Deep Learning Enables Design of Multifunctional Synthetic Human Gut Microbiome Dynamics

Mayank Baranwal, Ryan L. Clark, Jaron Thompson, Zeyu Sun, Alfred O. Hero, Ophelia S. VenturellibioRxiv 2021 · DOI

TL;DR

This paper introduces a Long Short-Term Memory (LSTM) deep learning framework to predict and design the complex dynamics of synthetic human gut microbiomes. By outperforming traditional models, this approach enables the identification of key microbial interactions and the design of communities with specific, health-relevant metabolite profiles.

What Problem Does This Paper Address?

Understanding and engineering microbial communities, particularly the human gut microbiome, is crucial for health and disease. These communities are incredibly complex, with intricate interactions between species and their environment that change over time. Traditional models, like the generalized Lotka-Volterra (gLV) model, struggle to capture these complex, non-linear dynamics, especially when dealing with many species and multiple functions simultaneously. They also don't scale well with increasing community complexity. This gap limits our ability to design synthetic microbiomes for specific beneficial outcomes.

Methods

The core of this work is the application of a Long Short-Term Memory (LSTM) neural network. LSTMs are a type of recurrent neural network well-suited for time-series data, capable of learning complex, non-linear relationships and temporal dependencies. The authors adapted LSTMs to predict both species abundance and metabolite concentrations over time.

Key methodological innovations include:

  • Data-driven Dynamical System Modeling: Instead of relying on predefined ecological theories, the LSTM learns a high-dimensional, data-driven model of the microbiome's dynamics.
  • Two-Stage Experimental Design: The LSTM was used to guide experimental planning. First, it predicted metabolite profiles for millions of potential communities, identifying "corner cases" (rare, extreme profiles). A second, more focused experiment then collected data around these corner cases to refine the model.
  • Interpretability Tools: To overcome the "black-box" nature of deep learning, the authors employed Local Interpretable Model-Agnostic Explanations (LIME) to decipher microbe-microbe and microbe-metabolite interactions from the LSTM's predictions. They also used gradient-based sensitivity analysis.

Datasets

The study utilized existing datasets from previous work on synthetic human gut communities. Specifically:

  • 12-member synthetic human gut community: Used for initial validation of the LSTM's ability to predict temporal changes in species abundance in response to dilution.
  • 25-member synthetic human gut community: Used for training and evaluating the model's performance in predicting species abundance and metabolite concentrations, including data from a pilot experiment and a subsequent, more detailed experiment.

Key Insights and Ideas

  • LSTM's Superiority: The LSTM framework significantly outperforms traditional gLV models, especially when dealing with higher-order interactions and complex community dynamics.
  • Data-Driven Design: The LSTM can navigate a vast "functional landscape" of potential communities to identify those with specific, desired metabolite profiles, including rare "corner cases" that would be missed by random sampling.
  • Uncovering Key Players: LIME and gradient-based analysis revealed that certain bacteria (e.g., Actinobacteria, Firmicutes, Proteobacteria) are crucial drivers of metabolite production, while others (e.g., Bacteroides) primarily shape community dynamics.
  • Temporal Dynamics Matter: End-point metabolite concentrations are not enough; understanding the temporal trajectories of metabolites and species is essential for a complete picture and for designing communities with specific dynamic behaviors.
  • Experimental Design Guidance: The LSTM can intelligently guide experimental planning, drastically reducing the number of experiments needed to explore complex community spaces.

Key Results

  • The LSTM model achieved high prediction accuracy for species abundance (Pearson R² scores up to 0.74 at 60 hours on hold-out test sets) and metabolite concentrations.
  • In silico experiments showed the LSTM significantly outperformed the gLV model when third-order interactions were present, with R² scores of 0.85 vs. 0.52 for mild interactions and even better for moderate interactions.
  • The LSTM framework enabled the design and experimental validation of communities with tailored metabolite profiles, outperforming a composite gLV and regression model in 59% of output variables.
  • Interpretability methods successfully identified key species and interactions driving metabolite production, aligning with known biological mechanisms and revealing new insights.
  • Clustering of dynamic metabolite trajectories revealed distinct community behaviors, with some driven by simple species presence/absence rules and others by more complex interactions.

Limitations and Future Work

  • Training Data Bias: The LSTM showed difficulty predicting the behavior of high-richness communities when trained only on low-richness data. This was addressed by augmenting the training set with more diverse communities.
  • Interpretability Nuances: While LIME and gradient analysis provide valuable insights, they are approximations. The authors suggest that model-aware interpretability methods could be further explored.
  • Environmental Context: The current model primarily focuses on species composition. Future work could integrate environmental factors and genomic information to create more comprehensive predictive models.
  • Broader Applications: The authors envision extending this approach to design communities for various applications like bioremediation and bioprocessing, and to predict the impact of environmental changes on microbiome function. They also suggest exploring the design of communities with varied metabolic capabilities based on genomic information.

Summary updated

Abstract

Abstract Predicting the dynamics and functions of microbiomes constructed from the bottom-up is a key challenge in exploiting them to our benefit. Current ordinary differential equation-based models fail to capture complex behaviors that fall outside of a predetermined ecological theory and do not scale well with increasing community complexity and in considering multiple functions. We develop and apply a long short-term memory (LSTM) framework to advance our understanding of community assembly and health-relevant metabolite production using a synthetic human gut community. A mainstay of deep learning, the LSTM learns a high dimensional data-driven non-linear dynamical system model used to design communities with desired metabolite profiles. We show that the LSTM model can outperform the widely used generalized Lotka-Volterra model. We build methods decipher microbe-microbe and microbe-metabolite interactions from an otherwise black-box model. These methods highlight that Actinobacteria, Firmicutes and Proteobacteria are significant drivers of metabolite production whereas Bacteroides shape community dynamics. We use the LSTM model to navigate a large multidimensional functional landscape to identify communities with unique health-relevant metabolite profiles and temporal behaviors. In sum, the accuracy of the LSTM model can be exploited for experimental planning and to guide the design of synthetic microbiomes with target dynamic functions.
BibTeX
@misc{baranwal2021deep,
  author = {Mayank Baranwal and Ryan L. Clark and Jaron Thompson and Zeyu Sun and Alfred O. Hero and Ophelia S. Venturelli},
  title = {{Deep Learning Enables Design of Multifunctional Synthetic Human Gut Microbiome Dynamics}},
  year = {2021},
  doi = {10.1101/2021.09.27.461983},
  howpublished = {bioRxiv},
}
Figure