Supervised-learning is an accurate method for network-based gene classification

Abstract

Background: Assigning every human gene to specific functions, diseases, and traits is a grand challenge in modern genetics. Key to addressing this challenge are computational methods such as supervised-learning and label-propagation that can leverage molecular interaction networks to predict gene attributes. In spite of being a popular machine learning technique across fields, supervised-learning has been applied only in a few network-based studies for predicting pathway-, phenotype-, or disease-associated genes. It is unknown how supervised-learning broadly performs across different networks and diverse gene classification tasks, and how it compares to label-propagation, the widely-benchmarked canonical approach for this problem. Results:In this study, we present a comprehensive benchmarking of supervised-learning for network-based gene classification, evaluating this approach and a state-of-the-art label-propagation technique on hundreds of diverse prediction tasks and multiple networks using stringent evaluation schemes. We demonstrate that supervised-learning on a gene's full network connectivity outperforms label-propagation and achieves high prediction accuracy by efficiently capturing local network properties, rivaling label-propagation's appeal for naturally using network topology. We further show that supervised-learning on the full network is also superior to learning on node-embeddings (derived using node2vec ), an increasingly popular approach for concisely representing network connectivity. Conclusion:These results show that supervised-learning is an accurate approach for prioritizing genes associated with diverse functions, diseases, and traits and should be considered a staple of network-based gene classification workflows. The datasets and the code used to reproduce the results and add new gene classification methods have been made freely available.

PecanPy: a fast, efficient, and parallelized Python implementation ofnode2vec

“…Some examples of this general task are: 1) classifying uncharacterized genes in a functional interaction network to cellular functions they might participate in (Liu et al , 2020) , and 3) classifying medical terms in a term co-occurrence network (mined from electronic health records) to semantic types (e.g. drug, disease, symptoms, etc.)…”

“…Large-scale molecular networks are powerful models that capture interactions between biomolecules (genes, proteins, metabolites) on a genome scale (McGillivray et al , 2018) and provide a basis for predicting novel associations between individual genes/proteins and various cellular functions, phenotypic traits, and complex diseases (Liu et al , 2020;Sharan et al , 2007) . An area of research that has gained rapid adoption in network science across disciplines is learning low-dimensional numerical representations, or "embeddings", of nodes in a network for easily leveraging machine-learning (ML) algorithms to analyze large networks (Goyal and Ferrara, 2018;Cai et al , 2018;Hamilton et al , 2018) .…”

“…Most popular among the numerous node embedding methods proposed so far is node2vec , a random-walk based network embedding approach (Grover and Leskovec, 2016). Recent work has shown that node2vec has superior performance in node classification tasks on biological networks (Liu et al , 2020;Yue et al , 2020;Nelson et al , 2019) .…”

Integrating multimodal data through interpretable heterogeneous ensembles

“…Moreover, even though the individual STRING datasets considered in this study were natively structured as networks, we used their adjacency matrices as sources of feature vectors for traditional classification algorithms such as SVM, LR and DT. Although this feature-based representation has been recently shown to be more effective for (gene) classification 37 , this assessment may need to be re-evaluated in the context of network integration. Additionally, in the current implementation of EI, we assigned all-zero feature vectors to proteins disconnected from an individual network.…”

“…For all the datasets, both individual and integrated, we used each protein's adjacency vector as its feature values vector for training and evaluating predictive models. This feature encoding has recently been shown to be effective for network-based gene classification 37 .…”