Utilizing longitudinal microbiome taxonomic profiles to predict food allergy via Long Short-Term Memory networks

Ahmed A. Metwally, Philip S. Yu, Derek Reiman, Yang Dai, Patricia W. Finn, David L. PerkinsPLOS Computational Biology 2019 · DOI · OpenReview

TL;DR

This paper introduces a deep learning framework using Long Short-Term Memory (LSTM) networks to predict food allergy in infants from longitudinal gut microbiome data. By leveraging the temporal dependencies in microbiome changes, their model significantly outperforms traditional machine learning approaches, suggesting a promising avenue for early allergy prediction and intervention.

What Problem Does This Paper Address?

Diagnosing food allergies in early childhood is challenging, and delayed diagnosis can lead to severe complications. While the gut microbiome is known to influence allergy development, effectively using longitudinal microbiome data for prediction has been difficult. Existing methods often struggle with the inherent variability in sample collection times and uneven data points across individuals in longitudinal studies. This paper aims to bridge this gap by developing a robust predictive model that can capture the dynamic nature of the gut microbiome over time to forecast food allergy development.

Methods

The core of the proposed framework is a Long Short-Term Memory (LSTM) network, a type of recurrent neural network adept at learning from sequential data. Unlike standard Recurrent Neural Networks (RNNs), LSTMs are designed to overcome the vanishing gradient problem, allowing them to capture long-term dependencies in data.

Before feeding data into the LSTM, the researchers explored two main strategies for feature selection/extraction:

  • Sparse Autoencoder: This unsupervised deep learning technique learns a compressed, latent representation of the microbial features. The architecture used had three hidden layers, with the middle layer acting as the latent representation. Sparsity was enforced to encourage more specific feature learning.
  • Traditional Feature Selection: Methods like Minimum Redundancy Maximum Relevance (mRMR) and selecting the top variable features were employed to identify a smaller, more interpretable subset of microbial features.

The framework then uses these selected/extracted features as input to the LSTM, which learns temporal patterns. The output of the last LSTM cell is fed into a softmax layer for classification into "allergic" or "non-allergic" categories.

Datasets

The study utilized the DIABIMMUNE dataset, a longitudinal study of infant gut microbiome profiles. This dataset includes stool samples collected from 222 infants across Russia, Finland, and Estonia from birth to 3 years of age. Food allergen-specific IgE levels were measured, and infants were classified as allergic or non-allergic to milk, eggs, or peanuts. After preprocessing (removing samples without allergy labels and subjects with fewer than 3 timepoints), the dataset comprised 658 samples from 148 subjects.

Key Insights and Ideas

  • Temporal Dynamics Matter: The study strongly supports the hypothesis that learning from the sequence of gut microbiome changes (temporal dependency) significantly improves food allergy prediction compared to analyzing individual time points in isolation.
  • LSTM's Superiority: LSTMs, by their nature, are well-suited for this task, outperforming traditional methods like Hidden Markov Models (HMMs) which, while also sequential, struggled with the sampling inconsistencies in the DIABIMMUNE dataset.
  • Feature Selection is Crucial: The performance of the LSTM model was highly dependent on the input features. Combining LSTMs with features selected by mRMR yielded the best results, suggesting that carefully chosen, interpretable features can be more effective than compressed latent representations from autoencoders for this specific problem.
  • Interpretability vs. Performance: While autoencoders offer compressed representations, mRMR-selected features, though less compressed, provided better predictive power and the advantage of interpretability (e.g., identifying specific bacteria like Faecalibacterium associated with allergy status).

Key Results

The proposed LSTM framework, particularly when using features selected by the mRMR method (LSTM-mRMR-25), achieved the highest performance metrics (auROC and MCC) among all tested models.

  • The LSTM-mRMR-25 model achieved an auROC of 0.69 (±0.09) and an MCC of 0.40 (±0.19).
  • This performance was significantly better than other models, including HMM, MLPNN, RF, SVM, and LASSO, across various feature types (latent features, mRMR features, variable features, and raw features).
  • While autoencoder-derived latent features showed some promise, they generally did not outperform mRMR-selected features, possibly because the autoencoder's training objective did not explicitly consider the phenotype.
  • The study also noted that the DIABIMMUNE dataset, while valuable, is relatively small with an average of only 6 timepoints per subject, which limits the ultimate predictive power achievable.

Limitations and Future Work

  • Dataset Size and Sparsity: The authors acknowledge that the DIABIMMUNE dataset is small and has sampling inconsistencies, which limits the current predictive accuracy. They suggest that larger, more comprehensive longitudinal microbiome datasets will be crucial for training more powerful models.
  • Predictive Threshold: The current model's performance, while superior to baselines, does not yet reach a level suitable for direct clinical utilization.
  • Transfer Learning: The authors propose exploring transfer learning as a future direction, where knowledge gained from models trained on one task could be applied to improve predictions for other diseases linked to host-microbiome interactions (e.g., asthma, diabetes mellitus).
  • Broader Applications: The study suggests that the LSTM network learning approach, combined with feature selection, could be beneficial for predicting other clinical outcomes from longitudinal microbiome data.

Summary updated

Abstract

Food allergy is usually difficult to diagnose in early life, and the inability to diagnose patients with atopic diseases at an early age may lead to severe complications. Numerous studies have suggested an association between the infant gut microbiome and development of allergy. In this work, we investigated the capacity of Long Short-Term Memory (LSTM) networks to predict food allergies in early life (0-3 years) from subjects’ longitudinal gut microbiome profiles. Using the DIABIMMUNE dataset, we show an increase in predictive power using our model compared to Hidden Markov Model, Multi-Layer Perceptron Neural Network, Support Vector Machine, Random Forest, and LASSO regression. We further evaluated whether the training of LSTM networks benefits from reduced representations of microbial features. We considered sparse autoencoder for extraction of potential latent representations in addition to standard feature selection procedures based on Minimum Redundancy Maximum Relevance (mRMR) and variance prior to the training of LSTM networks. The comprehensive evaluation reveals that LSTM networks with the mRMR selected features achieve significantly better performance compared to the other tested machine learning models.
BibTeX
@article{metwally2019utilizing,
  author = {Ahmed A. Metwally and Philip S. Yu and Derek Reiman and Yang Dai and Patricia W. Finn and David L. Perkins},
  title = {{Utilizing longitudinal microbiome taxonomic profiles to predict food allergy via Long Short-Term Memory networks}},
  journal = {PLOS Computational Biology},
  year = {2019},
  doi = {10.1371/journal.pcbi.1006693},
}
Figure