Deep Learning Enables Design of Multifunctional Synthetic Human Gut Microbiome Dynamics

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.

Predicting metabolomic profiles from microbial composition through neural ordinary differential equations

“…To tackle this lack-of-interpretability issue, we proposed a perturbation method for mNODE to decipher the relationships between microbes and metabolites by measuring “susceptibilities” based on the response of metabolite concentrations to perturbation in inputs such as microbial relative abundances. Our method is similar to but simpler than the previously developed LIME (Locally Interpretable Model-agnostic Explanations) [48, 49]. To reveal the effect of species i on metabolite α , we perturbed the relative abundance of species i ( x i ) by a small amount Δ x i for well-trained mNODE, re-predicted the concentration of metabolite α , and measured the deviation from the original prediction (we denoted the deviation as Δ y α ).…”

Predicting and prioritizing community assembly: learning outcomes via experiments

“…In scenarios where it is not time or cost effective to experimentally explore all possible communities (53), LOVE could allow rapid screening of a subset of assemblages which will likely have desirable properties (e.g. production of a certain metabolite (18,25,54) or ecosystem function (55, 56), presence of a certain subset of species (57, 58), similar to guided screening approaches in drug discovery (59). In this sense, LOVE is guiding reachability analyses - the control theory inverse problem of identifying the set of initial experimental conditions which will eventually come to a certain outcome (60).…”

“…LOVE provides a promising approach for predicting coexistence outcomes from experimental community assembly datasets, and complements a growing body of work on machine learning approaches in community ecology (7,(17)(18)(19)(24)(25)(26). Applications to experiment prioritization seem possible and could be useful if explored ethically.…”

“…Conceptually similar are machine learning methods such as random forest and neural networks that predict species abundance across space/time based on statistical relationships for observed species abundances. These approaches still have relatively high observational data requirements (18,(23)(24)(25)(26). For instance, accurate predictions of spatial variation in abundance for 23 plant species across a 1 km transect, required 37,240 observations in (26).…”

Model-guided design of the diversity of a synthetic human gut community

“…S12 ). The unexplained variance in the dataset could be attributed to differences in species richness in the training (10-species) and test data (2-4 species) 25,55 . In sum, these data demonstrate that the gLV model can guide the design of communities that exploit inter-species interactions to support the persistence of lower fitness species over longer timescales, as well as mitigate overgrowth of high fitness species.…”