Generative-discriminative hybrids for disease prediction and cell communication modelling

我们的目标是研究机器学习的新进展和专业知识，以提高从基因组和蛋白质组数据预测疾病的可靠性，并能够回答关于疾病发展的细胞-细胞通信机制的新的生物学问题。统计机器学习方法已经被证明在原则上对这些目标具有很大的前景。然而，高通量技术产生了越来越多的高维数据，而样本数量仍然有限。这些极端条件的影响在很大程度上被现有的技术水平所忽视。此外，新的生物学问题被提出，而目前的现有技术无法解决这些问题。机器学习的最新结果使解决这些问题成为可能。特别是，混合生成模型和区分模型可能会融合两者的优点，并可能减少所需的样本量。明智地选择距离函数和数据模型可能会减轻维度的诅咒。将以前开发的某些技术应用于社会网络推理，可能会为推断细胞间的通信机制提供所需的建模能力。通过探索这些技术的潜力，我们希望为生命科学家创造新的和改进的计算方法铺平道路。

We aim to investigate new advances and expertise in machine learning to improve the reliability of disease prediction from genomic and proteomic data, and to enable answering novel biological questions regarding cell-cell communication mechanisms that underlie the development of disease.Statistical machine learning methods have already been shown to hold a lot of promise towards these goals in principle. However, high-throughput technologies result in increasingly high dimensional data, while the number of samples remains limited. The implications of these extreme conditions are largely overlooked by the existing state of the art. In addition, new biological questions are being asked that currently existing techniques are unable to tackle.Recent results in machine learning make it possible to address these issues. In particular, hybridizing generative and discriminative models may blend the benefits of both and may reduce the required sample size. An informed choice of distance functions and data models may mitigate the curse of dimensionality. Adapting certain techniques previously developed for social network inference may provide the required modelling power for inferring cell-cell communication mechanisms.By exploring the potential of these techniques, we hope to pave the way towards creating novel and improved computational methods for life scientists.