Reciprocal Perspective Machine Learning to Identify Relationships in Sparse Biological Networks

我们的实验室专注于开发机器学习（ML）方法来解决生物医学信息学中的问题，特别是在班级不平衡的情况下，在所有数据中感兴趣的事件很少。最近，我们的实验室开发了许多半监督机器学习方法，可以从标记和未标记的数据中学习。这些方法与许多问题领域相关，特别是在生物信息学中，我们经常有大量未标记的数据（例如基因组或蛋白质组学序列数据），但获得标记数据需要昂贵的实验。我们的研究是由生物信息学中的一个常见故事驱动的：我们为给定的任务创建最好的预测器，给定有限的可用训练数据。在巨大的计算开销下，我们将预测器应用于所有可用的未标记数据，并为湿实验室合作者报告代表新颖可测试假设的“顶级预测”。但是我们不能利用这种巨大的计算投资来改进潜在的预测模型吗？这个故事导致了我们最近最令人兴奋的发现之一，互惠视角（RP）范式，它从大量未标记样本的预测分数分布结构中学习。通过这种方式，每个预测都在涉及每个元素的所有可能预测的上下文中进行评估。RP特别适合于配对预测问题，在这种情况下，人们的目标是预测稀疏网络中节点之间的链接。使用RP，我们已经证明了预测蛋白质相互作用和microRNA目标的显著性能改进。我们建议扩展和应用这些半监督机器学习方法，以继续我们的长期研究目标，即提高机器学习模型在受类别不平衡挑战的领域中的预测性能。该研究计划将扩展和推广RP范式，以创建一个强大的半监督ML框架，广泛适用于不同的领域。特别是，我们将发展RP与多视图协同训练的融合；探索RP作为一种结合多名专家以达成改进共识决策的手段；实现RP与半监督转导学习的概念合并；并将RP范式从成对预测扩展到n维情况。我们的实验室在生物医学信息学领域有着成功、创新和有影响力的跨学科研究记录。在这个研究项目中开发的新的机器学习方法将通过建立和有效的合作被翻译并应用于几个问题领域。这项研究将在基础机器学习研究、生物医学信息学等领域产生影响。

Our lab focuses on the development of machine learning (ML) approaches to solve problems in biomedical informatics, particularly in the presence of class imbalance where events of interest are rare among all data. Recently, our lab has developed a number of semi-supervised ML approaches that learn from both labeled and unlabeled data. Such approaches are relevant for many problem domains, especially within bioinformatics, where we often have large quantities of unlabeled data (e.g. genomic or proteomic sequence data), but acquiring labeled data requires costly experiments. Our research is motivated by a common story in bioinformatics: we create the best predictor possible for a given task, given the limited training data available. At great computational expense, we then apply the predictor to all available unlabeled data and report the "top predictions" representing novel testable hypotheses for wet-lab collaborators. But can't we leverage this enormous computational investment to improve the underlying prediction model? This story led to one of our most exciting recent discoveries, the Reciprocal Perspective (RP) paradigm, which learns from the structure of the distribution of prediction scores over a large body of unlabeled samples. In this way, each prediction is evaluated in the context of all possible predictions involving each element. RP is particularly well-suited to pairwise prediction problems where one aims to predict links between nodes in a sparse network. Using RP, we have demonstrated significant performance improvements for predicting protein-protein interactions and microRNA targets. We propose to extend and apply these semi-supervised ML approaches to continue our long-term research goals of improving the predictive performance of ML models applied to domains challenged by class imbalance. This research program will extend and generalize the RP paradigm to create a robust semi-supervised ML framework, broadly applicable to disparate domains. In particular, we will develop the fusion of RP with multi-view co-training; explore RP as a means to combine multiple experts to arrive at improved consensus decisions; achieve the conceptual merger of RP with semi-supervised transductive learning; and extend the RP paradigm beyond pairwise predictions to the N-dimensional case. Our lab has a demonstrated record of successful, innovative, and impactful interdisciplinary research in biomedical informatics. The novel ML methodology to be developed within this research program will be translated and applied to several problem domains via established and effective collaborations. This research will achieve impact in fundamental ML research, in biomedical informatics, and beyond.