A digital twin of the infant microbiome to predict neurodevelopmental deficits

Abstract

Despite the recognized gut-brain axis link, natural variations in microbial profiles between patients hinder definition of normal abundance ranges, confounding the impact of dysbiosis on infant neurodevelopment. We infer a digital twin of the infant microbiome, forecasting ecosystem trajectories from a few initial observations. Using 16S ribosomal RNA profiles from 88 preterm infants (398 fecal samples and 32,942 abundance estimates for 91 microbial classes), the model (Q-net) predicts abundance dynamics with R² = 0.69. Contrasting the fit to Q-nets of typical versus suboptimal development, we can reliably estimate individual deficit risk (M_δ) and identify infants achieving poor future head circumference growth with ≈76% area under the receiver operator characteristic curve, 95% ± 1.8% positive predictive value at 98% specificity at 30 weeks postmenstrual age. We find that early transplantation might mitigate risk for ≈45.2% of the cohort, with potentially negative effects from incorrect supplementation. Q-nets are generative artificial intelligence models for ecosystem dynamics, with broad potential applications.

Fig. 1.. Scheme of the study.

(A and B) Longitudinal fecal samples from infants born <35 weeks PMA subjected to 16S rRNA gene sequencing, along with infant head circumference growth (HCG) measurement. (C) Microbiome abundance data are quantized into 26 levels between measured maxima for each observed taxonomic class, and is used to infer a digital twin, reflecting complex emergent dependencies across the classes via learning a recursive forest of conditional inference trees (the Q-net). Component trees in the Q-net (two examples shown) are inter-dependent, where nonleaf node can recursively expand to its own tree. (D) Complex dependencies inferred for typically developing sub-cohort across classes and observation time points. Each inferred dependency rule probabilistically relates an a priori unspecified number of entities (>2), and together specify a generative model of ecosystem trajectory and its dynamical operation. (E) Applications enabled by the inferred digital twin, with out-of-sample validation and important mechanistic insights.

Fig. 2.. Trajectory forecasts.

Population-level forecasting of mean abundance trajectories of select taxonomic classes (defined by having mean relative abundance >0.01 in the training set) from a set of observations restricted to <28 weeks PMA. (A) Forecasts generated from initial conditions specified by the UChicago cohort (from which the Q-net was inferred). Average R² across these taxa is ≈0.856 at 28 weeks, and ≈0.896 at 31 weeks. (B) Forecasts generated using initial conditions specified by the Boston cohort (fully out-of-sample data that were not used for inference). Average R² across these taxa is ≈0.350 (which increases to ≈0.378 when allowing for a temporal shift of 1 week) at 28 weeks, and ≈0.689 at 31 weeks. In both cases, explained variance is typically high in these important classes, suggesting that the Q-net model successfully captures the complicated dynamical trajectories.

Fig. 3.. Classification performance out-of-sample.

(A) Classification performance to recognize infants with eventually suboptimal HCG. The AUC is maximum at 32 weeks PMA reaching 87.6% for the UChicago cohort (in-sample data), and 75.9% for the Boston cohort. (B) Precision-recall curves. (C) Trade-off between positive (LR+) and negative (LR−) likelihood ratios. (D) Change in LR+ with sensitivity or recall. (E) Comparison of PPV versus NPV at different PMA weeks. (F) Fitting the computed AUCs over time, we note that the AUC > 80% stabilizes approximately over 30 weeks PMA for the UChicago cohort. (G) Top positive SHAP values for the M_δ risk driving HCG classification at 36 weeks PMA, and (H) top negative SHAP values for risk. Positive (negative) SHAP values indicate if an individual’s abundance of a specific entity increases (decreases) the risk (compared to baseline samples) of a positive diagnosis of the target disease; thus, the observed levels of Gammaproteobacteria and Clostridia (among others) are often associated with increased individual risk while Bacteriodia and Actinobacteria (among others) are similarly implicated with decreased risk. Note, however, that several taxa appear on both lists, suggesting complex dependencies of risk on entity abundances that vary over time.

Fig. 4.. Network structure comparison between typical and suboptimal cohorts.

Change in inferred directional dependencies between prominent taxa in early development (≦31 weeks) between optimal and suboptimal HCG, visualized via computing LOMAR coefficients. (A) AHCG to (B) SHCG. The bold edges highlight some key structural changes in Actinobacteria interactions. The edges in red show that the key changes in Actinobacteria or Bacteroidia are supplanted as potential interventions in suboptimal HCG. Dashed edges for Actinobacteria are interactions that emerge later than their corresponding nondashed ones (tables S5 to S8). Notably, SHCG interactions are both more complex and strongly connected.

Fig. 5.. Impact of clinical variables and diet.

(A) Impact on M_δ risk of suboptimal HCG. We find that, on average, being male, use of antibiotics, and enteral feed in amount and number of days maximally increase such risk. (B) shows the distribution of SHCG patients by the types of intervention found to decrease their risk (microbiome-based supplantation intervention and/or feeding-based intervention). (C) depicts SHAP values for variables defining feeding and supplantation interventional categories in the SHCG cohort. (D) shows that supplantation features are associated with greater decreases in risk than feeding, but that among feeding interventions, total enteral feed is associated with the maximum decrease in risk.

Fig. 6.. Design of personalized early interventions to reduce risk of suboptimal HCG.

(A to C) SHAP profiles of three patients who have suboptimal HCG, but have three distinct intervention phenotypes, namely, where supplanting Bacteroidia reduces risk (A), supplanting Actinobacteria reduces risk (B), and where no time-independent consistent supplantation can be obtained from our SHAP analysis (C). Both Actinobacteria and Bacteroidia have opposing effects on risk at different time points. (D and E) The breakdown of these three intervention phenotypes in UChicago and Boston cohorts. (F) Breakdown of intervention phenotypes among all patients with suboptimal HCG, showing that 45.2% of patients have discernible interventions.