Bioinformatics Institute (BII), Agency for Science, Technology and Research (A*STAR), 30 Biopolis Street, Matrix #07-01, Singapore, 138761, Republic of Singapore.
School of Biological Sciences, Nanyang Technological University (NTU), 60 Nanyang Drive, SBS-01s-45, Singapore, 637551, Republic of Singapore.
BMC Biol. 2024 Oct 14;22(1):235. doi: 10.1186/s12915-024-02019-4.
The transition from explanative modeling of fitted data to the predictive modeling of unseen data for systems biology endeavors necessitates the effective recovery of reaction parameters. Yet, the relative efficacy of optimization algorithms in doing so remains under-studied, as to the specific reaction kinetics and the effect of measurement noises. To this end, we simulate the reactions of an artificial pathway using 4 kinetic formulations: generalized mass action (GMA), Michaelis-Menten, linear-logarithmic, and convenience kinetics. We then compare the effectiveness of 5 evolutionary algorithms (CMAES, DE, SRES, ISRES, G3PCX) for objective function optimization in kinetic parameter hyperspace to determine the corresponding estimated parameters.
We quickly dropped the DE algorithm due to its poor performance. Baring measurement noise, we find the CMAES algorithm to only require a fraction of the computational cost incurred by other EAs for both GMA and linear-logarithmic kinetics yet performing as well by other criteria. However, with increasing noise, SRES and ISRES perform more reliably for GMA kinetics, but at considerably higher computational cost. Conversely, G3PCX is among the most efficacious for estimating Michaelis-Menten parameters regardless of noise, while achieving numerous folds saving in computational cost. Cost aside, we find SRES to be versatilely applicable across GMA, Michaelis-Menten, and linear-logarithmic kinetics, with good resilience to noise. Nonetheless, we could not identify the parameters of convenience kinetics using any algorithm.
Altogether, we identify a protocol for predicting reaction parameters under marked measurement noise, as a step towards predictive modeling for systems biology endeavors.
从拟合数据的解释性建模到系统生物学研究中未见数据的预测建模的转变，需要有效地恢复反应参数。然而，优化算法在这方面的相对效果仍然研究不足，特别是对于特定的反应动力学和测量噪声的影响。为此，我们使用 4 种动力学模型（广义质量作用定律（GMA）、米氏动力学、线性对数动力学和便利动力学）模拟了一条人工途径的反应。然后，我们比较了 5 种进化算法（CMAES、DE、SRES、ISRES、G3PCX）在动力学参数超空间中的目标函数优化效果，以确定相应的估计参数。
我们很快放弃了 DE 算法，因为它的性能不佳。在没有测量噪声的情况下，我们发现 CMAES 算法在 GMA 和线性对数动力学方面仅需要其他 EA 计算成本的一小部分，而在其他标准下表现也一样好。然而，随着噪声的增加，SRES 和 ISRES 对于 GMA 动力学的可靠性更高，但计算成本也大大增加。相反，G3PCX 无论噪声如何，对于估计米氏动力学参数都非常有效，同时在计算成本方面节省了数倍。除了成本之外，我们发现 SRES 在 GMA、米氏动力学和线性对数动力学方面具有通用性，对噪声有很好的适应性。然而，我们无法使用任何算法来确定便利动力学的参数。
总之，我们确定了一种在显著测量噪声下预测反应参数的方案，这是迈向系统生物学研究预测建模的一步。
