HopPER: an adaptive model for probability estimation of influenza reassortment through host prediction

Abstract

Background: Influenza reassortment, a mechanism where influenza viruses exchange their RNA segments by co-infecting a single cell, has been implicated in several major pandemics since 19th century. Owing to the significant impact on public health and social stability, great attention has been received on the identification of influenza reassortment.

Methods: We proposed a novel computational method named HopPER (Host-prediction-based Probability Estimation of Reassortment), that sturdily estimates reassortment probabilities through host tropism prediction using 147 new features generated from seven physicochemical properties of amino acids. We conducted the experiments on a range of real and synthetic datasets and compared HopPER with several state-of-the-art methods.

Results: It is shown that 280 out of 318 candidate reassortants have been successfully identified. Additionally, not only can HopPER be applied to complete genomes but its effectiveness on incomplete genomes is also demonstrated. The analysis of evolutionary success of avian, human and swine viruses generated through reassortment across different years using HopPER further revealed the reassortment history of the influenza viruses.

Conclusions: Our study presents a novel method for the prediction of influenza reassortment. We hope this method could facilitate rapid reassortment detection and provide novel insights into the evolutionary patterns of influenza viruses.

Keywords: Host tropism; Influenza; Random forest; Reassortment estimation.

Fig. 1

Schematic overview of analysis workflow in HopPER. a The general diagram of the host prediction model based on seven physicochemical properties and reassortment probability estimation in the random forest. b Specific algorithmic steps for estimation probability model on influenza genome reassortment detection

Fig. 2

The structure of random forest T for probability estimation. θ_t is an independent random draw and f(θ_t,x₀) stands for the probability estimate by associated tree t at point x₀. $\mathbb{P}$(y_j|x₀) characterizes the aggregation of conditional probability of all trees for label y_i

Fig. 3

The RSR of three distinct host species of influenza strains across different years detected by HopPER. a The RSR of avian species from 1988 to 2017. b The RSR of human species from 1975 to 2017. c The RSR of swine species from 2000 to 2017