Optimisation of an exemplar oculomotor model using multi-objective genetic algorithms executed on a GPU-CPU combination.
Optimisation of an exemplar oculomotor model using multi-objective genetic algorithms executed on a GPU-CPU combination.
复制标题
DOI：
10.1186/s12918-017-0416-2
复制
发表时间：
2017-03-24
影响因子：
--
通讯作者：
Akman OE
中科院分区：
文献类型：
--
作者：
Avramidis E;Akman OE
参数优化是构建计算生物学模型的关键步骤。在眼动研究中，计算模型对于理解正常和异常行为的机制基础越来越重要。在这项研究中，我们考虑了一个现有的快速眼动(眼跳)的神经生物学模型，能够产生真实的模拟：(I)正常的水平眼跳；(Ii)婴儿眼球震颤--可以细分为不同波形类别的病理性眼球振荡。通过开发适当的适应度函数，我们使用成熟的多目标遗传算法，根据现有的实验眼跳和眼球震颤数据对模型进行了优化。该算法要求对大量参数组合的模型进行数值积分。为了解决这一计算瓶颈，我们实现了主从并行，在CPU的控制下，模型集成分布在GPU的计算单元上。虽然以前的眼震拟合是基于再现定性波形特征，但我们的优化方案使我们能够执行模型与实验记录的第一次直接拟合。对正常眼球运动的拟合表明，尽管不同幅度的眼跳可以通过单独的参数集准确地模拟，但无法确定能够同时拟合所有幅度的单一集合。对眼震振荡的拟合系统地识别了该模型能够高精度再现许多典型眼震波形的参数区域，同时还识别了该模型无法模拟的一些波形。与高端CPU相比，使用GPU执行模型集成可获得约20%的加速比。这两个优化问题的结果使我们能够量化模型的预测能力，建议进行具体的修改，以扩大其模拟行为的保留范围。此外，我们得到的最优参数分布与以前的计算研究一致，这些计算研究认为跳动制动信号是婴儿眼球震颤发展之前不稳定的根源。最后，我们开发的加速优化过程的主从并行方法可以很容易地适用于其他高度参数化的计算生物学模型来适应实验数据。本文的在线版本(doi：10.1186/s12918-0170416-2)包含补充材料，授权用户可以使用。
Parameter optimisation is a critical step in the construction of computational biology models. In eye movement research, computational models are increasingly important to understanding the mechanistic basis of normal and abnormal behaviour. In this study, we considered an existing neurobiological model of fast eye movements (saccades), capable of generating realistic simulations of: (i) normal horizontal saccades; and (ii) infantile nystagmus – pathological ocular oscillations that can be subdivided into different waveform classes. By developing appropriate fitness functions, we optimised the model to existing experimental saccade and nystagmus data, using a well-established multi-objective genetic algorithm. This algorithm required the model to be numerically integrated for very large numbers of parameter combinations. To address this computational bottleneck, we implemented a master-slave parallelisation, in which the model integrations were distributed across the compute units of a GPU, under the control of a CPU. While previous nystagmus fitting has been based on reproducing qualitative waveform characteristics, our optimisation protocol enabled us to perform the first direct fits of a model to experimental recordings. The fits to normal eye movements showed that although saccades of different amplitudes can be accurately simulated by individual parameter sets, a single set capable of fitting all amplitudes simultaneously cannot be determined. The fits to nystagmus oscillations systematically identified the parameter regimes in which the model can reproduce a number of canonical nystagmus waveforms to a high accuracy, whilst also identifying some waveforms that the model cannot simulate. Using a GPU to perform the model integrations yielded a speedup of around 20 compared to a high-end CPU. The results of both optimisation problems enabled us to quantify the predictive capacity of the model, suggesting specific modifications that could expand its repertoire of simulated behaviours. In addition, the optimal parameter distributions we obtained were consistent with previous computational studies that had proposed the saccadic braking signal to be the origin of the instability preceding the development of infantile nystagmus oscillations. Finally, the master-slave parallelisation method we developed to accelerate the optimisation process can be readily adapted to fit other highly parametrised computational biology models to experimental data. The online version of this article (doi:10.1186/s12918-017-0416-2) contains supplementary material, which is available to authorized users.
DOI：
10.1093/bioinformatics/btv062
发表时间：
2015-06-01
期刊：
Bioinformatics (Oxford, England)
影响因子：
--
作者：
Aitken S;Kilpatrick AM;Akman OE
通讯作者：
Akman OE