Qbiome

Reverse-engineering the Infant Microbiome To Predict Severe Neurodevelopmental Deficits

Last modified: 2022-11-01

Patient characteristics (UC MIND cohort)

AHCG: appropriate head circumference growth

SHCG: suboptimal head circumference growth

AHCG (\(n = 28\)) SHCG (\(n = 30\))
Age (EGA weeks), mean \(\pm\) SD 28.32 \(\pm\) 2.60 26.9 \(\pm\) 2.64
Male, no. (%) 13 (46.43) 15 (50)
BW, g, mean \(\pm\) SD 1021.96 \(\pm\) 382.91 998.3 \(\pm\) 423.45
BHC, cm, mean \(\pm\) SD 24.875 \(\pm\) 2.88 24.57 \(\pm\) 3.26
Vaginal delivery, no. (%) 12 (42.86) 4 (13.33)
Length of stay, days, mean \(\pm\) SD 77.89 \(\pm\) 34.77 103.27 \(\pm\) 61.21
PMA at discharge, weeks, mean \(\pm\) SD 39.04 \(\pm\) 3.80 41.13 \(\pm\) 6.80

The Q-net: A Linked Collection of Predictors

Goal: an ecosystem model expressive enough to learn nontrivial dynamics yet capable of generating actionable clinical insights

Direct modeling of the ecosystem is difficult due to limited understanding of microbial dynamics; standard/predictive machine learning methods are less interpretable and require more data than we have

The Q-net: A Linked Collection of Predictors

Node

\(\iff\)

Biome entity

\(\iff\)

Decision tree predictor of entity’s abundance

Node

\(\iff\)

Biome entity

\(\iff\)

Decision tree predictor of entity’s abundance

Forecasting evolutionary trajectories

We can sample (q-sample) the predictors to reconstruct or forecast biome trajectories

Realistic initial conditions (biome abundances measured over several weeks)

\(\implies\)

Sample the Q-net

\(\implies\)

Recover patterns in the ecosystem (even from small cohort sizes)

UChicago cohort

Boston cohort

Forecasting of Distinct Clinical Cohorts

Using distinct Q-nets allows us to forecast differences between phenotypes





Clinical Risk Assessment

The Q-net induces a novel risk measure that can highlight differences in phenotypes

Risk differences between cohorts

Distribution of risk

Risk computed weekly

Identifying a “Healthy Microbiome”

Goal: Want mapping of “microbiome profile” space, separated into healthy and unhealthy regions

Challenges:

  1. (Definition): How to define healthy?
  2. (Validity): How to identify valid microbiome profiles?

Claim: q-sampled microbiome profiles are more biologically likely/valid than random samples1

\(\implies\) q-samples help address the validity problem




[1] The Q-net predictors are nonparametric estimates of the full conditional distributions \(f(X_i \,|\, X_j, j\neq i)\) of each entity \(X_i\)

q-samples capture non-uniform structure in the space of microbiome profiles

Spatial distribution of biologically likely/valid microbiome profiles

High- and Low-Risk regions of biologically likely/valid profiles

High risk

Low risk

q-samples from high- and low-risk regions