A Joint Matrix Completion and Filtering Model for Influenza Serological Data Integration

Xiao–Tong Yuan, Tong Zhang, Xiu‐Feng WanPLoS 2013 · DOI

TL;DR

This paper introduces a novel "Joint Matrix Completion and Filtering" model to integrate noisy and incomplete influenza serological data (HI assays) from multiple experiments. By simultaneously addressing missing values and experimental variations, their method produces more accurate antigenic maps, crucial for vaccine strain selection and pandemic preparedness.

What Problem Does This Paper Address?

Influenza virus constantly evolves, necessitating regular updates to vaccine strains. Hemagglutination Inhibition (HI) assays are a standard method for characterizing the antigenic relationships between different influenza strains (antigens) and antibodies (antisera). However, performing all possible antigen-antiserum pairings is logistically impossible, leading to incomplete HI tables. Furthermore, measurements across different experiments can be inconsistent due to variations in experimental conditions (e.g., reagent batches, storage, personnel). This results in datasets with both missing values and noise, making direct integration and analysis challenging. Existing methods often struggle to handle both incompleteness and noise simultaneously, hindering accurate antigenic characterization.

Methods

The core of the paper is the proposed Joint Matrix Completion and Filtering (JMCF) model. This model tackles the integration problem by making two key assumptions:

  1. Low-Rank Assumption: The underlying merged HI data matrix is assumed to be low-rank. This means that the antigenic relationships between viruses can be represented by latent low-dimensional vectors for antigens and antisera, and their pairwise inner product approximates the HI titre. This is motivated by the idea that viruses with similar antigenic properties will have linearly dependent representations.
  2. Hierarchical Noise Modeling: The model explicitly accounts for different levels of noise. It incorporates a "bias" term for each antiserum within each experiment, capturing systematic variations due to experimental conditions. Gaussian noise is also modeled for the observations.

The JMCF model is optimized using a blockwise coordinate descent procedure. This involves iteratively updating two sets of parameters:

  • Latent Vectors (U, V): These represent the underlying antigenic profiles of antigens and antisera. This subproblem is framed as a standard matrix completion task.
  • Antiserum Biases (R): These capture the systematic deviations in measurements for each antiserum across different experiments.

The optimization aims to minimize the squared error between observed and predicted HI titres, while also accounting for the noise variance.

Datasets

The paper evaluates its method on two types of datasets:

  • Synthetic Influenza Datasets: Generated using the proposed model (Equation 1) with varying noise levels and ranks. This allows for controlled evaluation against known ground truth.
  • Real Influenza Datasets:
    • 2009 H1N1 Influenza A Viruses: Six HI tables collected from May to August 2009, comprising 253 viruses and 28 antisera.
  • 2000-2007 H3N2 Human Seasonal Influenza Viruses: 369 individual HI tables spanning 2000-2007, with a merged table of 11,383 viruses and 393 antisera.

Key Insights and Ideas

  • Simultaneous Handling of Noise and Missingness is Crucial: The paper strongly demonstrates that simply averaging data and then performing matrix completion (the "average model") is inferior to a method that explicitly models experimental biases. This highlights that experimental variations are not just random noise but systematic effects that need to be accounted for.
  • Low-Rank Assumption is Valid and Powerful: The success of the model, particularly in reconstructing antigenic cartographies, validates the intuition that influenza virus antigenic relationships can be effectively captured by low-dimensional latent representations.
  • Antigenic Cartography as a Biological Validation Tool: Visualizing the integrated data as antigenic cartographies provides an intuitive way to assess the biological plausibility of the integration. The JMCF model consistently produces cartographies that better reflect known biological patterns (e.g., mixing of viruses from different months in the 2009 H1N1 data) compared to the average model.
  • Generalizability Beyond Influenza: The authors propose that the JMCF model is a general framework for integrating noisy, incomplete biological data, applicable to other binding affinity datasets like DNA microarrays or protein microarrays.

Key Results

  • Improved Accuracy: The JMCF model consistently achieved lower Root Mean Squared Error (RMSE) compared to the "average model" on both simulated and real influenza datasets. For instance, on simulated data with noise configuration (r, σ) = (3,1), the JMCF model achieved an RMSE of 1.624 (±0.061) for HI titre, while the average model achieved 2.004 (±0.068).
  • Accurate Bias Recovery: The model was able to accurately recover the ground truth antiserum biases added to the simulated data, demonstrating its ability to disentangle systematic experimental effects.
  • Biologically Plausible Antigenic Maps: Antigenic cartographies generated by the JMCF model showed better clustering and mixing of related viruses, aligning better with biological expectations than those from the average model. For the 2009 H1N1 data, the JMCF model integrated viruses from different months more cohesively, which is biologically more sound than the strip-like separation observed with the average model.
  • Robustness to Noise: The model's performance showed resilience to increasing noise levels, particularly when the rank was close to the true underlying rank.

Limitations and Future Work

  • Parameter Selection: The rank parameter 'd' for matrix completion is a free parameter. While cross-validation was used to select it, it remains a point of tuning. The authors note that prediction performance deteriorates as the rank increases beyond the optimal value, reinforcing the importance of the low-rank assumption.
  • Computational Cost: While an efficient blockwise coordinate descent procedure is used, the computational complexity for very large datasets might still be a consideration.
  • Generalizability: While proposed as a general model, further validation on diverse biological datasets beyond influenza serology would be beneficial.
  • Future Work: The authors suggest that the model can be adapted for other biological data integration tasks, implying potential for broader application. They also mention the availability of Matlab code for reproducing their results, which is a good step for community adoption.

Summary updated

Abstract

Antigenic characterization based on serological data, such as Hemagglutination Inhibition (HI) assay, is one of the routine procedures for influenza vaccine strain selection. In many cases, it would be impossible to measure all pairwise antigenic correlations between testing antigens and reference antisera in each individual experiment. Thus, we have to combine and integrate the HI tables from a number of individual experiments. Measurements from different experiments may be inconsistent due to different experimental conditions. Consequently we will observe a matrix with missing data and possibly inconsistent measurements. In this paper, we develop a new mathematical model, which we refer to as Joint Matrix Completion and Filtering, for HI data integration. In this approach, we simultaneously handle the incompleteness and uncertainty of observations by assuming that the underlying merged HI data matrix has low rank, as well as carefully modeling different levels of noises in each individual table. An efficient blockwise coordinate descent procedure is developed for optimization. The performance of our approach is validated on synthetic and real influenza datasets. The proposed joint matrix completion and filtering model can be adapted as a general model for biological data integration, targeting data noises and missing values within and across experiments.
BibTeX
@inproceedings{yuan2013joint,
  author = {Xiao–Tong Yuan and Tong Zhang and Xiu‐Feng Wan},
  title = {{A Joint Matrix Completion and Filtering Model for Influenza Serological Data Integration}},
  booktitle = {PLoS},
  year = {2013},
  doi = {10.1371/journal.pone.0069842},
}
Figure