <p>Protein dephosphorylation is essential for regulating cellular processes, and accurate prediction of dephosphorylation sites is crucial for understanding protein function and signaling mechanisms. Existing predictive methods often lack flexibility and fail to capture complex sequence dependencies across diverse proteins.</p> Methods <p>A novel Termite Life Cycle Boltzmann Machine (TLCBM) integrated with Low-Rank Adaptation (LoRA), forming a protein language model that efficiently learns high-dimensional sequence representations. TLCBM–LoRA leverages a sliding window to extract local sequence contexts while simultaneously modeling long-range dependencies, focusing on key residues tyrosine (Y), serine (S), and threonine (T). The model achieves high predictive performance, with accuracy of 0.92 and 0.96, MCC of 0.91 and 0.94, specificity of 0.92 and 0.97, sensitivity of 0.93 and 0.95, and ROC_AUC of 0.95 and 0.98 for Y and ST residues, respectively. Compared to the traditional model, 3% performance was improved for the proposed system. The results demonstrate that TLCBM–LoRA is a flexible and powerful tool for enhanced protein sequence analysis and dephosphorylation site recognition.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimized deep intelligent parameter efficient fine-tuning protein language model system for predicting dephosphorylation site

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Anurag Singh,
  • P. K. Singh,
  • Rohit Kumar Tiwari

摘要

Context

Protein dephosphorylation is essential for regulating cellular processes, and accurate prediction of dephosphorylation sites is crucial for understanding protein function and signaling mechanisms. Existing predictive methods often lack flexibility and fail to capture complex sequence dependencies across diverse proteins.

Methods

A novel Termite Life Cycle Boltzmann Machine (TLCBM) integrated with Low-Rank Adaptation (LoRA), forming a protein language model that efficiently learns high-dimensional sequence representations. TLCBM–LoRA leverages a sliding window to extract local sequence contexts while simultaneously modeling long-range dependencies, focusing on key residues tyrosine (Y), serine (S), and threonine (T). The model achieves high predictive performance, with accuracy of 0.92 and 0.96, MCC of 0.91 and 0.94, specificity of 0.92 and 0.97, sensitivity of 0.93 and 0.95, and ROC_AUC of 0.95 and 0.98 for Y and ST residues, respectively. Compared to the traditional model, 3% performance was improved for the proposed system. The results demonstrate that TLCBM–LoRA is a flexible and powerful tool for enhanced protein sequence analysis and dephosphorylation site recognition.