[Paper Review] Imputing the mammalian virome with linear filtering and singular value decomposition

Timothée Poisot, Marie-Andrée Ouellet+7 more|arXiv (Cornell University)|May 31, 2021

Genomics and Phylogenetic Studies1 references11 citations

TL;DR

This study introduces a novel method combining Linear Filtering (LF) and Singular Value Decomposition (SVD) to impute missing host-virus associations in the mammalian virome, revealing a global hotspot of unsampled interactions in the Amazon rainforest and demonstrating that the imputed network enhances predictions of human viral infection risk from viral genome features.

ABSTRACT

At most 1-2% of the global virome has been sampled to date. Here, we develop a novel method that combines Linear Filtering (LF) and Singular Value Decomposition (SVD) to infer host-virus associations. Using this method, we recovered highly plausible undiscovered interactions with a strong signal of viral coevolutionary history, and revealed a global hotspot of unusually unique but unsampled (or unrealized) host-virus interactions in the Amazon rainforest. We finally show that graph embedding of the imputed network can be used to improve predictions of human infection from viral genome features, showing that the global structure of the mammal-virus network provides additional insights into human disease emergence.

Motivation & Objective

• To address the severe undersampling of the global virome, with less than 1-2% currently characterized.
• To develop a computational method that infers missing host-virus associations despite sparse and incomplete virome data.
• To uncover biologically plausible, coevolutionarily informed viral interactions that remain undetected in current sampling.
• To evaluate whether the global structure of the imputed mammal-virus network improves predictions of zoonotic disease emergence.

Proposed method

• Linear Filtering (LF) is applied to remove noise and stabilize the host-virus co-occurrence matrix by filtering out low-signal interactions.
• Singular Value Decomposition (SVD) is used to decompose the filtered matrix into low-rank components, capturing latent patterns of host-virus associations.
• The imputed matrix is reconstructed from the top singular values and vectors, enabling inference of missing but biologically plausible host-virus pairs.
• Graph embedding is applied to the imputed network to encode global network structure for downstream prediction tasks.
• The method leverages evolutionary signal in known associations to prioritize biologically plausible imputations.
• The approach is validated by assessing its ability to improve prediction of human infectivity from viral genome features.

Experimental results

Research questions

• RQ1 What host-virus associations are likely missing from current virome surveys, and where are they geographically concentrated?
• RQ2 Can latent patterns in known host-virus interactions, inferred via LF-SVD, reveal biologically meaningful, previously unsampled associations?
• RQ3 Does the global network structure of the imputed mammal-virus interaction network improve predictions of zoonotic potential?
• RQ4 Are there regions with unusually high levels of unique, unsampled host-virus interactions, suggesting hidden viral diversity?

Key findings

• The method successfully imputed a large number of highly plausible host-virus associations, with strong signals of coevolutionary history.
• A global hotspot of uniquely unsampled host-virus interactions was identified in the Amazon rainforest, indicating high viral diversity in this region.
• The imputed network revealed a significant enrichment of host-virus pairs with evolutionary signatures consistent with long-term coevolution.
• Graph embedding of the imputed network improved predictions of human infection risk from viral genome features, demonstrating the utility of global network structure for disease emergence modeling.