Gene Regulatory Network (GRN) is formed due to mutual transcriptional regulation within a set of protein coding genes in cellular context of an organism. Computational inference of GRN is important to understand the behavior of each gene in terms of change in its protein production rate (expression level). As Recurrent Neural Network (RNN) is efficient in GRN modeling, a bi-objective RNN formulation has been applied here. Based on Archived Multi Objective Simulated Annealing (AMOSA), four algorithms, namely, AMOSA Revised (AMOSAR), Modified Freezing based AMOSA (AMOFSA), Tabu based AMOSA (AMOTSA) and Modified Freezing and Tabu based AMOSA (AMOFTSA) have been proposed and applied to RNN (treated as GRN) for parameter learning taking four gene expression time series datasets. Comparative studies on the performance of the algorithms (based on each dataset) have been made in terms of the number of GRNs obtained in the final non-dominated front and the performance metrics, namely, recall, precision and f1 score. Two proposed variants, namely, AMOFSA and AMOTSA have been found competitive in performance. Experimental observations and statistical analysis show that, modified algorithms are better than AMOSAR and the state-of-the-art algorithms in respect of the above-mentioned metrics.

中文翻译：

---

推断基因调控网络的多目标模拟退火变体：一项比较研究

基因调控网络 (GRN) 是由于生物体细胞环境中一组蛋白质编码基因内的相互转录调控而形成的。GRN 的计算推断对于了解每个基因在其蛋白质生产速率（表达水平）变化方面的行为非常重要。由于循环神经网络 (RNN) 在 GRN 建模中非常有效，因此此处应用了双目标 RNN 公式。基于存档多目标模拟退火（AMOSA），提出了四种算法，即AMOSA修订（AMOSAR）、修正冷冻AMOSA（AMOFSA）、禁忌AMOSA（AMOTSA）和修正冷冻和禁忌AMOSA（AMOFTSA）。应用于RNN（视为GRN）进行参数学习，采用四个基因表达时间序列数据集。在最终非支配前沿获得的 GRN 数量和性能指标（即召回率、精度和 f1 分数）方面，对算法的性能（基于每个数据集）进行了比较研究。已发现两种提议的变体，即 AMOFSA 和 AMOTSA，在性能上具有竞争力。实验观察和统计分析表明，在上述指标方面，修改后的算法优于 AMOSAR 和最先进的算法。